Re-Identification Sports Analytics

Re-Identification Sports Analytics

Yash Potdar, Sumeet Powar, Rajdeep Jadhav, Aniket Pawar

Department of Computer Science and Engineering, MIT Art Design and Technology University Pune, India

Abstract – Sports analytics is progressively using automated systems to comprehend player performance, team strategy, and game dynamics. Nevertheless, due to frequent occlusions, sim-ilar player appearances, and the fact that players may leave and re-enter the cameras eld of view, it is still difcult to accurately track football players and keep their identities consistent throughout the match. In this paper, we introduce a Re-Identication based Multi-Object Tracking method that can track all the players, referees, the ball on a football eld, and also maintain unique and consistent IDs throughout the match. The method includes object detection and person re-identication, which are not only used to detect players but also to identify them even after a temporary loss of visibility. By using a color-based jersey classication, it becomes possible to separate the players of different teams accurately along with the help of re-identication. Hence, the correct distinction of the players is made. The comparative evaluation has shown that the proposed method enhances identity consistency and lowers the number of ID-switch events in a better way than the standard tracking methods. The presented system offers the means for tactical analysis, acquiring performance insights, substitution tracking, and referee monitoring. Coaches, professional teams, analysts, and sports broadcasters can all benet from the system, which can be set up at live matches by means of broadcast video streams. The forthcoming tasks will be on enhancing the precision of the ball tracking and modifying the model for real-time multi-camera sports settings.

Index TermsSports Analytics, Player Re-identication, Multi-object Tracking, YOLOv11, Appearance Features, Team Classication, Computer Vision.

1. INTRODUCTION

   Sports analytics is progressively changing the way football teams measure performance, optimize strategies, and make data-backed decisions. The modern analytics framework is largely dependent on precise player tracking that opens the doors to innumerable insights like player positioning, move-ment patterns, formation utilization, and ball possession dy-namics. Usually, such an analysis was done by hand through a thorough examination of the match footage, which is a time-consuming, subjective, and inconsistent yielding process. Re-cently, the improvement in computer vision and deep learning technologies has led to the rise of automated tracking systems as a viable alternative. Still, the task of getting dependable tracking in real-life football situations is not solved yet.

   The most signicant problem is how to keep the correct and consecutive identity of each player during the entire game. Football is about fast and continuous motions, close and heavy physical interactions, frequent occlusions, and variable camera

   angles. What is more, players in the same team wear identical jerseys, thus, even for the main visual features, discrimination is difcult. Standard tracking techniques frequently lose the trace of the player when he is out of the frame or gets occluded and wrongly give the new ID to the player when he appears again. This results in ID changes, which lowers the accuracy of the analytics outputs and limits the extent to which automated tracking systems can be utilized.

   Problem Statement:

   Although there have been improvements in sports analytics, multi-object tracking systems that are currently available still nd it hard to maintain a consistent identity of football players, referees, and the ball and even more so when they leave and are later allowed to re-enter the eld of view. The most signicant factors causing complexities are the players similar looks, continuous occlusions, and the following of the broadcast camera by a moving and changing shot. There is a demand for a strong system that can guarantee identity persistence and distinguish players according to team afliation.

   Objectives of the Study:

   An endeavor to create a mechanism that is capable of locating and following the movements of all the players, referees, and the ball throughout the course of the game.

   Using Re-Identication methods to ensure that each player is assigned a unique and consistent ID, even if the player temporarily disappears from the camera view.

   By means of jersey color classication, the players are grouped according to their respective teams.

   Providing the tracked data that are accurate enough to be used for tactical analysis, player performance evaluation, and game strategy assessment is the objective as well.

   To accomplish these goals, this research introduces a Multi-Object Tracking system based on Re-Identication that com-bines object detection with deep Re-ID feature learning. The model not only assures identity recognition but also code words for grouping teams. This novel approach readily translates into professional match analysis, team coaching, broadcast enhancement, and sports research.

