A Survey of VANET/V2X Routing From the Perspective of Non-Learning- and Learning-Based Approaches

Intelligent transportation systems (ITSs) have become increasingly popular because they support effective coordination in connected vehicles. ITSs present an integrated approach for exchanging relevant information in order to improve the safety, efficiency, and reliability of road transportation systems. A variant of mobile ad-hoc networks (MANETs) called vehicular ad-hoc networks (VANETs) are an integral component of ITSs. VANETs consist of interconnected vehicles with sensing abilities that exchange information related to traffic, positioning, weather, and emergency services. In general, vehicle-to-everything (V2X) refers to communications between any entity and a vehicle, where the entity may be a vehicle, a cloud-based network, a pedestrian, or equipment installed along a road. One of the crucial challenges in V2X is the reliable and timely circulation of information among vehicular nodes to allow drivers to make decisions that increase road safety. In this context, efficient V2X routing protocols play a key role in supporting reliability and safety, and enhancing the overall quality of service (QoS) in VANETs. However, VANETs have distinct characteristics, such as high vehicular node mobility, unsteady connectivity, rapid changes in network topology, and unbounded network size, that can significantly affect routing in the network. Various routing protocols for V2X communication exist in the open technical literature. In this survey, we categorize the routing mechanisms as non-learning- and learning-based approaches, and discuss existing V2X routing protocols and their contributions to and impacts on VANET performance. Here, the learning-based approach implies the use of machine learning algorithms. This survey also summarizes open challenges in designing effective V2X routing protocols and future research directions to consider when developing smart routing mechanisms for next-generation intelligent VANET technologies.


FIGURE 1. An illustration of V2X communications.
in communication that has recently been widely extended to challenging communication scenarios like communication among flying drones and ad-hoc communications among mobile devices. MANETs consist of mobile and fixed nodes that follow specific protocols to communicate with each other. Vehicular ad-hoc networks (VANETs) [4]- [6] are a specific kind of MANET, that are linked with ITSs to provide traffic analysis, road safety, collision avoidance, emergency alerts, vehicle speed management, and automatic traffic signal enforcement [7]. Thus, VANETs are ad-hoc networks in which information can be communicated among vehicles. As is rightly mentioned in [8], VANETs are a key component of ITSs, which have been extensively investigated, notably for improving efficiency in traffic management, developing future smart cities, and addressing future transportation safety-related issues. Furthermore, VANETs can be deployed in military environments or remote areas, where network infrastructure cannot be installed.
In general, vehicle-to-everything (V2X) [20]- [22] refers to communications between any entity and a vehicle. Thus, V2X technology mainly covers vehicle-to-infrastructure (V2I), vehicle-to-vehicle (V2V), vehicle-to-network (V2N), and vehicle-to-pedestrian (V2P) communications. Fig. 1 shows an example of V2X communications. In VANETs, vehicles can convey information to other vehicles using V2V communications through which a vehicle communicates directly with another vehicle and shares information related to traffic conditions, such as traffic jams or accidents. V2I communications rely on the use of road side units (RSUs) [23], [24], which are generally special wireless devices that are mounted on the roadside. An RSU is a fixed piece of infrastructure that is connected to the Internet and can communicate with other RSUs and roaming vehicles, which transmit data to the nearest RSU. RSUs ensure reliability in data communications, and are thus helpful in many life-saving mishaps [25]. When RSUs are not directly in the range of vehicles, V2V communication is considered mandatory in VANETs [14].
In the case of V2N communications, vehicles access cloud-based services and 3G/4G cellular networks are used to deliver and receive messages for vehicle monitoring due to their enhanced coverage and high data transfer rates [26]. In V2P communications, the exchange of data is between a vehicle and a pedestrian. In the case of in-vehicle communications, communication occurs inside a vehicle for the detection and analysis of both drivers and vehicles to minimize accidents due to driver drowsiness or vehicle malfunctioning [27].
Routing is the mechanism used to select packets and transmit them on the least-cost path between the source and destination nodes. Thus, successful data transmission needs to be ensured by the routing protocol, which helps forward data packets to the intended destination. If transmission on the selected route fails, an appropriate recovery mechanism needs to be executed to retransmit the data [10]. In VANETs, vehicles are positioned on the road and act as mobile nodes in the network. In this context, ITSs are important, where routing can be intelligently handled, leading to efficient communications in VANETs. Thus, V2X routing protocols need to be designed to address challenges of VANET environments, such as rapid changes in network topology, high vehicle mobility, network heterogeneity, and variable network size. The physical layout of VANETs changes repeatedly due to the frequent movement of nodes, and route maintenance consequently becomes very challenging [28]. In addition, since vehicles move at high speeds, infrastructure-based networks can suffer from rapid handovers and radio signal fluctuations, leading to unreliable network connectivity [29]. Moreover, safety-related applications of ITSs depend on the aggregation of information from secured messages that are transmitted by vehicles in VANETs. For instance, safety messages could indicate the location, acceleration, speed, or braking state of a vehicle. The timely delivery of such information can help to avoid accidents [30]. Emergency and warning messages are exchanged by VANET nodes through unicast or multicast routing by following one of the V2X communication mechanisms, where routing can be delayed due to a limited number of cellular gateways or access points in road traffic environments [4].
Considering the aforementioned challenges associated with routing in V2X, it is essential to develop efficient routing mechanisms that transmit packets from the source to the destination by dynamically selecting a route containing a set of nodes such that the quality of service (QoS) criteria are satisfied in VANETs [31], [32]. The QoS criteria include the reliable and timely delivery of information and ITS applications [9], routing overhead, end-to-end delay, packet loss rate, network bandwidth, etc. Over the last few years, several works, such as [6], [33]- [49], have addressed the challenges of V2X routing and accordingly proposed routing protocols to improve road safety, security issues, and QoS by considering high mobility, irregular traffic patterns, frequent changes in network topology, and repeat exchanges of different types of information. These routing protocols focus on discovering and maintaining routes between two VANET nodes using non-learning-or learning-based approaches. Non-learningbased mechanisms do not use any machine learning scheme and generally follow threshold-based approaches. These routing protocols have simple designs and apply knowledge based on network topology, position, cluster, broadcasting status, and so on. Learning-based mechanisms, on the other hand, utilize machine learning algorithms to dynamically adapt to VANET environments by intelligently selecting routing paths. Machine learning algorithms can be effective tools to enhance trust, reliability, and route detection accuracy in V2X, as it is reported in several existing works [50]- [52]. In V2X, the decision-making for routing becomes intelligent [49] due to the incorporation of online learning approaches.

A. WHY THIS SURVEY
V2X communication poses some specific challenges, particularly because of the dynamic nature of the communication environment and fast-changing network states. Although the topic of V2X routing has been thoroughly studied and there already exist surveys on this area (summary in Table 1), many recent works, as stated above, have focused on applying artificial intelligence and machine learning techniques to explore the communication environment and choose routing paths. Such learning-based techniques sometimes augment existing non-learning-based techniques for V2X routing; however, many of them put the problem in a different dimension. Therefore, it would be interesting to explore such techniques and compare them with non-learning-based V2X routing techniques to get a clear understanding of the problem's dimension. Table 1 shows existing surveys that deal with V2X routing protocols and that there is currently no thorough survey of both non-learningand learning-based V2X routing protocols. Existing nonlearning-and learning based routing protocols in V2X have strengths and limitations that need to be properly identified in order to strengthen future designs of V2X routing mechanisms. Therefore, in this survey, we present an extensive study of the existing works that deal with the design and deployment of both non-learning-and learning-based routing protocols in V2X. Based on this study, we highlight the impacts of these two types of routing approaches on overall network performance and their associated security and authentication issues. More specifically, the contributions of this survey are framed as follows: • We study existing works that deal with non-learningbased routing protocols in V2X and highlight their contributions. The impacts of such routing protocols on VANET performance are discussed in light of high node mobility, high vehicle speed, QoS improvement, network topology, security and privacy, etc. The routing approaches' strengths and weaknesses are also highlighted.
• We review existing works that consider learning-based routing protocols in V2X. We present the impacts of these routing protocols on VANET performance in light of high vehicle speed, high node mobility, irregular vehicle density, QoS, security and privacy, dynamic network management, etc. We also discuss the contributions, strengths, and weaknesses of existing learning-based routing mechanisms. 23024 VOLUME 10, 2022 • We highlight the findings from the extensive study of both non-learning-and learning-based V2X routing protocols and present a comparative study of a classification of directions in which the aforementioned two routing approaches are considered in the existing literature.
• We discuss the challenges of and future research directions for designing effective routing mechanisms that lead to smart and efficient routing protocols for V2X communications.

