ECMS: Energy-Efficient Collaborative Multi-UAV Surveillance System for Inaccessible Regions

The evolution and popular adaptation of drone technology in diverse applications has necessitated advancement of UAV communication framework. UAVs inherently support features like mobility, flexibility, adaptive altitude, which make them a preferable option for dynamic surveillance of remote locations. Multiple UAVs can cooperatively work to accomplish surveillance missions more efficiently. However, the intermittent network connectivity and the limited onboard energy storage impose a great challenge on UAV-assisted remote surveillance. This paper presents an Energy-efficient Collaborative Multi-UAV Surveillance (ECMS) system for surveillance of inaccessible regions. The system employs an optimal Multi-UAV Collaborative Monocular Vision (MCMV) topology to facilitate the surveillance with zero blind spot using minimum number of drones. We also propose an application-aware Multi-Path Weighted Load-balancing (MWL) routing protocol for handling congestion by distributing traffic among all available resources in UAV network and adaptively selecting the of source datarate (i.e. switching video resolution). The simulation results demonstrate that the proposed surveillance system achieves coverage with lesser number of UAVs compared to the existing systems. It also achieves higher throughput, higher packet-delivery ratio, higher residual energy of UAVs, and lower end-to-end delay.

The key contributions of this paper are listed below: 81 i) We propose an ECMS system to perform surveillance 82 in remote inaccessible area A. Therein, we develop 83 an optimal MCMV topology to decide the minimum 84 number of UAVs, at optimal flying height λ * , required 85 to conduct surveillance without any blind spot. 86 ii) We also propose an an efficient application-aware 87 MWL routing strategy that reduces the end-to-end 88 delay as well as the energy consumption, and improves 89 the throughput along with the energy efficiency in 90 UAV-Net based surveillance system by regulating con-91 gestion in the network. 92 iii) We perform in-depth simulations to evaluate the ECMS 93 system with respect to (w.r.t) its topology and routing 94 method. The simulation results show that the proposed 95 MCMV topology and the MWL routing perform effi-96 ciently when compared to alternative schemes.

97
The remaining paper is structured as follows. Section II 98 reviews the related works on the subject. Section III gives 99 an overview of ECMS system model. Section IV discusses 100 the MCMV optimal network topology solution for remote 101 surveillance. Section V highlights the MWL routing protocol 102 for efficient transmission of surveillance video from UAVs 103 to Ground Control station. Section VI presents the test-bed 104 details and performance analysis of the ECMS system and 105 Section VII concludes the paper.

107
Over the past few years, a significant amount of research 108 has been done on the UAV network for surveillance appli-109 cations. In [11], the authors propose a methodology for video 110 surveillance over 4G LTE network. The system makes use of 111 a multi-UAV network that performs video surveillance in the 112 area of interest by using the existing communication infras-113 tructure.
[6] develops a Wi-Fi-based emergency network to 114 conduct on-site surveillance and transmit the information to 115 the relief center for better rescue planning. Here, the coverage 116 is provided by creating a Wi-Fi zone over the area of interest. 117 In [12], the coverage issue of UAV-based surveillance in a 118 complex urban environment is addressed. The ideal number 119 of view points in the air to completely cover the target surface 120 is determined by using a polynomial-time greedy set cover 121 approach. All of the network models that have been discussed 122 so far require the communication infrastructure, thus cannot 123 be applied to monitor remote areas that lack communication 124 infrastructure and are completely inaccessible. Therefore, 125 in this work, we mainly focus on developing a network topol-126 ogy that can support the communication and surveillance over 127 inaccessible region. 128 There have been a number of solutions to determine the 129 UAV network topology to ensure coverage and connectivity 130 in remote location.
[13] provides a mathematical model to 131 find the optimal UAV position for maximum coverage. The 132 authors consider the average path loss between the UAV and 133 the ground user as a performance parameter and determine 134 VOLUME 10, 2022 the optimum operational height for maximum coverage. [14] 135 proposes a mathematical model to find the inter UAV dis-136 tances in a multi-UAV network to maximize the coverage. 137 In [15], the authors consider a network topology for monitor-  proposes an efficient topology-based routing mech-198 anism for faster delivery of messages to the destination. The 199 authors choose the best next-hop UAV based on the UAVs' 200 present locations and trajectory information. [31] illustrates 201 that the power consumption of UAV during data transmis-202 sion is proportional to the size of the transmitted data; thus, 203 the smaller the size of transmitted data, the smaller the 204 energy consumption. [32] develops an efficient load balanc-205 ing technique for reducing network congestion in wireless 206 LAN by using persistence weighted round-robin algorithm. 207 In [33], the authors provide a load balancing algorithm for 208 UAV-assisted wireless networks using SDN.