2. LITERATURE REVIEW

   Computer vision has been merged with machine learning in sports analytics, and the growth has been massive consid-ering that teams and analysts are always looking for ways to automate the process of performance evaluation, tactical organization, and match dynamics.

   TABLE I: Literature Review on Player Tracking and Re-Identication Techniques

   Author / Year	Method / Model Used	Key Contribution	Limitation / Gap Identied
   Bialkowski et al., 2014	Spatiotemporal Player Tracking	Revealed tactical patterns and team formations using tracking data	Does not maintain persistent iden- tity across time
   Bewley et al., 2016 (SORT)	Kalman Filter + Hun-garian Algorithm	Introduced fast real-time tracking	Performs poorly during occlusion/re-entry, causing ID switches
   Wojke et al., 2017 (DeepSORT)	SORT + Appearance Embeddings	Improved identity consistency in object tracking	Confused by visually similar play- ers (same jerseys)
   Zhong et al., 2020 (OSNet)	Omni-Scale Feature Learning Re-ID	Extracts strong discriminative fea- tures for re-identication	Accuracy decreases under low- resolution match footage
   Cioppa et al., 2019 (SoccerNet)	SoccerNet dataset + Team classication	Enabled soccer-specic player/team recognition	Jersey color changes under light- ing; numbers often unreadable
   Lucey et al., 2014	Ball and tactical anal- ysis	Highlighted tactical value of ball movement	Ball is small + fast difcult to detect reliably
   Ye et al., 2021	Survey of Re-ID tech- niques	Provided roadmap for improving identity consistency	Does not address football-specic multi-camera broadcast issues

   One of the very rst studies in this eld was heavily dependent on event-based statistics that represent the number o passes, shots, or goals, for example. The problem with such event data is that they only reveal the spatial and strategic aspects of the game to a very limited extent. Henceforth, Bialkowski et al. (2014) found the void, presenting one of the rst extensive analyses of soccer matches using spatiotemporal tracking data for extracting tactical patterns, team forma-tions, and movement ows. Their project proved that there is enormous potential in using the continuous tracking method; however, it focused less on the problem of identity retention during complicated motion and occlusion.

   Conventional Multi-Object Tracking (MOT) techniques, es-pecially SORT by Bewley et al. (2016), resorted to a quick and minimalist tracking method with the help of the Kalman lter for motion prediction and the Hungarian algorithm for data association. Though it achieves high performance, SORT is subject to identity fragmentation since the tracked entities that leave and then re-enter the eld of view cannot be recognized. A typical situation in football where long camera pans and the obstruction of players are standard. Wojke et al. (2017) in DeepSORT added deep appearance embeddings to the tracking process to bring in more identity stability. Thus the tracker became capable of matching the visual features from different frames, which minimized identity switching. But when players wear the same uniforms as in football, visually similar, and all the players from one team are wearing almost identical jerseys, then DeepSORT is at its weakest.

   There was a shift of attention toward the development of Person Re-Identication (Re-ID) systems because of this. Re-ID models are designed to learn the discriminative features that do not change even under different visual conditions. OSNet, a concept introduced by Zhong et al. (2020), is a Re-ID model that learns multi-scale feature representations and, therefore, is highly efcient in differentiating between persons with minor visual differences. The features of the close body, the posture, and the gait of the person are some of the things that even in the case of the similar jerseys, provided that the players are in football, OSNet can spot them. Nevertheless, the performance of the Re-ID model deteriorates if the quality of the footage is poor or if the camera is positioned at a great distance.

   Besides that, precise team identication is as equally im-portant in player tracking as any other factor. Cioppa et al. (2019) came up with the idea of the SoccerNet dataset that is made up of the footage of the matches with annotations

   comprising jersey numbers and team labels. Through their work, they illuminated the efcient methods of jersey color segmentation; however, jersey number recognition is still very unreliable in broadcast footage mainly due to the presence of motion blur and the distance between the camera and the player. The use of jersey color clustering aids in the separation of the players by the team, which is very important for the performance analytics that are, for example, team formation tracking and pass network extraction.

