Research On Vehicle Driving Behavior Analysis Based On Multi-sensor Information

The sensor information related to vehicle state(e.g.speed sensor,angle sensor and position sensor)is usually used to analyze of vehicle driving behaviors.On the other hand,the state information of vehicle drivers also is very important for recognition of abnormal driving behaviors in abnormal state of vehicle drivers(e.g.sensorimotor driving,cognitive driving and emotional driving),which can promote accuracy levels of recognition of abnormal driving behaviors.In the existing research of vehicle driving behavior analysis concerning vehicle state and driver state,the problems are as follows: the sensor data collected in the process of driving are completely random missing and continuous missing;the lack of analysis and judgment of driving behavior from entities which generate information leads to the unexplainable results and reduce the fault-tolerant ability;in the process of analysis,the fact that standards of abnormal judgment the driver’s external characteristics and physiological signals which will change with external environment is not considered;There is a lack of time-dependent processing of driving behavior data;many research take no account of the consequences that data collected by sensors with obvious fluctuation,which will affect the analysis and recognition of driving behavior.To solve the problem of data missing,this thesis proposes a strategy based on mixing DBN and LSTM to recovery data.On this basis,this thesis proposes an abnormal driving behavior recognition model named LSTM-d SCVAE.The model processes the sensor information from car and person respectively,and integrates the judgment results of the two modules to recognize the abnormal driving behaviors,which will enhance the interpretability of the recognition results and make the model have some ability in fault-tolerant;Using the LSTM unit in the model to capture the temporal characteristics of the sensor information,storing the historical information in the hidden variables;In order to improve the abnormal recognition performance of driver feature module,the vehicle stability which has a great influence on drivers’ state is added to the driver feature distribution model as an additional condition in CVAE;By making use of the Linear Gaussian State Space Model to model the potential variables and transfer the time series,and by making use of the potential variables containing the last moment information to measure the probability of the value of input at the current time under historical patterns learned by model,the impact of data fluctuation on the recognition performance is reduced.In this thesis,experiments are carried out on the driving behavior datasets of Houston University and Texas A & M University.Compared with Random Forest,EM and SVR which converge at 310,550 and 920 times respectively,the proposed data fusion strategy DBN+LSTM converges at about 160 times.Compared with LSTM-VAE,DAGMM and MAD-GAN,LSTM-d SCVAE has improved the F1 scores by + 3.44%,+ 5.25% and + 1.45% respectively.