Tracking People using YOLOv7 and DeepSORT

Table of Contents

1. Overview
2. Getting Started
3. Code
4. External Work

Overview

This repository implements a solution to the problem of tracking moving people in a low-quality video. It uses a state-of-the-art object detector YOLOv7 for detecting people in a frame, and fuses these robust detections with the bounding boxes of previously tracked people using the neural network version of SORT called DeepSORT tracker. The packages for YOLO and DeepSORT algorithms are located under yolo and deepsort folders, where the modules are adapted from the official repositories to fit the scripts here.

Getting Started

The necessary python environment required for working with this repository can be setup using the script setup_conda.sh. This script assumes that the system has a working conda distribution (Miniconda or Anaconda), and accepts one argument which will be used as the name for the newly created environment.
The installed conda environment will have the following major packages:

1. TensorFlow 2.8.1 (with GPU support)
2. CUDA 10.2 (with CuDNN 7.6.5)
3. PyTorch 1.10.1 (with TorchVision 0.11.2)
4. OpenCV 4.6.0
5. Matplotib 3.6.3
6. Pandas 1.5.3

Note (1): The script is not robust, and may fail when conda packages are updated in their respective channels.
Note (2): In case of import errors, it is recommended to have "channel_priority" set to "true" in ".condarc" file, with the channels priority being "pytorch" > "conda-forge" > "defaults", and rerun the setup script.

Code

After setting up the conda environment using the above script, the pipeline can be executed by -

python main.py --input-vid=/path/to/input/video/file --save-path=/path/to/output/video/file

This will read frames from the input video, detect people using a YOLO API, and track people using a DeepSORT API. These APIs can be found in api folder. The frames with people detected and tracked are saved to the specified video file.

Data and Results

Input video, versioned output videos, checkpoints and a README with versioning details can be found here. The checkpoints folder should be downloaded to the project root.

External Work

YOLOv7

Repository - Paper

All the modules related to this network can be found under yolo/, and the weights can be found at checkpoints/yolov7x.pt in the GDrive link. More details can be found in yolo/__init__.py.

DeepSORT

Repository - Paper

All the modules related to this network can be found under deepsort/, and the weights can be found at checkpoints/ReID.pb in the GDrive link. More details can be found in deepsort/__init__.py.