B. UNIQUENESS OF THIS SURVEY
To the best of our knowledge, the study of existing VANET/V2X routing protocols based on both learning-and non-learning-based approaches is what makes our survey unique. This survey clearly categorizes existing works based on the two aforesaid approaches and mentions their strengths and weaknesses.

C. ORGANIZATION OF THIS SURVEY
The rest of this survey is organized as follows. Section II presents an overview of VANET/V2X, focusing on routing protocols. Section III provides a discussion on existing non-learning-based routing protocols. Section IV describes the existing works that consider learning-based V2X routing protocols and their impacts on network performance. In Section V, we present overall observations on existing non-learning-and learning-based V2X routing protocols. Section VI summarizes the challenges of and future research directions for designing efficient V2X routing protocols, and finally Section VII presents the conclusion of this survey.

II. VANET/V2X OVERVIEW
In this section, we present an overview of VANETs and highlight different types of VANET/V2X routing protocols. In V2X communications, one of the most challenging issues is VANET routing, where all the participants are vehicles having high mobility, long inter-vehicle distance, and high vehicle speed, and the vehicle density varies. Vehicles also have time-varying topology, trajectory-based predictable locations, low power consumption requirements, and frequent topology partitioning [53]. Therefore, we specifically focus on VANET routing in V2X communications.
VANETs are decentralized, self-configuring, and costeffective networks that are important for traffic management, disaster rescue operations, road safety, multimedia content sharing, commerce on wheels [54], etc. VANETs are a type of ad-hoc network [13], i.e., an interim network that does not have any permanent infrastructure. Since VANETs are a sub-class of MANET architectures, they can be considered as multi-hop networks [11], [13]. They provide a quick and practical solution in emergencies, such as in the event of a disaster in which hydro poles, electrical services, and pre-established networks are damaged, causing the cellular network to fail. In this case, the disaster management team can create a simple ad-hoc network, like a VANET, for devices to communicate [55]. In addition, VANETs have been implemented to monitor traffic light behavior in the interest of curbing the risk on roads since road accidents are growing in number nowadays [56]. VANETs are characterized by: • High mobility: In VANETs, vehicles can have high speeds, and making decisions on routing and security measures based on the position of a node is thus very tricky and challenging [28].
• Unbounded network size: The number of vehicles is unbounded, which leads to an uncontrolled network size in VANETs [12].
• Rapid change of network topology: Due to the continuous movement and high speeds of vehicles, the position of the vehicles changes repeatedly. As a result, the network topology changes rapidly and becomes unstable.
• Frequent information exchange: In VANETs, vehicles need to exchange information frequently with RSUs and other vehicles in order to control and manage routing and other functionalities. Thus, the frequent transmission of information must be efficiently handled. Fig. 2 shows an example of a VANET that has two types of nodes: RSU and on-board unit (OBU). These nodes create an ad-hoc network and enable communications using single-hop and multi-hop links. The RSU [23], [24] is connected to the Internet and forms a fixed infrastructure that can communicate with other RSUs and roaming vehicles. RSUs provide distributed and cooperative applications in which they, other RSUs, and vehicles work together to process and share information in order to coordinate actions. An OBU [24] is a wave device that is mounted on-board a vehicle and exchanges and shares information with other OBUs and RSUs. An application unit (AU) [24] is a device within a vehicle, that uses applications provided by the end user. The AU device can be dedicated to safety applications or be a normal device like a personal digital assistant (PDA) that is capable of using the Internet. The AU is connected to the OBU using a wireless or wired connection and can reside with the OBU on a single physical unit.

A. TYPES OF COMMUNICATIONS
There are four types of communications in vehicular networks:   [56]. 2) Vehicle-to-infrastructure communications: When communication occurs between a vehicle and a fixed RSU, this is known as V2I communication [12]. The concept of V2I is not limited to life on land but also pertains to less-explored oceanic life [57]. Cyclones, ship accidents, accidental oil spills, and other dangerous situations can be prevented or mitigated by using V2I communication under water. V2I communications can help us know about life-threatening conditions prior to their occurrence such that we can be warned and fishermen can be notified in the event of emergencies. 3) In-vehicle communications: This type of communication happens inside a vehicle and is required to detect and analyze issues related to drivers and their vehicles in order to minimize accidents due to driver drowsiness, vehicle malfunctioning, etc. 4) Vehicle-to-broadband cloud communications: In the vehicle-to-broadband cloud communications, 3G/4G cellular networks are used to deliver and receive messages for vehicle monitoring and tracking due to their enhanced coverage, high data transfer rates, and reliability [26].
• Topology-based protocols: These protocols are based on the topology of the network. They maintain a routing table to store information related to the links on which data packets are sent from the source to destination vehicles. Existing algorithms attempt to minimize the control overhead in topology-based routing. However, in this routing mechanism, packets need more time to be delivered. Thus, topology-based protocols perform slower than other types of VANET routing protocols. Topologybased protocols are further classified into three types as follows. 1) Proactive protocols: Proactive protocols are proactive and reactive protocols. The network is divided into local and global regions. In the local region, proactive routing protocols are applied to reduce delay and overhead, whereas in the global region, reactive protocols are applied.
• Position-based protocols: This type of protocols is mostly adopted when the positions of the sender and receiver and the ideal path to the destination can be known with the help of the global positioning system (GPS). Position-based routing is also commonly known as geographic routing and does not require that a routing table be maintained.
• Broadcast-based protocols: This type of protocols is chosen when the vehicle that will receive the message is out of transmission range. Packet delivery is ensured by making use of flooding, where every node in the network receives the packet.
• Cluster-based protocols: In this type of protocols, vehicles are grouped into clusters, and every cluster has a leader that helps in bridging separate clusters. Leaders of two clusters communicate with each other and can make a decision on routing messages inside the clusters they control.
• Geocast-based protocols: This type of protocols is suited for multicast routing. Here, nodes have information about other nodes, and unlike in multicast routing, packets are delivered to a pre-known geographical region. The primary aim of these protocols is to guarantee the delivery of a message at a low cost. Routes do not need to be maintained, and route construction is also not required. Geocast-based routing protocols are effective when the reliable delivery of messages is a concern [54].

III. NON-LEARNING-BASED V2X ROUTING PROTOCOLS
In this section, we discuss existing works that consider nonlearning-based V2X routing protocols and their impacts on network performance. Tables 14-16 present works that discuss non-learning-based V2X routing protocols and highlight their strengths and weaknesses. Fig. 4 shows the number of works published each year since 2008, that consider nonlearning-based V2X routing. The authors of [60] propose a protocol in which common backbone vehicles are used for various types of traffic movement and interflow network coding to encode packets at the backbone vehicles. Here, vehicle movement and the quality of links are considered to form a backbone. A new class of routing protocols called road-based vehicular traffic (RBVT) protocols which perform better in city VANETs, is presented in [61]. In that work, reactive RBVT (RBVT-R) and proactive RBVT (RBVT-P) protocols were implemented and compared with MANET and VANET representative protocols. According to the simulation results, RBVT-R achieves a 40% increase in average delivery rate and RBVP-P achieves an 85% decrease in average delay compared to existing protocols. The protocols proposed in [62] have been implemented both in city and highway environments. The protocol perfomance analysis was performed considering different performance metrics. In [63], a routing algorithm based on Ad-hoc On-demand Distance Vector (AODV), which uses routing metrics including the length of each hop and the link remaining time, is proposed. The algorithm reduces routing overhead by using a receiver-based method. Moreover, a novel urban road scenario and mobility model are designed to describe car movement. In [64], the performances of the AODV and optimized link state routing (OLSR) VANET protocols are evaluated in VANET crossroad scenarios considering the end-to-end delay, packet delivery ratio (PDR), and throughput. Traffic lights affect VANET routing performance, and this was investigated in [65], which proposes shortest-path-based trafficlight aware routing (STAR). The way in which the packets are forwarded is determined by traffic patterns and traffic light signals on road intersections. STAR works over green light first forwarding where packets are forwarded along the path with green lights. V2I and V2V communications must cooperate to form V2X communications, which require a certain type of multi-hop routing design to enhance data delivery performance. By using a mathematical framework, the aforementioned issue is addressed in [66] for investigating the performance of data delivery in V2X networks. With theoretical analysis, the formulated global and distributed optimization problem helps improve latency and the data transfer rate. However, high node mobility and vehicle density are not highlighted. Table 2 highlights the contributions of the works that consider high node mobility-based routing protocols in VANETs. In [33], traffic signals are used to address the issue of irregular vehicle distribution and network congestion at peak traffic hours. In that work, a routing protocol called greedy traffic light and queue aware (GTLQR), which takes into consideration of channel quality, street connectivity, relative distance, and queuing delay, is proposed to diminish packet loss and balance traffic loads among vehicles. The goal is to decrease the end-to-end delay pertaining to multi-hop routing while guaranteeing complete data packet delivery. In response to the problem of variable node mobility in VANETs, the work in [42] proposes an architecture in which dynamic moving zones are formed to facilitate the dissemination of information. This work deals with moving object modeling with databases.

