Multi-vehicle Interaction Driving Intention Recognition And Trajectory Prediction In High-speed Mixed Traffic Environment

In multi-vehicle interaction scenarios,predicting the trajectory of surrounding vehicles considering driving intentions is the key to ensure the safe driving of self-driving vehicles.In this paper,we investigate the driving intention recognition and trajectory prediction of the surrounding vehicles in the high-speed mixed traffic multi-vehicle interaction scenario in conjunction with the National Natural Science Foundation of China project "Research on Automatic Driving Situation Perception and Robust Humanoid Decision Planning in High-speed Mixed Traffic Convergence Environment"(52072054):(1)Gaussian hybrid-support vector machine based driving style recognition.Firstly,the low-dimensional feature extraction is performed by Stacked Autoencoder(SAE)for the driving style feature parameters containing interaction features,and then the Gaussian Mixture Model(GMM)is used to cluster and analyze the low-dimensional feature parameters of driving style,and the clustering results are used as the basis to combine the speed and acceleration The results of validation using NGSIM dataset show that the model can accurately identify the driving style of weekly vehicles under the interaction scenario.(2)Driving intention recognition based on improved bidirectional long and short term memory neural network.Considering the problem that existing intention recognition models make less use of driving style and interaction features,a dataset containing multiscale interaction features is first constructed by combining the NGSIM dataset and the style types output from the driving style recognition model.Then,the Bi-directional Long Short-Term Memory(Bi LSTM)network is used to identify the driving intention of surrounding vehicles,and the whale optimization algorithm is used to find the optimal hyperparameters of the Bi LSTM to obtain the optimal learning rate and the optimal number of implicit layer nodes to avoid the negative impact of manual tuning on the model and improve the network’s After the verification of NGSIM dataset,the intention recognition model with the introduction of driving style and interaction features has a high recognition accuracy for weekly driving intention.(3)Trajectory prediction based on improved two-way gated recurrent neural network.Since the driving data in the interaction scenario is a high-dimensional time-series data containing features such as trajectory sequence,driving intention,driving style,and interaction behavior,we use Temporal Convolutional Neural Network(TCN)to extract its time-series features and introduce Multi-Head-Attention(MHA)mechanism.(MHA)to extract the interaction relations in the sequences and obtain the feature sequences containing the attention degree.To avoid information loss and network degradation,the outputs of the temporal convolutional network and the multihead-attention model are connected with residuals to obtain features containing important information and allaround information,and then input to the bidirectional gated recurrent network for decoding learning and generating prediction trajectories.The comparative analysis with other models and simulation experiments show that the composite prediction model can effectively extract temporal features,interaction relations,and have high prediction accuracy when predicting weekly vehicle trajectories.