Hierarchical vector transformer vehicle trajectories prediction with diffusion convolutional neural networks

In dynamic and interactive autonomous driving scenarios, accurately predicting the future movements of vehicle agents is crucial. However, current methods often fail to capture trajectory uncertainty, leading to limitations in trajectory prediction performance. To address these limitations, this paper introduces the hierarchical vector transformer diffusion model, a novel trajectory prediction method that prioritizes both speed and accuracy. The proposed model decomposes the traffic scene modeling into local patches and global interactions, allowing for the acquisition of relevant environmental and global information. Moreover, a local diffusion encoder is employed to effectively capture the aleatoric uncertainty. The proposed model utilizes an adaptive graph structure to exploit the spatial and temporal relationships inherent in the trajectory data. By employing a graph diffusion process, the model effectively captures dynamic features from the historical trajectory information. Moreover, the model demonstrates adaptability by dynamically adjusting to diverse trajectory data and scenarios, thereby enabling the generation of predicted trajectories that are uncertainty aware. This approach contributes to more effective and efficient modeling of dynamic autonomous driving scenarios. Experimental results demonstrate the superior speed and accuracy of the proposed method compared to existing approaches for trajectory prediction. The proposed method significantly enhances prediction accuracy, achieving results of ADE 0.68 and FDE 1.02 on the Argoverse dataset. In comparison to the baseline model, there are notable improvements in ADE and FDE by 0.03 and 0.06, respectively. It is noteworthy that this method also reduces the inference time by 7% when compared to the currently fastest method.