Mobility Analytics of Fans During the 2021 FIFA Arab Cup™ Football Tournament in Qatar

The FIFA Arab Cup, a test event for the 2022 FIFA World Cup, took place in Qatar from November to December 2021. The event showcased 32 matches across six venues that will also be utilized in the World Cup. This paper presents a groundbreaking spatiotemporal analysis of traffic mobility during the event, using data collected from WaveTraf road sensors. The sensors detect and track Bluetooth and WiFi-enabled user devices, allowing for an analysis of user mobility, including the origin of spectators, the time taken to reach the stadium, dwell time inside the stadium, and the time taken to return to their origin after leaving the stadium. The study processed tens of millions of records, overcoming challenges such as filtering data anomalies and validating and preparing the data for analysis. The findings offer a comprehensive understanding of user mobility patterns during the event, which is valuable information for event organizers, city planners, and transportation providers to optimize services and enhance the overall user experience. Furthermore, the study highlights the importance of leveraging advanced technology to improve the planning and execution of large-scale events and transportation systems. The research showcases the power of data analytics in providing key insights into human mobility.


I. INTRODUCTION
H UMAN mobility is a fascinating and essential field that has gained widespread attention from city planners, road operators, and event organizers worldwide. Recent advancements in technology and data collection methods have enabled researchers to gain unprecedented insights into how people move and interact with their environment. These insights are crucial in optimizing services and reducing operational costs, ultimately benefiting the end-users.
The 2021 FIFA Arab Cup TM held in Qatar provided a unique opportunity to study the mobility patterns of thousands of spectators who traveled from all over the world to support their favorite teams. With 32 matches hosted in 6 different venues and over 570,000 spectators, the tournament served as a testing ground for various facilities that will be used during the FIFA World Cup tournament scheduled in a few months. Mobility and transportation were identified as key enablers for a successful event, highlighting the need The review of this article was arranged by Associate Editor Jiwon Kim. to collect data about fans along their journey of attending matches in different venues.
The main data source used in this study was fixed Bluetooth sensing devices deployed in most of the junctions in Qatar, as well as around the stadium. The whole system was presented, and a deep analysis of the collected data was discussed, providing valuable insights for city planners, road operators, and event organizers. By understanding human mobility patterns, we can optimize services and reduce operational costs, ultimately benefiting both the end-users and the organizers.
QMIC (Qatar Mobility Innovations Center) has developed an innovative IoT detection and tracking solution called WaveTraf TM , which consists of a device and a back-end system. This non-intrusive technology can detect the wireless module identifiers, such as Bluetooth and WiFi, from user devices such as smartphones, in-car phone handsets, and navigation systems. The system reports this information in real-time to the back-end. Since January 2012, over 380 devices have been deployed in various cities across Qatar, making it a reliable and widespread solution for collecting data on human mobility.
The WaveTraf system consists mainly of two parts: • Embedded systems: Network of sensors that detects enabled Bluetooth or WiFi user devices and their information to the back-end part. • IOT platform: A back-end system that stores collected data from network of sensors for further processing. The objective of this paper is to leverage the data collected from the WaveTraf sensors to gain insights into the movement patterns of crowds of people attending the different matches. Specifically, we analyze the mobility of commuters before and after the games to and from the stadia to answer three questions: From which roads/zones do those people come and where do they go? How long does it take them to reach the stadium? What are the peak periods of movement? These insights can help organizers understand the impact of their measures, such as traffic diversions, parking facilities, and public transport, and help them adjust their plans to improve the overall experience for fans.
To achieve this, we detail the data collection process in Section III, where we explain how WaveTraf sensors were used to collect data about the movement of people in and around the stadium. We then describe the filtering steps applied to the noisy data in Section IV, where we explain how we processed the data to remove noise and ensure accuracy. In Section V, we present some hidden patterns in the raw data, such as estimating the origin-destination matrix and time taken by commuters before, during, and after the match. Finally, we conclude the paper in Section VI by summarizing our findings and suggesting future works.

II. BACKGROUND AND RELATED WORK
Throughout the past two decades, sensing technologies have been used to a variety of use cases based on their intrinsic properties, such as coverage range, form-factor, environmental sensitivity, power consumption, level of privacy protection, etc. Transportation, specifically people movement and logistics, is one of the primary industries in which sensors have enabled/enhanced a large variety of services. For instance, placed road sensors offered real-time services such as journey time calculation and forecast in traffic monitoring.

