Research on the Detection Method of Driving Fatigue Based on Anchor-Free And Attention Mechanism

Abstract

Driver fatigue is one of the significant causes of road safety hazards, making real-time monitoring of driver fatigue crucial. The changing network coverage in the vehicular environment, as vehicles move, can result in degraded response times and affect detection efficiency. To incorporate a lightweight detection model in the vehicular environment, this paper proposes a fatigue detection method based on anchor-free detection and attention mechanism network model. The improved single-stage face detector, CenterFace, is used to localize facial landmarks and divide the driver's eye and mouth regions proportionally. A fatigue recognition model, built using the attention mechanism from Ghost modules and Sknet, detects the states of the eyes and mouth. Fatigue assessment is achieved by calculating the PERCLOS (Percentage of Eye Closure) values of the eyes and mouth. Experimental results on the NHTU-DDD dataset and a self-built dataset demonstrate that the proposed network reduces the single-frame eye-mouth state recognition time to 6.4 ms while achieving an accuracy of 98.87%. The model size is reduced to 2.96 MB, achieving a balance between accuracy and real-time performance.