Research on the Optimization of an Edge-side System for Fatigue Detection Based on an Adaptive Threshold Tuning Algorithm

To address the challenges of poor environmental adaptability, sensitivity to individual physiological differences, and high computational demands on edge devices in traditional computer vision-based fatigue detection systems during practical implementation, this study proposes a systematic optimization approach through architectural improvements. Firstly, the replacement of Dlib framework with MediaPipe Face Mesh keypoint detection model achieves a significant reduction in model footprint while maintaining detection accuracy, doubling processing speed while enhancing robustness against illumination variations, head rotations, and occlusions. Secondly, the establishment of a driver physiological characteristics database and adaptive threshold tuning algorithm effectively mitigates false detection caused by individual differences. By integrating MobileNet lightweight architecture with hybrid quantization techniques, combined with spatiotemporal optimization of detection logic, the system reduces GPU memory usage by two-thirds on edge devices and significantly decreases latency. Experimental results demonstrate that the optimized system achieves high detection accuracy with ultra-low latency on automotive-grade embedded platforms, while maintaining adaptability across diverse driving scenarios. This research provides an engineering paradigm for low-cost, high-reliability active safety solutions.