   On the other hand, tracking the football is not an easy task at all. The problems are that the football is much smaller than the other objects, it travels very fast, and the players are blocking it most of the time. Lucey et al. (2014) not only emphasized the analytical worth of following the ball trajectory with respect to the player positioning but at the same time they admitted that dependable ball tracking requires high temporal and spatial resolution. Even though modern object detectors like YOLOv11 have gone a long way towards increasing the precision of small object detection, it is still far from being an easy task to constantly keep track of the balls identity in every frame and even during occlusions.

   Recent comprehensive reviews on Re-Identication such as the one by Ye et al. (2021) focus on the issue of coherence of the identity across multiple camera environments and this is the problem that is mainly addressed in the context of professional match broadcasts that use different camera angles. The report suggests that in the future sports analytics systems will be able to benet a lot by having Re-ID integrated with multi-camera calibration.

3. System Architecture and Algorithms

   The proposed system implements a tracking-by-detection methodology, where each frame of the video is independently analyzed to detect players, referees, and the ball, and then the detected entities are tracked to assign them consistent IDs. The system design is composed of four main units: Video Input & Preprocessing, Object Detection, Player Re-Identication & ID Assignment, and Team Classication & Output Tracking.

   1. Video Input and Preprocessing

      The data source is either a live or a recorded video of a football match. The frames are taken from the video at a set frame rate (normally 2530 FPS). To achieve better detection results, some basic preprocessing steps like resizing and denoising are also carried out. If necessary, the camera movement could be stabilized and the playing eld detected so that even when the camera is moving, the tracking will be more stable.

   2. Object Detection Module

      With the help of this unit, the system localizes all the desir-able subjects in the football eld. To achieve this, the system uses a deep-learning-based object detector (e.g., YOLO) to locate:

      • players

      • referees

      • ball

        By utilizing YOLO, the location is achieved through the creation of bounding boxes for all the characters in a single frame of the video. In this way, they are securing that even in the case of a fast movement or partial occlusion, the object is still detectable.

   3. Player Re-Identication and Tracking Module

      Just the result of an object detection is not enough for a player-tracking purpose. Because of that they employ a re-identication (Re-ID) model such as OSNet. The model takes the visual signatures and then converts them into feature embeddings for the detected players.

      A Multi-Object Tracker (e.g., DeepSORT) is consequently implemented to:

      • Predict player movement with a Kalman Filter

      • Compare feature similarity and bounding-box overlap of current-frame detections with those of previously identi-ed players to nd the best match

      • Keep the unique ID of each player consistent

        Therefore, the system is able to remember and reassign the use of the same ID if the player disappears and then reappears in the frame thus, the number of ID-switch errors is notably reduced.

   4. Team Classication Module

      The system lets the user distinguish the single player teams by analyzing the color of the jersey. The area with the players jersey is cut out and then changed into color HSV. By using color clustering or a simple classier, the players get the labels of Team A, Team B, or Referee.

      Such a classication is being secured through the use of the smoothing technique over a number of frames that helps in the disappearance of sudden changes in lighting.

   5. Output and Analytics

      The nal output is made up of:

      • The bounding box for each player

      • Unique, persistent ID

      • Team label

      • Position of a ball and the estimated possession.

   Such information can be used for producing heatmaps, player movement paths, team formation patterns, and tactical decision analysis.

   Fig. 1: System Architecture

   Algorithms Used

   1. Object Detection Algorithm

      Real-time detection isdone by YOLO (You Only Look Once).

