Existing fatigue driving detection models often suffer from limitations such as single evaluation criteria, insufficient adaptability, high computational demands, and performance degradation under low-light conditions at night. A fatigue state evaluation system based on adaptive threshold optimization is proposed. The system leverages linear polarization lighting technology to enhance image quality in low-light environments and integrates the YOLO-GM model with boundary constraint optimization to improve the selection and recognition of eye and mouth ROI under occlusion scenarios, thereby enhancing feature recognition accuracy. A multi-feature fusion-based fatigue state evaluation model is constructed, and an adaptive fatigue threshold determination method based on a binary decision tree is proposed to dynamically adjust evaluation thresholds, further improving classification accuracy. Experiments conducted on the YawDD and a self-constructed dataset demonstrate that the proposed model reduces the parameter size by 3.95MB compared to the original model, achieves a feature recognition accuracy of 95.08%, and reaches a fatigue state evaluation accuracy of 95%, with an average processing time of 88.5ms per frame. Given its ability to perform effectively under low-light conditions at night, combined with high real-time performance and strong adaptability to individual driver differences, the system is well-suited for integration into in-vehicle platforms with limited computational resources.