A. ORIGIN-DESTINATION MATRICES
Origin-destination (OD) matrices are utilized to depict and analyze inter-locational flows of people, goods, and information. A OD matrix depicts the number of journeys or other flow between all conceivable origin-destination pairings in a particular location.
Planning and analyzing transportation requires OD matrix calculation. For traffic planning in congested cities, precise and up-to-date OD matrices are necessary. Using traffic counts, surveys, GPS monitoring, and mobile phone data, OD matrices are computed.
Methods for computing OD matrices have pros and cons. Traffic assignment models, survey techniques, and passive data collecting techniques, which use automated sensors and tracking equipment to gather movement patterns, are the most often employed methodologies.
For example in [1], the authors presented a study using mobile phone-based origin-destination data to examine human mobility in Seoul, Korea. The authors suggest a graph-based method for analyzing the data, which entails the construction of a directed graph to depict the flow of people between various areas in the city. In addition, graph theory measures such as node centrality, community detection, and network density are used to examine the graph. The study's findings provide insights into the patterns of human mobility in Seoul, including the identification of popular destinations and transit hubs. The study relies on data acquired from mobile phone users, who may not be representative of the overall population. Individuals who do not use mobile phones or do not disclose their location information may be excluded, leading to the possibility of sampling bias.
Reference [2] proposed a method for calculating Origin-Destination (OD) matrices by assimilating sub-path flows inside a bilevel optimization framework. The suggested method utilizes sub-path travel time data from various sources to estimate the sub-path flows, and then incorporates them into a bi-level OD matrix estimation procedure to increase the accuracy of the estimated OD matrix. On a real-world traffic network, the method is evaluated and compared to two other state-of-the-art methods. The outcomes demonstrate that the proposed method outperforms the alternatives in terms of estimating precision and computing time. Transportation planners and traffic engineers can use the suggested method to increase the accuracy of OD matrix calculation and enable improved transportation planning and management decisions. The analysis suggests the existence of a few potential constraints. Among these are the reliance on a restricted number of datasets and the assumption of user behavior homogeneity. In addition, the proposed method may not be applicable to scenarios with a significant number of sub-paths, which could result in computational complexity and scalability concerns. In addition, the proposed method implies a complete understanding of the network's topology, which may not be possible in real-world settings.
In recent years, data analytics and machine learning have produced novel OD matrix computing algorithms. Big data, machine learning algorithms, and data fusion techniques analyze and integrate GPS, cell phone, and social media traffic data.
Despite these advancements, it is challenging to compute accurate and reliable OD matrices in big, complicated urban regions with high traffic volumes and fluctuating demand. Improving the accuracy and utility of OD matrices in transportation planning and traffic management will require continual study and improvement.

B. WORK OBJECTIVE AND MAIN CONTRIBUTIONS
The aim of this work is to make use of already deployed infrastructure of road sensors in Qatar to conduct a deep study of the mobility of people during the 2021 FIFA Arab Cup TM .
Boudabous et al. did a prior study comparable to ours titled [3]. Their investigation was based on Bluetooth passive scanning, while ours is based on Bluetooth active scanning. In contrast to our methodology, this work relied on an approximation of a model to give online estimation. Nevertheless, noisy data impeded the suggested model, making it difficult to implement and resulting in mistake propagation. Abbas et al. [4] also suggested a comprehensive implementation of a graph-based online estimation of traffic flow utilizing Apache Kafka and Spark Streaming technologies. Nonetheless, their methodology does not capture and/or assess human movement. Our suggested methodology, on the other hand, solves the deficiencies of current approaches and delivers precise and efficient estimates of human mobility and traffic flow.
Our method, in contrast to those that have come before it, makes use of Bluetooth active scanning, which enables a more exact detection of devices that are in the immediate vicinity and supplies information regarding location that is more accurate. In addition to this, we make use of machine learning techniques in order to eliminate noise and recognize trends in the data that was collected. Because of this, we are able to generate origin-destination matrices that are more dependable and accurate in their estimations, which in turn may be utilized to improve the planning and management of transportation. In addition, our system was developed to be expandable, and it is simple enough to be implemented in a wide variety of urban settings. We feel that our method has the potential to make a significant contribution to the existing body of research as well as provide a number of benefits that are immediately applicable to transportation authorities and urban planners.
Weppner and Lukowicz in his work on crowd density estimate [5], employed a time series feature engineering based strategy, which gave his model estimation more interoperability. This is a slight departure from the approach that we took in our study. To estimate the density of human crowds, our research on crowd density estimation with Bluetooth technology focuses on the use of machine learning models, more specifically the Random Forest method. This approach was taken since it is more accurate. This methodology provides a way for estimating crowd density that is more flexible and scalable than time series-based methods since it does not require the same level of feature engineering as those methods. In addition, the utilization of Bluetooth technology enables the estimation of the crowd density in real time, making it appropriate for deployment in a variety of settings, such as stadiums, shopping centers, and public places, among other possible locations.

