Yolov5 + Deep Sort with PyTorch

Introduction

This repository contains a moded version of PyTorch YOLOv5 (https://github.com/ultralytics/yolov5). It filters out every detection that is not a person. The detections of persons are then passed to a Deep Sort algorithm (https://github.com/ZQPei/deep_sort_pytorch) which tracks the persons. The reason behind the fact that it just tracks persons is that the deep association metric is trained on a person ONLY datatset.

Description

The implementation is based on two papers:

• Simple Online and Realtime Tracking with a Deep Association Metric https://arxiv.org/abs/1703.07402
• YOLOv4: Optimal Speed and Accuracy of Object Detection https://arxiv.org/pdf/2004.10934.pdf

Requirements

Python 3.8 or later with all requirements.txt dependencies installed, including torch>=1.7. To install run:

pip install -U -r requirements.txt

All dependencies are included in the associated docker images. Docker requirements are:

• nvidia-docker
• Nvidia Driver Version >= 440.44

Before you run the tracker

1. Clone the repository recursively:

git clone --recurse-submodules https://github.com/mikel-brostrom/Yolov5_DeepSort_Pytorch.git

If you already cloned and forgot to use --recurse-submodules you can run git submodule update --init

2. Github block pushes of files larger than 100 MB (https://help.github.com/en/github/managing-large-files/conditions-for-large-files). Hence you need to download two different weights: the ones for yolo and the ones for deep sort

• download the yolov5 weight from the latest realease https://github.com/ultralytics/yolov5/releases. Place the downlaoded .pt file under yolov5/weights/
• download the deep sort weights from https://drive.google.com/drive/folders/1xhG0kRH1EX5B9_Iz8gQJb7UNnn_riXi6. Place ckpt.t7 file underdeep_sort/deep/checkpoint/

Tracking

Tracking can be run on most video formats

python3 track.py --source ...

• Video: --source file.mp4
• Webcam: --source 0
• RTSP stream: --source rtsp://170.93.143.139/rtplive/470011e600ef003a004ee33696235daa
• HTTP stream: --source http://wmccpinetop.axiscam.net/mjpg/video.mjpg

MOT compliant results can be saved to inference/output by

python3 track.py --source ... --save-txt

Other information

For more detailed information about the algorithms and their corresponding lisences used in this project access their official github implementations.