A. TOPOLOGY MANAGEMENT AND HIGH NODE MOBILITY
The work in [67] proposes parking-area-assisted spiderweb routing protocol (PASRP), which is a spider-web model for the transmission of data in urban VANETs. It is built on the basis of parking areas using digital maps and geographic information. In VANETs, there is restricted WiFi coverage, and the topology changes frequently due to highly irregular node mobility. The work in [68] defines the link state routing (LSR) optimization problem. Many protocols use the concept of flooding to know the destination position, which increases network overhead. If we consider less populated regions, the hop count would be increased due to repeated usage of recovery strategies. Lastly, hop-greedy routing protocol that selects the path with the lowest number of in-between intersection nodes while considering good connectivity is proposed in [54].
Man-made and natural obstacles, irregular node mobility, and variable topology cause repeat link disconnections in communications that are not in a line-of-sight configuration. However, three-dimensional (3D) movement can be used by unmanned aerial vehicles (UAVs) to improve routing in VANETs. As it is correctly stated in [16], UAVs can increase the line-of-sight probability and the efficiency of the store-carry-forward mechanism. Routing protocols designed for VANETs mostly address the issues of high node mobility and variable topology, which can lead to data loss, frequent link disconnections, and high end-to-end delays. Geographybased routing protocols are the most efficient of the existing protocols due to their low network overhead and the fact that they can control vehicular environments very well. The work in [69] demonstrates how high mobility, data loss, and link disconnections must be carefully dealt with when working with geographic routing protocols. That work proposes Beaconless Traffic-Aware Geographical Routing Protocol (BTA-GRP) that considers inter-vehicle distances, traffic density, and the direction of vehicles when selecting the next forwarding node and the corresponding route. Many existing research works consider variable node mobility; however, the data bit rate, efficient multi-hop data transmission, and route length are important issues that need to be addressed efficiently in real VANET environments. In light of this, the work in [47] discusses challenges associated with VANET routing protocols on the basis of real-world experiments. Dynamic node mobility is a challenging issue in multi-hop data transmission because the connecting links are not stable and the packets are thus not always delivered to the destination. Due to the absence of global information, conventional broadcasting methods cannot accurately determine link dynamicity in VANETs. However, if global network information is available, link stability can be better scrutinized with the help of the software-defined vehicular network (SDVN). The authors of [70] analyzed link stability in route selection and formulated it as the least-cost-enabled flow problem. The time complexity is decreased by proposing an incremental packet allocation plan. Variable node mobility is the reason why video-on-demand (VOD) sessions cannot guarantee QoS. A fresh adaptive geographic routing plan is proposed in [43] to establish simplex VOD transfer in urban areas. That scheme involves finding numerous independent routes between the source and destination instead of just one. The number of routes is dependent on the size of the video file to be sent and the time interval over which the route is fixed.
Highly dynamic node mobility in VANETs is a serious issue, and thus a reliable and stable network connection is needed by existing secure payment solutions like drivethru Internet. Moreover, service providers may fail to deliver proper services in order to keep messages confidential. The work in [71] takes this issue into account and addresses highly dynamic node mobility in VANETs using an optimized mechanism. The work in [72] proposes a secure installment payment framework for drive-thru Internet in VANETs. VANETs are prone to errors, mainly because of high node mobility and frequent disconnections. Therefore, the work in [73] proposes an opportunistic routing (OR) protocol that can enhance the VANET's performance by prioritizing node transmission using a relay set. The source vehicle uses packet advancement, packet delivery possibility, and vehicle density to determine the vehicle for the next hop.
Communication link failure is very common in VANETs because of highly dynamic node mobility. The work in [74] proposes a mechanism in which a parallel route to the destination is established to reduce losses in real-time applications. The work in [75] proposes a dynamically evolving networking model that takes input of real vehicular traces to get results which can be more practical. The effects of node addition/deletion and loss of link nodes due to mobility are considered. A link compensation process and preferential attachment are also proposed to maintain an evolving network.

B. GEOGRAPHIC ROUTING
Recently, geographic routing protocols have been prioritized due to their scalability; however, some metrics, like vehicle mobility, bandwidth availability, and link quality, show trade-offs in them. The forwarding geographic routing strategy selects the nodes that are closest to the destination, which could be located near the border of the communication range, where the possibility of link breakage is aggravated. Continuously selecting these nodes without considering the available bandwidth results in more losses and high packet delays. The work in [76] addresses this by proposing a new fuzzy logic-based protocol. Successful VANET packet delivery is challenging because of high node mobility and frequent disconnections. Thus, the work in [77] proposes a spider geographic routing for urban VANETs. VANETs consider car nodes to be relays; however, VANET nodes are highly mobile, which makes it an ordeal to maintain persistent communication between neighboring nodes. The work in [78] proposes a protocol that helps to divide urban spaces into multiple zones, of which one with high geographical circumstances is selected to be the next zone, and relay node selection is applied to forward packets to the next zone.
In [34], geographical source routing is enhanced using an ant-based algorithm to discover the route with optimum network connectivity. Vehicles choose the path with the lowest total weight for the entire route. In [79], simulation platforms are used to compare position-based and geographical protocols with the integration of mobility models and realistic physical layer. Multipath routing and protocol tuning are also discussed to show their potential to increase the reliability of data transmissions. In [35], the authors propose a multilevel plot directed greedy opportunity routing protocol and investigate geographic routing protocols for multilevel VANET scenarios. Their results indicate that the possibility of connectivity is deteriorated in multilevel environments due to the degradation of the wireless transmission range. In VANETs, one of the widely studied routing protocols is geographic routing, mainly due to the prevalence of wireless devices with GPS. Selecting the next relay node is a challenging task, and the work in [80] thus proposes path aware geographic perimeter stateless routing (GPSR) following an analysis of the shortcomings of GPSR. Beacons are small HELLO messages that are periodically broadcast by every vehicle and are used by many VANET protocols to select the next neighbor. Since beacons have a smaller payload size than data messages do, they can easily be delivered over very weak links, which means protocols that rely on beacons can select routes with low signal strength. Most of the existing works incorporate a huge number of broadcasts for forwarding data packets, which results in wasted bandwidth. The authors of [81] propose beacon information independent routing (or geographic routing) that incorporates fewer broadcasts and intelligently uses information collected from vehicles during previous attempts to discover routes to the destination to forward data packets. Table 3 highlights the contributions of the works that consider geographic routing-based routing protocols in VANETs. Table 4 highlights the contributions of the works that consider position-based routing protocols in VANETs. Existing topology-based works are not quite efficient due to high vehicle mobility, which is why the authors of [14] worked on position-based routing (PBR) protocols. The consideration of PBR in city environments and connectivity-aware routing protocols is highlighted in [14]. That work presents a fog-based VANET that supports PBR to use parked vehicles when utilizing road intersections to select a path. Vehicles in parking areas are also involved in packet transmission. PBR has been a conventional routing protocol of choice in VANETs; however, it does not ensure that there is a routing path between the source and destination before the transmission or provide details about connection duration. The work in [82] proposes connectivity-aware routing based on infrastructure (iCAR-II), which helps in Internet service, mobile data offloading, and multihop vehicle applications. To construct a global network topology, iCAR-II predicts the connectivity of local networks and updates the location servers with real-time networking information. In 3D VANETs, dynamic vehicles and path loss gradually uplift the possibility of link interruptions that deteriorate packet reception probability. In [83], a packet reception reliable routing protocol is proposed to improve link reliability. The work attempts to find the links with the highest packet reception possibility to use them as relay links. In position-based routing protocols, VANETs use positions acquired from satellite systems, like GPS, Compass and Galileo, and global navigation satellites. However, they are prone to technical and environmental issues that affect routing performance. Thus, the work in [84] proposes a location prediction technique that uses a Kalman filter to improve routing performance by reducing location errors. According to [85], routing protocols based on position are considered to be more reasonable while comparing with other VANET routing protocols. That work proposes a movement-predictionbased joint routing and hierarchical location based service. The primary objective of the survey in [86] is to identify the most acceptable topology-based routing protocols in a dynamic mobile VANET environment. The work in [87] proposes a position-based routing protocol that is an improvised global state routing protocol. The protocol uses the ant colony algorithm in an enhanced manner such that a road is found with an optimal connection in the network.

