Object Tracking using YOLO-NAS and DeepSort

This repository contains code for object tracking in videos using the YOLO-NAS object detection model and the DeepSORT algorithm. The code processes each frame of a video, performs object detection using YOLO-NAS, and tracks the detected objects across frames using DeepSort.

Demo of Object Tracker

Demo of Object Tracker Demo with object blurring

Prerequisites

• Python 3.10
• OpenCV
• PyTorch
• deep-sort-realtime
• super-gradients

Installation

1. Clone the repository:

   git clone https://github.com/sujanshresstha/YOLO-NAS_DeepSORT.git
   cd YOLO-NAS_DeepSORT
   

2. Create new environment using conda

   conda env create -f conda.yml
   conda activate yolonas-deepsort
   

Usage

1. Prepare the video file:

   • Place the video file in the desired location.
   • Update the video flag in the code to the path of the video file or set it to 0 to use the webcam as the input.

2. Configure the YOLO-NAS model:

   • Update the model flag in the code to select the YOLO-NAS model variant (yolo_nas_l, yolo_nas_m, or yolo_nas_s).
   • Make sure the corresponding model weights are available.

3. Configure the output video:

   • Update the output flag in the code to specify the path and filename of the output video file.

4. Set the confidence threshold:

   • Adjust the conf flag in the code to set the confidence threshold for object detection. Objects with confidence below this threshold will be filtered out.

5. If you want to detect and track certain object on video

   • Modify the class_id flag in the code to specify the class ID for detection. The default value of the flag is set to None. If you wish to detect and track only persons, set it to 0, or refer to the coco.names file for other options.

6. If you want to blur certain object while tracking

   • Modify the bulr_id flag in the code to specify the class ID for detection. The default value of the flag is set to None.

7. Run the code:

   # Run object tracking using YOLO-NAS and DeepSort on a video
   python object_tracking.py --video ./data/video/test.mp4 --output ./output/output.mp4 --model yolo_nas_l
   
   # Run object tracking using YOLO-NAS and DeepSort on webcam (set video flag to 0)
   python object_tracking.py --video 0 --output ./output/webcam.mp4 --model yolo_nas_l
   
   # Run person tracking using YOLO-NAS and DeepSort on a video (set class_id flag to 0 for person)
   python object_tracking.py --video ./data/video/test.mp4 --output ./output/output.mp4 --model yolo_nas_l --class_id 0
   
   # Run tracking on a video with burring certain objects (set blur_id flag to 0 for person)
   python object_tracking.py --video ./data/video/test.mp4 --output ./output/output.mp4 --model yolo_nas_l --blur_id 0
   

8. During runtime:

   • The processed video frames will be displayed in a window.
   • The frames per second (FPS) will be printed in the console.
   • Press the 'q' key to stop the processing and close the window.

9. After completion:

   • The output video file will be saved inside output dir.

Model and Data

• The YOLO-NAS model used in this code is pre-trained on the COCO dataset.
• The COCO class labels are loaded from a file named coco.names located in the specified path.
• The DeepSort algorithm is used for tracking the detected objects. It associates unique track IDs with each object and maintains their states across frames.

License

This project is licensed under the MIT License.

Acknowledgements

• This code is built upon the YOLO-NAS model and the DeepSort algorithm.
• Credits to the authors and contributors of the respective repositories used in this project.

References

• YOLO-NAS: Towards Real-time Multi-object Detection with Neural Architecture Search
• Simple Online and Realtime Tracking with a Deep Association Metric