III. SENSING DEVICES AND DATA COLLECTION PROCESS
Qatar Mobility Innovations Center (QMIC) developed and deployed a revolutionary IoT detection and tracking solution, branded as WaveTraf TM [6], to gather valuable data on user mobility patterns. By detecting MAC addresses of Bluetooth and WiFi-enabled user devices, WaveTraf TM system provides insights into user movement patterns in congested areas. The system comprises two main components: a detector that identifies moving objects and a back-end platform that processes the real-time information collected from nearly 385 deployed detectors as shown in Fig. 1. This section provides a comprehensive overview of the WaveTraf TM system, including its architecture, design, and deployment in Qatar. With over millions of user devices detected per day, the WaveTraf TM system offers a powerful tool for analyzing crowd mobility patterns and making data-driven decisions.
Each WaveTraf detector is designed to connect to the back-end system through various communication technologies, including Ethernet, optical fiber, and cellular networks, ensuring reliable and flexible data transmission. The capability to connect via different communication technologies is a fundamental characteristic of WaveTraf devices, which guarantees seamless data transmission in different environments. The WaveTraf sensors are equipped with a solar panel that provides a continuous power supply with a capacity of 20 watts, enabling the devices to be deployed in various locations with ease. This feature also eliminates the need for any significant modifications to existing infrastructure, as the detectors can be installed on existing structures like poles or lampposts designated for other assets. With the solar-powered and cellular-connected WaveTraf sensors, users can efficiently deploy the detectors with minimal effort, enabling them to collect real-time data on user mobility with ease.
The WaveTraf back-end system is a critical component of the overall WaveTraf system, allowing for the collection, processing, and analysis of data from over 380 sensors deployed throughout Qatar (see Fig. 2). With each sensor able to collect millions of records per day, the backend system handles a massive amount of data on a daily basis. In addition to data from the WaveTraf sensors, other sources of data are also integrated into the system, such as weather data, social media feeds, and transportation schedules.
Using this wealth of data, the WaveTraf system is able to provide a range of mobility and traffic services through QMIC's Masarak product. These services include real-time traffic flow monitoring, dynamic route planning, and incident management. The system can also provide historical data analysis, allowing for the identification of long-term trends and patterns in mobility and traffic behavior. With the ability to monitor and analyze traffic in real-time, the WaveTraf system helps to improve the efficiency and safety of transportation networks, ultimately leading to a better quality of life for commuters and residents alike.
The WaveTraf system detects user devices using a tracking and scanning system that identifies the Medium Access Control (MAC) addresses embedded in the devices. For each user device, the system provides the identifier of the WaveTraf device that detected it, along with the detection timestamp and the dwell time of the device within the detection range. The detection range is approximately 60 meters around the location of the WaveTraf device, as it uses omnidirectional antennas for both Bluetooth and WiFi detection. This information is essential for understanding user mobility patterns and analyzing traffic flow in different areas.
While the WaveTraf system is designed to detect as many user devices as possible, occasional missed detections can occur due to factors such as weak communication or signal interference. However, this is not a significant challenge as the focus of our analysis is on the overall trajectories of users rather than individual observations.
The WaveTraf system can face a second challenge of losing connection to an already detected device while it is still in the detection zone and its re-detection in the second time, leading to an excess of detection. However, this challenge can be overcome by clustering those observations into a single trajectory. In this way, the trajectory of the user device can be accurately captured and analyzed, despite occasional loss of connection or excess detection. The clustering of observations also allows for a more comprehensive analysis of user mobility and behavior, which can be useful for various applications in mobility and transportation. Therefore, this challenge does not significantly affect the overall performance of the WaveTraf system.

