MTPI: Multimodal Trajectory Prediction for Autonomous Driving via Informer

The multimodal prediction of surrounding vehicle trajectories serves as a valuable tool in assisting autonomous vehicles in making safe and rational decisions. The complexity and partial observability of autonomous driving scenarios result in the generation of dense data between target vehicles and moving entities as well as traffic scenes, which includes static data such as the number of lanes and dynamic data with timing information. Due to the impact of dense data on the accuracy of trajectory prediction, it becomes crucial to model different sets of input features. Modeling different input feature sets can provide a more comprehensive and accurate representation of the data and ultimately improve the accuracy of trajectory prediction. In this paper, we propose a multimodal trajectory prediction method for autonomous driving based on Informer, called MTPI. The MTPI model uses an encoder-decoder structure based on Informer for multimodal trajectory prediction. The proposed model enhances trajectory prediction accuracy by modeling various input feature sets, enabling a better understanding and capturing of the relationship between the target vehicle, moving entities, and the traffic scene. To optimize internal memory utilization and reduce computational complexity, we incorporate self-attention distilling and probsparse self-attention operations based on Informer into the model. Experimental results on the Argoverse dataset demonstrate that MTPI achieves outstanding performance in multimodal trajectory prediction.