A Confidence-Guided Multi-Object Tracking Method for Sports Videos Using Jersey Semantic Fusion

Summary

This protocol describes a confidence-guided multi-object tracking workflow for sports videos that integrates jersey color and player number information to improve trajectory association and identity consistency under confidence fluctuations, occlusion, and visually similar athlete appearances.

Abstract

Multi-Object Tracking (MOT) in sports videos provides trajectory information for tactical analysis, player behavior interpretation, and automated statistical analysis. However, sports scenarios frequently involve rapid direction changes, abrupt stops, frequent occlusion, and visually similar teammates, resulting in fluctuations in detection confidence and identity confusion during inter-frame association. To address these challenges, this study presents JerseyTrack, a confidence-guided sports MOT method based on jersey semantic fusion. The workflow adaptively partitions detection boxes into high-, medium-, and low-confidence intervals according to the confidence distribution of each frame. High-confidence detections are associated primarily using motion similarity, whereas medium- and low-confidence detections additionally incorporate jersey color and player number features extracted through color clustering and a lightweight Optical Character Recognition (OCR) model. These semantic features are fused with motion information during supplementary matching to improve trajectory continuity and identity consistency. The method was evaluated on the SportsMOT dataset. JerseyTrack achieved 88.9% Multiple Object Tracking Accuracy (MOTA), 74.3% IDentity F1 Score (IDF1), and 69.9% Higher Order Tracking Accuracy (HOTA), demonstrating improved tracking performance under the evaluated sports tracking setting. This protocol provides a reproducible workflow for integrating confidence-guided association with jersey semantic information to improve multi-object tracking in sports videos.

Introduction

Multi-object tracking (MOT) provides continuous object localization and identity maintenance in video sequences and supports athlete trajectory extraction, tactical analysis, and automated statistical analysis in sports-video applications^1,2,3. Reliable visual information is also related to sport-specific decision-making and performance assessment in sports science, further emphasizing the value of stable video-derived movement data for downstream sports analysis⁴. In sports scenarios, the reliability of tactical data depends on the continuity of tracking trajectories and the consistency of target identities.

Compared with general MOT scenarios, sports videos contain rapid direction changes, abrupt stops, jumps, dense interactions, and frequent occlusion. These factors cause substantial fluctuations in detection-box confidence and increase the frequency of low-confidence detections, which may interrupt trajectory association^5,6. Uniform team jerseys further reduce the discriminative ability of general appearance features and increase identity confusion among visually similar teammates^7,8. When occlusion and appearance similarity occur in the same sequence, detection confidence decreases and target appearance information becomes incomplete, increasing the difficulty of trajectory association and identity maintenance^9,10. In MOT, re-identification (Re-ID) refers to the process of associating the same target identity across frames, occlusion periods, or re-entry cases using identity-related visual evidence. In team-sport videos, general Re-ID cues may be weakened because athletes from the same team often share similar uniforms and body appearances. The technical literature review indicates that trajectory fragmentation and identity confusion in sports MOT are commonly addressed through two complementary approaches: detection-confidence utilization and identity-cue enhancement. JerseyTrack is positioned at the intersection of these approaches by regulating the use of jersey semantic cues according to detection quality rather than applying color and player-number information uniformly to all detections.

This study presents JerseyTrack, a confidence-guided sports MOT method based on jersey semantic fusion. The method is designed for team-sport videos in which athletes wear visible jerseys and detection results provide bounding boxes with confidence scores. JerseyTrack combines four main components: athlete detection, confidence-level partitioning, jersey semantic feature extraction, and cascaded inter-frame association. Detection confidence is used to adaptively partition detections into high-, medium-, and low-confidence groups, which are processed using different association strategies. High-confidence detections are associated primarily through motion information, whereas medium- and low-confidence detections additionally incorporate jersey color and player-number information to improve identity maintenance when visual evidence is weak. The method is intended for sports video analysis tasks that require stable athlete trajectories, such as tactical analysis and player behavior interpretation.

The proposed approach also has operational limitations. Jersey semantic extraction may be affected by severe occlusion, motion blur, low image resolution, abnormal jersey poses, or invisible player numbers. In these cases, jersey-based semantic cues may become incomplete, and the association process relies more heavily on motion consistency and multi-frame verification. Therefore, the performance claims in this study are restricted to the evaluated datasets and experimental settings. Experiments on the SportsMOT dataset show that JerseyTrack improves the evaluated tracking metrics and reduces identity-switching events under the tested sports tracking conditions. The main difficulty of MOT in sports videos lies in maintaining trajectory continuity and identity consistency under rapid motion, occlusion, and appearance similarity. Existing studies relevant to JerseyTrack can be broadly categorized into two research directions: the use of detection confidence for trajectory association and the enhancement of identity cues through sports-specific semantic features.

Detection confidence has been widely used to estimate the reliability of detection boxes and to guide trajectory association in MOT. Early tracking methods usually treated confidence as a binary filtering criterion, in which detections below a fixed threshold were removed to reduce false positives^11,12. This strategy reduces noisy detections but may also discard true targets under occlusion, rapid motion, or low image quality, resulting in trajectory fragmentation^13,14,15. Similar filtering strategies have also shown that fixed confidence thresholds are sensitive to scene variation and detection quality^16,17. To improve the use of low-confidence detections, ByteTrack introduced an association strategy that retains low-confidence boxes as candidate targets for secondary matching and links them with existing trajectories through motion similarity and trajectory history¹⁸. McByte further extends sports MOT association by incorporating temporally propagated segmentation masks as an association cue, improving robustness under fast motion, occlusion, and camera shift conditions without additional per-video training¹⁹. Related studies have also adjusted the use of low-confidence detections to retain weak but potentially valid targets during association^20,21,22. Confidence-adaptive association strategies subsequently refined matching criteria according to motion consistency and detection reliability^23,24,25. Trajectory refinement methods based on detection-box regression and trajectory smoothness optimization have also been used to reduce trajectory fragmentation^26,27,28. Additional refinement strategies provide complementary support for maintaining trajectory continuity under complex motion conditions²⁹. Temporal coherent object flow further models object-level temporal consistency across frames, providing another association-oriented approach for reducing trajectory interruption in complex MOT scenarios³⁰. Tracker fusion methods also learn from the outputs of multiple trackers and use learner-based selection or fusion strategies to improve tracking robustness across different MOT benchmarks³¹. Collectively, these studies indicate that confidence-aware association helps recover interrupted trajectories; however, confidence and motion cues provide limited identity information when players from the same team have similar visual appearances. Although these methods effectively alleviate trajectory fragmentation in general scenarios, they face inherent limitations in sports settings because players on the same team often have highly similar visual appearances, and relying solely on confidence and motion information cannot provide a sufficient basis for identity differentiation^32,33,34. Even when low-confidence boxes are effectively recalled, feature similarity can still lead to identity switching, making it difficult to satisfy the stringent identity consistency requirements of sports analysis.

Sports-specific identity cues have also been used to improve identity discrimination in athlete tracking. Jersey semantic information, such as team color and player number, provides identity cues that are more directly related to sports scenarios than general appearance embeddings^35,36,37. Jersey color has been used to support team-level discrimination and reduce cross-team confusion, whereas player number recognition provides a finer identity cue when the number region is visible and readable^38,39. Existing sports tracking studies have incorporated domain-specific visual features and jersey color recognition to improve player association under crowded sports conditions^40,41. Related work on jersey number recognition and synthetic number data further supports identity modeling under low-resolution or partially occluded conditions^42,43. These studies indicate that jersey semantics can strengthen identity representation in sports MOT. However, most semantic-enhanced tracking methods do not sufficiently model the relationship between detection confidence and semantic feature quality. High-confidence detections may already contain reliable spatial and appearance information, making repeated semantic extraction less efficient. Low-confidence detections often suffer from occlusion, blur, truncation, or low resolution, reducing the reliability of color and player-number cues. This limitation motivates a confidence-guided association design in which jersey semantics are introduced according to detection quality rather than applied uniformly to all detections. The reviewed studies show that confidence-aware association and jersey semantic modeling address different aspects of sports MOT. Confidence-based strategies mainly determine whether a detection should participate in association and how it should be matched with existing trajectories, whereas jersey semantic strategies primarily enhance identity representation through sports-specific cues. However, these two mechanisms are often designed as separate components. As a result, semantic cues are not always adjusted according to detection quality, and confidence levels do not directly regulate the use of color and player-number information during association. This separation limits the joint handling of trajectory fragmentation and identity confusion in team-sport videos. The remaining research gap lies in linking detection-confidence quality assessment with jersey semantic association within the same tracking workflow.