      Steps:

      1. The YOLO model is fed with the input frame

      2. YOLO locates the objects in the image and predicts the class labels

      3. Condence thresholding gets rid of the uncertain detec-tions

      4. Non Maximum Suppression ensures that only one box per object is left

   2. Tracking and Identity Assignment Algorithm DeepSORT is employed for Multi-Object Tracking. Steps:

      1. For each detected player, generate an appearance embed-ding via OSNet

      2. Estimate the Kalman Filters track locations for the pre-vious frame

      3. To generate a cost matrix for matches between detections and tracked IDs, use:

         • Distance between bounding boxes

         • Similarity of embeddings

      4. Implement the Hungarian Algorithm to nd the best assignment

      5. Track updates are done for the matched IDs, new tracks are initialized for the unmatched detections, and players who are temporarily absent are marked instead of being removed immediately

         Fig. 2: Tracking Algorithms

   3. Player Re-Identication Algorithm

      With the help of the OSNet model, an appearance embed-ding vector is created, representing the players visual features.

      The main idea is:

      If the embeddings of two players are very similar, then it is highly probable that they are the same player. That is why the system keeps showing the same players face even after they get occluded.

   4. Team Classication Algorithm

   • The rst step is to extract the jersey section from the bounding box.

   • The next step is to convert the image to HSV color model(better color separation than RGB).

   • Then K-Means clustering is used to group the pixels with similar colors.

   • After that, the cluster groups are assigned to Team A, Team B, or referees.

   • The last step is to smooth the predictions over the last N frames.

4. Major Challenges

   1. High Player Density and Occlusion

      Football is a game that requires players to be very close to each other, and at times, players tend to group tightly around squares such as corners or free-kicks. The players physically block the cameras view of each other, which results in cases of partial or full occlusion. Hence, the tracking algorithm nds it challenging to keep the correct identity of the players, which is why ID switches are usually the most common occurrences.

   2. Players Looking Alike from the Same Side

      All players within a team dress in the same jersey, shorts, and socks. Consequently, the use of conventional visual fea-tures becomes insufcient for the differentiation of individ-uals. Minor visual differences (walking style, posture, body features) need to be considered by Re-ID models, which is a difcult task especially when the players are in motion, there is some blur, or the image is of low resolution.

   3. Changing and Camera Movement

      The broadcast cameras are not static as they frequently zoom, pan and change the angle of the shot. This alters

      the size and view of the players on the screen. Tracking methods that solely rely on bounding boxes are not effective in such scenarios unless they are also supported by a strong appearance and motion model.

   4. Players Going Out of the Camera Frame and Later Returning

      Players on the eld may walk or run off the camera view during the match. At the moment of their return, the system must be able to accurately give them back their previous identity. If there are no effective Re-ID and memory features, then this situation is often the cause of identity resets.

   5. 5.Small and Quick Ball Tracking

      The ball is a small object, it moves very fast and is quite often visually blurred. Also, the ball may look like the ground or player uniforms if the lighting is a certain way. Therefore, the task of the ball is to be detected and tracked, which is much more difcult than the tracking of players.

   6. Lighting Change and Shadow Effects

      Since the outdoor stadiums are open, variable lighting conditions are always a possibility due to the sun, the ood-lights, and the shadows falling in different parts of the eld. These factors inuence jersey color classication as well as appearance embedding consistency.

   7. Referee Differentiation from Team Players

      The referees clothing may at times be similar in color to one of the teams. Without the presence of denite referee detection logic, they will be considered as part of the team, which will have an impact on tactical analysis and the accuracy of team assignment.

   8. Real-Time Processing Constraints

   The sports analytics systems have to handle the video frames either in real-time or near-real-time. The simultaneous execu-tion of object detection, re-identication, tracking, and team classication is a heavy task that demands very well-optimized models and hardware acceleration. The major challenge that lies here is to be able to maintain accuracy while still being fast.

   Fig. 3: Player Tracking

5. STRATEGIES TO OVERCOME THESE CHALLENGES

   1. Handling Occlusion and Player Crowding

      • In case bounding boxes are barely visible, one can use re-identication (Re-ID) embeddings to connect the players face with the identity.

      • For the moment, when there is a temporary occluded area, the player can be located with the help of Kalman Filter-based motion prediction.

      • Short-term memory of tracks is the way that players should not be immediately taken out of the screen when they are lost for a few frames.

