🚗 Automatic Vehicle Accident Detection from Images During Adverse Weather Conditions Using YOLOv10 and VGG19

Research Paper | Computer Vision | Deep Learning | Road Safety

---

📄 Documentation

• Research Paper: Research_Paper.pdf
• Dataset Information: DATASET.md

---

📌 Overview

This research presents a hybrid deep learning framework for detecting vehicle accidents under adverse weather conditions, including rain, fog, snow, and low-light environments.

Poor visibility and environmental noise make accident detection particularly challenging for intelligent transportation systems. To address this problem, this work combines the strengths of YOLOv10 for real-time object detection and VGG19 for deep feature extraction and classification, improving accident detection accuracy while reducing false positives.

The proposed framework contributes to safer roads by supporting autonomous driving systems, traffic monitoring, and faster emergency response.

---

🔄 Proposed Framework

┌─────────────────────────────┐
│         Input Image         │
└─────────────────────────────┘
               │
               ▼
┌─────────────────────────────┐
│    Image Preprocessing      │
│ • Dehazing                  │
│ • Contrast Enhancement      │
└─────────────────────────────┘
               │
               ▼
┌─────────────────────────────┐
│          YOLOv10            │
│      Object Detection       │
└─────────────────────────────┘
               │
               ▼
┌─────────────────────────────┐
│ Detected Vehicle/Accident   │
│          Region             │
└─────────────────────────────┘
               │
               ▼
┌─────────────────────────────┐
│           VGG19             │
│       Classification        │
└─────────────────────────────┘
               │
               ▼
┌─────────────────────────────┐
│ Accident / Non-Accident     │
└─────────────────────────────┘

The proposed framework integrates image preprocessing, YOLOv10 object detection, and VGG19 classification to improve accident detection performance under adverse weather conditions.

---

🎯 Key Highlights

Feature	Description
Dataset	3,480 annotated road images collected under different weather conditions
Detection Model	YOLOv10
Classification Model	VGG19
Annotation Tool	Roboflow
Training Platform	Google Colab
Best YOLOv10 Precision	81.8% (Epoch 200)
Key Contribution	Hybrid framework combining object detection and image classification

---

🧠 Model Architecture

YOLOv10

• Real-time object detection
• Fast single-stage detector
• Robust under adverse weather conditions
• Best precision achieved at 81.8%

YOLOv10 Performance

Epoch	Precision	Recall	mAP@50
50	54.2%	45.7%	44.6%
100	63.2%	56.8%	62.9%
200	81.8%	61.0%	72.5%

---

VGG19

VGG19 was used to classify accident and non-accident images through deep feature extraction.

Compared with baseline CNN architectures, VGG19 consistently achieved the highest accuracy, precision, recall, and F1-score.

---

📈 Results

Sample Prediction

Example predictions produced by the YOLOv10 model for accident detection under different weather conditions.

---

Confusion Matrix (Epoch 200)

The confusion matrix illustrates the final classification performance of the best-performing YOLOv10 model.

---

Training Curves

Training curves showing model convergence and performance improvements over 200 training epochs.

---

🛠️ Technical Stack

Category	Technology
Programming Language	Python
Detection Model	YOLOv10
Classification Model	VGG19
Frameworks	PyTorch, TensorFlow
Libraries	OpenCV, NumPy, Matplotlib, Scikit-learn
Annotation	Roboflow
Training Platform	Google Colab

---

👩‍💻 My Contributions

As a researcher and developer, I participated in:

• Designing the hybrid deep learning framework.
• Collecting and annotating the dataset.
• Training and fine-tuning YOLOv10 and VGG19.
• Performing image preprocessing and data augmentation.
• Evaluating model performance using Precision, Recall, Accuracy, mAP, and F1-score.
• Analyzing experimental results.
• Writing the complete research paper.

---

🔮 Future Work

• Integrate LiDAR and radar data for multi-modal accident detection.
• Improve robustness under extreme weather conditions.
• Optimize the framework for deployment on edge devices.
• Evaluate newer object detection architectures such as YOLOv11 and YOLO-NAS.

---

🏆 Research Impact

This work demonstrates the effectiveness of combining object detection and deep image classification to improve vehicle accident detection under adverse weather conditions.

Potential applications include:

• 🚑 Intelligent emergency response systems
• 🚗 Autonomous driving
• 📹 Smart traffic surveillance
• 🌧️ Weather-aware road safety monitoring

---

📫 Contact

• GitHub: YaraAlansari-CS
• LinkedIn: Yara Alansari
• Email: yara.alansari01@gmail.com

---

⭐ If you found this research interesting, consider giving this repository a star!