Point-yolo: A fatigue driving detection method for edge detection devices

Due to the computational limitations of edge devices, it is a challenge to apply AI-based fatigue detection algorithms at high speeds and achieve accurate results. We propose a lightweight network model, Point-YOLO, based on the YOLOv5 algorithm, aiming to achieve fatigue driving detection. We designed the GLiteConv convolution module and optimized the backbone network, introducing CA-BiFPN to improve the model's feature extraction capability. Additionally, we improved the feature fusion network and effectively optimized the multi-scale feature fusion module based on the driver's feature scene, reducing the number of model parameters while maintaining accuracy. We modified the head module to output 12 landmarks of the eyes to judge whether the driver is fatigued while driving. The Point-YOLO model size is 2.1M, with a CPU speed of 43ms and an average accuracy of 88.5%. Experimental results show that the Point-YOLOv5 model is effective in recognizing fatigue driving features and achieves a balance between model lightness and accuracy. The algorithm can run stably on edge devices with low power consumption.