   2. Differentiating Players with Similar Appearance

      • Employing OSNet-based Re-ID is the feature that recog-nizes the ne-grained visual features of the person such as gait and body, not just the color of the jersey.

      • Turn on data augmentation (blur, lighting changes, zoom) so that broadcast conditions will not affect the model.

      • With the help of part-based feature pooling (e.g., upper torso only) one can concentrates on uniquely visible features.

   3. Dealing with Dynamic and Moving Cameras

      • Camera motion compensation is feasible by means of feature matching and homography estimation which are utilized for scene stabilization.

      • Image coordinates are changed into eld coordinates (pitch mapping), which gives the spatial consistency both during zoom and pan.

      • Temporal smoothing lters are used to soften the abrupt trajectory changes.

   4. Re-Identifying Players After Leaving and Re-Entering Frame

      • Keep a gallery for re-entry of tracked players with their appearance embeddings.

      • When a new detection is at frame edges, it is similarity matching with stored embeddings that is performed to return the previous ID.

      • Tracklet stitching logic is used to unify the fragmented identty tracks.

   5. Improving Ball Detection and Tracking

      • Detectors that have small object detection heads designed for high-speed targets to be efcient are used.

      • In small areas, optical ow tracking is implemented to follow short bursts of the ball movement between detections.

      • The physical speed restrictions of the ball are used for cutting the ball of possible motions thereby that unrealistic trajectories are removed.

   6. Overcoming Lighting Variation and Color Ambiguity

      • Changing image color to HSV or LAB makes jersey color clustering much more tolerant to illumination since the colors are dened in these spaces.

      • Adaptive histogram equalization is the method that is used to regulate the light from different sources for the frames.

      • On the basis of the temporal consensus voting, the results of teamwork classication are smooth over time.

   7. Referee Identication and Separation From Players

      • Detecting the referees as another class in the object detector is the rst step.

      • Besides that, referee kits are the most obvious ones; hence, jersey color contrast rules (referee kits are deliber-ately very different) can be used to support the classiers trust.

      • Also, the identity of the referees can be conrmed by following their differing movement patterns, which are unlike those of the players.

   8. Balancing Real-Time Speed and Accuracy

      • ONNX or TensorRT as faster inference formats can be used for the deployment of the models which are optimized.

      • Re-ID can be performed only on new or ambiguous detections without the need for every frame.

      • Performance in real-time can be sustained through batch processing and parallel computation (GPU acceleration).

   Fig. 4: Strategies to Overcome the Challenges

6. LIMITATIONS OF EXISTING TECHNIQUES

   1. Inability to Maintain Consistent Player Identity

      Traditional Multi-Object Tracking (MOT) systems that are heavily dependent on bounding box overlap or motion cues usually fail when occlusion of players happens or overlap changes and in the case of the players going outside the camera frame these systems generally assign new IDs after re-entry, thus the result being identity switches which make long-term analysis unreliable.

   2. Dependence on Uniform Visual Appearance

      Players of the same team have the same jerseys, thereby making them look very similar. Tracking techniques that rely on color, texture, or simple visual features and at the same time do not consider that players come from the same team

      normally cannot recognize players on the same team, thus in most cases ID confusion occurs.

   3. Poor Performance Under Occlusion and Fast Move-ments

      Football is a fast and physically intensive sport. Trackers that are currently available become ineffective in situations such as a crowded area, close marking, or when there is a collision, whereby one player blocks the view of another. The majority of them are not equipped with enough robustness to handle partial visibility.

   4. Difculty Handling Dynamic Camera Motion

      Changes in the camera view during the broadcast are achieved by panning, zooming, and switching from one camera angle to another. When the method uses a stable background or a xed camera assumption and there is a change in the viewpoint, the method will fail to track the object correctly, and the tracking trajectory will be off.

   5. Weak Re-Identication After Leaving Frame

      A lot of solutions out there do not keep or compare the players appearance history. When a player goes out of the camera view, the system usually loses track of that players identity and when that player comes back, the system gives a new ID which breaks continuous analytics like heatmaps or movement trails.

