Lightweight YOLOv8n with Cross-Scale Feature Fusion for Driver Fatigue Detection

Currently, most fatigue driving detection methods based on deep learning meet accuracy requirements. However, these models have a high number of parameters and require significant computation, resulting in low real-time speed. Hence, they cannot be applied to on-board equipment with limited memory and computing power. To address these issues, a lightweight network model was designed based on the YOLOv8n object detection network. The Down sampling module, which reduces the spatial resolution of feature maps to decrease the model's computational load and number of parameters, replaces the Convolution module in the traditional YOLOv8n model series. This module captures image features at a higher level using a specific Down sampling strategy. Additionally, a lightweight Cross-Scale Feature Fusion Module framework is further adopted in the neck to replace the original neck structure, reducing giga floating-point operations per second and enhancing the model's detection capabilities. The Focal Modulation technology was used to replace the original Spatial Pyramid Pooling Fast module, improving the model's recognition ability for these areas and solving the problem of reduced model accuracy caused by the lightweight design. The experimental results demonstrate that the proposed model effectively achieves lightweight effects. The improved model shows a 1.1 % increase in accuracy, a 42.2 % decrease in the number of parameters, a 27.1 % decrease in computation, and a 40.6% reduction in model size. Moreover, the model accurately extracts the driver's eyes and mouth opening characteristics. The eye closure index percentage is used to objectively and accurately assess a driver's fatigue, providing an early warning to the driver.