IV. PRE-PROCESSING OF VEHICLE TRAJECTORY
To extract accurate vehicle trajectories from the collected data, the hashed Bluetooth device identifier is stored in the database along with all associated records. However, due to the possibility of mixing trajectories in one group, we employed advanced filtering and clustering techniques. These methods helped to differentiate between trajectories and extract precise vehicle trips information to and from the stadium, ensuring that the data is reliable for further analysis and use.
There are several clustering techniques used in machine learning, but here are four commonly used ones: • K-Means Clustering: This is a well-known unsupervised learning algorithm that divides a dataset into K clusters. It is an iterative algorithm that minimizes the sum of distances between each point and the cluster center to which it is assigned. • Hierarchical Clustering: This is yet another unsupervised learning algorithm that clusters similar data points based on their distances. It generates a dendrogram that represents the hierarchical relationships between clusters and can be agglomerative (bottom-up) or divisive (top-down).
• DBSCAN (Density-Based Spatial Clustering of Applications with Noise): This is a density-based clustering algorithm that groups together nearby points while separating them from distant points. It can recognize clusters of arbitrary shapes and is noise resistant. • Gaussian Mixture Models (GMM): This is a probabilistic clustering algorithm that uses a mixture of Gaussian distributions to model data distribution. It can recognize clusters of various shapes and sizes and estimate the likelihood of each point belonging to each cluster. It is worth noting that the performance and suitability of each clustering algorithm may vary depending on the specific dataset and problem at hand. DBSCAN has several advantages over other clustering techniques: • Can identify clusters of arbitrary shapes: DBSCAN can identify clusters of arbitrary shapes because it makes no assumptions about the shape of the cluster. This makes it ideal for datasets with non-spherical or circular clusters. • Robust to noise: DBSCAN is noise-resistant because it distinguishes between noise points and clusters. This helps to avoid assigning noise points to any cluster and affecting the overall clustering quality. • No need to specify the number of clusters: Unlike K-means clustering, which requires a predefined number of clusters, DBSCAN does not require a predefined number of clusters. • Works well with large datasets: DBSCAN can efficiently handle large datasets because it does not require the entire dataset to be loaded into memory. It processes data points one at a time, making it appropriate for datasets containing millions of data points. • Fewer assumptions about the data: In comparison to other clustering algorithms, DBSCAN makes fewer assumptions about data distribution. This increases its flexibility and ability to detect clusters in datasets with complex structures.

A. SPACE FILTERING
Each detected device is identified by a unique ID derived from its Bluetooth MAC address, and as it moves through the network of WaveTraf sensors deployed in the country, it creates its own trajectory. However, due to the randomization techniques employed by mobile phone operating systems such as iOS and Android, two different devices may share the same Bluetooth MAC address for a certain period, which could lead to misinterpretation of their actual respective trajectories. This issue is resolved by using the Density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm [7]. DBSCAN helps to capture continuous trajectories by separating them based on the Euclidean distance between sensors locations. Additionally, detecting devices in distant places can lead to space discontinuity, which results in a combination of different trips. The use of DBSCAN addresses this challenge, making it possible to separate trajectories and obtain accurate trip data. Our case does not align with clustering algorithms such as K-means due to the lack of prior knowledge regarding the number of clusters. Although Hierarchical Clustering is capable of addressing the task, it poses a computational burden.
To compute the distance between two locations, we utilized the GeoPy Python library [8] which provides various geocoding APIs to convert addresses into geographic coordinates, calculate distances between two locations, and more. We found this library to be efficient and accurate for our purposes. Additionally, we utilized the widely-used scikit-learn library for the DBSCAN algorithm, which is a powerful clustering method that can handle noisy data and detect clusters of arbitrary shapes. The scikit-learn library provided us with a robust and customizable implementation of the DBSCAN algorithm that we could easily integrate into our workflow. Overall, these two libraries played a critical role in our analysis pipeline and helped us achieve accurate and reliable results.

