Fatigue Driving Behavior Detection Based on PiPnet Improvement

Traffic accidents are a major social problem worldwide, resulting in a large number of casualties and property damage. In this paper, the driver fatigue detection system monitors the coordinates of the driver's eyes, eyebrows, mouth, and other facial landmarks of the head, and evaluates the fatigue state of the driver by analyzing the blinking frequency, yawning frequency, and head nodding frequency, which helps to improve road safety and reduce traffic accidents. In this paper, a facial landmark detection model CA-Mobilevitv2 is proposed on the basis of PIPNet, which is applied to the driver fatigue detection system in combination with face detection (Retinaface). CA-Mobilevitv2 uses a lightweight Coordinate Attention module embedded in MobileVitv2 as a backbone CA-Mobilevitv2 uses a lightweight Coordinate Attention module embedded into MobileVitv2 as the backbone network, which reduces the computational and parameter counts by 0.5GFLOPs and 5M, respectively; meanwhile, it adopts the strategy of fusing multi-layer auxiliary detection heads to fully utilize feature maps at different levels to improve the model's adaptability and robustness. Finally, adding the credibility prediction of each facial landmark neighbor to the PiP Regression detection header reduces the normalized mean error (NME) of the model by 0.51% on the challenge set of 300W dataset.