222
Stochastic centralized Multipath UAV Routing protocol for 223 FANETs (SMURF) computes the most reliable route [39] 224 for transmission. Enhanced Optimized Link-State Routing 225 protocol for FANETs (OLSRF) prevents communication 226 interruptions due to rapid topology changes [40]. Robust 227 multi-path communication (RMPC) in UAV systems can con-228 trol network congestion by dynamically selecting the best 229 performing path from multiple wireless multihop paths [41]. 230 Multipath TCP (MTCP) can provide stable traffic flow con-231 trol and coordination of drones [42]. Stochastic packet for-232 warding algorithm (SPA) [43], [44] selects the forwarding 233 drone based on network metrics and provides efficient data 234 transmission with improved throughput in FANETs. Success 235 Ratio-based Routing (SRR) is a light-weight protocol for 236 dynamic opportunistic networks that improves the packet 237 delivery ratio [45].

238
In our paper, we propose an application-aware Multi-path 239 Weighted Load balancing (MWL) routing protocol that is 240 light-weight and achieves congestion control in the network. 241 The routing algorithm distributes the load (video surveil-242 lance data packets) over multiple node disjoint paths based 243 on the existing traffic in the path. We have summarised 244 the comparison of MWL with other state-of-the-art rout-245 ing protocols in terms of distinct features in Table 1 aware. 252 We have summarized the major notations used in the rest 253 of the paper in Table 2.

255
The ECMS system consists of three major components: the 256 UAV-Net, the Anchor UAV (AU), and the GCS, as depicted 257 in Fig. 2. The UAV-Net is a cooperative multi-UAV network 258 consisting of N identical UAVs that performs the surveillance 259 operation. Let N = {1, 2, . . . , . . . , N } represent the set of 260 UAVs present in the UAV-Net. The UAVs in the UAV-Net 261 hover at a predetermined altitude over the mission area, 262 record the surveillance video, and then transmit it to the 263 GCS over Wi-Fi for further processing. The AU serves as a 264 relay, connecting the UAV-Net to the GCS. The system may 265 include either one AU or a chain of AUs depending on the 266 distance between the mission area and GCS. The GCS is the 267 on-ground control unit for managing the surveillance opera-268 tion. It also plays a pivotal role in connecting the communi-269 cation framework setup by UAV-Net in the remote location 270 to the 5G/6G communication core network. With the right 271 negotiation method, we can modify the system architecture 272 to complete the mission.

273
While conducting surveillance over the mission area, it is 274 essential that all the UAVs assist in maintaining the topology. 275 This involves maintaining inter-UAV distances, flying forma-276 tion and flying height of the UAV grid. The synchronisation 277 and coordination is achieved through continuous flow of 278 Message Queuing Telemetry Transport (MQTT, [46], [47]) 279 commands and instructions between UAVs over wireless 280 network.