B. TIME-BASED FILTERING
To ensure that multiple trips made by the same vehicle with the same device ID are detected, we employed the DBSCAN clustering algorithm [7] to split the combined trajectories for the same detected user device. This approach enables the identification of trips made at different times of the day, which may have the same or different paths from source to destination. Unlike other clustering algorithms, DBSCAN considers the density of points within a region rather than just the proximity of the points. We defined the distance between two observations as the difference between the detection times, allowing for the detection of multiple trips made by the same vehicle with the same device ID. By utilizing this approach, we are able to accurately capture the multiple trips made by the same vehicle over time, providing a more comprehensive analysis of traffic patterns in the area of interest. By the way, this method can also allow for a more precise understanding of traffic flow and congestion, which can be utilized for future planning and decision-making.

C. VELOCITY-BASED FILTERING
In the case of two different vehicles having the same randomly-generated Bluetooth MAC address, they may be detected at two different places by two different WaveTraf sensors at close timestamps. This is not a typical scenario for a single vehicle, as it would generate a significant gap in velocity and a discontinuity in velocity trajectories. To address this issue, we employ the DBSCAN clustering algorithm [7] which takes the distance between two observations as the Euclidean distance between the fixed locations of the WaveTraf sensors divided by the difference in detection timestamps. This approach allows us to split the trajectories of the two different vehicles back to their original paths. By applying this technique, we can accurately identify the trips of each vehicle, enabling us to extract more precise and reliable traffic data from the sensor network.

D. DEALING WITH MULTIPLE DEVICES PER USER
In our analysis pipeline, we have devised a method that leverages intelligence to identify situations where devices exhibit a collective movement pattern. This enables us to identify instances where multiple devices belonging to the same individual are repeatedly detected by the same sensors. By considering the potential influence of multiple devices per person on our estimation of origin-destination (OD) volumes, we account for any associated impact in our analysis.
This approach employs the DBSCAN algorithm for clustering device trajectories. The proximity between two device samples is established by assessing the dissimilarity between their individual trajectories. More precisely, this dissimilarity measure is calculated as the ratio of distinct detections to the combined count of distinct and overlapping detections within the trajectories.
This allowed us to account for the potential impact that having multiple devices per person could have on our estimation of OD volumes. For instance, if a person is carrying a smartphone in addition to a wearable device like a fitness tracker, then it is possible for both of these devices to be detected by the same sensor multiple times while they are in close proximity to one another. This would indicate that the person carrying both of these devices is most likely the same person. In a similar vein, if a family decides to take a vacation together, it is possible for their electronic devices to be found in close proximity to one another, which would indicate that they are most likely vacationing together as a group. We are able to separate signals from multiple devices that belong to the same person or group and ensure that they are not counted as multiple individuals in our estimation of OD volumes by identifying and tracking these instances. This allows us to ensure that OD volumes are not underestimated. We are able to obtain more accurate estimates of travel time and OD matrices by using this approach because it allows us to take into account the movements of individuals who carry multiple devices. We are aware that our method has some shortcomings, such as the possibility of an insufficient detection of situations in which a single person is carrying multiple devices but these devices are not detected as moving "together." Nevertheless, we are of the opinion that our method is a significant improvement over the practice of counting all devices as independent individuals and that it makes a contribution to the fields of travel time and OD matrix estimation.

V. RESULTS, ANALYSIS, AND DISCUSSIONS
To analyze the behavior of commuters around the stadium during the FIFA Arab Cup TM held in Qatar in November 2021, we extracted traffic trajectories from the WaveTraf raw tabular data records, which naturally form a graph connecting zones, districts, and WaveTraf devices. We then filtered the trajectories to select only those that had detections around the stadium. This allowed us to focus specifically on the traffic data and trajectories in this high-traffic area during the event.
In the analysis below, the focus is given to the final match which was held on 18 th of December 2021 at 6pm local time which was attended by around 60,456 spectators [11].

