Driving Behavior Analysis Based On Machine Learning

Intelligent vehicles have become the strategic development direction of the global automotive industry.The driving assistance system is an important part of the intelligent vehicles.To realize intelligent driving assistance,an important prerequisite is the effective evaluation of the driving behavior.Besides,the knowledge of artificial driving process can also promote the intelligent level of the driving assistance system.Therefore,it is of great significance to study the internal mechanism of driving behavior,and to develope more accurate and reasonable driving behavior analysis methods.So,in this paper,the distracted driving behavior,driving style,and driving risk perception are studied by using several machine learning methods.The main contents are as follows:(1)To improve the recognition accuracy of the inattention driving image,a distracted driving recognition method based on improved residual network and spatio-temporal fusion convolutional network is proposed.The experimental results show that the proposed method for recognizing distracted driving behaviors is superior to basic residual network and the identification accuracy can reach to 91.35%.Then the vehicle state and road environment information are analyzed through the information divergence analysis method to construct the traffic situation.By fusing the distracted driving state and traffic situation,an end-to-end distraction behavior risk assessment method are constructed,which can implement fine-grained analysis to distracted driving.(2)For driving style analysis,natural driving data are taken as the research object.By using mathematical model and data mining methods,the driving manipulation style,driving economic style and driving concentration style are studied respectively.Then based on the Analytic Hierarchy Process,the independent driving style results are integrated as single comprehensive result.In detail research process,firstly,the car-following behavior is modeled based on the Gaussian mixture model,and then based on the model,the longitudinal driving style is evaluated according to the safe driving distance.Secondly,based on the convolutional social deep learning network,the lateral driving safety area is predicted.A quantitative evaluation method for the rationality of lateral driving is proposed according to the lateral driving safety area.Thirdly,based on the parallel spectral clustering method,the six types of vehicle status data are analyzed to obtain the driver's macro driving economic style.The experimental results show that the accuracy of parallel spectral clustering for economic style classification is improved by 8.1% and 3.9% compared to k-means and fuzzy kernel clustering methods,respectively.Finally,to realize the effective fusion of multiple driving styles,the Analytic Hierarchy Process are carried out to integrate multi-source driving style evaluation results.(3)Most risk perception research focuses on the assessment of risk perception ability,while few studies pay their attention on constructing the behavior model which can reflect the risk perception behavior feature in real-world driving process.As a result,a prediction model for the driver's risk perception behavior is constructed based on the Long Short Term Memory network.The road environment information and the driver 's risk perception data server as the learning data to the LSTM-based model.The experimental results show that the LSTM-based risk perception model can accurately predict the driver's risk perception behavior,the recognition AUC value is0.81,and the recognition accuracy is better than the method based on support vector machine and multi-layer neural network.Finally,the effectiveness of the proposed risk perception model is studied based on the Ability of Potential Risk Perception assessment model.Compared with the evaluation results based on video clips,the proposed risk perception model can simulate the driver's risk perception ability more accurately.(4)By using the architecture and scheduling mechanism of AUTOSAR,a forward collision warning mechanism is proposed based on the distraction behavior recognition and risk perception model.Firstly,a naturalistic driving data collection system is designed for the collision prevention assistance system to obtain driving behavior characteristics.Then,based on the AUTOSAR architecture,the behavior analysis method is abstracted into independent functional components,and a forward collision warning mechanism based on the adaptive detection of distracted driving is designed according to the AUTOSAR scheduling mechanism.Finally,the simulation is performed according to the simulated dangerous traffic scenario to verify the Effectiveness of the proposed mechanism.Experimental results show that the fusion of distraction recognition and forward collision warning can increase the time for the driver in distracted state avoid the collision risk.At the same time,the adaptive detection of distraction behavior based on the risk perception model can reduce the energy consumption form the distraction behavior detecting.