Protocol

This study involved secondary analysis of publicly available benchmark video datasets, namely SportsMOT, TeamTrack, SoccerNet Tracking, MOT17, and MOT20. No participants were recruited, no intervention was performed, and no new human-subject data were collected. Ethics committee approval was not required for this study under the applicable institutional requirements of Bangkok Thonburi University.

The following protocol describes the workflow used to reproduce JerseyTrack from data preparation through tracking evaluation. The workflow includes dataset preparation, detector preparation, jersey semantic extraction, confidence-level partitioning, confidence-guided association, tracking inference, and performance evaluation, as summarized in Figure 1. The required software, datasets, and hardware resources are listed in the Table of Materials.

Figure 1. Overall workflow of the JerseyTrack method. The workflow consists of jersey semantic feature extraction, initial object matching, confidence-guided supplementary matching, and feature fusion. Team-color and player-number features are extracted from detected athlete regions and incorporated into the association process to improve trajectory matching under uncertain detection conditions. Please click here to view a larger version of this figure.

1. Prepare the datasets and directory structure

1. Download the public benchmark datasets listed in the Table of Materials. Use SportsMOT as the primary dataset for detector preparation and tracking evaluation. Retain TeamTrack, SoccerNet Tracking, MOT17, and MOT20 for cross-dataset evaluation.
2. Store all datasets under a unified project directory. Ensure that the detector, semantic extraction module, the tracking module, and the evaluation toolkit use identical sequence identifiers.
3. Convert the official SportsMOT annotations into the detector-training format using the data-preparation script employed in this study. Execute the following command:
   ..\venv\Scripts\python.exe scripts\prepare_data.py --config configs\datasets.local.json --dataset SportsMOT --out data\processed\SportsMOT_yolo --splits train val.
4. The data-preparation script (scripts/prepare_data.py) is available in the JerseyTrack source code repository (https://doi.org/10.5281/zenodo.21274352). The required software dependencies are specified in environment.yml, including Python 3.12, PyTorch 2.11.0, and OpenCV 4.10 or later. Configure the software environment using the following command: conda env create -f environment.yml. Execute the data-preparation script using the following command: python scripts/prepare_data.py --config configs/datasets.local.json --dataset SportsMOT --out data/processed/SportsMOT_yolo --splits train val.
5. Verify that the conversion generates the detector-training configuration file: data\processed\SportsMOT_yolo\data.yaml and the conversion manifest: data\processed\SportsMOT_yolo\conversion_manifest.csv.
   NOTE: In this study, the converted SportsMOT training and validation dataset contained 90 sequences, 55,544 frames, and 608,152 annotated objects.
6. Inspect representative sequences to verify that the frame order, bounding-box coordinates, and identity labels remain consistent with the original benchmark annotations.
7. Preserve the converted annotation files without modification throughout detector preparation, tracking inference, and evaluation so that all methods use the same dataset partition and evaluation protocol.
   NOTE: The expected outcome of this section is a standardized SportsMOT detector-training directory containing the converted annotations, the conversion manifest, and the dataset split files required for detector preparation and tracking evaluation.

2. Prepare detector outputs

1. Fine-tune the object detector using the prepared SportsMOT training split. Optimize the detector using the classification loss and bounding-box regression loss defined in Equation 1. Use the detector input resolution, batch size, learning rate, number of training epochs, and other training parameters reported in the implementation setup and the Table of Materials
   L_det = L_cls + L_box   (1)
   Here, L_det  denotes the total detector loss, L_cls  denotes the classification loss, L_box and denotes the bounding-box regression loss.
   NOTE: The detector checkpoint used in the primary experiments (internal experiment identifier e3) was trained using the Ultralytics YOLO framework with the SportsMOT YOLO-format dataset configuration (data/processed/SportsMOT_yolo/data.yaml). Execute detector training using the following command: yolo detect train data=data/processed/SportsMOT_yolo/data.yaml model=yolo11x.pt epochs=80 imgsz=960 batch=16 lr0=0.01 project=outputs/formal name=checkpoints/detector device=0. After training, use the resulting best-performing checkpoint to generate detector outputs for the validation sequences using the following command: python scripts/formal_detect_yolo.py --dataset-root datasets/SportsMOT/dataset --split val --model outputs/formal/checkpoints/detector/weights/best.pt --out-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_threshold_sweep/t005/detections --imgsz 960 --conf 0.05 --device 0 --batch 16.
2. Save the trained detector checkpoint, the corresponding metadata file, and the training log after training.
   NOTE: In this study, the detector checkpoint used for the formal experiments was saved as:
   outputs\formal\checkpoints\detector\sportsmot_yolo11x_i960_e3_best.pt. The corresponding metadata file was saved as: outputs\formal\checkpoints\detector\sportsmot_yolo11x_i960_e3_best.json. The detector-output directory used for the main SportsMOT validation experiments was: outputs\formal_runs\sportsmot_val_yolo11x_i960_e3_threshold_sweep\t005\detections.
3. Run the trained detector on each validation sequence to generate athlete bounding boxes and confidence scores. Export one detection file for each sequence containing the frame index, bounding-box coordinates, detection confidence, and class label.
   NOTE: In the SportsMOT validation run used for the main experiments, a confidence threshold of 0.05 generated 343,990 athlete detections.
4. Inspect the detector-output directory before tracking inference. Confirm that every sequence has a corresponding detection file, that the frame indices match the prepared image sequence, and that every detection record contains a confidence score.
   NOTE: The expected outcome of this section is a complete detector-output directory that serves as the input for jersey semantic extraction, confidence-level partitioning, and tracking association.

3. Extract jersey semantic features

1. Use the detector-output files to crop athlete regions from each video frame. Save each cropped athlete image with its sequence identifier, frame index, and detection index. Exclude invalid crops with missing image content or bounding boxes extending outside the image boundary. Maintain the linkage between each crop filename and its original detection record so that the extracted semantic features can be matched to the corresponding detection during tracking association.
2. Extract jersey color features from the cropped athlete images using the semantic extraction script employed in this study.
   NOTE: The semantic feature extraction scripts (scripts/extract_fast_color_semantics.py and scripts/formal_extract_semantics.py) are available in the JerseyTrack source code repository (https://doi.org/10.5281/zenodo.21274352). No separate configuration file is required; all runtime parameters are provided through the command-line arguments.
   Generate jersey-color descriptors using the following command: python scripts/extract_fast_color_semantics.py --detections-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_threshold_sweep/t005/detections --frames-root datasets/SportsMOT/dataset/val --out-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_t005_color. Generate player-number semantic features using the OCR model with the following command: python scripts/formal_extract_semantics.py --detections-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_threshold_sweep/t005/detections --frames-root datasets/SportsMOT/dataset/val --out-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_t005_ocr. The generated jersey-color descriptors and player-number probability outputs are linked by sequence identifier and combined into the semantic-feature files used during tracking association.
3. Convert each cropped athlete image to the Hue, Saturation, Value (HSV) color space. Extract foreground pixels from the jersey region and summarize the dominant jersey color using K-means clustering with K = 3. Normalize the resulting color descriptor using L2 normalization before feature comparison.
4. Train the player-number recognition branch using cropped athlete images with an input size of 128 × 64 pixels. Generate player-number pseudo-labels using the weakly supervised labeling procedure employed in this study. Exclude crops with invisible, unreadable, severely occluded, or low-confidence player-number regions from Optical Character Recognition (OCR) loss calculation.
5. Optimize the OCR branch using the cross-entropy loss defined in Equation 2.
   (2)
   Here, L_ocr denotes the OCR loss, y_i denotes the supervised player-number label, and denotes the predicted player-number category.
6. Optimize the semantic feature extraction module using the adaptive optimizer, learning rate, batch size, weight decay, and training duration listed in the Table of Materials. Combine the detector loss and OCR loss using the total training objective defined in Equation 3.
   L_total = L_det + γL_ocr   (3)
   Here, L_total denotes the total training loss, L_det denotes the detector loss defined in Equation 1, L_ocr denotes the OCR loss defined in Equation 2, and denotes the user-defined weighting coefficient for the OCR loss.
7. Apply the trained OCR branch to each cropped athlete image. Save the predicted player-number category or probability vector for each crop. If the player number is not visible or the OCR confidence is below the threshold defined in the implementation, record the player-number feature as unavailable rather than forcing a prediction.
8. Combine the jersey color descriptor and player-number output into the semantic-feature file used by the tracking module.
   NOTE: In the main SportsMOT validation run, the semantic extraction script processed 343,990 detections without missing frames or invalid crops. In the current revision package, the semantic extraction summary reported an overall color-and-OCR semantic extraction accuracy of 86.7% under the evaluated SportsMOT setting. The expected outcome of this section is a semantic-feature file in which each valid detection record is linked to a jersey color descriptor, a player-number output when available, and a flag indicating whether semantic information is available for tracking association.