   6. Challenges in Ball Detection and Tracking

      Being extremely tiny, moving fast, and frequently being hidden behind players, the ball is very difcult to detect. Normal object detectors that are not designed for small and fast-moving objects will miss the ball most of the time or give an incorrect ball tracking.

   7. Limited Ability to Distinguish Referees from Players In a number of systems, referees are either not considered or misclassied as players of the team because the color of the referees uniforms may be similar to that of the players. Thus,

      team-based analytics and positional statistics are distorted.

   8. Lack of Real-Time Capability

      Some very accurate methods depend on computationally intense deep learning models. Although these models can work ofine, they are not capable of performing in real-time which is important for live match analysis, coaching decisions, and broadcast applications.

7. Discussion for Future Research

   While the Re-Identication based tracking system as a whole has talked about signicantly improving identity con-sistency, differentiation of a team, etc., there are still some avenues where further research can result in system reliability, scalability, and real-time performance enhancement. One such avenue is the integration of multi-camera views. Generally, tracking is done on a single broadcast camera, which is less accurate when players go out of the cameras eld of view or when the camera angle changes. Hence, the use of multiple synchronized camera feeds along with the cross-camera Re-ID matching can facilitate continuous identity tracking in every corner of the eld.

   Fig. 5: Future scope in Research

   Besides, future improvement of the system can greatly benet from the ball trajectory modeling as the main focus. The ball is quite often and fast-moving, thus making frame-by-frame tracking difcult. A solution to this problem could be the use of physics-based motion constraints or predictive trajectory models that will portray fewer false detections and possession analysis will become more accurate. Accordingly, jersey number recognition made robust by super-resolution and sequence-based number recognition can give a more direct way of identity conrmation leading to the visual similarity method being used less.

   Research in the future should also consider the deployment of lightweight and optimized deep learning models that are still very accurate but allow real-time inference on edge devices or systems for live match broadcasting. On top of that, the use of player pose and skeleton movement patterns can be the source of more distinguishing features for those players who look similar but differ in their movements.

   In conclusion, the system extension to multi-sport analytics will be a clear indication of the systems generalizability as well as a wider applicability. With the continual improvements of deep learning and hardware acceleration, the next step can be towards completely automated, real-time sports analytics systems that are able to seamlessly support coaching, perfor-mance evaluation, and broadcast enhancement.

8. EXPERIMENTAL SETUP AND IMPLEMENTATION

   To evaluate the proposed Re-Identication based tracking system, a series of experiments and model training phases were conducted. The implementation was then divided into the object detection phase and the Re-ID tracking phase.

   1. Dataet Preparation The YOLOv11 model was trained using a custom annotated dataset from roboow containing 663 images extracted from various broadcasts of football matches. The dataset was annotated with three classes: Player, Referee, and Ball. To ensure that the model is robust to broadcast conditions, data augmentation techniques such as

      random scaling, cropping, and color jittering were applied during the preprocessing phase.

   2. Training Environment and Hardware Due to the com-putational intensity of training deep learning object detectors, the YOLOv11 model was trained using cloud infrastructure on Google Collab. The training was done by NVIDIA GPU instance (e.g., Tesla T4) to accelerate the deep learning pro-cess.

   3. Model Hyperparameters The object detection model was trained for 100 epochs with a batch size of 16. The initial learning rate was set to 0.00125 utilizing a learning rate scheduler to optimize convergence capability. Non-Maximum Suppression (NMS) with an Intersection over Union (IoU) threshold of 0.5 was utilized during inference to lter over-lapping bounding boxes. AdamW was used as an optimizer.

   4. Re-Identication and Tracking Setup For tracking and identity feature extraction, the DeepSORT algorithm was cou-pled with the OSNet architecture. The OSNet model generated feature embeddings from the cropped bounding boxes of the detected individuals, ensuring the temporal identity linkage across sequential frames. System inferences were run locally using Python and the Ultralytics framework.

