Understanding Abnormal Driver's Behaviours Detection

Abstract

Road safety is a critical issue as traffic accidents cause economic losses and fatalities. Many of these accidents result from driver errors, often due to distracted behaviors that divert attention from the road. Our research addresses this by detecting and classifying distracted driving behaviors using the Drive&Act dataset, which includes actions such as grabbing something from behind, using a phone, and reading. The goal is to reduce accidents and improve road safety. Data is captured from an inner mirror view, to provide clear visibility of the driver's face and upper body. Keypoint extraction is then performed with MediaPipe for facial and body landmarks and while YOLO focuses on human detection, considering only the detection with the highest confidence. We compared six deep learning models, finding that the CNN achieved 95% accuracy in classifying driver behaviors on the benchmark dataset, outperforming others. In real-life scenarios, the CNN-RNN model showed 89% accuracy, demonstrating how combining spatial and temporal feature extraction effectively detects complex driving actions. These findings underscore the potential of using advanced models to detect distractions, contributing to safer roads.