Research And Design Of Fatigue Driving Detection Algorithm Based On Deep Learning

With the gradual development of the economy and the increasing demand of people’s consumption,automobiles as consumer goods have gradually become an irreplaceable means of transportation.However,the rapid growth in household car ownership has also led to a rise in traffic accidents,traffic congestion,and other issues.According to statistics,fatigue driving and the resulting improper driving behavior are one of the important reasons for highway and urban highway traffic accidents.Intelligent devices and technologies have great potential in monitoring,analyzing,and preventing driver fatigue driving behavior.Currently,fatigue driving detection has problems such as insufficient real-time performance,poor sensitivity,and reliability.In this thesis,driver fatigue detection system is researched and designed based on deep learning method,and the single-stage Retina Net algorithm is used as the benchmark model.By integrating the self attention mechanism,lightweight and loss function improvement,the accuracy and calculation efficiency of the algorithm are improved,providing effective help for real-time monitoring of driver fatigue status.The research content and main contributions of this thesis are as follows:(1)In view of the problem of low detection accuracy in traditional single stage algorithms,this thesis proposes an improved algorithm for fatigue driving state detection based on the Retina Net model.The proposed algorithm abandons the commonly used Res Net backbone network and designs a hybrid neural network incorporating self attention mechanism.By effectively utilizing global information,it establishes long-range dependency relationships.Secondly,based on the application field and analysis of the dataset,some high-level output feature maps for predicting large targets were discarded,and smaller anchor box sizes were used to obtain facial eye and mouth state information,reducing the computational complexity of the model.The experiment shows that compared to the original Retina Net model,the improved detection algorithm in this thesis has an accuracy improvement of about 2.6% on the fatigue driving detection dataset,and a recall increase of about 3.3%.(2)In view of the problem that the linear layer parameters in the traditional attention model are large and the real-time detection is poor,this thesis conducts lightweight design based on the improved Retina Net model,and optimizes the loss function.Firstly,a deep separable convolution and inverse residual structure are introduced in the design of the backbone network to lightweight the improved attention model.Secondly,combined with the geometric information of the target box,the loss function of the boundary box coordinates is improved and optimized.The experiment shows that the improved lightweight model not only has high accuracy,but also reduces the parameter quantity of the DW deep separable convolutional layer model to 1/9 of the original,improving the computational efficiency of the model.After introducing the improved loss function,the convergence speed of the improved model and the accuracy of the boundary box coordinate prediction are also verified.Based on the above research,this thesis designs a fatigue driving detection system.The system monitors the changes in human eye and mouth states per unit time,and qualitatively analyzes the fatigue status of drivers based on fatigue assessment indicators.The thesis includes 40 figures,6 tables,and 82 references.