Research On Intention Recognition And Trajectory Prediction Method For Vehicles Based On Deep Learning

Automated driving technology is one of the current research hotspots.In order to guarantee the safety and reliability of decision planning for autonomous driving systems in complex traffic scenarios,it is crucial to implement trajectory prediction methods that can effectively cope with complex scenarios.Typically,both vehicle intention recognition and trajectory prediction methods are based on the target vehicle’s historical behavior information and its current environment information to infer the vehicle’s future behavior,intention recognition results can provide reference for longterm trajectory prediction and improve the prediction performance.Current trajectory prediction research has made significant progress,but the application scenarios are mostly limited to sparse traffic flow or closed scenarios.In complex scenarios,the modeling of traffic participant characteristics is not accurate enough,focusing too much on single-vehicle prediction and ignoring multi-vehicle interaction game,poor longterm prediction,low quality of generated trajectories.This research has been supported by the China National Key R&D Program during the 14 th Five-year Plan Period(Grant No.2021YFB2500703),to address the above problems,this paper proposes an intention recognition model based on recurrent neural network method and a Transformer-based multi-vehicle interactive trajectory prediction network,the main research content includes the following aspects:(1)Designing an intention recognition model based on recurrent neural network.Firstly,the driving intention recognition dataset is configured to select the state information of the target vehicle itself and the state information of traffic participants within a certain distance around it as joint inputs,combined vehicle heading angle and lane changes to distinguish vehicle driving behavior.Secondly,build DAE model to control the input feature scale,and build a Bi-LSTM-based intention recognition network.The experimental results show that the network achieves better accuracy even in long-term recognition tasks compared to networks involved in cross-sectional.(2)Designing a Transformer-based multi-vehicle interactive trajectory prediction network.Firstly,data pre-processing work is performed and an expert trajectory database is formed based on an improved K-means clustering algorithm.Secondly,design the target vehicle and the surrounding vehicle state information encoder,complete the accurate modeling of the vehicle’s own state.Design the interaction feature extraction module to fully extract the potential interaction between vehicles and vehicles,vehicles and roads in complex scenes.Fusing expert trajectory database to enhance the ability of models to learn potential rules of operation in the real world.Fusing the intention recognition modules to enhance long-term predictive decision making,and design a multimodal decoder to improve the final output trajectory quality.Finally,the mean square loss function and the negative log-likelihood loss function are combined as the loss function in this paper,and the mean distance error and the final distance error are selected as the evaluation functions to measure the performance of the prediction model.(3)Validating the effectiveness of trajectory prediction network based on Waymo dataset.Firstly,the network performance is analyzed quantitatively using ADE,FDE,min ADE and min FDE metrics.A visualization program is developed to qualitatively analyze the network performance and to visually determine the error of the predicted trajectory as well as the feasibility of the predicted trajectory.The design of ablation study demonstrates that the innovative incorporation of the interaction feature extraction module,expert trajectory database and intention recognition module have effectively improved the model performance.These operations make the predicted trajectory more closely match the real trajectory,improve the overall trajectory quality of the multimodal trajectory,and reduce the generation of bad trajectories.The experiments show that the ADE and FDE evaluation metrics of the trajectory prediction network are 8.5909 m and 11.4638 m in the long-term prediction task of 1.1 s predicting the future 8s.Compared with the existing prediction networks involved in the comparison,the error is reduced by more than 5.9%;in the medium-term prediction task of 5 seconds,the error reduction is also mostly above 4.4%;in the short-term prediction of 3 seconds,the prediction network FDE and min FDE errors are larger,and ADE and min ADE maintain more than 2.7% advantage.