D. SOFTWARE-DEFINED NETWORKING-BASED APPROACHES
Software-defined networking (SDN) provides a global network topology and can introduce programmability VOLUME 10, 2022 to VANETs. Thus, SDN-based architectures can manage highly complex and dynamic VANETs by decoupling the control plane from the data plane. An SDN-enabled connectivity-aware geographical routing protocol (SCGRP) is introduced in [88] for optimal data transmission. VANET nodes are highly dynamic, and consequently, there are frequent link disconnections that lead to high end-to-end delays. Thus, selecting a stable route is a key concern in VANETs. In [89], a routing protocol based on SDN is proposed for stable route discovery between the source and the destination. The SDN main controller (MC) finds the global network view, whereas numerous local controllers take charge of localized views of multiple zones under the MC. In [41], the authors propose a routing protocol called ROAMER that combines geographic and carry-and-forward protocols and uses RSUs to route messages in VANETs. ROAMER was later upgraded to shift routing operations that use a distributed SDN architecture as well as blockchain system in RSU networks in order to route packets securely and efficiently [41]. Vehicular communications have improved due to the introduction of SDN in VANETs. In [90], a concise summary of VANET structures is provided, with details of the SDN control. This survey also discusses the issues that must be addressed and the technologies that are required to design a robust protocol by maintaining latency, connectivity, and safety bounds in VANETs. The majority of existing VANET routing protocols take a distributed approach that leads to a local optimum and results in network congestion; SDN is thus quite popular because of its centralized control. In [38], a hierarchical geography-based SDVN routing protocol is proposed. First, the large region is divided into numerous small grids based on geographical location to find a set of grids that have good connectivity. Then, the path cost function is constructed with load balancing. Table 5 highlights the contributions of the works that consider SDN-based routing protocols in VANETs. Table 6 highlights the contributions of the works that consider cluster-based routing protocols in VANETs. In [91], every vehicle estimates a trustworthy low-overhead path using a cluster-based QoS algorithm known as the cluster-based adept cooperative algorithm. It helps maintain long-lived routes for which the most sustainable path is adaptively selected on the basis of signal strength, beacon, and node mobility. It reduces control message overhead and the routing table recalculation process. When transmitting data in VANETs, reliability cannot be guaranteed because of highly dynamic node mobility. Vehicle density is generally higher at junctions than on straight roads, and thus the packet loss rate tends to increase as the velocity of a vehicle increases in junctions. Vehicular routing assisted by clustering is introduced in [92] for network scalability. The moving clusters enhance the connection establishment time and message delivery ratio. To optimize network performance, VANET routing protocols used clustering and multipoint relaying (MPR) until it was discovered that MPR, which was initially designed for open areas, does not really benefit from road section configurations. The authors of [93] propose a chain branch leaf clustering scheme to enhance broadcast traffic and help with routing actions.

E. CLUSTER-BASED APPROACHES
VANETs are used for gathering and distributing safety-related information in order to reduce road accidents. It is in this context that the work in [94] proposes an adapted routing protocol for road safety in hybrid VANETs. An important contribution of that work is the introduction of a rider optimization algorithm to select neighbor nodes in order to ensure network lifetime as well as lossless connections. This protocol enhances road safety by forwarding valid data from the source to the destination. Attack detection using cluster-based data compilation is considered in [95], with the decisions related to message delivery executed at the cluster head. As it is discussed in [15], malicious nodes can be very dangerous and may send a substantial number of data packets to vehicles or RSUs. To address this issue, the authors of [95] propose a stream position performance analysis (SPPA) approach that considers distributed denial of service (DDOS) to monitor the position of the sender field station. SPPA tests the trustworthiness of data packets that are used for decision-making. VANETs have three possible architectures, namely, pure ad-hoc, hybrid, and wireless local area network. There are several real-life traffic scenarios that need to be tackled during the design of a routing protocol in VANETs. Sybil attacks are crucial attacks in VANETs, as it is discussed in [15], [95], [97], and thus need to be handled efficiently to improve security aspects in VANETs. The works [15], [97] clearly describe how Sybil attacks can affect VANET routing protocols and briefly mention prevention mechanisms. The work in [98] reviews communication models, like RSU, cluster-based, and V2V communications. According to its authors, the topic will forever be research-oriented and will be further enhanced in the future. Cluster-based approaches have the ability to use web signals to disseminate messages among vehicles. The probable distribution of V2V packet delays is analytically derived in [99] assuming bi-directional highways. The interesting thing in this work is that it discusses how related works miscalculated the maximum distance between two UAVs, which are aircraft that operate without the involvement of a human pilot on board. In [96], a cluster-based routing mechanism is proposed, that is based on SDN-enabled shortrange communication and cellular-V2X hybrid vehicular networks. In [96], a heuristic algorithm is designed, that involves a two-stage approach in which the aggregated data can be delivered via cellular networks or V2V transmissions. However, the centralized one-hop based approach can suffer from transmission bottleneck and network delay.

F. NAMED DATA NETWORK-BASED APPROACH
In VANETs, named data networks (NDNs) use contentcentered mechanisms to name content instead of the host, which is associated with a broadcast problem. The work in [100] addresses this by proposing a Bayesian receiver forwarding decision to resolve the broadcast problem. Vehicles receiving interest packets make forwarding decisions with the help of Bayesian decision theory pertaining to present network conditions, which are determined by interacting with neighbors. Sometimes drivers in vehicles want to acquire data whose location is unknown. NDN, which focuses on delivering mechanisms on the basis of the content of the message rather than the host address of the data, can be helpful in addressing this issue. In [101], an RSU assisted NDN mechanism is proposed that shows improved network connectivity in VANETs. Table 7 highlights the contributions of the works that consider NDN-based routing protocols in VANETs.

G. INTERSECTION-BASED DISTRIBUTION ROUTING
VANETs are affected by traffic lights, road intersections, and variable traffic conditions. The authors of [102] take these issues into account and propose an intersection-based distribution routing (IDR) concept. Vehicles are supposed to stop at intersections because of traffic lights, and thus an intersection vehicle fog (IVF) approach is proposed in which the stopped vehicles form a dynamic collection or vehicle fog. To avoid rerouting, the IVF approach adjusts the routing direction in real time on the basis of destination. When the IVF approach makes a routing decision, the IDR model uses ant colony optimization (ACO), which is discussed in [34], to identify the optimal routing path. In that case, the connectivity for multi-hop links between the intersections is based on traffic conditions. Table 8 highlights the contributions of the works that target the improvement of QoS in V2X routing protocols. In VANETs, multi-hop communication is an ordeal, particularly due to irregular topology and frequent disconnections. To resolve this issue, the work in [103] considers multi-constrained QoS and interestingly proposes an optimized cost model to solve routing problems like road congestion, collisions, quality of the network, QoS, cost, etc. An optimum route is found using a mean computing Jaya algorithm and it is observed that this mechanism performs better than existing ones in terms of cost and convergence analysis. The work in [104] addresses the issues of large VANET network size, irregular topology, and frequent disconnections by proposing adaptive QoS-based VANET routing, which adaptively chooses the intersections through which data packets reach their destinations. The proposed scheme improves QoS on the basis of delay, connection probability, and PDR. In that work, an ACO-based algorithm is proposed to reduce routing search time and the impact of network congestion. Video content streaming between vehicles is possible in VANETs, but because of frequent disconnections, dynamic node mobility, and spaced vehicle distribution, video data packets are transferred via multiple discovered routes in order to increase the chances of error-free packet delivery. The work in [105] discusses packet loss ratio minimization by optimally distributing video packets on multiple routes. In that work, video playback and reconstruction are achieved while also ensuring QoS.