A. ORIGIN-DESTINATION MATRICES
The computation of the O-D matrix is a fundamental task in transportation analysis and can be used to analyze travel patterns, plan for infrastructure improvements, and estimate demand for public transportation.
To compute the number of trips between a set of n locations, including fixed locations, zones, districts, or userdefined geofences, an origin-destination (O-D) matrix T is calculated. Each element T i,j of the n × n matrix represents the number of trips from location L i to location L j . This matrix is computed for a specified time period between t start and t end . Zones are typically divided into smaller districts and are used to define the locations. We use the following notations: • T Z to refer to the O-D Matrix between zones. Assume n Z to be the total number of zones. • T D to refer to the O-D matrix between districts. Assume n D to be the total number of districts. • T W to refer to the O-D matrix between WaveTraf locations. Assume n W to be the total number of WaveTraf sensors.
To analyze traffic patterns between different locations, we compute an origin-destination (O-D) matrix between the locations of WaveTraf sensors, denoted as T W . The computation is based on Bluetooth device detections recorded during a user-defined time period, [t start , t end ]. We extract all detections from all sensors and sort them in ascending order of their detection timestamp. Each detection is described by the following fields: • t detection is the detection timestamp by the WaveTraf sensor. • w id is the identifier of the WaveTraf sensor. • d id is the identifier assigned to the Bluetooth device that was detected, which could be the original MAC address of the device, a hashed value of the MAC address, or a unique identifier from the database where the detection data is stored.
Let the variable T (wt 1 ,wt 2 ) max be the maximum possible travel time in seconds between each pair of WaveTraf devices w 1 and w 2 during the observation period. If real-time traffic information in terms of average speed or travel time is available, this value can be obtained from the system. Otherwise, the worst travel time can be estimated based on the time of day. The free flow travel time between the two sensors is denoted by T (wt 1 ,wt 2 ) ff , and the maximum travel time can be assumed to be equal to: • θ peak * T (wt 1 ,wt 2 ) ff : in on-peak hours, where θ peak is set to 5.0 in our data analysis. • θ offpeak * T (wt 1 ,wt 2 ) ff : in off-peak hours, where θ offpeak is set to 2.0 in our data analysis.
In order to analyze the behavior of commuters during the FIFA Arab Cup TM held in Qatar in November 2021, we first defined two matrices to represent the trips coming to the stadium and those leaving the stadium. The data for these matrices is derived from the almost 380+ WaveTraf sensors that are installed in 89 different districts in Qatar, creating a very dense network that enables the detection and tracking of each trip anonymously. By grouping the detections by districts, we can get a macro view of the trips in terms of their origins and destinations, allowing us to analyze traffic patterns and make more informed decisions regarding traffic management during events. Furthermore, this data can also be used to improve overall transportation infrastructure in the future, as it provides valuable insights into the movement of people across different parts of the city.
The histogram in Fig. 4 depicts the distribution of the origin districts for trips coming to the Al-Bayt stadium. The data suggests that the distribution is highly skewed, indicating that a vast majority of commuters come from only a handful of districts. Specifically, the highest number of trips originate from Al-Heedan, which is the district where the stadium is located. The next most frequent origin districts are Al-Gharrafa and Al-Khor, as well as other districts located in close proximity to the Al-Bayt stadium.
Another representation of the origin destination matrix is plotted with a Chord graph using Bokeh Python library in Fig. 5 and Fig. 6. Fig. 5 represents the distribution of trips to Al-Bayt stadium before the match. The upper half-circle represents the destination: Al-Bayt stadium. The lower half-circle represents the main origins of these trips like Al-Heedan, Al-Gharrafa, and Al-Khor. The links between the sources  and the destination are presented in different colors and sizes according to the number of trips from each source to final destination (Al-Bayt Stadium). Fig. 6 represents the distribution of trips originated from Al-Bayt stadium after the match. The upper half-circle represents the source: Al-Bayt stadium. The lower half-circle represents the main destinations of these trips such as Al-Heedan, Al-Gharrafa, and Al-Khor.
The fact that the majority of the trips that are going to Al-Bayt stadium before the match have the same destinations in mind for when they leave the stadium after the  match indicates that the overwhelming majority of people make their way back to where they started. Transportation planners need to take this into account in order to develop and implement strategies that will improve the flow of traffic not only on the way to the stadium but also on the way away from it. During the match, the expected volume on the way back can be determined, and the planners can then adjust their real-time scenario testing accordingly. This allows them to make decisions regarding the parameters of the network, such as the optimal dynamic segments of the road segments and signal phase timings. This information could help transportation planners and traffic managers in gaining a deeper understanding of travel demand. It can enable them to efficiently coordinate vehicles that share similar destinations, forming platoons that maximize various benefits. These advantages include enhanced travel times and significant energy savings as reported in [14], [15], and [16].
Another presentation of the O-D information is shown in Fig. 7 where a snapshot of a video playing the trips as detected by the system from different locations in Qatar. As expected trips are being originated from different zones, but most of the users are taking two of the main highways leading to the Al-Bayt stadium namely Al-Khor and Al-Shamal highways. Fig. 8 displays the distribution of destination districts of trips leaving Al-Bayt stadium after the match. We observe that commuters mainly travel to 7 districts, which are similar to the ones observed in the origin matrix. This inverse relationship between the destination and origin matrices can be attributed to the closure of commuters' trajectories, meaning that most of the commuters return to their point of origin after the event.

