LSTM-Based Intelligent Vehicle Human-Like Driving Intention Recognition And Trajectory Prediction

The rapid growth of domestic vehicle ownership has increased the occurrence of traffic congestion and traffic accidents.As an important way to improve the safety of vehicle driving,autonomous driving system can process historical information through prediction modules to predict the movements and trajectories of surrounding vehicles,avoid possible traffic accidents or activate passive safety functions in advance to achieve active-passive safety integration.Traditional prediction models aim to improve prediction accuracy,but rarely consider the influence of driver characteristics,and the driving styles studied in the past are assumed to be stable and unchanging,and rarely consider the influence of the environment on driving styles.In order to study the characteristics of driving styles changing with the environment,improve the recognition accuracy of the current prediction module,and seek a prediction output more in line with driver characteristics,this paper proposes an intention recognition and trajectory prediction model considering dynamic driving styles to achieve human-like intention recognition and trajectory prediction.The specific research of this paper is as follows:(1)Aiming at the problems that traditional driving style recognition rarely consider the impact of environmental changes and one driver not only has one driving style.a subjective driving style questionnaire is designed first.The data dimension is reduced through principal component analysis,and the driving styles of respondents are divided into three categories by K-Means clustering to qualitatively analyze the influence of road environment on drivers’ driving styles.Secondly,simulation experiments of different experimental scenarios are designed.Drivers are selected from the questionnaire to collect simulation driving data.Through data processing and clustering,driving characteristics of all kinds of people are analyzed in detail according to the data statistics,as well as the changes of drivers’ characteristic values under different environments,the transformation of driving style and the sensitivity to different changes.Finally,a driving style neural network classification model is designed and established,and objective driving data is used for training and verification,so as to realize dynamic recognition of drivers’ driving styles.(2)In view of the current intention recognition model that rarely considers driver characteristics and low recognition accuracy,this paper select data of different roads and time points in High D dataset as the original data set,select vehicle characteristics information as input,and take TTC of surrounding vehicles as consideration of the impact of surrounding vehicles,divide lane changing into intention generation-change of lane-change of lane success in three stages.The lateral movement threshold of the vehicle is taken as the criterion to determine the start of lane changing,and the data set of time window size is intercepted based on it.Build a human-like Long Short-Term Memory Network(LSTM)intention recognition model and compare and analyze the role of the human-like module in intention recognition.The driving habits of drivers with different driving styles are summarized through differences in their driving intentions,and the effects of different time windows of intercepted data on recognition accuracy are also verified qualitatively and quantitatively.(3)An LSTM-based human-like trajectory prediction model with improved attention mechanism is proposed to solve the problems of low long-term prediction accuracy of the mainstream prediction model and shallow combination of model and driver.The new database structure is redesigned to introduce driving style and driving intention data in the aspect of human factors.Based on the convolution calculation of surrounding vehicles by CS-LSTM,the feature extraction of surrounding vehicles in space is realized.At the same time,dynamic driving style recognition model and human-like driving intention recognition model are introduced.In terms of time,the decoder attention mechanism is introduced to strengthen the emphasis on historical key information,and the attention mechanism is improved based on driving style to achieve a weight allocation in accordance with driver characteristics.By lateral comparison with the prediction errors of other models,the DACS-LSTM model proposed in this paper has certain advantages in the human-like trajectory prediction of long-term prediction tasks.The prediction trajectories of the ACS-LSTM and the proposed model in this paper are visualized,and the influence of anthropomorphic module on the predicted trajectory is analyzed.Compared the performance of the two models,the trajectory,predicted by DACSLSTM,is more consistent with the actual trajectory.The weight allocation of traditional attention mechanism and improved attention mechanism is intercepted and the trajectory changes are analyzed by combining driving styles.