Research on the Analysis and Recognition System for Dangerous Driving Behaviors Based on Convolutional Neural Networks

Abstract

To address the critical need for real-time monitoring of hazardous behaviors that compromise road safety, this study presents an intelligent detection system centered on the lightweight YOLOv11n model. The system integrates multiple functional modules, including object detection, multi-object tracking using ByteTrack and BoTSORT, and behavior recognition logic based on temporal windows, thereby forming a comprehensive technical workflow encompassing data acquisition, preprocessing, inference, tracking, behavior analysis, and early warning. It is capable of accurately detecting six common dangerous behaviors-yawning, eye closure, smoking, drinking, mobile phone usage, and in-vehicle conversations-while providing timely alerts to mitigate potential risks. Extensive experiments demonstrate that the system achieves a validation accuracy of 65.55%, a recall rate of 79.71%, and an mAP50 of 69.26%. The inference speed for individual images is maintained within 30-60 milliseconds, and video processing can consistently reach 20-30 frames per second, ensuring real-time performance suitable for practical deployment. The proposed framework not only enhances detection precision but also supports continuous monitoring and risk prevention in real-world driving environments, providing a technically feasible and operationally efficient solution for improving driver safety.