282
In this section, we discuss the framework for multi-UAV net-  the coverage area of a UAV can be given as: where f h and f v are the horizontal and vertical FOVs of the 303 on-board camera, respectively, and can be defined as: dimension (S v ) and can be expressed as: By substituting (2) and (3) in (1), we obtain the on-ground 312 coverage area of the UAV as: It is evident from equation (4) that for a given camera 315 specifications, the coverage area of UAV depends on the 316 flying height. As the flying height increases, the coverage 317 area also increases. However, at the same time, the quality of 318 surveillance declines. Therefore, it is essential to determine 319 a threshold height that will maintain a balance between the 320 UAV coverage and the quality of surveillance (QoS). The QoS 321 of a UAV at a flying height λ can be defined as: where L(λ) is the packet loss ratio between the UAV and the 325 target location (i.e. GCS, another UAV, or anchor drone), and 326 can be expressed as an exponential function of flying height 327 λ (that is also the threshold inter-UAV distance) and packet 328 loss exponent µ. The parameter a is a constant (a > 0) and 329 µ is the packet loss exponent (µ > 0) that depends on the 330 channel characteristics (channel capacity) and The UAVs form a wireless network called UAV-Net to 333 perform surveillance. We assume that all the UAVs in 334 the UAV-Net have identical camera specifications. As a 335 result, at a given operating height, all UAVs have the same 336 on-ground coverage area. We can establish a minimum num-337 ber of UAVs to cover the entire monitoring region by carefully 338 placing each UAV. UAVs, on the other hand, are not station-339 ary. A small deviation from the desired position may result 340 in a ground coverage hole. Therefore, when computing the 341 minimum number of UAVs, we consider the coverage area 342 overlapping to ensure zero blind spots. Let χ i,j denotes the 343 coverage overlapping between UAV i and UAV j. Then, χ i,j 344 can be defined as: where (x i , y i ) and (x j , y j ) are the horizontal positions of UAV 347 i and j, respectively. The overlap area (covered on ground) for 348 two drones i and j, o A i,j , is given as: i.e, Q(λ) = Q th , where Q th indicates the threshold QoS value.