B. TRAVEL TIME & DWELL TIME
It is necessary for transportation planners and traffic management systems to have a solid understanding of the conditions of the traffic as well as the behavior of commuters while an event is taking place. The trip time is a piece of information that is required for this evaluation, and it is one of the most important elements.
Trips to the stadium, trips while at the stadium, and journeys away from the stadium can be broken down into their respective components in terms of the amount of time spent traveling for each one. The first group of people reflects the amount of time it takes commuters to get to the stadium, while the second group of people relates to the amount of time spent within or close to the stadium. This time is also referred to as the dwell time. The final group represents the amount of time it took for commuters to go to their respective locations after exiting the stadium.
We begin the process of calculating the journey time for these groups by first generating the trajectories of trips. To do this, we apply filtering and clustering techniques on the raw detections of commuters made by our WaveTraf sensors across the network. Since we have the trip trajectories in hand, we can determine the travel time by subtracting the time of the most recent detection from the time of the first detection along the trajectory.
The behavior of passengers before, during, and after the event can be better understood by doing an analysis of the distribution of journey times for each set of trips. For instance, we are able to recognize patterns of congestion and delays during particular times, such as the rush hour before the event or the peak hour after the event has concluded. On top of that, we are able to analyze how much of an impact things like road closures, disruptions in public transportation, and adverse weather have on travel times.
Transportation planners and traffic management systems are able to make informed decisions to optimize traffic flow and reduce congestion by gaining a deeper understanding of the travel behavior of commuters during an event. This results in the travel experience being more efficient and enjoyable for everyone. Fig. 9 presents a transparent bar plot that shows the relationship between traffic travel time and the detection time of commuters who are coming to the stadium before reaching it. The transparency of the bars is used to reflect the instantaneous number of trips. The distribution of travel times is not even, with more trips during the match taking longer times.
The stay duration is calculated as the difference between the detection time of the trip's last sensor before reaching the stadium and the detection time of the trip's first sensor after leaving the stadium. Fig. 10 shows that visitors tend to stay longer around the stadium during the match compared to before and after the match. The average duration of stay during the match is approximately 4 hours, which is longer than the average duration of stay before and after the match. Fig. 11 depicts the travel time from the Al-Bayt stadium to the final destinations of fans during at after the match from 4pm to 2am the next day. It shows that the average travel time is around. The histogram in Fig. 12 provides a macro view of the number of stadium visitors in relation to their dwell time, while the instantaneous distribution of their dwell time provides a micro view of their behavior. It shows that the majority of visitors have short dwell times and are just passing by. Fig. 13 provides a relative representation of the number of commuters as a function of their dwell time, revealing that the 90th percentile of commuters have a dwell time of approximately 30 minutes.

VI. CONCLUSION AND FUTURE WORKS
In this work, we dive into the exciting world of people movement at the FIFA Arab Cup TM held in Qatar in November 2021. By harnessing the power of fixed Bluetooth sensing devices deployed throughout the city and around the Al-Bayt stadium, we were able to capture a wealth of data and derive valuable insights into commuter behavior.
Our comprehensive analysis of this data offers a unique glimpse into the dynamics of human mobility in a highstakes, high-pressure environment. We discuss the challenges and opportunities presented by this type of research and highlight the potential impact it can have on city planning, public transport, and event management.
There are various potential avenues for continued exploration in the coming months. First, the group is working with a road operator in Qatar to install extra WaveTraf sensors around the stadiums. This initiative aims to expand the coverage area and improve the precision of vehicle recognition along the road to the stadium. Second, the aim is to ensure that systems operated by other suppliers have access to stadium detections. This would allow the team to employ sophisticated data filtering techniques to ensure that only user devices with the stadium as their final destination are considered. Lastly, historical data collected thus far can be used to construct forecast models for stadium visits during future events. The team is continuously enhancing the system's functioning and plans to provide it as a service for the eagerly anticipated FIFA World Cup Qatar 2022 TM .