YOLO Deep-Learning Based Driver Behaviors Detection and Effective Gaze Estimation by Head Poses for Driver Monitor System
Yi-Chiao Fang, Xi-Liang Zhao, Hsuan-Yu Lin, Yu-Cheng Yang, Jiun-In Guo, Chih‐Peng Fan · 2023
This work develops a non-contact driver behavior monitoring system based on intelligent visual sensing to improve the driving safety. By the deep learning technology with YOLO, a car-specification near-infrared (NIR) camera is installed to detect the driver’s behaviors and gaze directions. The YOLO-based head pose inference method is developed, and the driver’s gaze directions are predicted with the simplified calibration. In experiments, the input size of YOLOv4-tiny based model is set to 416x416 pixels. After functional tests, the proposed method performs average precision (AP) to be 86.58% for detecting eleven classes including driver’s objects and behaviors. Besides, the proposed gaze estimation technology by driver’s head poses performs average detection accuracy up to 83% to estimate twelve driver’s gaze directions.