H. IMPROVING QoS
In [106], the authors propose a spider-web-based transmission mechanism that combines electronic maps and geographic information systems (GISs) to deal with emergency data in VANETs. When transmitting emergency data, confirmed spiders and request spiders are used to obtain the transmission path between the source and destination in order to improve packet delivery and average transmission delay. Speed and QoS are the two main characteristics of VANETs. The work in [107] studies the proactive fisheye source VANET protocol on the basis of speed, altitude, QoS, etc., in real traffic scenarios.
To effectively predict link reliability between vehicles and design a routing protocol that reliably meets QoS requirements, the work in [112] analyzes the reasons behind communication link instability and discusses vehicle motion characteristics. In that work, a time-based link duration model is shown to help find the optimal end-to-end path during rapid topology changes. 5th Generation (5G)enabled VANETs help enable time-sensitive services with a designated QoS and better bandwidth usage. A review of various management protocols is presented in [113], along with issues pertaining to 5G networks. In that work, a brief description of existing VANET models, applications, classifications, benefits, and drawbacks are also presented.
The connected dominating set is employed in [114] to broadcast video content and design an efficient routing protocol in VANETs in order to provide high-quality video streaming at various rates in urban environments. Interleaving is applied to spread loss bursts and reduce the impact of loss distributions. Single-packet losses are also taken into consideration, and a store-carry-forward protocol is used to retransmit packets stored in local buffers. In [108], routing reliability is ensured by selecting trustworthy and compatible routes considering minimum additional communications. Two very important goals are tackled by [109]: latency reduction delivery ratio enhancement, and the minimization of network overhead. The authors of [109] propose a routing protocol for urban environments that shows promising success ratios, negligible communication overheads, and low delivery delays. In [115], two new AODV-based protocols are designed to broadcast and filter route reply (RREP) messages and improve network performance. AODV is studied in [116] for 5G VANET routing optimization.
When there are rapid topology variations and frequent loss of links between vehicles, it is challenging to achieve QoS while delivering emergency messages. This issue is addressed in [110] by designing QoS guaranteed routing in device-todevice (D2D) assisted cellular-V2X (C-V2X) communications for emergency message dissemination. The proposed mechanism follows a stable matching-based routing approach that minimizes end-to-end delay by using the best forwarder device. If the requested data is not found in the neighbor device, a 5G base station is requested through a pedestrian. Due to the 5G-based environment, the time complexity of the proposed mechanism is crucial and thus needs to be analyzed. In V2X communications, route discovery can be done using reactive or proactive approaches, which are compared in [111]. The results of that work highlight that destinationsequenced distance-vector routing (DSDV) must be preferred over AODV when V2X communications are in the participating candidates. However, DSDV cannot be used directly; it is feasible only as a starting point to develop other exclusive routing protocols.

I. HIGH VEHICULAR SPEED
In VANETs, maintaining end-to-end connections is indeed difficult, and thus the store-and-forward concept, which depends solely on node cooperation, is used for vehicular communications. The authors of [53] propose a hidden Markov model for predictive VANET routing that monitors vehicle movements to enhance transmission performance. The model predicts the future location of vehicles by considering where a vehicle regularly visits or who a vehicle regularly contacts, i.e., historical data is considered. To enhance the reliability and efficiency of path selection, a routing protocol is designed in [117] that considers speed, traffic loads, the number of nodes, and error rate. Channel parameters are also considered to increase reliability. VANET nodes are highly dynamic, and thus dynamic link stability is a challenging issue. V2V works without RSUs or gateways, but the possibility of successful link establishment gradually decreases as the number of red traffic lights, crossroads, sparsely distributed vehicles, and variations in vehicle speeds increases. RSUs are considered in [118] for ensuring route establishment and broadcast control route requests in sparse VANETs with crossed road layouts. RSUs also help with route discovery in gateway zones when the next hop for relaying route request packets is absent. Since there are chances of link failure, handoff overhead, and link re-establishment because RSUs are static and vehicles are dynamic. Thus, relying solely on RSU forwarding is not effective. To minimize the loss of buffered packets along the route and maintain a stable link, a local route repair mechanism that incorporates RSUs is designed in [118].
The work in [119] proposes a scheme in which short video recordings can be sent to any city infrastructure through VANETs, which helps provide quick emergency services. In this work, a multimedia multimetric map-aware routing protocol is proposed to send video complaint using VANETs in smart cities. Due to technological developments, the most commonly used single-hop multicast communications are gradually being overpowered by multi-hop unicast communications in VANETs. The work in [120] addresses this by using a predicted value for calculating route lifetime, and the prediction scheme is implemented using AODV routing. Self-organization in wireless mutil-hop networks leads to inefficient and unreliable decision schemes for data routing. To address this, the authors of [121] propose the Manhattan mobility model, which integrates RSUs into both wired and wireless models for transmitting data and optimizing route decisions. Irregular vehicle mobility is analyzed using motion information, and a greedy algorithm that incorporates wired RSUs is used for wireless vehicle communication. The work in [122] proposes an Harris hawks optimization-based algorithm to select two rather than multiple hops and thereby ensure route stability and a high data transmission success rate. Table 9 highlights the contributions of the works that consider high vehicular speed.

J. HYBRID PROTOCOL
The authors of [123] propose direction-based cache agent location-aided routing, which is a hybrid routing protocol. In that work, geocast and position-based routing are combined, which significantly improves the routing performance of VANETs. Data delivery performance in V2X communications is addressed in [124] by developing a mathematical framework that considers throughput and latency. Moreover, an optimization problem is formulated for maximizing throughput and latency for hybrid V2I-V2V networks. An algorithm is proposed in [124] that uses a multi-hop routing approach to select the optimized path to maximize the weighted sum of throughput and latency. The solution provided in [124] achieves low latency and high throughput; however, the frequent change of topology in V2X is not highlighted.

K. MAINTAINING SECURITY AND PRIVACY
As per [125], VANETs protect nodes from outside observers by using unidentified routing protocols; however, existing protocols depend on unnecessary traffic or encryption, which makes them highly costly and unable to provide complete sender, receiver, route, and data privacy. The work in [125] proposes a cost-effective and efficient location-based routing protocol to address this. Multilateral security is guaranteed by using public key infrastructures (PKIs) and digital signatures. In [126], DDOS attacks are prominently discussed and a secured minimum delay routing protocol is proposed. In [127], a novel anonymous onion-based routing protocol introduces dynamic relay groups, which maintains privacy in VANETs. In this work, groups are formed of the vehicles around specific locations, and the vehicles are used for cryptographic onion relay. This protocol keeps the source, destination, route, and real vehicle identity anonymous. To ensure privacy in VANETs, the authors of [39] propose an efficient routing protocol that exploits handshakeless communications, a competent collision resolution mechanism, and acknowledgment (ACK) decoupling. Real-world experiments, computer simulations, and theoretical modeling are used to analyze the mechanism. Security and a global consensus are very much needed to cope with the increased use of cryptocurrencies for real-time payments. Payment channel networks present a promising solution to cryptocurrencies. In that work, transactions are offloaded from blockchain and then directly handled using a minimum-involvement-required payment channel. Despite their importance for successfully dealing with path-based transactions, incentive mechanisms pertaining to payment service providers (PSPs) have not been efficiently studied. PayGo, a routing protocol that is proposed in [40], does not find the best path in terms of budget, but derives an optimal incentive for the PSP using contact theory. The PSP is induced by PayGo to make a deal with counterparties that guarantee throughput and penalize payment latency.
VANETs are vulnerable to several types of malicious attacks, such as Sybil attacks [15], which are able to disrupt the normal functioning of most VANET protocols and are not easily detectable. In Sybil attacks, a legitimate vehicle with ill intentions (Sybil vehicle) gets into the network and mischievously launches virtual camouflaged vehicles to affect the VANET routing mechanism. In [15], the authors explore position-based routing protocols that are attacked by Sybil vehicles and consequently propose greedy perimeter coordinator routing (GPCR). Table 10 highlights the contributions of the works that consider security and privacy in V2X routing protocols.

L. FAULT DETECTION
VANET routing protocols must receive valid information from neighboring vehicles as beacons; if an OBU is soft faulty, it will reduce system performance. Thus, a fresh selfsoft fault observation routing protocol is proposed in [128] to send data to the destination without delays and detect soft faulty vehicles. A vehicle is assumed to know whether or not decisions made by neighboring vehicles are faulty and whether or not neighboring vehicles should be excluded from the routing process. Data forwarding is done via fault-free vehicles on almost completely connected paths, and the best path at every junction is selected by fetching the path values through which data is forwarded. False alarm and fault detection rates are used to test the performance of the fault detection method.

M. PROBABILISTIC ROUTING
An efficient VANET routing protocol shows good performance in terms of packet loss, end-to-end delays, etc. In the interest of achieving these promising results, the work in [37] proposes a probabilistic multimetric routing protocol (ProMRP) that is especially designed for VANETs. It uses probability to estimate the percentage of packets successfully delivered to the destination on the basis of four metrics, namely, distance to the destination, position of a node, available bandwidth, and density of the nodes. An improved version of ProMRP is also proposed that helps to estimate the node's current position before sending the data instead of utilizing the last known position from the previous beacon message.
Considering the intersection-based routing mechanism, a microscopic scheme is proposed in [129], where intersecting nodes help maintain and update records by collecting data from passing vehicles. Vehicles' current positions are recorded to compute the connectivity probability and estimate the delivery delay. Both microscopic and individual data are used for enhancing the accuracy of the results. To decrease the complexity associated with position estimation, vehicles are arranged in a queue. In this context, the network overhead is also low because there is no generation of dedicated control packets in the proposed scheme. Table 11 highlights the contributions of the works that consider probabilistic routing in VANETs.