4. Partition detection confidence levels

1. Load the detector-output file for each video sequence. Collect the confidence scores of all athlete detections in each frame.
2. Partition the frame-level confidence scores into high-, medium-, and low-confidence intervals using K-means clustering with K = 3, following the confidence-guided association workflow shown in Figure 2. Determine the adaptive confidence thresholds by minimizing the within-cluster variance according to Equation 4.
     (4)
   Here, s_d denotes the confidence score of detection d; D₁, D₂, and D₃ denote the high-, medium-, and low-confidence intervals, respectively; μ₁, μ₂, and μ₃ denote the mean confidence scores of the three intervals; and τ₁ and τ₂ denote the adaptive confidence thresholds obtained from the K-means clustering results.
   NOTE: The confidence-partitioning script (scripts/formal_partition_confidence.py) is available in the JerseyTrack source code repository (https://doi.org/10.5281/zenodo.21274352). The confidence-partitioning module uses a NumPy-based one-dimensional K-means clustering procedure with K = 3. No separate configuration file is required; the input directory, output directory, and clustering parameter are specified through the command-line arguments. Execute the confidence-partitioning procedure using the following command: python scripts/formal_partition_confidence.py --detections-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_threshold_sweep/t005/detections --out-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_threshold_sweep/t005/confidence --k 3. The required software dependencies, including the NumPy version, are specified in environment.yml.
3. Run the confidence-partitioning procedure using the detector-output directory as the input.
   NOTE: In this study, the detector-output directory was: outputs\formal_runs\sportsmot_val_yolo11x_i960_e3_threshold_sweep\t005\detections. The confidence-partition output directory was: outputs\formal_runs\sportsmot_val_yolo11x_i960_e3_threshold_sweep\t005\confidence. The generated output files included the per-sequence confidence CSV files, _confidence_frame_summary.csv, and _confidence_runtime.csv.
4. Assign a confidence-level label to every detection. Label high-confidence detections for motion-based association, medium-confidence detections for motion-semantic association, and low-confidence detections for trajectory recovery only. Preserve the original confidence score together with the assigned confidence-level label.
5. Inspect the confidence-score distributions and representative frames, as illustrated in Figure 3. Verify that high-confidence detections correspond predominantly to complete and reliable athlete bounding boxes, whereas medium- and low-confidence detections mainly contain partial, occluded, blurred, or weak detections.
   NOTE: The expected outcome of this section is a confidence-partition file containing the detection index, original confidence score, assigned confidence level, and frame-level adaptive confidence thresholds for every detection.

Figure 2. Confidence-guided association strategy. The workflow partitions detection confidence into high-, medium-, and low-confidence intervals using adaptive confidence clustering. High-confidence detections are associated using motion similarity, medium-confidence detections are associated using combined motion and jersey semantic similarity, and low-confidence detections are used for semantic-assisted trajectory recovery with multi-frame verification. The right panel summarizes the mathematical formulation used for confidence partitioning and association. Please click here to view a larger version of this figure.

Figure 3. Distribution of detection confidence scores. The histogram shows the number of athlete detection boxes within five confidence intervals for soccer, basketball, and volleyball sequences in the SportsMOT dataset. The distribution illustrates differences in detector confidence across the evaluated sport categories. Please click here to view a larger version of this figure.

5. Run confidence-guided association

1. Load the detector-output file, semantic-feature file, and confidence-partition file for each video sequence. Initialize the tracker using the first valid detections and maintain one trajectory state for each tracked athlete. For each subsequent frame, predict the position of every existing trajectory using the Kalman filter motion model.
2. Calculate the motion similarity between each predicted trajectory and candidate detection using the Mahalanobis distance defined in Equation 5.
     (5)
   Here, denotes the I-th trajectory in frame k, denotes the Kalman-predicted trajectory state, d_j denotes the candidate detection, denotes the inverse covariance matrix of the predicted trajectory state, and S_mot denotes the motion distance between the predicted trajectory and the detection.
   NOTE: The core association workflow is implemented in jerseytrack/formal_tracker.py and executed through scripts/formal_track.py, both of which are available in the JerseyTrack source code repository (https://doi.org/10.5281/zenodo.21274352). No separate configuration file is required. All runtime parameters, including the confidence thresholds, maximum track age, motion-weight coefficients, and center-distance weight, are supplied through command-line arguments. The complete execution command and parameter settings are provided in Step 6.2. The implementation uses SciPy's linear sum assignment algorithm for Hungarian matching and NumPy for cost-based association.
3. Run the tracking workflow using the detector-output file, semantic-feature file, and confidence-partition file as inputs.
   NOTE: In this study, the core tracker was implemented in jerseytrack\formal_tracker.py, and the tracking workflow was executed using scripts\formal_track.py.
4. Associate high-confidence detections with existing trajectories using motion consistency. Update matched trajectories and initialize new trajectories only when unmatched high-confidence detections satisfy the trajectory initialization criteria.
5. Inspect the jersey semantic extraction outputs illustrated in Figure 4, and construct the jersey semantic feature vector for each detection by combining the team-color descriptor and player-number feature according to Equation 6.
   f_sem = [f_col;f_id] (6)
   Here, f_sem denotes the fused semantic feature vector, f_col denotes the team-color descriptor, and f_id denotes the player-number feature generated by the OCR branch.
6. Associate medium-confidence detections by combining motion similarity with jersey semantic similarity. Calculate the fused matching score according to Equation 7.
   (7)
   Here, S_sem denotes the cosine similarity between the semantic feature vectors of the trajectory and candidate detection, S_mot denotes the motion distance defined in Equation 5, and λ denotes the adaptive fusion coefficient balancing the motion and semantic similarity terms.
7. Calculate the adaptive fusion coefficient according to Equation 8.
   (8)
   Here, λ₀ denotes the initial motion-weight coefficient, σ_sd denotes the standard deviation of the detection-confidence scores in the current frame, and σ_max denotes the maximum confidence-score standard deviation used for normalization.
8. Use low-confidence detections only for trajectory recovery. Match low-confidence detections with interrupted trajectories only when semantic information is available and the recovery criteria are satisfied. Apply multi-frame verification before restoring a trajectory identity, and discard detections that fail verification.
   NOTE: When the player-number feature is unavailable, perform association using the available semantic information and motion consistency rather than forcing a player-number match. The expected outcome of this section is a frame-by-frame tracking record containing trajectory identities, bounding boxes, confidence-level labels, motion costs, semantic information, and final association decisions. In the revision evidence package, the tracking results were stored under: outputs\formal_runs\sportsmot_val_yolo11x_i960_e3_center_motion_sweep_round2\age45_h0p95_m0p94_l0p58_mw0p65_lw0p5_va0p25_hsp0p0_cdw0p1\tracking.

Figure 4. Jersey semantic feature extraction. Representative athlete detections are annotated with the extracted team-color descriptors and player-number information generated by the jersey semantic extraction module. The extracted semantic features are subsequently incorporated into confidence-guided data association. Please click here to view a larger version of this figure.

