Visual Recognition Method For Illegal Driving Behavior Based On Deep Learning

Illegal driving behavior is one of the main reasons that cause traffic accidents.To reduce the relevant safety hazards,it is an urgent need to accurately detect the illegal driving behavior and give warning.Nowadays,the analysis of the illegal driving behavior based on visual sense is the method of the highest cost performance and the most convenient deployment with wide application foreground.With the rapid development of deep learning technology,the existing research on driving behavior recognition still has significant room for improvement in the efficiency,accuracy,and system of fatigue driving state detection methods.Therefore,this dissertation focuses on the study of the visual recognition method for illegal driving behavior based on deep learning,providing the a theoretical foundation and technical assistance for driving warning system building.The main work of this dissertation includes the following five parts.(1)From the perspective of systems engineering and human factor engineering,this dissertation analyzes the definition of the illegal driving behaviors such as dangerous driving and tired driving and its formation,in addition,its relationship with road traffic accidents is discussed as well,proposing the definition of visual identification of illegal driving behavior,and its whole-process visual recognition model.While,the building of the system framework of visual recognition method for illegal driving behavior,and the analysis of the deep learning method system for illegal driving behavior identification,the visual attention mechanism method system for improving the accuracy of illegal driving behavior identification,and the network lightweight design method system for improving the efficiency of illegal driving behavior identification lay a theoretical and methodological foundation for the follow-up study.(2)A precise recognition method for dangerous driving behavior based on Res Net and visual attention module is proposed in this dissertation.To improve the accuracy of the existing risky driving behavior recognition module,and overcome the problem of feature loss caused by a single pooling scheme,and in consideration of the characteristics of max pooling,average pooling,and random pooling methods in the dimension reduction of driving image features.We designed four visual attention models,and embedded them in the residual error branch of Res Net model respectively.Through image collection and data augmentation,we compare the recognition accuracy of dangerous driving behavior of Res Net modules with four visual attention modules embedded at different network depths.Furthermore,we also conduct confusion matrix analysis and gradient-weighted class activation mapping visualization analysis to evaluate the recognition performance of the proposed model from different perspectives.(3)A high-efficiency method for identifying dangerous driving behaviors based on an improved Shuffle Net lightweight model is presented.The Leaky Re LU activation function is used to optimize the feature mapping process of the Shuffle Net lightweight deep learning network,improving the model’s ability to learn from partial feature values(<0)and avoiding the problem of feature learning failure caused by the activation function’s unchanged values.At the same time,the feature extraction module structure of the Shuffle Net network is optimized,using a spatial downsampling module with a stride of 2,and the number of its output channels of the feature image is twice as many as the number of its input channels.This effectively increases the network width without significantly increasing the computational complexity,further optimizing the feature extraction of input image,revealing the recognition process of the proposed model by channel visualization and classification weight visualization techniques.(4)A fatigue driving state recognition method based on an improved Efficient Det deep learning network is proposed.The computational complexity of the feature extraction network was reduced and the efficiency of fatigue driving behavior image feature extraction was improved by using depth separable convolution and mobile inverted bottleneck convolution modules.A visual attention mechanism was applied to construct squeeze and excitation modules,further enhancing the feature extraction capability of the entire detection network.The low-level location information and highlevel semantic information in the detection model were fused to increase the overall information flow.The K-means clustering algorithm was used to statistically analyze the real boxes in the dataset,and the prior boxes that can better meet the requirements of fatigue driving behavior were set to optimize the direction of feature extraction and reduce the model’s attention to redundant information.And based on PERCLOS,we classified the 3 categories of fatigue driving degree,providing the identification method for fatigue driving status.(5)A visual recognition system for detecting illegal driving behavior has been developed.This system has been optimized to minimize the impact of environmental factors through design of the visual recognition hardware system,and validated it through experimental testing.To assess the consistency and effectiveness of the facial videobased fatigue detection method,we have introduced validity and reliability evaluation methods for quantitative analysis.Moreover,based on the real-world data collected from illegal driving incidents,we have demonstrated the effectiveness and superiority of the proposed methods in this dissertation,providing guidance for the widespread use of the visual recognition technology for illegal driving behavior.