Toward Human-Like Prediction: Vehicle Trajectory Prediction via Velocity-Aware Complementary Interaction Transformer

Vehicle trajectory prediction (VTP) poses unique spatial-temporal modeling challenges, as a human-like prediction requires considering fine-grained interactions. Prior models have often used various attention mechanisms to extract key spatial-temporal interactions from the foreground, thus missing background information processing. However, this may result in the model’s insufficient generalization ability in a specific modality. Here this paper presents VCIFormer, a velocity-aware complementary interaction Transformer designed to enhance vehicle trajectory prediction by capturing complex spatial-temporal interactions between foreground and background with context awareness. VCIFormer combines inverse attention with traditional spatial-temporal attention as a complementary mechanism, applying bidirectional optimization to capture foreground and background attention flows. An adaptive visual mask is also developed to align attention allocation with human visual patterns at varying velocities. It enables the model to prioritize critical regions analogous to human driving behavior. Moreover, a context-aware encoder, consisting of a surround-aware module and a motion-enhancement module, is incorporated to provide additional interaction cues and spatial information. VCIFormer is evaluated on six real-world datasets (NGSIM, HighD, RounD, ExiD, Argoverse, and nuScenes) and attains state-of-the-art performance in critical metrics. For example, in comparison to baseline models, there are significant improvements in ADE and FDE by 2.84-18.18% and 9.88-13.33% on the NGSIM, HighD, RounD, and ExiD datasets, respectively. In sum, compared with previous architectures, VCIFormer presents a more effective combination of spatial-temporal interaction layers and context awareness for VTP.