Pedestrian Detection and Tracking

This project implements a pedestrian tracker from a side-view camera.

• For object detection in each frame, pretrained YOLO-v3(on MS-COCO dataset) is used. The number of detections are reduced to keep only person class and filter out others.
• For tracking the detected objects, a kalman filter based SORT(Simple Online and Realtime Tracking) algorithm is used https://arxiv.org/abs/1602.00763

Code Walkthrough

• main.py ----> main file for the detect_and_track function
• sort.py -----> SORT algorithm implementation
• YOLO ------> YOLO-v3 related config files(icluding pretrained-weights)
• utils.py -------> utility methods

How to run the code

pip install -r requirements.txt
python main.py

This would return the list of tracks. In order to see the whole trajectory, one need to plot the returned points on the video frames and assemble them in a video using cv2.VideoWriter.

References

• Simple Online and Realtime Tracking https://arxiv.org/abs/1602.00763
• SORT implementation https://github.com/abewley/sort
• YOlOv3 An Incremental Improvement Paper: https://arxiv.org/abs/1804.02767