This project evaluates the performance of different tracking algorithms for player tracking in volleyball match videos. The focus is on comparing the YOLOv8 Default Tracker, ByteTrack, and BoT-SORT in terms of tracking accuracy, identity consistency, and robustness.
- YOLOv8 Default Tracker: Baseline tracker using YOLOv8’s default tracking mechanism.
- ByteTrack: Combines strong detection with a simple association mechanism to improve tracking accuracy.
- BoT-SORT: Enhances tracking with appearance features and advanced motion models for improved identity preservation.
- Volleyball match videos featuring multiple players.
- Ensured videos captured complex player interactions to test tracker robustness.
- YOLOv8 was used for object detection across all trackers.
- Each tracker (YOLOv8 Default, ByteTrack, BoT-SORT) was applied to the videos.
- Player bounding boxes and identity assignments were generated for each frame.
- BoT-SORT Tracker:
- Maintained consistent IDs and bounding boxes.
- YOLOv8 Tracker:
- Efficient but exhibited lower confidence scores.
- ByteTrack:
- Accurate bounding boxes and high confidence scores.
- BoT-SORT:
- Best at maintaining player identities due to appearance features and advanced motion models.
- Identity Switches:
- Reduced in BoT-SORT but still present during complex interactions.
- Computational Load:
- Advanced trackers like BoT-SORT require more resources, impacting real-time feasibility.
- False Positives:
- Low-confidence detections led to occasional tracking of non-player objects.
- BoT-SORT: Best performance in maintaining player identities.
- ByteTrack: Accurate bounding boxes with fewer identity switches.
- YOLOv8 Default Tracker: Efficient but less accurate in maintaining consistent IDs.
- Optimize BoT-SORT for real-time performance.
- Reduce false positives by refining confidence thresholds.
- Further minimize identity switches during complex interactions.