This project was developed in collaboration with Francesco Zampirollo and Vincenzo Macellaro.
| 📑 Final Report | 🖼️ Slides |
Distraction, is the major cause of road accidents, using Computer Vision,Machine Learning and Pose Estimation we are able to detect drivers attention and distraction.
The system classifies driver behavior into five risk levels and uses a Graph Convolutional Network (GCN) for enhanced analysis. Preliminary results show 90% accuracy, suggesting significant potential to improve road safety by alerting drivers or initiating corrective actions.
- Mediapipe: Used for keypoint detection to analyze the driver's state.
- YOLO (You Only Look Once): Employed for detecting potential distraction objects such as phones.
- Graph Neural Network (GNN): Developed by us, that combine Mediapipe's output with the YOLO bounding box's coordinates, for the classification of the driver's state.
- Faiss: Library used to Retrieval part. Retrieval system returns the embedding images most similar a specified query. The process involves comparing the ground-truth classification (GNN) with the classification obtained through K-nearest neighbors (KNN) during the retrieval process with the K-embeddings.