by Jason Sohn (website: jasonsohn.com)
ClassySORT is a simple real-time multi-object tracker (MOT) that works for any kind of object class (not just people).
ClassySORT is designed to be a state-of-the-art (SOTA) multi-object tracker (MOT) for use on your own projects. And bcause the You-only-look-once algorithm (YOLO) detector is pretrained on COCO dataset, ClassySORT can detect and count and track 80 different kinds of common objects 'out of the box'.
Tested on Pop_OS! 20.04 (similar to Ubuntu) and NVIDIA RTX 2070s.
Modifying it is exactly the same process as training YOLOv5 with your own dataset. How do I do that?
ClassySORT implements
This repository uses a fixed version of YOLOv5 to ensure compatbility. Replacing the YOLOv5 code to the updated ultralytics/YOLOv5 code may result in breaking changes. If you are able to this without issues, please submit a pull request.
If you only need to track people, or have the resources to train a model from scratch with your own dataset, see 'More Complex MOTs' section below.
Clone this repository
git clone https://github.com/tensorturtle/classy-sort-yolov5.git
cd classy-sort-yolov5
Python 3.8 or later with all requirements.txt. To install run:
pip install -r requirements.txt
./download_weights.sh
This script will save yolov5 weights to yolov5/weights/
directory.
To run the tracker on your own video and view the tracked bounding boxes, run:
python classy_track.py --source /path/to/video.mp4 --view-img
To get a summary of arguments run:
python classy_track.py -h
The text results are saved to /inference/output/
from the array above in the following format. That location in the script is also a good point to plug your own programs into.
The saved text file contains the following information:
[frame_index, x_left_top, y_left_top, x_right_bottom, y_right_bottom, object_category, u_dot, v_dot, s_dot, object_id]
where
The original implementation of SORT threw away YOLO's object class information (0: person, 1: bike, etc.).
I wanted to keep that information, so I added a detclass
attribute to KalmanBoxTracker
object in sort.py
:
I found that for my own dataset in which bounding boxes change size fairly quickly, the default Q value (process covariance) was too low. I recommend you try experimenting with them.
If you only need to track people, or have the resources to train a model from scratch with your own dataset, then I recommend bostom/Yolov5_DeepSort_PyTorch. DeepSORT adds a separately trained neural network on top of SORT, which increases accuracy for human detections but slightly decreases performance. It also means that using your custom dataset involves training both YOLO and DeepSORT's 'deep association metric'
For a 'bag of tricks' optimized version of YOLOv5 + DeepSORT, see GeekAlexis/FastMOT
ClassySORT is released under the GPL License version 3 to promote the open use of the tracker and future improvements. Among other things, this means that code from this repository cannot be used for closed source distributions, and you must license any derived code as GPL3 as well.