| The world’s increasing transportation demands have brought about significant traffic safety issues,which have had major impacts on society’s economy and stability.To address this,the Advanced Driving Assistance System(ADAS)has been developed to help drivers detect possible dangers and improve driving safety.Driving behavior recognition plays a crucial role in this system,as it allows for the identification and analysis of dangerous driving patterns,and ultimately helps drivers reduce risks and prevent accidents.On-board sensors provide a convenient means of collecting real-time driving data that is not easily affected by environmental factors.As a result,the driving behavior recognition method based on on-board sensors has seen rapid development in recent years,with a focus on achieving high-precision recognition using these sensors.This thesis proposes a driving behavior recognition model based on on-board sensors through deep learning methods,with a focus on modeling in-vehicle sensor data,spatio-temporal feature extraction,and driving recognition.The proposed model constructs on-board sensor data from the perspective of graph(Graph)and extracts spatio-temporal correlation information of the sensor sequence.The research in this thesis includes the following:(1)In terms of space,based on the data characteristics of in-vehicle sensors,this thesis proposes a sensor graph structure prior representation method by mining the relevance of each sensor;secondly,in order to obtain better sparse directed interaction,a self-attention mechanism is used to construct learnable sensors’ graph structure,and based on the graph convolution,graph learning branches for prior graphs and learnable graphs is proposed,which integrates the prior graph information and the learnable graph information,and obtains better data interaction at the spatial level;finally,in this thesis,The effectiveness of the method is verified on three public real driving behavior data sets.The experiments show that the method of explicit graph modeling can make full use of the spatial information of in-vehicle sensors,obtain higher recognition accuracy results,and the recognition effect outperforms existing methods.(2)In terms of time,in order to further extract the time information of on-board sensors,this thesis designs a Encoder-Decoder structure that fuses historical information and future information based on LSTM,so that the future information of on-board sensors can be used,thereby helping to better identify driving behavior at the current moment.Based on the above method,this thesis builds an adaptive future space-time graph convolutional network(SF-STGCN).In further experiments,this thesis verifies the effectiveness of fusing future information,and the network obtains better recognition accuracy. |