364
After solving, the operational threshold height, λ th = λ * , 365 is found to be Assuming that all the UAVs operate at a height λ = λ th , 368 the MCMV topology problem for a surveillance area A can 369 be formulated as: Note that for the above optimization problem, both the objec- can be expressed as where ψ is the Lagrange multiplier. By solving the KKT 402 conditions i.e, ∂L(N , ψ) ∂N = 0, and ∂L(N , ψ) ∂ψ = 0 , we obtain 403 the optimal number of UAVs as where · represents the ceiling function that gives the nearest 406 integer value for N .
For a given number of UAVs, the Placement of UAVs can 409 be optimized to achieve zero blind spot on the ground. The 410 placement optimization problem can be formulated as Since the position of UAVs are not static, the zero blind spot 417 can be ensured by considering maximum possible overlap-418 With 419 this, we propose a successive placement strategy to find a near 420 optimal solution to the problem P3. The fundamental idea is 421 to place the UAVs sequentially in a grid structure starting 422 at one end of the surveillance region and making the way 423 to the other until the entire surveillance region is covered. 424 Each UAV is guaranteed to have an overlapping of αC(λ). 425 Assuming that all the UAVs have equal overlapping areas, 426 the overlapping constraint can be written as where h and v represent the horizontal and vertical over-429 lapping of a UAV, respectively. To simplify the analysis, 430 we consider h = v. Then, h and v can be expressed 431 as With this overlapping, the UAVs are arranged in a grid struc-436 ture. Fig. 4 shows the placement of UAVs over the geograph-437 ical area under surveillance. The grid consists of rectangular 438 zones, each containing a UAV at its center, with the dimension 439 f h × f v . We approximate the closest possible rectangular area 440 with the end points (x min , y min ), (x min , y max ), (x max , y max ), and 441 (x max , y min ), and perform the placement of UAVs over this 442 rectangular area. The number of rows and columns are given 443 as in the i th row and j th column are as follows: are the x-and y-position of UAV in the  for i = 1 to K do 8: for j = 1 to L do The primary requirement of the routing protocol is to find 489 various node disjoint minimum hop pathways from source 490 nodes (UAVs in the grid) to the destination node (AU). 491 In order to accomplish this, the source node transmits a route 492 discovery control message. The format of the route discovery 493 control message is displayed in Fig. 5. The message is flooded 494 (broadcast) network-wide from source node to the destination 495 node. During the process, the intermediate node may receive 496 multiple copies of the same control message through different 497 pathways. In case a few route discovery messages are lost 498 or some possible paths are missing, route discovery control 499 message broadcast helps in finding the next best feasible set 500 of paths from the source to the destination. The intermediate 501 node selects the one with least hop count value and discards 502 the others. The AU examines all the received control mes-503 sages and selects the one with the fewest hops. The selected 504 message is then compared to other control messages for node 505 disjointness. This results in selection of two node disjoint 506    In MWL routing, the source node can either be an edge 534 node or an inner node. We use the variable P i to characterize 535 the source node i. P i can take two possible values: 0 or 1. 536 P i = 0 for edge node and P i = 1 for inner node. Let 537 P = {P i , ∀i ∈ N } be the description of all the source 538 nodes. The data traffic for an edge node must be sent over 539 two node disjoint paths, whereas it must be distributed over 540 three node disjoint paths for the inner node. The amount of 541 data flow along each path is decided by utilising the path 542 weight, which can be calculated by adding the associated 543 link weights. The link weights are set as the link capacity 544 (i.e. the maximum traffic the link can support). Assume that 545 the network has M links, denoted as l 0 , l 1 , . . . , l M , which 546 can be given as M = {1, 2, . . . , M }. The collection of link 547 weights can be expressed as W = {w k , ∀k ∈ M}, where w k 548 denotes the weight corresponding to the link k. Assume that 549 the association between the source node i and the link k is 550 represented by σ i j,k . Then, σ i j,k can be defined as Let A = {σ i j,k , ∀i ∈ N , j ∈ {1, 2, 3}, and k ∈ M} be 554 the association matrix reflecting the association between the 555 nodes and links in the network. Then, the path weight of the 556 node disjoint path j of the source node i can be evaluated as 557 The traffic ratio corresponding to path j can be calculated as 559 The UAVs in the UAV-Net are the sources that are gen-561 erating data traffic, R, the volume of which depends on 562 the surveillance video acquisition resolution. A high reso-563 lution video corresponds to more number of packets to be 564 transmitted over the network. The source video consists of 565 frames. A set of frames, also known as the group of pictures 566 (GOP), that are captured per second get encoded into a video 567 stream. The video stream is transmitted in the UAV-Net in the 568 form of packets. The traffic distribution can be performed by 569 following load balancing mechanism, in which more traffic 570 is forwarded over the path with less weight and less traffic 571 is forwarded over the path with higher weight. The details 572 of data traffic distribution is provided in Algorithm 2. This 573 algorithm is periodically executed after transmission of each 574 surveillance video GOP thereby governing re-routes and traf-575 fic distribution updates in each GOP transmission period. The 576 association matrix A is used in updating path weight, W i j , 577 according to (20). Every time there is a traffic flow, associated 578 with the surveillance video GOP, on the path, the weight of 579 the path increases. To update the path weight, the source node 580 broadcasts a link update control message at the end of each 581   Before transmitting through any path, the source node 597 inspects the path condition. For this, it has to check two 598 threshold conditions: (a) Path break threshold, and (b) Con-599 gestion threshold. The path break threshold and the con-600 gestion threshold for the edge node are set to n i T and 2T , 601 respectively, whereas the path break threshold and congestion 602 threshold for the inner node are (n i +1)T and 3T , respectively, 603 where T indicates the highest data rate supported by each 604 link. If a source node finds the weight of any node disjoint 605 path above the path break threshold value, it forwards the 606 entire traffic through other available paths. A re-route discov-607 ery is conducted if the node finds all its disjoint paths have 608 weights above the path break value. In case of congestion, 609 the nodes wait for the directions from higher layer and store 610 the video in the buffer in the meantime.

611
In such a scenario, when congestion occurs, the video 612 acquisition module on the UAVs switch to a lower resolution. 613 This reduces the amount of data (i.e. number of packets) to 614 be transmitted from a given source, i.e. R. 615 Video transmission is tolerant to packet losses to a certain 616 extent [51], [52]. A packet loss of upto 4% is acceptable for 617 video streaming application in terms of the perceptual video 618 quality [53], [54]. The MCMV topology has the inter-UAV 619 distance and UAV-GCS distance selected to be less than the 620 threshold that prevents the packet loss to be greater than the 621 acceptable limits. Furthermore, in the event of congestion, 622 with MWL routing scheme, the source data (and packet) rate 623 is reduced to prevent the packet loss from exceeding the 624 acceptable limit.