N. UAV IN VANETs
With the introduction of flying VANETs (FANETs), costeffective UAVs can be used to accomplish complex missions in the sky [130]. In FANETs, the reliability of information exchange is still in its initial stages. UAVs stand apart from traditional networks as they are deployed in the sky and the mobility model adopted is determined by the nature of the mission. A comprehensive study that critically compares FANETs and considers architecture, constraints, mobility models, simulation tools, routing techniques, and future challenges is presented in [130].

O. HANDLING OF CONGESTION AND SHADOWING EFFECTS
One of the most crucial challenges in VANETs is addressing the end-to-end delay problem. This problem can be exacerbated by congestion in the network. Greedy-based techniques are generally used in this context, but they result in the local maximum problem and data congestion, which lead to even higher end-to-end delays. The work in [131] proposes a mechanism that helps to evaluate the end-to-end delay of the entire path before messages are sent. To this end, stable backbones are built on road segments and connected via bridge nodes at intersections. Road segments are assigned weights by the bridge nodes on the basis of delay and connectivity-related information. The routes with the lowest average weights are then selected to forward data packets.
The time delay and the quality of the communication link must be considered when designing VANET routing mechanisms. Yet, many communication protocol theories ignore obstacles, such as buildings and trees, that are bound to be present in an urban environment. These obstacles can cause signal fading or completely block signals. The vehicles parked on roadsides can be considered relays to reduce shadowing effects. The work in [36] proposes a parked vehicle relay routing algorithm to improve vehicular communications in VANET routing.

P. NETWORK OVERHEAD
The work in [132] compares proactive, reactive, and hybrid routing protocols on the basis of behavior and performance in urban environments. It also studies the performance of User Datagram Protocol (UDP) and Transmission Control Protocol (TCP) with variable car density. It was observed that the proactive OLSR protocol has the least network overhead, the reactive AODV protocol performs best in terms of transfer rate at the TCP level, and the hybrid zone routing protocol (ZRP) has the lowest latency and increases the delay rate. A novel routing protocol is proposed in [133] to reduce the network overhead generated due to the periodic exchange of beacon messages when routing. To achieve this, twodirectional antennas are used to decrease the size, rather than the number, of beacons. Greedy forwarding is modified at road segments to calculate the distance between the packet sender and the direct neighbors instead of using their present geographical locations. Table 12 highlights the contributions of the works that consider network overhead in V2X routing.

Q. NETWORK CONNECTIVITY
Guaranteeing connectivity in VANETs is challenging because of topological aspects that depend on the flow of vehicles, which is dynamic. The utilization of fixed infrastructures could be a promising option, although the costs involved are prohibitive. In urban cities, metropolitan buses can be used as the backbone (metropolitan VANETs). The work in [134] addresses this by proposing a live mobile ad-hoc backbone that can increase network connectivity and provide an effective infrastructure for VANETs. Statistical analysis shows that the proposed backbone can ensure good connectivity in the network. A new routing protocol is also designed in that work to have the proposed backbone exploit intrinsic connectivity. This protocol can improve network performance by giving priority to nodes during the route discovery process.
To choose the next forwarding street, routing protocols generally consider the density or number of vehicles. However, unequal spatial vehicle distributions significantly change the connectivity for multi-level or single-level streets. The authors of [135] propose a connectivity-aware routing protocol based on spatial distribution that helps to divide streets into multiple segments and gather vehicle numbers and positions in every segment in order to monitor the connectivity quality of the segments. Table 13 presents the contributions of the works that consider network connectivity in V2X routing.

R. 3D VANET SCENARIOS
Traditional VANET routing protocols assume planar scenarios but are actually applied in 3D. Many routing issues for 3D VANET scenarios are addressed in [136]. More specifically, that work analyzes the features of 3D urban road networks and develops 3D routing protocols. It also proposes an improved two-part 3D oriented routing protocol, the first part of which draws attention to issues in simple 3D city structures and the second part of which addresses complex 3D city structures.

IV. LEARNING-BASED V2X ROUTING PROTOCOLS
In this section, we discuss the existing works that consider learning-based V2X routing protocols and their impacts on network performance. Tables 22 and 23 present the works and highlight their strengths and weaknesses. Fig. 5 shows the number of existing works that consider learning-based V2X routing, where machine learning has been employed since 2009.

A. IRREGULAR VEHICLE DENSITY
Vehicle density and traffic are irregularly distributed in metropolitan cities. Moreover, man-made constructions can block vehicular traffic, which makes designing protocols for VANETs challenging. In [48], a Q-learning-based routing protocol, known as Q-learning-based RSU-aided trafficaware routing (QTAR), is introduced for urban VANETs. VANETs connect highly mobile vehicles with dynamic inter-vehicle spacing and fluctuating vehicle density on roads. These characteristics affect VANET performance; however, if the mobility of vehicles can be predicted, continuous service becomes possible. In [6], the authors propose a central routing scheme that can predict mobility using a software network controller that is powered by artificial intelligence. Moreover, it is difficult to provide trustworthy multi-hop communications because of vehicle movement, a shortage of wireless channels, and the lossy attributes of wireless channels. The work in [137] thus investigates a VANET routing protocol that employs the Q-learning algorithm with a fuzzy constraint that is portable and can learn optimal routes on the basis of AODV.
Due to unstable link connections, reliable communication between vehicular nodes is still not achievable. The work in [138] proposes a reinforcement learning-based routing protocol for clustered electric vehicle (EV)-VANETs, that uses a K-harmonic means algorithm to upgrade the cluster structure's stability. Distributed ITSs require that information be stored independent of fixed infrastructure like RSUs. Therefore, [46] proposes a protocol to collect data VOLUME 10, 2022   in VANETs by considering vehicle velocity and bandwidth efficiency. The work in [139] addresses VANET data aggregation and proposes an adaptive forwarding delay control system that is based on a distributed learning. The proposed scheme dynamically changes the speed of the nearby sensors' reports to improve their chances of connecting with each other, which leads to efficient aggregation of data. The vehicle variable ratio analysis in [52] focuses on analyzing V2X communication capabilities and addresses how well the predictive rerouting algorithm performs when many vehicles (a high percentage of them) do not get to communicate due to traffic density. The existing V2X communication channel does not include crucial properties like suspicious message transfers and latencies. Table 17 highlights the contributions of the works that consider irregular vehicle density in V2X routing.

B. UNCERTAIN NODE MOBILITY
Aside from vehicle mobility, we see that route length, data transmission rate, and data transmission efficiency in multi-hop environments also play a vital role in performance of VANETs. The work in [47] discusses issues in VANET routing and proposes a routing protocol that has the ability to learn the best suited transmission parameters from interactions with the environment. The proposed approach specifically takes into account the data transmission rate, route length, and vehicle movement. In [45], the bus rapid transit (BRT) system is proposed, whose buses have regular mobility patterns. It has the capability to tackle uncontrolled node mobility and select the most suitable routing paths to deliver data. It uses knowledge of the predictable and periodic movements of buses to learn the temporal distance to transmit every piece of data to RSUs using a dedicated busbased foundation. The authors of [140] propose a superior form of BRT and use simulations to show the performance improvement achieved.
Machine learning (ML) approaches are quite beneficial, and the work in [141] thus proposes a trustworthy best-path prediction mechanism that has a fog node-based architecture and a longer compatibility time. The proposal achieves a packet delivery ratio of 16%, 99% accuracy, and longer connectivity than existing comparable solutions. The work in [142] considers the City section, INVENT, and Manhattan mobility models to learn about vehicle mobility and analyze the impact of mobility on VANET performance. The performance of these models is recorded and correlated using suitable mobility metrics, as they play a very important role in network design. Furthermore, reinforcement learning is used to meet the requirements of VANETs. The authors of [51] consider beam alignment difficulties and routing stability issues in millimeter Wave (mmWave)-based V2X transmissions, and accordingly propose a scheme for detecting the position of vehicles in 3D for beam selection or alignment purposes. A group-based routing approach is followed to choose a secure path that can achieve trustworthy data transmissions. Table 18 highlights the contributions of the works that consider uncertain node mobility in V2X routing.

C. UAV IN VANETs
In [143], an adaptive UAV-assisted VANET routing protocol is proposed that is based on geographic routing and a  Q-learning mechanism. In that work, the global route in the aerial routing module is calculated using the depth-first search (DFS) and fuzzy-logic algorithms and considers UAVcollected information, such as the global road condition, that is forwarded to vehicles on the ground.