6. Run tracking inference and export results

1. Run tracking inference for each video sequence using the prepared detector-output files, semantic-feature files, and confidence-partition files. Process the video frames in chronological order so that trajectory states, semantic history, and trajectory recovery decisions are updated consistently throughout the sequence. Use the same inference configuration for all evaluated sequences unless a dataset-specific setting is explicitly reported.
   NOTE: Tracking inference was executed using scripts/formal_track.py, which is available in the JerseyTrack source code repository (https://doi.org/10.5281/zenodo.21274352). No separate configuration file is required. Execute tracking inference using the following command: python scripts/formal_track.py --detections-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_threshold_sweep/t005/detections --semantic-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_t005_color --confidence-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_threshold_sweep/t005/confidence --out-dir outputs/formal_runs/sportsmot_val_yolo11x_i960_e3_center_motion_sweep_round2/age45_h0p95_m0p94_l0p58_mw0p65_lw0p5_va0p25_hsp0p0_cdw0p1/tracking --max-age 45 --high-threshold 0.95 --medium-threshold 0.94 --low-threshold 0.58 --medium-motion-weight 0.65 --low-motion-weight 0.5 --velocity-alpha 0.25 --center-distance-weight 0.1. All unspecified parameters retained the default values recorded in the archived source code.
2. Following inference, apply the primary post-processing sequence using the following commands: python scripts/merge_tracklets.py and python scripts/sweep_track_filter.py --min-len 95.
3. Export the tracking results in the format required by the evaluation toolkit. Include the frame index, trajectory identity, bounding-box coordinates, and all fields required by the benchmark evaluation format. Save one tracking-result file for each sequence using a consistent file-naming convention so that each result file can be matched with its corresponding ground-truth annotation file.
4. Apply only the post-processing operations used in this study. Perform tracklet merging and track filtering using the post-processing scripts described in the revision package.
   NOTE: In this study, the post-processing workflow used the following scripts:
   ⋅ scripts\merge_tracklets.py
   ⋅ scripts\sweep_track_filter.py
   ⋅ scripts\correct_identity_swaps.py
   ⋅ scripts\sweep_identity_swap_correction.py
   ⋅ scripts\interpolate_tracks.py
   ⋅ scripts\sweep_track_interpolation.py
5. Inspect the exported tracking results before quantitative evaluation. Confirm that trajectory identities remain numeric and consistent within each sequence, that no frame indices are missing, and that all bounding boxes remain within the image boundaries.
   NOTE: The expected outcome of this section is a complete set of sequence-level tracking-result files ready for quantitative evaluation.

7. Evaluate tracking performance

1. Evaluate the exported tracking-result files using the evaluation toolkit listed in the Table of Materials. Match each tracking-result file with the corresponding ground-truth annotation file and dataset split. Use the same evaluation configuration for all compared methods to ensure that performance differences arise from the tracking results rather than from the evaluation settings.
2. Run the evaluation using the following command
   \.venv\Scripts\python.exe evaluation\trackeval\scripts\nun_mot_challenge.py --GT_FOLDER datasets\SportsMOT\dataset\val --RESULTS_DIR outputs\formal_runs\sportsmot_val_yolo11x_i960_e3_center_motion_filter_fine
   _round1\minlen95 --OUTPUT_FOLDER outputs\formal_runs\sportsmot_val_yolo11x_i960_e3_center_motion_filter_fine_
   round1\minlen95\trackeval --TRACKERS_TO_EVAL JerseyTrack_i960_e3_center_motion_minlen95 --BENCHMARK SportsMOT --SPLIT_TO_EVAL val
   NOTE: Evaluation was performed using the standard TrackEval (Version 1.3.0) metric implementation archived within the JerseyTrack reproducibility package. No modifications were made to the Higher Order Tracking Accuracy (HOTA), CLEAR, or Identity metric implementations. The repository includes the MOTChallenge-style execution wrapper located at evaluation/trackeval/scripts/run_mot_challenge.py, which configures the dataset and result paths and invokes the standard TrackEval evaluation metrics.
3. Record the evaluation metrics reported in the manuscript, including Multiple Object Tracking Accuracy (MOTA), IDentity F1 Score (IDF1), Higher Order Tracking Accuracy (HOTA), Detection Accuracy (DetA), Association Accuracy (AssA), false positives (FPs), false negatives (FNs), and identity-switching events. Export the evaluation summary as a table and retain the per-sequence evaluation files for verification.
4. When comparing JerseyTrack with baseline methods, use the same dataset split, detector outputs, and evaluation configuration for all methods. Record the dataset, detector-output source, evaluation toolkit configuration, and metric values for each comparison.
5. Inspect the evaluation logs to verify that all sequences are evaluated successfully and that no tracking-result files are missing.
   NOTE: In this study, the primary TrackEval summary file was stored at: outputs\formal_runs\sportsmot_val_yolo11x_i960_e3_center_motion_filter_fine_
   round1\minlen95\trackeval\JerseyTrack_i960_
   e3_center_motion_minlen95\pedestrian_summary.txt. The expected outcome of this section is a complete evaluation summary table together with the per-sequence metric files supporting the reported Results and subsequent verification.

Results

This section reports the quantitative and qualitative results obtained from the evaluated sports multi-object tracking experiments. The results focus on tracking accuracy, identity preservation, association quality, and robustness under the tested dataset settings. Performance claims are restricted to the evaluated methods, datasets, detector outputs, and evaluation toolkit configuration described in the Protocol and Implementation Setup.

Experimental setup and evaluation design

Dataset and preprocessing:

SportsMOT was used as the primary benchmark for detector preparation, tracking inference, and quantitative evaluation. The dataset contains 240 video sequences and more than 150,000 image frames across football, basketball, and volleyball scenarios (Figure 5). In the formal evaluation, the main SportsMOT results were calculated using the validation split together with the detector outputs, tracking configuration, and TrackEval settings described in the Protocol. Cross-dataset evaluation was conducted on TeamTrack, SoccerNet Tracking, MOT17, and MOT20 to examine performance transfer across different dataset types rather than to make direct metric-level comparisons between datasets.

Figure 5. Representative datasets used for evaluation. Example frames illustrate representative soccer, basketball, and volleyball sequences used for experimental evaluation. The figure highlights variations in camera viewpoint, player density, and scene complexity across the evaluated datasets. Please click here to view a larger version of this figure.

The SportsMOT data were organized according to the official dataset structure and converted into the detector-training format described in the Protocol. The converted SportsMOT training and validation data contained 90 sequences, 55,544 frames, and 608,152 annotated objects. The training partition was used for detector preparation and parameter tuning, whereas the validation partition was used for the formal tracking evaluation reported in this study. All input frames were resized according to the detector configuration reported in the Protocol and the Table of Materials. The same preprocessing procedure was applied during detector preparation, semantic feature extraction, tracking inference, and TrackEval evaluation so that the reported differences primarily reflected the tracking association strategy rather than differences in data processing.

Evaluation indicators:

Tracking performance was evaluated using a MOTChallenge-compatible evaluation protocol implemented with the evaluation toolkit listed in the Table of Materials. The reported metrics describe three aspects of sports MOT performance: overall tracking accuracy, identity preservation, and detection-association quality. MOTA, IDF1, and HOTA were used as the primary evaluation metrics^44,45. MOTA summarizes false positives, false negatives, and identity switches into an overall tracking accuracy score. IDF1 measures the consistency between predicted trajectories and ground-truth identities. HOTA jointly evaluates detection accuracy and association accuracy and therefore characterizes the balance between target localization and trajectory association. DetA and AssA were reported as complementary metrics. FPs, FNs, and Identity Switches (IDs) were used to characterize error sources related to false detections, missed detections, and identity changes. For comparisons among methods on SportsMOT, the same detector outputs and evaluation toolkit configuration were used to reduce the influence of detector variation and evaluation-setting differences. For comparisons across datasets, the metric values were interpreted as within-dataset evaluation results rather than as evidence of absolute performance superiority across different datasets.