626
A. TESTBED DETAILS 627 A Raspberry Pi 3 B+ companion computer is used on board 628 to assist in the surveillance along with mission planning soft-629 ware. The UAV's movement is guided by the flight controller. 630 In order to facilitate packet forwarding, the UAV includes two 631 WLAN adapters, one internal and one external, which are 632 shown in Fig. 8. The internal WLAN adapter is used to create 633 the Access Point (AP), while the external WLAN adapter 634 is utilized to connect the UAV as a client to the AP created 635 by an adjacent UAV. The hardware and software components 636 necessary to set up a UAV are listed in Table 3. It is important 637 to note that, we have have used two power supplies (as listed 638 in Table 3) in our UAV assembly, one for the UAV flying oper-639 ation (motor, controller), and the other for communication 640 framework (Raspberry Pi board). The proposed MWL routing 641 scheme focuses to efficiently transmit the surveillance video 642 from the UAV to the GCS with a reduced energy consumption 643 in the associated communication that effectively improves the 644 UAV-Net lifetime. 645 We use wireless LAN (WLAN) as the primary technique 646 for connecting UAVs to the GCS. Fig. 9 shows an experimen-647 tal setup to illustrate WLAN-based communication between 648 the UAV 1 and the GCS. The AU includes two WLAN 649 adapters: internal WLAN (WLAN 0) and external WLAN 650   UAV with on-board camera specifications of F l = 3.60 mm, 662 S h = 3.76 mm, and S v = 2.74 mm. The UAV is configured 663 to fly at three different flying heights of 5 m, 10 m, and 664 15 m, and the coverage area is measured for each variation. 665 A comparison between the actual covered area and mathe-666 matically calculated values is given in Table 4. The result 667 shows that the actual covered area is always greater than 668 the calculated value. Therefore, if we keep the inter UAV 669 distances in both horizontal and vertical directions below f h 670 and f v , respectively, there will always be overlap between 671 the two contiguous fields of view of adjacent drones. This 672 guarantees zero blind spots even if drones deviate slightly 673 from their intended positions.

674
After validating zero blind spot, we investigate the effect 675 of flying altitude on topology development. For this, we con-676 sider different geographic regions and evaluate the average 677 number of UAVs required at three different flying heights of 678 50 m, 75 m, and 100 m. Fig. 10 shows the coverage area vs 679 average number of UAVs required at different flying heights. 680 It is observed that higher the flying height of the UAVs, more 681 is the on ground coverage area and hence, lesser number of 682 UAVs are needed to cover a geographical area. On the other 683 hand, UAVs flying at lower height cover lesser geographical 684 area on ground, hence more UAVs are needed to cover the 685 same area. So, by customizing the UAVs to fly at a larger 686 height, we can achieve the coverage with minimum possible 687 UAVs. However, larger height degrades the service quality. 688 Therefore, we decide a threshold height (λ th ) in order to 689 optimize the number of UAVs required to cover a mission 690 area.  UAVs required for surveillance. We consider two different 725 simulation scenarios. In Fig. 13 (a), we consider a mission   Table 5. We have performed Monte-Carlo 738 simulation with more than 100 instances of network scenario 739 and 95% confidence interval. 740 We have evaluated the performance of our proposed MWL 741 routing scheme in comparison with SMURF [39], SPA [43], 742 [44], and LMP-DSR [34]. In addition to being light-weight, 743 link-existence based, and load-balanced, similar to SPA, our 744 FIGURE 14. Topology for simulation of MWL routing.   This is because in LMP-DSR and SPA, the source only 773 transmits via the path selected (by the routing scheme) from 774 amongst the multiple paths available. This results in cer-775 tain links like l 1 , l 7 , and l 13 (which serve as the interface 776 between the AU and UAV-Net) to support a maximum of three 777 simultaneous video transmission by three UAVs. This puts a 778 limitation on the system throughput and PDR. However, such 779 limitation is avoided in MWL routing through load balancing. 780 Here, the source node balances the load among multiple node 781 disjoint paths, thus allowing links l 1 , l 7 , and l 13 to accommo-782 date more than three simultaneous video transmissions. Since 783 SMURF performs packet replication on the multiple paths to 784 improve PDR, it is not effective in our scenario with limited 785 capacity links and many transmitting nodes. In MWL, the 786 network can allow multiple simultaneous video transmissions 787 with its source adaptive congestion control strategy. When the 788 number of active drones increases, congestion occurs which 789 is avoided by reducing the video resolution (data/ packet 790 rate). In this situation, the network supports multiple UAVs 791 to transmit simultaneously. Thus, the throughput and PDR 792 achieved by the network to be higher (19.38% and 23.74%, 793 on average, respectively) with MWL scheme as compared to 794 the other routing protocols. The effectiveness of MWL routing is also evaluated by com-797 paring the end-to-end delay performance of MWL in com-798 parison with ith SMURF, SPA, and LMP-DSR. We consider 799 the source video resolution set as {1080p, 720p, 480p}, where 800 transmission of a 1080p corresponds to a true colour HD reso-801 lution from source node to the Anchor node. The transmission 802 is based on group of picture (GOP) that is a set of video 803 frames encoded collectively in a second. We compute the 804 delay at the destination end for the GOP (30 frames in a GOP 805 for frame rate = 30fps). The simulation scenario is shown in 806 Fig. 14 and the parameters used for simulation settings are 807 listed in Table 5. 808 With the above parameters, the free space packet propaga-809 tion delay for a link distance of 100 m is found to be 4ms. 810 Then, the propagation delay for a video frame to reach the 811 destination can be calculated as where n h is the number of hop count from source UAV to 816 the Anchor UAV, n i h is the number of hop count along the 817 path selected for packet i, P d is the packet propagation delay 818 for each hop (4 ms), F no is the number of packets in the 819 video GOP (per sec), and n disjoint is the number of node 820 VOLUME 10, 2022  path is lower, which ultimately reduces the end-to-end delay. 843 We also notice that nodes at the same distance from the anchor 844 drone experience the same amount of delay. As the distance 845 increases, the end-to-end delay also increases. The experimental packet loss analysis (Fig. 11) in the topol- with the other simulation settings given in Table 5. routing has lesser packet loss percentage by 76.91%, 84.74%, 857 and 86.89%, on average in comparison to SMURF, SPA, and 858 LMP-DSR routing schemes, respectively. This is attributed to 859 the application-aware and simultaneous multi-path load bal-860 anced packet transmission in MWL routing. Load-balancing 861 and source adaptive congestion control enables MWL routing 862 to dynamically balance traffic throughout the network and 863 reduce average packet loss percentage in the UAV-Net. 864

