DeepSORT Vehicle Tracker

!!! This project is created for learning purposes do not use in production directly !!!

Introduction

This is a 2D Vehicle Tracker based on works Simple Online and Realtime Tracking with a Deep Association Metric (Deep SORT) and Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks.

For DeepSORT, official implementation of authors is used. The official repository contains deprecated modules, so I fixed them. The implementation is using a feature extractor(CNN) trained on pedestrian images. I trained a new Siamese Network using the vehicle dataset shared by user @John1liu and used it as feature extractor for association. Dataset Link

For Faster-RCNN, I used the official torch implementation with a backbone MobileNet V3 trained on COCO Dataset. Tracker only tracks object with classes "car", "truck" and "bus".

Prerequisites

• Python 3.8+
• Poetry 1.1.11(offical installation guide)

Installation

Clone the vehicle-tracker repo:

git clone https://github.com/mucozcan/deepsort-vehicle-tracker.git
cd deepsort-vehicle-tracker/

Clone the official DeepSORT implementation with the dependency fix made by me:

git clone -b fix/sklearn https://github.com/mucozcan/deep_sort.git

Install dependencies:

poetry install

Run the demo:

poetry run python3 tracker.py --source test.mp4

or

poetry run python3 tracker.py --source [RTSP Stream Link]

TODO

• detector trained on custom vehicle dataset
• onnx and TVM support
• tuning on DeepSORT parameters