Abstract
This article addresses the problem of improving face detection reliability in video streams used in driver state monitoring systems. Reliable localization of the driver’s face is considered a preliminary computer vision stage required for the subsequent analysis of eye state, gaze direction, facial behavior, and fatigue-related visual cues. The purpose of this study is to develop and experimentally evaluate an adaptive face detection method that maintains stable performance under variations in illumination, head pose, image noise, and partial face occlusion while operating with a fixed, externally calibrated camera. The proposed method combines initial face-region detection using the Viola–Jones algorithm with SVM-based verification, external camera calibration, and incremental parameter updating based on high-confidence detections used as pseudo-labels. The method was evaluated using 18,000 video frames obtained from 12 recordings involving seven volunteers under different shooting conditions. The experimental results showed that the integrated method provided higher face detection performance than the standalone Viola–Jones algorithm and the Viola–Jones method combined with a conventional SVM classifier. The proposed method achieved a Precision of 0.91 and a Recall of 0.90 while maintaining a processing speed of 22 frames per second, which is sufficient for real-time operation. The study evaluates face detection and does not directly classify driver fatigue, drowsiness, eye state, or facial expressions. The obtained results demonstrate that the proposed method can be used as a lightweight preliminary component of driver state monitoring systems operating on embedded hardware with limited computational resources.