🚗 Car Accident Detection and Deformation Classification Using YOLOv5

This project uses YOLOv5 to detect car accidents from images and classify the severity of damage into five categories based on visual deformation. The model is trained on a custom annotated dataset of real-world car accidents.

📂 Dataset Structure

The dataset contains two main folders: images/ and labels/, each with train/ and val/ subfolders:

dataset/ ├── images/ │ ├── train/ │ └── val/ ├── labels/ │ ├── train/ │ └── val/

Each .txt label file follows YOLO format:
<class_id> <x_center> <y_center> <width> <height> (all normalized)

Example: 4 0.560967 0.598094 0.829009 0.629283

🏷️ Class Labels

| Class ID | Description | Estimated Deformity | |----------|--------------------------|----------------------| | 0 | No Accident | 0% | | 1 | Minor Accident | ~30% | | 2 | Moderate Accident | ~50% | | 3 | Severe Accident | ~70% | | 4 | Totaled Vehicle | ~100% (on fire, flipped, crushed) |

🛠️ Model Training

• Base Model: YOLOv5 (s, m, l as needed)
• Framework: PyTorch
• Training Image Size: 640x640
• Batch Size: 16
• Best Weights: weights/best.pt

🔧 Training Command

bash python train.py --img 640 --batch 16 --epochs 100 \ --data dataset.yaml --weights yolov5s.pt --name accident_detector

🔍 Inference Example

bash python detect.py --weights weights/best.pt --img 640 --source data/test_image.jpg

Output will be saved in the runs/detect/ folder.

📈 Results

Sample detections and training performance plots are included in the results/ directory. Add your visuals here to showcase model performance.

📦 Dataset Collection

The dataset was collected using frames extracted from publicly available online videos (YouTube, Facebook, etc.). A Python script was used to extract frames, followed by manual annotation of bounding boxes and classification into five deformity levels based on visual inspection. please download the dataset : https://www.kaggle.com/datasets/marslanarshad/car-accidents-and-deformation-datasetannotated after downloading dataset please update the paths in custom.yaml:

📚 Requirements

bash pip install -r requirements.txt

Dependencies include: - torch - torchvision - opencv-python - numpy - matplotlib - PyYAML

📁 Repo Contents

• weights/last.pt – Trained YOLOv5 weights
• dataset.yaml – Training config
• images/, labels/ – Dataset directories
• detect.py, train.py – Inference and training scripts
• results/ – Inference outputs
• README.md – Project documentation

📜 License

Open for academic and non-commercial use. If you reuse this dataset or model, please give credit to the original source.

👤 Author

Developed by Muhammad Arslan Arshad
Custom YOLOv5 model for accident detection and deformation severity classification.