Vehicle Orientation Detection (Based on YOLOv5)

YOLOv5: ultralytics/yolov5: YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite

A vehicle-orientation detection system based on YOLOv5 for real-time vehicle location detection. It uses the relative location of tires to obtain the direction. The code is compatible with PyTorch 1.7+ and supports GPU acceleration.

Due to various objective factors and the author's own limitations in coding ability, please forgive any shortcomings. Your valuable suggestions for improvement are sincerely hoped for.

!!!Disclaimer: The functionalities provided by this project have not been fully tested. The authors do not take any responsibility for issues arising from its use!!!

🚀 Major Functions

• ✅ Vehicle detection: High quality YOLOv5 object detection.
• 🧭 Orientation: Method of detecting orientations of complex objects based on key points.
• 📊 Visualized output: Output marked with ellipses and arrow lines.

📦 Quick start

Environment

```bash

clone project

git clone https://github.com/Steven8686/vehicleorientationdetect.git cd vehicle-orientation-detection

installation (Python 3.8+)

pip install -r requirements.txt ```

Model test

bash python detect.py --weights ./orientation/vehicle_orientation.pt --img 640 --conf 0.25 --source ./orientation/test_picture/test.jpg --ori_mode e-lsm

pre-trained model (based on dataset of our own) is saved as ./orientation/vehicle_orientation.pt

Parameters

This project has 3 more parameters compared to original YOLOv5. Description: (original parameters of YOLOv5 are not mentioned here)

bash --ori_mode # Modes for ellipses(tires) detection, Including "e-lsm", "lsm", "e-hough", "hough",default="e-lsm" --ori_detect # Enables orientation detection. If False, there's no difference between original yolov5 for detect.py --ori_savepath # Save path for orientation detection. Default=runs/detect/expx/orientation.jpg/mp4.

Extra information

In this version, due to some historic reasons, classes.txt (./orientation/classes.txt) defined 6 categories, representing:

| label | meaning | | :---- | :----------------------------------------------------------- | | sr | static right (static tires on the right side of vehicle, usually back wheel) | | mr | movable right (movable tires on the right side of vehicle, usually front wheel) | | ml | movable left | | sl | static left | | l | left (Left side of vehicle) | | r | right (Right side of vehicle) |

In fact, 3 categories (car, front, back) are enough for detection. Currently used 6 categories have decreased the accuracy of YOLO detection and resulted in duplicate detections. We plan to correct and optimize this problem later.

Future Plans

• Modify the way of labeling data and correspondent code.
• Optimize the save process and keep it the same as original YOLOv5.
• Optimize the tire detection logic. Plan to add a direct way of plotting the arrow using the detection boxes as well as optimize performance.