Experimental environment and parameter configuration:

The experiments were conducted using the hardware and software resources listed in the Table of Materials. The same computational environment, detector framework, detector outputs, and evaluation toolkit were used throughout the primary SportsMOT experiments to maintain consistency during detector preparation, tracking inference, and performance evaluation. The detector framework listed in the Table of Materials was trained using a batch size of 16, an initial learning rate of 0.01, and 80 training epochs. In the jersey semantic extraction module, color clustering used K-means with K = 3, and the OCR branch used a lightweight convolutional recurrent neural network with an input size of 128 × 64 pixels. The OCR branch was optimized using the adaptive optimizer listed in the Table of Materials with a learning rate of 1 × 10⁻³, a weight decay of 1 × 10⁻⁵, a batch size of 64, and 40 training epochs. No additional learning-rate scheduling was used during training of the semantic feature extraction module. The OCR-loss weighting coefficient (γ) was treated as a user-defined hyperparameter and selected according to validation-set performance. Confidence-level stratification used adaptive K-means partitioning, in which τ₁ and τ₂ were calculated from the detection-confidence distribution of each frame rather than being manually predefined before inference.

Tracking performance across sports and scenario complexity

Quantitative performance under different sports and scene conditions:

Tracking performance was evaluated under two grouping factors: sport category and scene complexity. The sport-category analysis included soccer, basketball, and volleyball sequences. The scene-complexity analysis considered occlusion level, motion speed, and target density using the same grouping criteria across the evaluated sequences. The reported metrics followed the evaluation protocol described in Protocol Section 7. 

Figure 6 summarizes the quantitative tracking results under different sport categories and scene-complexity conditions. Figure 6A reports MOTA and DetA across soccer, basketball, and volleyball sequences. Figure 6B reports MOTA variation under different levels of occlusion, motion speed, and target density. Figure 6C reports stacked FP and FN counts for each scene-complexity dimension.

Figure 6. Tracking performance across sport categories and scene conditions. (A) Multiple Object Tracking Accuracy (MOTA) and Detection Accuracy (DetA) for soccer, basketball, volleyball, and the overall average. (B) MOTA across representative scene conditions with different levels of occlusion, player density, and motion complexity. (C) Numbers of False Positives (FPs) and False Negatives (FNs) across the evaluated scene conditions. Please click here to view a larger version of this figure.

Across the three sport categories, JerseyTrack achieved an average MOTA of 88.9% and an average DetA of 80.2%. The difference in MOTA among the sport categories was 1.3 percentage points, with the highest MOTA observed for volleyball sequences (89.2%) and the lowest observed for basketball sequences (87.9%). The lower MOTA observed for basketball sequences is consistent with the higher frequency of close-range interactions and dense occlusion in the evaluated sequences. Under less complex scene conditions, including light occlusion, low motion speed, and sparse target distribution, MOTA reached 91.8%, 93.1%, and 93.8%, respectively. Under more complex scene conditions, MOTA decreased to 84.9% under heavy occlusion, 83.6% under high-speed motion, and 83.1% under dense target distribution. These results show that tracking performance decreased as scene complexity increased while remaining within the reported range across the evaluated sports scenarios.

Tracking consistency across representative sports scenarios:

Figure 7 presents representative tracking examples illustrating the temporal consistency of JerseyTrack under three challenging sports scenarios: dense player interactions, prolonged partial occlusion, and rapid directional changes. Across these representative sequences, the tracker maintained consistent trajectory identities despite frequent athlete overlap and dynamic motion. Identity fragmentation was observed primarily during severe occlusion or when jersey semantics became temporarily unavailable; however, the confidence-guided association strategy enabled trajectory recovery after reliable detections reappeared. These representative examples qualitatively support the quantitative results shown in Figure 6 by demonstrating stable identity preservation under the evaluated sports-tracking conditions.

Figure 7. Representative tracking results produced by JerseyTrack. Consecutive frames from basketball, soccer, and volleyball sequences illustrate the maintenance of athlete identities and trajectories during tracking. Colored bounding boxes represent tracked identities maintained across successive video frames. Please click here to view a larger version of this figure.

Ablation results for identity preservation:

An ablation analysis was conducted to examine the contributions of the confidence-guided association strategy and the jersey semantic fusion module to identity preservation. Four model variants were compared: V0, the baseline without confidence-guided association or jersey semantic fusion; V1, the variant using only confidence-guided association; V2, the variant using only jersey semantic fusion; and V3, the complete JerseyTrack configuration incorporating both components. The evaluation focused on identity-related metrics, including IDs, IDF1, Identity Precision (IDP), and Identity Recall (IDR). The representative identity-preservation patterns are shown in Figure 8.

Figure 8. Ablation analysis of confidence-guided jersey semantic association. (A) Overall Identity Switches (IDs) and IDentity F1 Score (IDF1) for the evaluated model variants. (B) Comparison of IDs between the complete model (V3) and the baseline configuration (V0) under representative challenging tracking scenarios. (C) IDF1, ID Precision (IDP), and ID Recall (IDR) of the complete model across different tracking scenarios. Please click here to view a larger version of this figure.

In the overall SportsMOT validation setting, the complete V3 configuration produced the lowest number of IDs and the highest IDF1 among the four model variants. V3 recorded 2,768 IDs, representing 477 fewer identity switches than V0, and achieved an IDF1 of 74.3%, representing an increase of 7.1 percentage points over V0. These results indicate that the two modules jointly contributed to identity preservation under the evaluated sports-tracking conditions. Comparison of the single-module variants with the complete configuration further showed that the two modules affected different sources of tracking error. The confidence-guided association strategy reduced matching errors associated with unstable detection confidence and localization uncertainty, whereas the jersey semantic fusion module provided additional identity information when motion information alone was insufficient. The complete V3 configuration achieved better identity-related performance than either single-module variant, suggesting that the two modules provided complementary rather than redundant contributions.

Scenario-level results showed larger differences under more challenging identity-preservation conditions. During long-duration occlusion, V3 recorded 215 IDs compared with 482 for V0. In dense same-team scenarios containing 6–10 players, V3 recorded 328 IDs compared with 615 for V0. Under high-speed directional changes, V3 recorded 482 IDs compared with 960 for V0. These reductions are consistent with the intended use of semantic cues during occlusion and dense interactions and with confidence-guided association under fluctuating detection confidence during rapid motion. The IDP and IDR trends shown in Figure 8C indicate that the reduction in IDs was not accompanied by a substantial reduction in either identity precision or identity recall. The complete configuration maintained IDF1 values above 71% across the evaluated challenging scenarios, with lower values observed under dense same-team interactions and high-speed directional changes. These results indicate improved identity consistency under the evaluated conditions while identifying dense interactions and rapid motion as the principal remaining sources of performance degradation.

Cross-dataset evaluation results:

Cross-dataset evaluation was conducted to examine whether the tracking behavior observed on SportsMOT could be maintained under different dataset conditions. The evaluation included two sports-specific datasets, SoccerNet Tracking and TeamTrack, and two general MOT datasets, MOT17 and MOT20. The purpose of this experiment was to examine performance transfer across different dataset types. The results were not used to claim direct metric-level superiority across datasets because the annotation protocols, scene composition, camera viewpoints, and target categories differ.

Table 1 summarizes the cross-dataset evaluation results. On the sports-specific datasets, JerseyTrack maintained relatively stable MOTA and IDF1 values compared with the main SportsMOT validation setting. This trend suggests that the confidence-guided association strategy and jersey-related semantic cues remained effective when the tracking scenes retained team-sport visual characteristics, including repeated uniforms, dense athlete interactions, and frequent occlusion. On the general MOT datasets, the method maintained competitive MOTA and DetA values, whereas IDF1 was lower than that observed for the sports-specific datasets. This pattern is consistent with the absence of jersey color and player-number cues in MOT17 and MOT20, which are used by the semantic fusion module. Under these conditions, the confidence-guided association component remained applicable, whereas the semantic branch contributed less identity-specific information than it did for team-sport videos. Overall, these results indicate that JerseyTrack transferred more effectively to datasets containing sports-specific visual semantics than to general pedestrian-tracking datasets. The cross-dataset evaluation therefore supports the intended application of the method as a sports-oriented tracker while also identifying the absence of explicit jersey semantics as a limiting factor for non-sports MOT datasets.

