Adaptive Fatigue Driving Monitoring Algorithm Based on Neural Network

Abstract

In practical application scenarios, drastic changes in ambient light and driver posture offset can easily lead to a decrease in facial image quality, resulting in weakened key fatigue related features such as blinking and yawning, severely limiting the robustness of existing visual detection models. To address this issue, this paper proposes a feature reconstruction YOLO detection method for image degradation and constructs two core modules: dual convolution feature decoupling and multi-scale reconstruction. The dual convolution feature decoupling module achieves collaborative modeling of facial structured information and degraded region responses by constructing parallel local detail feature branches, which significantly enhances the model's ability to characterize highlight-interfered regions and weak-texture fatigue features. The multi-scale reconstruction module relies on cross-layer feature reconstruction and semantic alignment mechanisms to accurately compensate for the tiny fatigue feature details lost during downsampling, effectively improving the stability of fatigue detection in low-illumination and dynamic scenarios. The experimental results show that the proposed method outperforms the benchmark model in terms of detection performance and robustness, and can effectively alleviate the problem of insufficient fatigue feature expression caused by image degradation in actual monitoring environments. This algorithm has certain practical value in the field of intelligent transportation.