Personalized Driver Model And Its Application To Driving Behavior Evaluation

Vehicle test is the key link of the development of vehicle control system.Establishing an appropriate driver model to perform the vehicle test can greatly shorten the development cycle of the onboard control system.Because the driving styles can directly affect the vehicle test result,it is desirable to establish personalized driver models.In this paper,a novel method for establishing style-oriented driver models by imitating human driving behaviors is proposed.The styles of the original human drivers can be retained while accomplishing the speed following task.Considering the divergence and local mutability of the real-world data,the neural network algorithm is utilized together with the direct inverse model approach to accomplish this model,which is used to replace the real driver or PID-like model for FTP test.The effectiveness of the proposed scheme is verified by VTD and FTP based test,as well as the practical applications.Furthermore,a preliminary exploration on qualitative evaluation of driving style and abnormal driving behavior detection is proposed based on the established driver model.The detailed research work and contribution are shown as follows:(1)Personalized driver modelling by imitating human driver behavior.In order to integrate different driving styles into the vehicle test system and make the results much closer to the reality,a novel method for driver modeling based on real-world data and neural network is proposed.Considering the divergence and local mutability of the real-world data,the cerebellar model articulation controller(CMAC),a locally designed neural network model,is utilized together with the direct inverse model approach to accomplish this model,which is used to replace the real driver or PID-like model for laboratory federal test procedure(FTP)test.On the other hand,a vehicle test data(VTD)and neural network based vehicle model is established and employed for simulation test.VTD and FTP based test are conducted to verify the effectiveness of the proposed scheme.The personalized driver model is able to retain and reproduce the driving styles of the original human drivers while accomplish the speed of the following task.(2)Quantitative evaluation of driving styles based on phase space reconstruction.Firstly,the driving behavior is normalized to eliminate the impact of driving environment.Secondly,the normalized driving behavior is analyzed by phase space reconstruction and a Driving Style Index is proposed based on the correlation dimension for quantitatively evaluating the aggressiveness of the driving style.Finally,the index is applied to the recognition of driving styles.Simulation results are conducted to verify the effectiveness of the proposed scheme.(3)Abnormal driving behavior detection based on deep learning network of sparse auto-encoder.Firstly,the abnormal driving behaviors are characterized and simulated based on the driving behavior normalization and the present research.Secondly,the abnormal driving behavior detection system is established based on the deep neural network of sparse auto-encoder.In addition,to increase the robustness of feature expression,the method of denoising auto-encoder is employed in the input of sparse auto-encoder.The "Dropout"technique is also introduced into the entire training process to reduce the prediction error caused by "overfitting".Finally,the comparison tests are conducted to show the effectiveness of the proposed scheme.