Dataset	MOTA↑	IDF1↑	DetA↑	FPS↑
SoccerNet Tracking	86.8	71.3	77.9	32.1
TeamTrack	86.2	70.7	77.3	32.8
MOT17	84.7	69.5	76.1	33.9
MOT20	84.1	68.9	75.3	33.2

Table 1: Cross-dataset evaluation results. Tracking performance of JerseyTrack evaluated on four public benchmark datasets. Multiple Object Tracking Accuracy (MOTA), IDentity F1 Score (IDF1), Detection Accuracy (DetA), and processing speed measured as Frames Per Second (FPS) are reported. Higher values indicate better performance for MOTA, IDF1, DetA, and FPS.

Robustness under controlled perturbation conditions

Controlled perturbation experiments were conducted to examine the stability of JerseyTrack under common visual and detection-quality disturbances encountered in sports videos. The evaluated perturbations were grouped into three categories: semantic-feature perturbation, detection-box perturbation, and environmental perturbation. Semantic-feature perturbations included player-number occlusion, image blur, resolution degradation, and similar jersey colors. Detection-box perturbations included bounding-box offset, detection-confidence fluctuation, and false detection boxes. Environmental perturbations included rain-like noise, fog-like degradation, strong illumination, and shadow interference. Figure 9 summarizes the tracking performance under these perturbation conditions. Under semantic-feature perturbation, tracking performance decreased as player-number visibility and crop quality deteriorated. When 40% of the player-number region was occluded, MOTA decreased by 3.2 percentage points and IDF1 decreased by 5.3 percentage points relative to the no-perturbation condition, whereas the semantic extraction accuracy remained 86.7%. These results indicate that the jersey color and player-number cues provided complementary information when one semantic cue became less reliable. Under detection-box perturbation, the method exhibited moderate performance degradation rather than abrupt failure. When the detection boxes were shifted by ±10 pixels and the confidence scores were perturbed with Gaussian noise, the decrease in MOTA remained below 3 percentage points, and the increase in IDs remained within 16%. This trend is consistent with the confidence-guided association strategy, in which medium- and low-confidence detections are processed using additional association constraints rather than the same strategy applied to high-confidence detections.

Figure 9. Robustness evaluation under controlled perturbations. (A) Changes in Multiple Object Tracking Accuracy (MOTA), IDentity F1 Score (IDF1), and Identity Switches (IDs) with increasing jersey-number occlusion. (B) Performance degradation under representative interference conditions. (C) Growth in IDs under different interference types. (D) Jersey semantic extraction accuracy under the evaluated perturbation conditions. Please click here to view a larger version of this figure.

Under environmental perturbation, decreases in the primary tracking metrics remained below 5 percentage points under the evaluated conditions, and the semantic extraction accuracy remained 87.2%. The use of HSV-based color descriptors reduced sensitivity to illumination changes, whereas the OCR branch retained usable player-number information for a subset of blurred or degraded image crops. These results indicate that the semantic module remained usable under the evaluated perturbation conditions, although its reliability continued to depend on jersey visibility and crop quality. Overall, the controlled perturbation experiments showed that JerseyTrack maintained measurable tracking performance under the evaluated semantic, detection-quality, and environmental perturbations. These findings should be interpreted within the tested perturbation range. Systematic out-of-view re-identification, severe background clutter, and long-term complete occlusion were not evaluated separately in this experiment and are discussed as limitations in the Discussion section.

Module contribution and feature discrimination analysis

Module contribution and feature discrimination were further analyzed to examine how the two proposed components affected identity-related tracking performance. The module contribution analysis used the four model variants defined in the ablation analysis, whereas the feature discrimination analysis compared traditional Re-ID features, color-only features, player-number-only features, and fused jersey semantic features. The inter-class/intra-class distance ratio was used to describe feature separability, with larger values indicating greater separation between different identities relative to variation within the same identity. Table 2 summarizes the module contribution analysis. In the overall evaluation, the estimated contribution of the confidence-guided association strategy was 41.7%, the estimated contribution of jersey semantic fusion was 47.6%, and the remaining 10.7% was attributed to the combined use of the two components. These values indicate that both components contributed to the observed improvement, whereas the complete configuration benefited from their combined use rather than from either component alone. The contribution pattern varied across scene types. Under heavy occlusion, the estimated contribution of jersey semantic fusion increased to 69.4%, which is consistent with the reduced reliability of motion cues when athletes were partially or temporarily occluded. Under high-speed motion, the estimated contribution of confidence-guided association reached 44.3%, and the combined gain reached 20.6%, indicating that confidence-level partitioning became more relevant when detection confidence fluctuated because of rapid motion or localization uncertainty.

Model variant	Test scenario	MOTA↑	IDF1↑	IDs↓	Module contribution	Synergistic gain
V0 (Baseline)	Overall scene	83.5	67.2	3245	–	–
	Severely occluded scenes	77.8	61.5	3862	–	–
	High-speed motion scene	77.1	62.3	3689	–	–
V1 (Confidence-guided association)	Overall scene	86.3	70.9	2893	41.7	–
	Severely occluded scenes	80.9	65.9	3415	29.8	–
	High-speed motion scene	81.4	67.9	3024	44.3	–
V2 (Jersey semantic association)	Overall scene	85.2	71.8	2937	47.6	–
	Severely occluded scenes	82.7	72.8	2753	69.4	–
	High-speed motion scene	80.1	66.8	3157	35.1	–
V3 (Complete JerseyTrack)	Overall scene	88.9	74.3	2768	–	10.7
	Severely occluded scenes	84.9	75.6	2689	–	0.8
	High-speed motion scene	83.6	71.5	2842	–	20.6

Table 2: Ablation analysis of module contributions. Performance comparison of different JerseyTrack model variants under overall, severely occluded, and high-speed motion scenarios. Multiple Object Tracking Accuracy (MOTA), IDentity F1 Score (IDF1), and the number of Identity Switches (IDs) are reported together with the estimated module contribution and synergistic gain. Higher values indicate better performance for MOTA, IDF1, module contribution, and synergistic gain, whereas lower values indicate better performance for IDs.

Table 3 summarizes the feature discrimination analysis. The fused jersey semantic feature achieved an inter-class/intra-class distance ratio of 5.63, compared with 1.82 for traditional Re-ID features, corresponding to a relative improvement of 209.3% in the distance-ratio measure. Color-only and player-number-only features also improved feature separability relative to traditional Re-ID features, although each individual cue remained susceptible to specific failure cases, including similar team colors, player-number occlusion, motion blur, and unreadable jersey regions. Among the evaluated feature representations, the fused semantic representation achieved the highest feature separability and corresponded to the highest IDF1 and the lowest ID count observed in the ablation analysis. These findings support the use of jersey-specific semantic cues as complementary identity information for sports MOT when athletes have similar appearances and motion information alone is insufficient. These observations are limited to the evaluated SportsMOT setting and should not be interpreted as evidence that jersey semantics resolve all identity ambiguity in every sports video.

Feature type	Inter-class/Intra-class Distance Ratio	Relative Improvement (%)	IDF1↑	IDs↓
Traditional Re-ID features	1.82	–	65.7	3482
Team-color features	3.17	74.2	69.8	3125
Player-number features	3.49	91.8	70.5	3018
Fused jersey semantic features	5.63	209.3	74.3	2768

Table 3: Discrimination analysis of different feature representations. Comparison of feature discrimination using traditional re-identification (Re-ID) features, jersey color features, player-number features, and fused semantic features. The inter-class/intra-class distance ratio, relative improvement in feature discrimination, IDentity F1 Score (IDF1), and the number of Identity Switches (IDs) are reported. Higher values indicate better performance for the distance ratio, relative improvement, and IDF1, whereas lower values indicate better performance for IDs.

