Yolov5 StrongSORT Demo

Introduction

Reproduce the Strong Sort algorithm

Simple implementation of the Strong Sort algorithm so that it can run private video samples. The detector uses Yolov5 instead of the official detector Faster-RCNN.

StrongSORT: Make DeepSORT Great Again

Official github address: https://github.com/dyhBUPT/StrongSORT

Before you run

1. Download Re-ID model weights from the following web site. The model is trained from light-reid and obtained by distillation of the light model.

   • 百度云
   • Google Driver

2. Download detector yolov5 and download yolo wights from here.

3. Set the parameters for Strong Sort in the file opt.py

4. Set the Re-ID model, yolov5 detectors and their weights, etc. paths in the config file.

5. Put the test video into others/input

My Environment

• python 3.8
• pytorch 1.8.2 + cu111
• torchvision 0.9.2+cu111
• cuda 10.2
• numpy
• opencv
• scipy

Tracking

• DeepSORT

  python strong_sort.py

• StrongSORT

  python strong_sort.py --NSA --EMA --MC --woC

Results

DeepSORT

StrongSORT

Note

AFLink and GSI are offline algorithms and cannot be applied in video.

Acknowledgement

A large part of the codes and results are borrowed from DeepSORT, YOLOv5, light-reid and StrongSort. Thanks for their excellent work!