Research On Key Technologies For Detection Of Typical Dangerous Driving Behavior Based On Human-Vehicle Data

As the level of urbanization and the number of civilian vehicles increase,the impact of traffic accidents on human safety becomes increasingly prominent.Drivers are often the main factor in causing road traffic accidents,with dangerous driving behavior caused by anger,distraction,and fatigue being the most common.Therefore,this paper aims to design a model that accurately identifies these three main dangerous driving behaviors and to construct a corresponding warning system for dangerous driving behavior.This approach can reduce road traffic safety hazards during driving and provide reference and experience for future human-machine driving technology.This study conducted simulation driving experiments in a simulated driving laboratory.The virtual simulation scene was built based on multi-person and multimachine interactive driving simulators,UC-win/Road simulation software,Psy LAB human factor equipment,and dual-angle cameras,which together formed a multisource data collection system.By designing reasonable experimental plans and using the multi-source data collection system,the authors collected data from participating drivers to build driver facial data sets,side-driving data sets,simulated driving vehicle data sets,and human factor data sets.Based on the driver’s facial data set,an improved multi-layer Convolutional Neural Networks(CNN)model was constructed.Due to the relative simplicity of some network model structures,it cannot accurately and efficiently extract and process driver facial feature information.Therefore,this paper adopted a continuous convolution idea to construct the model,which achieved a recognition accuracy rate of 92.8%,a loss value of 0.0347,an F1-Score of 94.0%,and a Macro F1 of 95.0%.Based on the driver’s side-dangerous driving data set,a recognition model combining You Only Look Once X(YOLOX)and Back Propagation Neural Network(BPNN)was constructed.In this section,the yolox＿l recognition model in YOLOX was used to identify dangerous driving behavior features,and then BPNN was used for classification.By referencing the two-stage recognition model idea of BPNN,this model achieved an average recognition accuracy rate of 93.5%,an F1-Score of 93.6%,a Macro F1 of 93.3%,and a loss value of 0.0728.After conducting Principal Component Analysis(PCA),normalization,and feature fusion data processing on the driver’s simulated driving vehicle data set and human factor data set,a combined recognition model based on Artificial Neural Network(ANN)was constructed to identify feature fusion data.In order to overcome the shortcomings of traditional ANN,the authors used the ELU function as the activation function in the input layer and hidden layer.This model achieved a recognition accuracy rate of 95.3% and a loss value of 0.064.Based on previous research,a warning system for dangerous driving behavior was constructed in this section,which issues warnings to drivers based on different situations to protect their driving safety.The identification models established in this study have improved the recognition rate of dangerous driving behaviors among drivers to some extent,and provide a theoretical basis for personalized automotive driving assistance safety systems and human-machine driving technology.