Research On Driving Behavior Recognition And Prediction And Early Warning Methods For Intelligent Connected Environment

In recent years,technologies and functions related to intelligent & connected vehicles have been developed rapidly because of the support of national policies and the effective applications of technological innovation achievement such as communication and artificial intelligence.The rise of intelligent & connected vehicles has brought new development opportunities and challenges to improve the stability and safety of road transportation system.In the development process of gradual realization of vehicle-road cooperation and autonomous driving,the heterogeneous mixed traffic flow consisting of human-driven vehicles and intelligent &connected vehicles with insufficient network connectivity,unequal intelligence and low penetration rate has become the mainstream scenario of road traffic and will be maintained for a long time.In the context of human-machine co-driving,the study of driving behavior for the purpose of ensuring the safe driving of intelligent & connected vehicles has become a popular direction in the field of transportation engineering and vehicle engineering.However,so far,due to the diversity of scenarios in mixed traffic environments and the uncertainty of humandriven vehicles,there are still many problems in related research: how to define and mine the abnormal driving behaviors of vehicles in lateral as well as longitudinal movements,and establish effective,multiple and massive training sets of driving behaviors;how to predict the potential driving intentions and behaviors of human-driven vehicles in complex traffic environments;how to make accurate predictions of driving trajectories in a connected multivehicles environment,and evaluate,analyze and warn about the dangerous situation of vehicles.This paper makes full use of the advantages of deep learning models in feature mining,aims to improve the driving safety of intelligent & connected vehicles,and focus on the recognition,prediction and early warning needs of driving behaviors.This paper takes the heterogeneous and mixed traffic environment as the research background,and uses the highquality vehicle driving state data as the research basis.This paper expects to make a breakthrough on the key question "How to improve the safety performance of intelligent &connected vehicles through driving behavior research?".The main innovations are:(1)Aiming at the abnormal driving behavior that seriously affects the stability of traffic flow,a deep learning-based abnormal driving behavior recognition method is proposed.A behavior-aware real-vehicle test environment consisting of a multi-source data collection platform and a closed test site is built to realize abnormal driving behavior samples collection,different behavior data characterization patterns are analyzed and corresponding definition methods are proposed.A multi-dimensional multivariate abnormal driving behavior training set is built based on open source data sets.A deep residual network model with self-attention mechanism is proposed to realize the organic combination of heterogeneous models,and the training set is used to optimize the model parameters.The validation experimental results show that the proposed method can recognize abnormal driving behaviors up to 97% and on average up to about 95%,and realize the mining of abnormal driving behaviors in historical trips,which can effectively and accurately identify abnormal driving behaviors.(2)Based on the uncertainty of driving behavior of traditional human-driven vehicles and the importance of accurate prediction for driving intentions and behaviors in complex environment for the active safety warning of intelligent & connected vehicles,driving behavior prediction research is carried out.The research introduces finite state machine theory to simplify the incentive-behavior-state transformation mechanism,and a driving behavior prediction method based on game theory and deep learning is proposed.The prediction model for lane-changing decision based on Stackelberg game is built.The common behaviors of highway scenario are analyzed to build a typical driving behavior sample set.In response to the need for both accuracy and timeliness in the process of driving behavior prediction,the deep learning model is designed to be lightweight while achieving the prediction of typical driving behaviors.The validation experimental results show that the game theory model achieves the purpose of outputting the potential lane-changing probability of vehicles,the typical driving behaviors can be effectively predicted by lightweight deep learning algorithm model,and the organic combination of the two models achieves a prediction accuracy of about 94% for typical driving behaviors within 0.2-0.4 seconds.(3)Based on the demand of real-time collision risk warning for intelligent & connected vehicles,a dangerous situation warning method based on driving behavior perception is proposed.For the human-machine co-driving characteristics in heterogeneous mixed traffic environment,a human-driven vehicle trajectory prediction model is constructed under the premise of predictable driving behavior,with long short-term memory network as the encoderdecoder and the social tensor pool is used to realize scene data coupling and achieve effective prediction of trajectories.The local candidate path generation method based on third-order Bézier curve is proposed for the target vehicle,which combines the scene boundary to cover the potential trajectory of the target vehicle at the fine-grained level.The collision risk of vehicles under different scenarios is analyzed by using the artificial potential field method.The validation experimental results show that the trajectory prediction algorithm model has high accuracy for human-driven vehicles,and it can achieve accurate warning of the dangerous situation by combining with the local candidate path of the target vehicle,which can be used for the active safety warning of intelligent & connected vehicles.This paper conducts an in-depth study on the theoretical,methodological and technical aspects of the driving behavior related to the driving safety of intelligent and connected environment under heterogeneous and mixed traffic flows.The models and algorithms proposed in this paper can provide strong theoretical and technical support for the active safety warning system of the intelligent & connected vehicles and effectively improve the driving safety performance of the vehicles.