Scribd
Upload
8 views2 pages

Player Reid Report

The document discusses a methodology for detecting and re-identifying football players in broadcast footage using the YOLOv8 model for object detection and Deep SORT for tracking. Challenges included high inference times and ID switching due to player proximity, while future improvements aim to enhance identity consistency and runtime performance. Overall, the system is effective but requires further optimization and feature enhancements.

Uploaded by

gulimcherla hemaharshithreddy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
8 views2 pages

Player Reid Report

The document discusses a methodology for detecting and re-identifying football players in broadcast footage using the YOLOv8 model for object detection and Deep SORT for tracking. Challenges included high inference times and ID switching due to player proximity, while future improvements aim to enhance identity consistency and runtime performance. Overall, the system is effective but requires further optimization and feature enhancements.

Uploaded by

gulimcherla hemaharshithreddy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Player Re-identification in Broadcast Sports Footage

Approach and Methodology:

--------------------------

The objective was to detect and re-identify football players across frames in broadcast video footage. The

approach utilized a combination of the YOLOv8 object detection model for identifying players and Deep

SORT for tracking identities over time.

YOLOv8 was used for its real-time inference capability and high accuracy in object detection. Only detections

labeled as "player" were passed to the tracker.

The Deep SORT algorithm handled the temporal tracking of detected players, assigning consistent IDs to

them across consecutive frames.

Techniques Tried and Their Outcomes:

------------------------------------

1. Lowering frame resolution to 320x192 to reduce latency and enable faster inference.

2. Filtering only relevant classes (i.e., "player") to reduce tracking noise.

3. Visualization with bounding boxes and player IDs using OpenCV.

4. Tested different YOLO confidence thresholds for optimal balance between false positives and misses.

Challenges Encountered:

------------------------

- Real-time processing was challenging with CPU-only execution; inference times were high.

- Model size (YOLO weights and environment dependencies) exceeded GitHub upload limits, requiring use of

external storage.

- Occasional ID switching for players close together or overlapping.

- Ensuring modular and clear code structure within a tight deadline.

Page 1
Player Re-identification in Broadcast Sports Footage

Remaining Work and Future Improvements:

----------------------------------------

- Incorporate re-identification embeddings to improve identity consistency across occlusions and camera

cuts.

- Further optimize runtime performance using a GPU setup.

- Improve robustness under poor lighting or low-resolution scenarios.

- Add player jersey color/number recognition for better re-ID.

Conclusion:

-----------

The system successfully detects and tracks players in broadcast footage using YOLOv8 and Deep SORT.

The modular code allows for easy improvements, and despite some latency issues, the solution is effective

and extensible.

Page 2

You might also like