Research On Driver Style Recognition And Vehicle Behavior Prediction For Intelligent Driving

With the in-depth development of the fourth industrial revolution,the intelligent driving technology with advanced driver assistance system(ADAS)as the core has been developed at a high speed.Both academia and industry have done much theoretical research and productization research on it.Compared with the traditional driving mode,the intelligent driving system under the "host-assisted machine" can cooperate with the driver or replace the driver to develop a driving strategy with good safety performance and high efficiency based on multi-source data and intelligent algorithms,which has many advantages.However,the key technologies of smart cars are not yet fully mature.Especially for decision-making and pre-planning systems that require high algorithm real-time,accuracy,and robustness,there is still a lot of room to improve.This article focuses on the decision-making and pre-planning system of intelligent vehicles under the "host-assisted machine",and adopts scientific and cutting-edge technical routes to carry out research,and realizes driver style recognition and behavior prediction with early warning,real-time,and reproducibility where vehicle behavior prediction includes two stages: driving intention recognition and trajectory prediction.In view of the shortcomings of existing research,three directions are improved and realized based on the deep time series network: driver style recognition,driving intention recognition and trajectory prediction combined with driving intention,which are compared and verified on the NGSIM data set.The specific work and results of this paper are as follows:1)Aiming at the problems that the existing unsupervised clustering research methods of driver style do not or inefficiently use the time dimension information of driver style,cannot completely obtain the driver’s dynamic driving style,only consider the statistical value in a period of time,and use traditional methods to extract features,such as PCA linear dimensionality reduction,this paper is based on the principle of autoencoder based on deep learning,A feature extraction and dimension reduction method of LSTM denoising autoencoder suitable for driving high-dimensional time series is proposed to extract low-dimensional features for style clustering and analysis.Compared with other dimensionality reduction methods and data analysis,it is proved that this method can capture the micro dynamic information on the driver’s time dimension and focus on the driver’s "dynamic change" to define the driver’s style,which is more suitable for real application scenarios and more reasonable.Then,based on the clustering results and the difference of feature distribution of different clusters,drivers are divided into three styles: radical,general and conservative.Finally,the speed,acceleration,headway and lane changing frequency of different styles are analyzed to verify the correctness and rationality of style division.2)Aiming at the problems of insufficient real-time prediction,insufficient accuracy,not based on data driving,and inefficient use of peripheral feature information in the existing driving intention recognition research methods,this paper builds on the problem based on the lane-changing behavior data extracted from the NGSIM data set.The model has been revised to improve the rationality and reproducibility of the study of intention scenarios.The feature level takes into account the neighbor information of surrounding vehicles that is ignored in most current studies and the driver’s style features generated in Chapter 2,and proposes a method based on Bi-GRU’s vehicle driving intention recognition method.Compared with other existing algorithms in the test set,the algorithm proposed in this paper outperforms other existing algorithms with the highest accuracy and F1-Score.The confusion matrix of driving intention recognition is analyzed,and the intention recognition process is visualized on the real left and right lane changing vehicle behaviors.Finally,based on the results of the model output,the early warning of driving intention in vehicle behavior is analyzed,and the change of driving intention recognition is analyzed through different predictive time windows to the successful lane change point,which proves the influence of predictive time on driving intention.3)Aiming at the problems of insufficient prediction accuracy and insufficient use of vehicle information features in the existing research methods of vehicle future trajectory prediction,a multi-information fusion trajectory prediction algorithm based on the attention mechanism is proposed.In the existing trajectory prediction scene,based on the Convolutional Social LSTM network module,the characteristics of motion information,the inherent characteristics of the vehicle and the characteristics of the driver’s style are introduced.Then the attention mechanism is introduced to obtain the surrounding vehicle information that has a greater impact on the subject’s vehicle trajectory and the information at different moments in the subject’s vehicle history from the level of the attention mechanism.Finally,the convolutional interaction layer is optimized.The interaction layer completes the main vehicle feature embedding,and the convolutional layer adds the avg-pooling module to further extract interactive information.Compared with the results of other models,it is proved that the trajectory prediction network based on the attention mechanism has certain advantages.Based on the results,a detailed quantitative analysis of the role of the attention mechanism in the model is made,and the performance of the single-modal model and the multi-modal model based on driving intention are compared.Finally,the experimental results of different historical trajectory durations and different future trajectory durations are analyzed,and the influence of different time windows on the results is analyzed quantitatively and qualitatively.