Research On Microscopic Abnormal Driving Behavior Recognition Algorithm Based On Deep Learning

Traffic accidents have caused a large number of personnel and property safety losses for a long time.Most traffic accidents are related to human factors.Therefore,it is important to avoid the danger caused by improper operation of the driver.With the development of data acquisition technology and computing resources,the analysis of driving behavior by using real-time vehicle motion state has become an important direction of driving behavior research.However,the process of pushing back driving behavior from vehicle motion parameters is very complex.For simple models,it is difficult to identify microscopic abnormal driving behavior adaptively.Studying the driver behavior patterns implied in the motion state of the vehicle requires a more analytical model.Based on the physical characteristics of microscopic driving behavior,this research studies the time series data extraction algorithm for abnormal event recognition.Then,three data modeling components for abnormal driving behavior are discussed,the driving behavior recognition framework,the recognition model based on deep learning,and the driving behavior transfer learning.The deep learning and transfer learning are applied to the microscopic driving behavior research.Firstly,the analysis is based on the vehicle motion state data from real-car driving experiment.Based on the vehicle motion time series data and the idea that motion state remains stable under normal driving conditions,this study puts forward the technique of recognizing abnormal driving behavior based on residuals without time window.Secondly,the CNN-LSTM cascade deep learning model is constructed for microscopic abnormal driving behavior recognition.The microscopic abnormal driving behavior recognition model was trained for normal,aggressive and fatigue driven style.When the training sample is large enough,the accuracy and recall rate of the worst models can exceed 75%.Finally,on the basis of the cascading model,the effect of transfer learning in microscopic driving behavior recognition was tested.Similar but different datasets are used as source and target domains for model transferring.The study also evaluates and analyzes the performance of transfer learning in different target domain sizes and types.The results show that transfer learning can help improve the recognition results of smaller target domains,and when the target domain reaches one-half of the source domain size,the model transferring is basically no improvement.To sum up,this paper explores the process of microscopic driving behavior recognition in the time series of vehicle motion state.Then a deep-transfer learning system was established to analyze driving behavior data.Then the model knowledge transfer was implemented in different scenarios.