Benchmark comparison with existing MOT methods

A benchmark comparison was conducted to compare JerseyTrack with representative MOT methods under the same SportsMOT validation setting. The compared methods included QDTrack, DeepSORT, ByteTrack, FairMOT, Deep OC-SORT, and MOTR. All methods used the same detector outputs generated by the detector framework listed in the Table of Materials so that the comparison focused on tracking association performance rather than detector variation. The evaluation used the metrics defined in Protocol Section 7. The primary evaluation metrics included MOTA, HOTA, and IDF1. The auxiliary metrics included DetA, AssA, FPs, FNs, and IDs. Table 4 summarizes the quantitative comparison on the SportsMOT validation split, and Figure 10 presents a visual comparison of tracking accuracy, inference speed, and HOTA distribution across the evaluated sports scenes. JerseyTrack achieved the highest MOTA among the compared methods, reaching 88.9%. This value was 1.1 percentage points higher than that of Deep OC-SORT (87.8%) and 2.5 percentage points higher than that of ByteTrack (86.4%). Under the evaluated SportsMOT validation setting, JerseyTrack therefore achieved the highest overall tracking accuracy among the compared methods. For HOTA, JerseyTrack achieved 69.9%, compared with 64.4% for Deep OC-SORT and 62.8% for ByteTrack. These differences correspond to improvements of 5.5 and 7.1 percentage points, respectively, indicating a higher balance between detection and association performance under the same evaluation setting.

Method	MOTA↑	HOTA↑	IDF1↑	DetA↑	AssA↑	FP↓	FN↓	IDs↓
QDTrack	75.9	53.7	51.6	65.6	39.7	96,324	89,871	8,216
DeepSORT	77.6	54.1	53.2	67.3	44	76,228	86,319	6,836
ByteTrack	86.4	62.8	67.7	73.8	50.9	33,752	78,698	4,192
FairMOT	84.1	54.7	62.5	77.8	47.8	45,187	83,620	4,817
Deep OC-SORT	87.8	64.4	69.9	78.2	52.1	25,012	74,385	3,211
MOTR	82.9	57.2	58.4	68.9	46.1	54,931	85,063	5,137
JerseyTrack	88.9	69.9	74.3	80.2	54.3	21,953	67,160	2,768

Table 4: Performance comparison of JerseyTrack and representative multi-object tracking methods on the SportsMOT validation split. Comparison of JerseyTrack with representative multi-object tracking (MOT) methods on the SportsMOT validation dataset. Reported metrics include Multiple Object Tracking Accuracy (MOTA), Higher Order Tracking Accuracy (HOTA), IDentity F1 Score (IDF1), Detection Accuracy (DetA), Association Accuracy (AssA), False Positives (FP), False Negatives (FN), and the number of Identity Switches (IDs). Higher values indicate better performance for MOTA, HOTA, IDF1, DetA, and AssA, whereas lower values indicate better performance for FP, FN, and IDs.

Figure 10. Performance comparison with representative multi-object tracking methods. (A) Comparison of Multiple Object Tracking Accuracy (MOTA) and Frames Per Second (FPS) for JerseyTrack and representative multi-object tracking methods evaluated on the SportsMOT dataset. (B) Distribution of Higher Order Tracking Accuracy (HOTA) across the evaluated tracking methods. Please click here to view a larger version of this figure.

For identity preservation, JerseyTrack achieved an IDF1 of 74.3%, compared with 69.9% for Deep OC-SORT and 67.7% for ByteTrack. JerseyTrack also recorded 2,768 IDs, fewer than the 3,211 IDs recorded by Deep OC-SORT and the 4,192 IDs recorded by ByteTrack. These observations are consistent withthe ablation results presented in Figure 8 and Table 2, in which the complete JerseyTrack configuration reduced identity switching relative to the baseline model variants. For detection and association quality, JerseyTrack achieved a DetA of 80.2% and an AssA of 54.3%. Compared with Deep OC-SORT, JerseyTrack improved DetA by 2.0 percentage points and AssA by 2.2 percentage points. JerseyTrack also produced fewer FP and FN than the other compared methods reported in Table 4, recording 21,953 FP and 67,160 FN. Overall, the benchmark comparison showed that JerseyTrack achieved the highest values for the primary tracking metrics among the evaluated methods under the SportsMOT validation setting. These results are consistent with the observed improvements in identity preservation and association stability obtained using confidence-guided association and jersey semantic fusion. The comparison is limited to the shared detector outputs and SportsMOT validation configuration used in this study.

Parameter sensitivity results

Parameter sensitivity analysis was conducted to examine how key implementation parameters affected tracking performance under the SportsMOT validation setting. Three parameters were evaluated: the OCR loss weighting coefficient (γ), the confidence-level hierarchical cluster number (K), and the OCR network learning rate (η). Table 5 summarizes the MOTA, IDF1, HOTA, and IDs obtained under different parameter settings.

Parameter	Value	MOTA↑	IDF1↑	HOTA↑	IDs↓
OCR loss weight (γ)	0.2	84.7	69.4	66.2	2,944
	0.4	86.3	71.5	68.2	2,893
	0.6	87.9	73.1	68.9	2,815
	0.8 (optimal)	88.9	74.3	69.9	2,768
	1	88.2	73.8	69.3	2,792
Number of confidence clusters (K)	2	86.7	72.1	68.5	2,876
	3 (optimal)	88.9	74.3	69.9	2,768
	4	88.1	73.6	69.7	2,803
	5	87.3	72.8	69.2	2,851
OCR learning rate (η)	1 × 10⁻⁴	85.9	70.8	67.3	2,934
	5 × 10⁻⁴	87.6	72.7	68.7	2,832
	1 × 10⁻³ (optimal)	88.9	74.3	69.9	2,768
	5 × 10⁻³	87.8	73.2	69	2,817
	1 × 10⁻²	86.5	71.6	68.7	2,889

Table 5: Parameter sensitivity analysis. Tracking performance obtained using different parameter settings for JerseyTrack. Multiple Object Tracking Accuracy (MOTA), IDentity F1 Score (IDF1), Higher Order Tracking Accuracy (HOTA), and the number of Identity Switches (IDs) are reported for different values of the Optical Character Recognition (OCR) loss weighting coefficient (γ), the number of confidence clusters (K), and the OCR learning rate (η). Parameter values identified as optimal correspond to the settings used in the reported experiments. Higher values indicate better performance for MOTA, IDF1, and HOTA, whereas lower values indicate better performance for IDs.

For the OCR-loss weighting coefficient (γ), the best evaluated performance was obtained at γ = 0.8. Under this setting, JerseyTrack achieved a MOTA of 88.9%, an IDF1 of 74.3%, a HOTA of 69.9%, and 2,768 IDs. When γ was reduced to 0.2, MOTA, IDF1, and HOTA decreased to 84.7%, 69.4%, and 66.2%, respectively, whereas the number of IDs increased to 2,944. When γ was increased to 1.0, the performance metrics decreased slightly relative to γ = 0.8, with a MOTA of 88.2%, an IDF1 of 73.8%, a HOTA of 69.3%, and 2,792 IDs. These results indicate that insufficient OCR supervision reduced player-number discrimination, whereas increasing the OCR-loss weighting beyond the evaluated optimum did not improve tracking performance under the tested conditions.

For the confidence-level hierarchical cluster number (K), the best evaluated performance was obtained at K = 3. This setting corresponds to the high-, medium-, and low-confidence association strategy described in the method. When K = 2, confidence partitioning was less granular, and MOTA, IDF1, and HOTA decreased to 86.7%, 72.1%, and 68.5%, respectively. When K increased to 4 or 5, performance also decreased relative to K = 3, suggesting that excessive partitioning divided detections into overly narrow confidence groups and reduced the stability of the association strategy. These results support the use of three confidence levels in the evaluated JerseyTrack configuration.

For the OCR network learning rate (η), the best evaluated performance was obtained at η = 1 × 10^-3. At a lower learning rate of 1 × 10^-4, MOTA, IDF1, and HOTA decreased to 85.9%, 70.8%, and 67.3%, respectively, whereas the number of IDs increased to 2,934. At a higher learning rate of 1 × 10^-2, MOTA, IDF1, and HOTA decreased to 86.5%, 71.6%, and 68.7%, respectively, whereas the number of IDs increased to 2,889. These results indicate that the OCR branch was sensitive to learning-rate selection, with 1 × 10^-3 providing the best balance between player-number recognition and identity-preservation performance under the evaluated conditions.

