This repository contains a Python script that demonstrates a collision detection system using YOLOv5 (You Only Look Once) for object detection and DeepSORT for object tracking. The system is designed to detect collisions between cars in video footage.
- Add the copy of this colab to your drive - Colab link
- Add the shortcut of this shared folder to your drive Folder shared with you
- Put the setting to T4 and do run all
- Clone the repository and navigate to the project folder.
- Install the required dependencies using pip:
pip install -r requirements.txt - Download the DeepSORT model weights using the following command:
!git clone https://github.com/ZQPei/deep_sort_pytorch.git && cd deep_sort_pytorch && !pip install -r requirements.txt - Run the
main.pyscript using Python:python main.py
The script uses several parameters that can be adjusted to fine-tune the collision detection system. These parameters include:
collision_threshold: The distance threshold below which two vehicles are considered to be colliding.iou_threshold: The Intersection over Union (IoU) threshold above which two bounding boxes are considered to be overlapping.size_ratio_threshold: A bounding box's minimum size ratio to the frame's largest bounding box.
The script generates several output files, including:
detections.json: A JSON file containing the detection results for each frame.collisions.json: A JSON file containing the collision detection results.output_vid.mp4: A video file with annotated bounding boxes and collision detection results.frames_of_collision: A folder containing frames where collisions were detected.
The current implementation has some limitations, including:
- Currently, it is based on a defined method; we could also train the model to detect crash if large amount of data is there
- The system may detect false positives or false negatives depending on the video quality, camera angle, and lighting conditions.
- The system may not generalize well to different scenarios or environments.
Several improvements can be made to the system, including:
- Could use any advanced object detection model to detect cars (YOLOv8)
- Could use any advanced tracking algorithm to track cars (BotSORT, DeepocsSORT)
- Could use object detectors that give quadrilateral output rather than rectangles parallel to x and y axis (YOLO-R)