D. ENERGY EFFICIENCY AND SECURITY
The work in [144] considers energy efficiency, reliability, and connectivity in data transmission and designs a new routing protocol that combines multicopy, connection lifetime prediction, and a forwarding scheme with link state information. Security issues in VANETs are considered in several existing works since ITS services are adversely affected by malicious vehicles. A secure intelligent routing protocol is proposed in [145] to send data via the quickest path through authenticated vehicles. The authors show system performance is improved when the selection of authenticated vehicles using the quickest path keeps the network free from malevolent attacks. The proposed mechanism considers safe message transmission, delay, link connectivity, and vehicle position models. In [146], an approach is proposed that uses a hash function and a digital signature to ensure security in implementations of VANET routing protocols. Table 19 highlights the contributions of the works that consider energy efficiency and security in V2X routing.

E. SDN-BASED APPROACHES, TOPOLOGY, AND TRAFFIC MANAGEMENT
Dynamic topology, variable bandwidth, and network changes make it harder for VANETs to support stable data transmission at any moment in time. Hybrid routing algorithms VOLUME 10, 2022  have been used to tackle this issue; however, they are known to cause the blind path problem. The authors of [147] use a reinforcement learning-based hybrid routing algorithm to track the available paths and their status online. VANETs can manage traffic in real time and send emergency messages to base stations, and are thus an emerging part of wireless networks. All vehicles behave as sensor nodes to collect information from their surroundings and send it to base stations. In [148], artificial neural networks (ANNs) are used to test VANETs' packet loss and throughput performance. The M-estimators performance function is used instead of the conventional mean squared error performance function. The work in [44] proposes a framework known as road data enrichment that collects data from different sources to elevate ITS services, including traffic event detection and vehicle routing. The Twitter MAPS of location-based social media data is presented for route services, while Twitter incidents (T-incidents) and low-budget learning-based road detection incidents are built with the help of variable data fusion for event services.
In [149], the authors discuss using the metabolic processes involved in cell survival for the first time in VANETs. A cognitive routing optimization agent is proposed for autonomous vehicles in urban cities. The agent helps enhance traffic flow over the network to minimize road clogging and maximize the number of automatic vehicles that reach their destination. Applications like traffic management and transport efficiency are important in VANETs. However, VANETs are sensitive to malicious nodes. SDN has recently been proven able to manage VANETs in a dynamic way and address this issue. In [150], a trust-based software-defined VANET architecture is proposed in which a central SDN controller serves as the learning agent to achieve an acceptable communication policy using the Q-learning approach. Packet routing is a challenging issue in VANETs due to the network topology and traffic management. In [151], the authors explore deep learning to propose an algorithm and a routing architecture that makes packet forwarding decisions on the basis of the current surroundings' conditions. Table 20 highlights the contributions of the works that consider SDNbased approaches, topology, and traffic management in V2X routing.

F. NETWORK MANAGEMENT AND CONGESTION HANDLING
The reinforcement learning-based adaptive routing protocol (ARPRL) is proposed in [152], where it learns using distributed Q-learning and proactively gets fresh network link status information from periodic HELLO packets. Simulation results show that ARPRL achieves better performance than existing protocols with respect to end-to-end delay, route path hops, and packet delivery ratio while keeping network overhead in a permissible range. A congestion game-based routing algorithm is proposed in [153] to resolve the issue of network congestion in VANETs and provide Internet access paths. The routing algorithms proposed in that work are more feasible and effective than baseline mechanisms for communicating in VANETs. Furthermore, efficient message delivery is a challenge in VANETs. The work in [154] addresses this by introducing a Q-learning-based protocol for routing that considers microscopic as well as macroscopic issues when making a routing decision.
The authors of [155] present an ML-based approach that can successfully detect local and global invasions in VANETs. The model is claimed to be more resilient to changes in the environment when cluster heads (CHs) and RSUs are considered. To make sure that road conditions are reported properly, the work in [156] introduces data forwarding using location information from roadside sensors and K-shortest course routing. Q-learning is combined with this routing approach to achieve sensor convergence faster than is possible with techniques that find the shortest path using simple Q-learning. Multi-hop routing is combined with Huffman and arithmetic coding to compress data packet payloads. The work in [157] addresses communication proficiency and routing in VANETs by using a support vector machine (SVM)-based ML scheme to study and process data collected from cars.
In [158], it is noted that learning automata and the leapfrog method improve multipath routing with particle swarm optimization (PSO). The work in [159] proposes an efficient charging details transmission strategy for spatiotemporally coordinated V2V charging services. Spatial systems (areas that are measured as GPS point systems, like malls, market areas, and housing areas), in which traffic density is generally high, are considered in [160]. The system proposed in that work predicts a routing strategy to control traffic congestion using VANETs and a GIS architecture. The routing protocol proposed in [161] is based on clustering and the K-means algorithm. In that work, the proposed algorithm is applied to a vehicular environment on a highway to improve performance compared to baselines. Moreover, traffic congestion is reduced, throughput is significantly increased, the time delay is decreased, and mobility is enhanced. In [162], a routing scheme is proposed that is based on Penicillium reproduction and is able to dynamically select routing strategies using edge server computational power. To this end, the proposed mechanism learns the network traffic pattern in VANETs.
The V2X routing protocol proposed in [50] is intersectionbased and takes into consideration the learning mechanism of historical traffic flows, network status monitored in realtime, and Q-learning. The proposed mechanism is efficient at detecting network loads and making timely routing decision adjustments so that network congestion can ultimately be prevented. In the proposed approach, packets are reliably transmitted with reduced latency and communication overhead. However, the network resources are not optimally utilized, and an increase in hop count is observed due to the addition of RSUs. The hidden Markov model is used in [53] to propose a predictive routing-based protocol, that predicts vehicle movement patterns based on past route traces such that transmission performance can be increased. The proposed model helps in V2I and V2V communications by improving network performance, the delivery ratio, traffic overhead, buffer occupancy, and delay. However, real-world implementation is required to verify the protocol's performance in real VANET scenarios. The work in [163] addresses the issue of inefficient data dissemination in V2X communications, which is caused by a shortage of bandwidth in very dense networks. The clustering system proposed in that work has two levels -in level 1, the CHs are selected using the fuzzy logic algorithm, and in level 2, the CHs are selected by applying a Q-learning mechanism such that the gateway selection iteration count is reduced. However, in a very dense environment, multiple actions and agents are present, which leads to increased iterations being discovered in a gateway. Furthermore, temporal connectivity calculations are very time-intensive. Table 21 highlights the contributions of the works that consider network management and congestion in V2X routing.

V. OVERALL OBSERVATIONS ON EXISTING NON-LEARNING-AND LEARNING-BASED V2X ROUTING PROTOCOLS
In this section, we present our overall observations on the existing works that consider non-learning-and learningbased V2X routing approaches. Fig. 6 shows the number of works published each year (since 2009) that consider non-learning-and learning-based routing in V2X communications. From Fig. 6, it is noted that the number of works that consider learning-based routing approaches is quite a bit lower than the number that consider non-learning-based routing approaches. Therefore, there is a huge opportunity related to the application of ML in designing intelligent V2X routing mechanisms, where efficient dynamic routing can be achieved by incorporating a learning agent in VANET systems. Furthermore, Table 24 highlights the various primary directions that are considered in existing works when designing routing mechanisms in V2X communications. From Table 24, it is noted that several directions, such as geographic and position-based routing, intersection-based distribution, fault detection, probability-based approaches, 3D scenarios, and network connectivity, need to be explored when designing an efficient and intelligent V2X routing mechanism.

VI. CHALLENGES AND FUTURE RESEARCH DIRECTIONS
In this section, we discuss the challenges of and future research directions for designing V2X routing protocols. We first present the challenges, and then highlight possible research directions that could lead to a new era of smart V2X routing mechanisms.

A. CHALLENGES
Based on the discussion of existing V2X routing techniques and their limitations, it can be noted that there are certain areas that need to be further developed to build user trust [151], [156], [165] in VANETs. Several factors that lead to challenges when designing effective V2X routing protocols are: Findings: Non-learning-based V2X routing protocols. Several existing works consider non-learningbased V2X routing approaches. Works specifically focus on different directions, such as high node mobility, geographic and position-based routing, SDNbased approaches, clustering, named data networkbased approaches, intersection-based distribution routing, QoS, high vehicular speed, security and privacy, probabilistic routing, and congestion. Existing routing mechanisms improve V2X routing performance on the basis of the aforementioned directions. However, since the proposed approaches are primarily threshold-or heuristic-based, these routing schemes' adaptability in different network scenarios, where vehicular network conditions, such as congestion and vehicular movement patterns, can change frequently remains to be determined, and efficient dynamic network management is required in routing.