Overall, Table 5 shows that the parameter combination γ = 0.8, K = 3, and η = 1 × 10^-3 produced the highest evaluated MOTA, IDF1, and HOTA values together with the lowest number of IDs. The sensitivity analysis also showed that moderate deviations from these parameter values resulted in measurable, but not abrupt, changes in tracking performance. Accordingly, these parameter settings were used for the benchmark comparison and the primary SportsMOT validation experiments reported in this study.

Data Availability

The raw evaluation summaries, final tracking outputs, environment records, dataset mapping files, and intermediate summary files supporting the findings of this study are available in a public repository at https://doi.org/10.5281/zenodo.21274353. The original public benchmark datasets used in this study, including SportsMOT, TeamTrack, SoccerNet Tracking, MOT17, and MOT20, are available from their respective providers and are not redistributed in the repository.

Discussion

This study evaluated a confidence-guided sports MOT workflow that combines detection-confidence partitioning with jersey semantic cues. Confidence-guided association and semantic feature fusion have been investigated as complementary strategies for improving data association in multi-object tracking^{12,20,23,24,46}. The results show that the complete JerseyTrack configuration improved identity preservation and association stability under the evaluated SportsMOT validation setting. The main SportsMOT validation result reached a MOTA of 88.9%, a HOTA of 69.9%, an IDF1 of 74.3%, a DetA of 80.2%, and an AssA of 54.3%. These values should be interpreted within the detector-output source, dataset split, and TrackEval configuration described in the Protocol.

A critical step in the workflow is confidence-level partitioning, which enables different association strategies to be applied according to the reliability of detector outputs^20,22,23,24. By separating detections into high-, medium-, and low-confidence groups, JerseyTrack applies different association rules to detections with different reliability levels. High-confidence detections are mainly associated through motion consistency, while medium-confidence detections use motion and jersey semantics together. Low-confidence detections are not used to initialize new trajectories and are only considered during trajectory recovery. This design reduced the risk that weak detections or false positives would create unstable identities while still allowing partial detections to contribute to recovery when additional semantic evidence was available^8,11,20. The second critical step is jersey semantic extraction. Team-color features provide a team-level constraint, whereas player-number features provide a finer identity cue when the number region is visible and readable^35,41,42,43. The ablation results indicate that these cues are most useful in dense same-team interaction and occlusion cases, where motion-only association is less reliable. However, jersey semantics should be treated as auxiliary evidence rather than as a complete replacement for motion association. Player numbers can be unreadable under blur, truncation, back-view poses, severe occlusion, or low image resolution. Team-color features also cannot distinguish athletes from the same team when number information is unavailable^35,42,43. The cross-dataset results further clarify the scope of the method. JerseyTrack transferred more naturally to team-sport datasets than to general pedestrian MOT datasets because the semantic branch was designed around jersey color and player-number cues^{19,33,34,35,36,37}. On general MOT datasets, the confidence-guided association component remained applicable, but the sports-specific semantic branch contributed less identity information. Therefore, the method is most suitable for team-sport videos in which athletes wear distinguishable uniforms and the jersey region is sufficiently visible for semantic extraction^{19,33,34,35,36,37}.

Several limitations remain. First, the method depends on the quality of the detector outputs; severe missed detections or inaccurate bounding boxes reduce the reliability of both motion prediction and semantic cropping^14,17,46. Second, long-term out-of-view re-identification was not separately optimized in the current implementation. When an athlete leaves the camera view for a prolonged period and later re-enters, the current trajectory recovery strategy may not be sufficient to restore the original identity^19,34. Third, severe background clutter, overlapping players, motion blur, and unreadable jersey numbers can reduce the reliability of semantic features^14,35,42,46. Fourth, the current evaluation focused on public benchmark datasets and controlled perturbation settings; deployment in broadcast videos with camera cuts, zoom changes, and non-standard viewpoints may require additional preprocessing or re-identification modules^19,33,34.

The main practical implication is that confidence-guided association and jersey-specific semantic cues can be integrated into a modular MOT pipeline without changing the detector structure. This capability is applicable to sports-analysis tasks such as athlete trajectory extraction, team movement analysis, tactical event review, and semi-automatic video annotation^1,4,33,36. Future work should focus on stronger out-of-view re-identification, background-clutter handling, temporal feature aggregation, and more robust player-number recognition under severe blur or occlusion^14,19,34,46.

In summary, JerseyTrack is a sports MOT method that combines confidence-guided association with jersey semantic fusion. The method partitions detections into high-, medium-, and low-confidence groups and applies different association rules according to detection reliability while using jersey color and player-number cues as auxiliary identity information during medium-confidence association and trajectory recovery. Under the evaluated SportsMOT validation setting, JerseyTrack demonstrated improved tracking accuracy and identity preservation relative to the evaluated baseline configurations. The ablation results showed that the complete configuration reduced identity switches compared with the baseline and single-module variants, and the benchmark comparison showed higher values for the primary tracking metrics under the shared detector-output and evaluation configuration. These findings support the use of confidence-level information and jersey-specific semantic cues for athlete tracking in team-sport videos, particularly when jersey regions are visible and team-color or player-number information provides supplementary identity evidence^20,35,46. Addressing the identified limitations may further improve the applicability of the proposed workflow to more challenging sports-tracking scenarios.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
CUDA Runtime	NVIDIA	Version 12.8	GPU computing runtime used for detector training, semantic feature extraction, and tracking inference.
EasyOCR	Jaided AI	Version 1.7.2	Optical character recognition toolkit used for player-number recognition.
JerseyTrack source code	Authors	N/A	Custom implementation containing the data preparation, semantic feature extraction, confidence partitioning, tracking, post-processing, and evaluation scripts. Available at: https://doi.org/10.5281/zenodo.21274352
Lightweight CRNN OCR model	Authors	N/A	Custom convolutional recurrent neural network (CRNN) used for player-number recognition.
MOT17 dataset	MOTChallenge	N/A	Public benchmark dataset used for cross-dataset evaluation. Available from: https://motchallenge.net/data/MOT17/
MOT20 dataset	MOTChallenge	N/A	Public benchmark dataset used for cross-dataset evaluation. Available from: https://motchallenge.net/data/MOT20/
motmetrics	motmetrics Developers	Version 1.4.0	Library used for diagnostic multi-object tracking metric calculation.
NVIDIA GPU	NVIDIA	GeForce RTX 5090 D V2	Graphics processing unit used for detector training and tracking inference.
NVIDIA GPU Driver	NVIDIA	Version 610.62	GPU driver used in the experimental environment.
NumPy	NumPy Developers	Version 2.4.4	Numerical computing library used for feature processing and data handling.
OpenCV	OpenCV	Version 4.10.0	Computer vision library used for image loading, cropping, preprocessing, and visualization.
Python	Python Software Foundation	Version 3.12.2	Programming language used throughout the workflow.
PyTorch	PyTorch Foundation	Version 2.11.0+cu128	Deep learning framework used for detector and semantic model implementation.
scikit-learn	scikit-learn Developers	Version 1.9.0	Machine-learning library used for K-means clustering during jersey-color extraction and confidence partitioning.
SoccerNet Tracking dataset	SoccerNet	N/A	Public football tracking benchmark used for cross-dataset evaluation. Available from: https://www.soccer-net.org/tasks/tracking
SportsMOT dataset	SportsMOT Benchmark	N/A	Primary benchmark dataset used for detector preparation and tracking evaluation. Available from: https://deeperaction.github.io/datasets/sportsmot.html
TeamTrack dataset	TeamTrack Benchmark	N/A	Public sports tracking benchmark used for cross-dataset evaluation. Available from: https://atomscott.github.io/TeamTrack/
Torchvision	PyTorch Foundation	Version 0.26.0+cu128	Computer vision library used with PyTorch for image transformation and model support.
TrackEval	TrackEval Developers	Version 1.3.0	Evaluation toolkit used to calculate Multiple Object Tracking Accuracy (MOTA), IDentity F1 Score (IDF1), Higher Order Tracking Accuracy (HOTA), Detection Accuracy (DetA), Association Accuracy (AssA), false positives (FP), false negatives (FN), and identity switches (IDs).
Ultralytics	Ultralytics	Version 8.4.90	Object detection framework used to train the detector and generate athlete detections.