Trajectory Prediction For Autonomous Driving In Complex Environments

Trajectory prediction in complex environments is a key task in autonomous driving systems,predicting trajectories can help driving systems complete future planning and decision-making,and is of great significance for improving the safety performance and landing applications of autonomous driving.In recent years,with the efforts of a number of research institutions and technology companies,there has been some progress in the field of trajectory prediction,usually based on deep learning time series models to solve such problems.However,due to the complexity of the prediction scenario and the existence of the long-term dependence problem of time series model,there is no more universal and recognized solution at present,and this paper hopes to further study and experiment on the basis of the predecessors.In this paper,starting from the basic needs of trajectory prediction tasks,combining relevant literature and previous methods for theoretical analysis,different research strategies are adopted for different categories of prediction targets.The model design focuses on pedestrians’ historical behavior,social interaction considerations and intention estimation.The model design focuses on vehicles’ observations information,road semantics,and dynamic constraints.By combining the above theoretical analysis,predicting models for different categories of targets are designed.Both pedestrian and vehicle trajectory prediction models are based on an encoder-decoder network architecture.For pedestrian trajectory prediction,a multi-view transformation prediction model MVT-GAN based on adversarial generation and multi-head attention is proposed.It considers social interaction through the multi-view transformation method,adds an end point prediction module to complete the intention estimation,and uses the multi-trajectories generation method to output.For vehicle trajectory prediction,a prediction module DMVF based on multi-view fusion and high-definition map is proposed.The multi-view fusion method is used to increase the observation information,and an efficient road semantic feature extraction module is designed.Finally,the dynamic model is used to iteratively output the trajectory.The MVT-GAN and DMVF models achieve competitive prediction performance through experiments on the corresponding datasets,and the actual utility of each module was verified by ablation experiments,and the prediction results were visually analyzed.