Findings: Learning-based V2X routing protocols.
When it comes to learning-based routing protocols, several existing works consider different ML mechanisms to automate the routing process in V2X communications. Learning-based routing approaches handle irregular vehicle density, uncertain node mobility, energy efficiency, security and privacy, topology, congestion, etc. The use of ML algorithms makes the proposed routing mechanisms intelligent and able to dynamically perform routing according to current network conditions and vehicle requirements. Consequently, existing learning-based routing mechanisms can significantly improve overall network performance, for example, by decreasing the transmission delay or increasing the throughput. However, existing learning-based routing approaches need to thoroughly analyze learning overhead and convergence, which are key issues in ML algorithms. In addition, a huge volume of real-life data is required to train supervised learning-based routing schemes. Moreover, the adaptability of ML-based routing models in real-life implementations needs to be explored.
• Obstacles: Man-made obstacles (like buildings and other constructions) and natural obstacles (like trees) reduce signal strength, packet reception rate, route discovery, etc.
• Lack of centralization: It is very challenging to design an effective central node that can practically cover a wide range of communications and monitor outgoing and incoming messages.
• Limited resource availability: Resources such as power and bandwidth are limited.
• Irregular topology: Due to high mobility and changes in topology, there are frequent communication outages in the network, which significantly affect the routing mechanism.
• Connectivity issues: Due to the small network range of VANETs, communication between the nodes is weak. Consequently, routing requires that the distance between two vehicles be short.
• Lack of authentication: Since messages are broadcast by vehicular nodes, authentication is required to ensure that malicious nodes cannot become part of the network and tamper with crucial routing-related information. This information could be used by malicious nodes to forward the messages to themselves instead of transmitting them to the actual destinations.
• Packet delivery: A packet must be delivered in the shortest period of time possible with a minimum packet loss rate such that trust can be ensured in times of crisis.
• Control overhead: VANET architectures need to handle high mobility, which leads to frequent topological changes in the network. As a result, the control overhead and the frequency of communication link failures are increased. Therefore, such overhead and losses must be minimized during V2X routing protocol design.
• Error analysis: It is indeed challenging but essential to perform error analysis on a huge real-life dataset to predict the actual performance of a routing mechanism. The minimization of errors in routing helps improve the selection of the correct path for packet forwarding.
• Routing model: Channel noise, physical layer attributes, and the signal propagation model must be included in the design of a packet routing technique. In addition, malevolent nodes should be classified by their attack type [156], [166] such that routing decisions can be made in advance to avoid attacks that can affect the security of the information transmitted through a routing path.
• Traffic control: Long-or short-term traffic prediction is vital in ITSs to consider complex and multi-road intersections. This prediction can reduce the packet loss rate by using a virtual network and efficiently selecting the routing path. Different vehicles have different sizes and speed limits. Therefore, the type of vehicle must be considered when analyzing traffic.
• Information disclosure principle: In [167], PKI-based or identity (ID)-based cryptography is used; however, this work does not consider the least information disclosure principle wherein a driver gets to know about its neighbors while keeping their real identity and address under cover.  • Vulnerability to worm infections: In VANETs, vehicles can transport, sense, and process information. An attacker can affect these normal vehicle functionalities in VANETs [168]. Therefore, vehicles are more vulnerable to worm infections than conventional hosts.
• Management of SDN-enabled VANETs: A VANET that is accompanied by SDN controllers has more control over vehicles with the use of software-based management. However, SDN architectures require a high number of interactions among nodes and efficient control mechanisms. Therefore, the routing mechanism needs to be effectively controlled and managed to perform routing efficiently in SDN-based VANETs.
• Video streaming quality: VANETs are characterized by variable node mobility, which makes continuous video streaming not yet possible. Stable network protocols must be developed to enhance the quality of video files while routing them in the network.
• OceanNet VANET architecture: VANET used in marine environments has been explored; however, systems in such scenarios face several challenges when it comes to maintaining QoS considering sparsely connected nodes.

B. FUTURE RESEARCH DIRECTIONS
In V2X communications, the routing mechanism can be greatly improved. In the last decade, several advancements related to the development of routing protocols in VANETs have been made, which also revealed the associated challenges. For instance, on urban roads, the concept of 'automated vehicles' demands a smart routing approach. Pre-known natural calamities under water can be prevented by VANETs [165], and in this case, we also need effective V2X routing techniques. Possible future research directions for V2X routing are listed below.
• Driver safety: There are several areas that require ML-based intervention to ensure the safety of vehicles and drivers of automated or semi-automated vehicles. These areas include driver drowsiness, weather prognosis, accident prediction and prevention, driving habits, obstacle detection, lane detection, and road condition prediction.
• Routing mechanism scalability: The scalability and calculative complexity of routing mechanisms must be significantly improved to cope with different networking scenarios in VANETs.
• Feedback mechanism: High priority and emergency messages must be routed with a feedback mechanism to ensure successful transmission of messages.
• Traffic maintenance: When designing V2X routing protocols, traffic signals and GPS data must be provided in the virtual wireless network to help predict traffic flow and congestion, and increase the efficiency of the routing approach.
• Trust-based techniques: Robust security systems can be incorporated in the design of V2X routing mechanisms to ensure authenticated message delivery and increase end users' confidence in the system.
• Vehicular cloud computing (VCC): VCC has been proven to be competent enough to provide optimum solutions to protect networks against various threats and increase the efficiency of traffic control and management systems. Therefore, VCC-based routing can enrich V2X routing mechanisms.
• Time restriction of VANET nodes: Broadcasting messages in VANETs requires an efficient and reliable system in which emergency message alerts reach their destination at the right time and not ahead of or behind schedule. Thus, the time delay of message transmission can be minimized by using optimized routing mechanisms.
• VANET for crossroads: Existing routing protocols provide information about the shortest or least crowded path in VANETs; however, in the case of crossroads, vehicles are more prone to accidents. Therefore, a protocol should be developed to quickly and accurately warn drivers of the potential crashes that may happen due to vehicles that are not within their eyesight or people that are crossing the road.
• Parking management: Drivers generally learn about parking spot vacancies upon arriving at a parking area, which is somewhat of a hassle. By using a methodical and intelligent V2X routing protocol, drivers could be alerted in advance and choose another route based on parking availability.
• Automated toll collection: VANETs can be used to automate toll collection, which is a major reason behind heavy traffic on bridges and highways. Such an automated approach could be incorporated in V2X routing mechanisms as a part of the routing process, which would help accelerate the routing decisions at a breakpoint. However, crucial information like people's address and contact information should not be made available to any unauthorized parties.
• Improvement of VANET architecture: In designing V2X routing protocols, SDN can be combined with the internet of things (IoT). SDN helps centralize and automate network management control. The IoT provides enhanced data collection and assists in managing widely distributed network connections.
• Blockchain in VANETs: VANETs can be integrated with blockchain, which is a very recent research topic.
To ensure security during routing, blockchains can be used to impose a security mechanism in a credit-based approach in which vehicles collectively adhere to a routing decision to communicate among themselves and validate new blocks in blockchains.
• 5G and 6G in VANETs: 5G and 6G can be used in VANETs to improve network functionality and peformance by providing more speed and intelligence. Therefore, a V2X routing mechanism that incorporates features of 5G and 6G needs to be designed to exploit the strengths and advantages of these technologies.
• OceanNet VANET architecture: V2X networks can benefit from ML to predict link quality and provide location-based services.
• Consideration of urban scenarios: Many cloud-based architectures in VANETs have not been tested in complex real-life urban scenarios. Therefore, the routing protocols proposed in existing works need to be implemented in real life to analyze their effectiveness in practical VANET scenarios.

VII. CONCLUSION
VANETs are a crucial component of ITSs that have the potential to handle traffic management, accident prevention, and emergency services. In this context, routing in V2X communications plays a key role in the reliable and timely delivery of messages. Automated cars, real-time news reports, weather forecasts, file sharing over the Internet, and other entertainment services can be promoted via VANETs. However, due to the high mobility of vehicular nodes, unsteady network connectivity, frequent changes in network topology, and unbounded network size, routing becomes a challenging issue in V2X. Consequently, several researchers have focused on increasing the coordination efficiency of vehicular nodes by designing appropriate routing protocols. Existing routing mechanisms use non-learning-or learning-based approaches. This survey presents a thorough study of the existing routing protocols by categorizing them as one of the two aforesaid approaches and summarizing their strengths and weaknesses. This paper highlights not only the impacts of these routing schemes on the overall performance of VANETs, but also the limitations of existing research works and future research directions for designing and developing intelligent V2X routing mechanisms. As a result, next-generation VANET technologies will be enriched with more efficient coordination among vehicular nodes.