Readme.md

# 🚦 Accident Detection using YOLOv8

A real-time traffic accident evaluator and severity classifier built using the **YOLOv8x** object detection model, trained on the [Roboflow Accident Evaluator Dataset](https://universe.roboflow.com/accident-evaluator).

---

## 📌 Project Overview

This project aims to detect and classify traffic accident-related scenarios in surveillance footage or images. The model can identify:

- 🚑 Detected Injury  
- 🔥 Fire  
- 📈 High Severity  
- ⚠️ Low Severity  
- ⚙️ Medium Severity  
- 💨 Smoke  

The model is fine-tuned using **YOLOv8x** and trained in stages up to **120 epochs** for optimal accuracy.

---

## 🔍 Dataset Details

- 📦 **Source**: Roboflow Universe  
- 🖼️ **Images**: 3,000+  
- 🏷️ **Classes**: 6  
- 📁 Format: YOLO with `.txt` annotations

We used the dataset in the following folder structure:

accident-evaluator-3/ ├── train/ ├── valid/ ├── test/ └── data.yaml

---

## 🚀 Model Training

We trained the model in **incremental steps of 20–30 epochs** with manual control to resume/stop as needed. Key configurations:

```python
from ultralytics import YOLO

model = YOLO("yolov8x.pt")
model.train(
    data="data.yaml",
    epochs=50,
    imgsz=640,
    batch=16,
    optimizer="SGD",
    lr0=0.001,
    device=0,
    warmup_epochs=3,
    patience=15,
    cache=True,
    mosaic=1.0,
    mixup=0.1,
    fliplr=0.5,
)

📌 Early stopping was enabled to avoid overfitting.

---

📈 Validation Results (Epoch 120)

Metric	Score
[email protected]	0.697
[email protected]:0.95	0.438
Precision	0.764
Recall	0.646

Per-Class Accuracy:

Class	[email protected]
Detected-Injury	0.741
Fire	0.611
High	0.869
Low	0.834
Medium	0.795
Smoke	0.331

---

🧪 Inference on Custom Image

from ultralytics import YOLO

model = YOLO("best.pt")  # path to the final trained model
results = model("your-image.jpg", save=True, imgsz=640)
results.show()

---

📁 Folder Structure

project/
├── accident-evaluator-3/     # Dataset
├── training_stages/          # Saved model weights per stage
├── runs/detect/train/        # YOLO output folders
├── best.pt                   # Final trained model
└── README.md

---

🛠️ Future Work

• ✅ Improve underperforming classes (like smoke)
• ✅ Deploy as a Flask web app or streamlit dashboard
• ✅ Integrate into live CCTV/RTSP camera feeds
• ✅ Export to ONNX or TensorRT for real-time deployment

---

👨‍💻 Author

Developed by [Your Name] – passionate about AI for social impact. Reach out on LinkedIn or email at [email protected].

---

📄 License

This project is licensed under the MIT License.