Detection And Recognition Of Abnormal Driving Behavior Based On Deep Learning

China’s urbanization is progressing rapidly,the population is gradually concentrated in cities,and the number of vehicles has increased dramatically,with the consequent violations and traffic accidents,etc.How to effectively reduce the occurrence of traffic accidents has become a research hotspot in the field of transportation.In real life,the occurrence of traffic accidents is closely related to the driving behavior of vehicles.In order to reduce the occurrence of traffic accidents and ensure road traffic safety,this thesis conducts an in-depth study on the rapid and accurate detection and identification of the driving behavior of vehicles.This thesis deeply investigate the data characteristics of driving behavior based on the motion state data of moving vehicles obtained by sensors in real scenes;on this basis,this thesis proposed a neural network-based abnormal driving behavior detection method.The method uses isolated forests to detect abnormal points in the data,obtains sample data and builds up a dataset;using the advantages of neural networks in identifying time series data,proposes a practical extended network framework for abnormal driving behavior recognition and detection.Finally,a sensitivity analysis is performed on the input parameters in the model to calculate the sensitivity coefficients of the relevant variables,identify the parameters that have a greater impact on the classification results,and achieve the goal of simplifying the model to improve the training efficiency by discarding useless parameters.The main work of the thesis is as follows.(1)Use the vehicle driving data in real environment as the data source for this study.Before establishing the data set,the data is pre-processed to ensure the quality for the problems that occur in the collection process of driving status data.Moreover,the driving behavior of the driver is analyzed,and abnormal driving behaviors such as rapid acceleration,rapid deceleration,rapid leftward turn and rapid rightward turn,which appear more frequently,are selected together with normal driving behaviors as the objects of driving behavior classification in this chapter.(2)After pre-processing the data,the data is segmented using sliding windows to obtain their feature data separately,and then the data are reconstructed using autoencoders,and finally the isolated forest algorithm is used to identify the abnormal points in the driving data.According to the distribution of anomalies,the data samples of the above-mentioned types of driving behaviors are obtained separately to establish a training set for the recognition of abnormal driving behaviors.(3)Analyze and compare the advantages and disadvantages of different neural networks in the processing of time-series data.According to the data characteristics,it is proposed to use a network with memory function such as Recurrent Neural Network(RNN)to cascade with convolutional neural network to achieve the classification of abnormal driving behavior.The models are built according to the corresponding algorithms mentioned above,and each model is trained separately using the same data set.Analyzed the performance of each network in the classification results after the training is completed,and the algorithm with the best time efficiency and accuracy is selected as the main algorithmic model for this chapter.(4)Sensitivity analysis is performed for the input attributes in the network,mainly using the variable perturbation method to calculate the sensitivity coefficients of each attribute variable,find the input variables most relevant to abnormal driving behavior,and discard useless variables to simplify the model.In summary,the thesis uses neural networks to build an abnormal driving behavior recognition model to achieve the classification detection of driving behavior.The performance of the algorithm is also verified using the established dataset,and the accuracy reaches 98.69%,with a significant improvement in training speed as well as accuracy.