Research On Vehicle Trajectory Prediction Based On The Improved Transformer Model

Autonomous vehicles need to make safe and efficient decisions to pass through the complex traffic scenes,such as changing lanes,overtaking or decelerating.Therefore,autonomous vehicles need to reasonably infer the future movement of surrounding agents.It is significant to research on vehicle trajectory prediction.The prediction results can be used for vehicle path planning,autonomous navigation,traffic prediction and congestion management.In this paper,a research based on improved Transformer is carried on to predict vehicle trajectory in complex traffic scene.The main work is summarized as the following aspects:(1)Firstly,starting from the three technical difficulties of trajectory sequence modeling,multi-agent interaction and multi-modal trajectory,the existing trajectory prediction models and methods are comprehensively investigated.It is found that most of the classical trajectory prediction methods are based on pedestrian datasets.Especially in trajectory sequence modeling,the trajectory prediction method based on Transformer model still not verifies on the effectiveness of the model in traffic datasets.Therefore,this paper selects the NGSIM dataset with multiple traffic flow for preprocessing.Experiments are carried out based on the Transformer model,which verifies the feasibility of the application on vehicle trajectory prediction.(2)A vehicle trajectory prediction method combining channel attention mechanism and Transformer model is proposed.Channel attention mechanism is added to model the interaction between vehicles.Based on the structure of Transformer model,it has advantages of capturing long-time dependency and dealing with absence input data.The historical trajectory data of multi-agent are used to predict the position information under the implicit interaction.The results of NGSIM and INTERACTION show that Transformer is suitable for vehicle trajectory prediction in complex traffic scenes.The interactive modeling method with channel attention mechanism also further improves the prediction results.(3)In order to achieve multimodal trajectory prediction,this paper improves the decoder structure of Transformer model combined with conditional variational auto encoder(CVAE).Different from the multimodal trajectory prediction method that generates all future trajectory sequences,this paper uses CVAE to only infer the potential distribution of the endpoint of trajectory and predicts the position of remaining trajectory sequences combined with the context information of the historical trajectory and the future endpoint.It solves the problem of large endpoint deviation caused by error accumulation.This paper focuses on the bidirectional interaction modeling between vehicles.The self-attention mechanism and multi-head attention mechanism are combined to learn the behavior representation between different vehicles in traffic scene.The inverse reinforcement learning method is set to learn the evaluation method of multimodal model,which takes the ground truth data as teaching example and learns the reward function by neural network in decision refinement module.The experiment analyzes the qualitative and quantitative experiment of the baseline models,which shows that the improved model achieves promising performance on NGSIM dataset and INTERACTION dataset over the other existing methods.In conclusion,this paper researches on the vehicle trajectory prediction based on the improved Transformer model in complex traffic scene.The model structure and evaluation method were optimized,which improved the prediction result of vehicle trajectories.