865
The energy efficiency of a routing protocol can be evaluated 866 in terms of nodes' residual energy and network lifetime. The 867 residue energy of a node is calculated using the equation 868 below.
where E ini is the initial energy available at the node and 871 E con is the total energy consumed by the node. E ini can be 872 determined from the battery specifications using the relation 873 Energy(J ) = Voltage(V ) × Charge(mA.h) × 3.6. To evaluate 874 E con , we use the first order energy consumption model [55]. 875 This is a popular model used for evaluating the WSN routing 876 protocols. According to this model, the energy consumed by 877 a node for sending and receiving m bit data is given as: where E Tx and E Rx are the energy consumed during data 881 transmission and reception respectively, E elec denotes the 882 single bit energy consumption at the transmitter and receiver 883 circuits, E amp represents the single bit energy consumption 884 at the amplifier circuit, and d is the distance over which 885 information is transmitted. The total energy consumed by the 886 node is calculated as  For evaluation, we consider the scenario shown in Fig. 14   889 with settings as given in Table 5 and we evaluate the residue

916
In this work, we have introduced an efficient UAV-assisted 917 surveillance system to facilitate surveillance in remote inac-918 cessible locations. In particular, we have developed a QoS-919 aware topology, that minimizes the number of UAVs required 920 to conduct surveillance with zero blind spots. We have 921 also determined the threshold operational height in order to 922 keep the packet loss within the acceptable limit. Then, for 923 reducing network congestion, we have proposed an effective 924 and efficient application-aware load balancing-based routing 925 strategy. Finally, we assess the performance effectiveness of 926 the suggested surveillance system through extensive simula-927 tions. We have shown that the proposed surveillance system 928 accomplishes zero blind spot surveillance with lesser number 929 of UAVs and reduced energy consumption, while offering 930 higher throughput and lesser end-to-end delay when com-931 pared to the existing state-of-the-art mechanisms in literature. 932 Further extension of the proposed framework will incorpo-933 rate synchronisation aspects of UAV swarm in order to assure 934 collision avoidance during surveillance. The interference can 935 be further reduced through dynamic channel allocation in the 936 UAV-Net.