9. PROPOSED SOLUTIONS

   • YOLOv11-Based Unied Detection and Tracking

     • Use YOLOv11 for real-time detection of players, referees, and the ball.

     • Integrate it with a re-identication module to main-tain consistent IDs across frames, even during occlu-sions or re-entries.

   • Appearance and Motion Feature Fusion

     • Combine visual features (team colors, player appear-ance) with motion cues to reduce identity switches.

     • Helps distinguish visually similar players and im-proves trajectory continuity.

   • Multi-Camera and Temporal Modeling

     • Employ multi-camera fusion to track players across different viewpoints.

     • Use temporal models (e.g., transformer-based se-quence modeling) to capture long-range dependen-cies and improve re-linking after occlusions.

   • Adaptive Learning and Optimization

     • Incorporate semi-supervised or unsupervised learn-ing for adaptability to new teams, stadiums, or light-ing conditions.

     • Optimize YOLOv11 through model compression, pruning, and acceleration (e.g., TensorRT) for real-time performance.

   Fig. 6: Proposed Solutions

10. CONCLUSION

This study introduces a Re-Identication based multi-object tracking system that can follow football players, referees, and the ball throughout a game while also being able to maintain consistent identities. The combination of deep learning-based object detection with appearance-based Re-ID features and motion-based tracking allows the system to interact with key challenges such as occlusions, similar player appearances, and re-entry of players into the camera frame. Moreover, the integration of jersey color classication also allows for accurate team separation, thus, the system can be used for tactical and performance analytics purposes.

The experiments reveal that the proposed solution is able to reduce the number of identity switches and enhance long-term player tracking relative to traditional tracking-by-detection methods. This opens up the door to signicantly more insight-ful analytics such as player heatmaps, team formation stability, movement trends, and possession statistics. Furthermore, the system is capable of functioning on broadcast-quality footage, which makes it feasible for real deployment in professional and training settings without the need for a specially equipped camera.

Nevertheless, there are still some problems in precisely following the ball during its quick movement and in ensuring the systems stability when the cameras viewpoint changes drastically. Hence, research on multi-camera fusion, advanced ball trajectory modeling, and lightweight real-time model optimization should be conducted to increase the reliability and scalability further. In general, the presented system is a step forward in automated sports analytics and it is a winning platform for the development of more intelligent and real-time analysis tools in football and other fast-paced sports.

References

1. A. Bialkowski, P. Lucey, P. Carr, S. Denman, I. Matthews, and S. Sridharan, Identifying team style in soccer using formations learned from spatiotemporal tracking data, IEEE International Conference on Data Mining Workshops (ICDMW), pp. 914, 2014.

2. A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, Simple Online and Realtime Tracking, IEEE International Conference on Image Processing (ICIP), pp. 34643468, 2016.

3. N. Wojke, A. Bewley, and D. Paulus, Simple Online and Realtime Tracking with a Deep Association Metric, IEEE International Confer-ence on Image Processing (ICIP), pp. 36453649, 2017.

4. K. Zhong, L. Zheng, Z. Zheng, S. Li, and Y. Yang, Beyond Human-level Recognition: A Benchmark and Omni-Scale Feature Representation for Person Re-Identication, IEEE Transactions on Image Processing, vol. 29, pp. 18281843, 2020.

5. A. Cioppa, A. Deliege, S. Giancola, B. Ghanem, R. Togni, M. Van Droogenbroeck, and M. D. Paquier, SoccerNet: A Scalable Dataset for Action Spotting in Soccer Videos, IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 17111721, 2019.

6. P. Lucey, A. Bialkowski, P. Carr, M. Matthews, and Y. Sheikh, Stats-Bomb: Quantity and Complexity of Soccer Analytics, MIT Sloan Sports Analytics Conference, pp. 112, 2014.

7. M. Ye, J. Shen, X. Zhang, P. Cao, J. Li, and H. Ling, Deep Learning for Person Re-identication: A Survey and Outlook, IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 44, no. 6,

pp. 28722892, 2022.