A non-local multi-head spatiotemporal attention LSTM for vehicle trajectory prediction

Abstract

Vehicle trajectory prediction (VTP) is the primary part of the perception-planning-control pipeline in autonomous driving. The inter-vehicle interactions strongly shape the trajectory of vehicles in complex traffic scenarios. Neural networks that use recurrent or convolutional architectures often exhibit performance degradation in longer prediction horizons. Many existing approaches confine interaction modeling to a predefined spatial neighborhood and a narrow temporal window, while implicitly assuming that all neighboring vehicles have equal influence on the future trajectory of target vehicle. This results in error accumulation and the inability to capture long-range spatiotemporal dependencies. To overcome these limitations, a non-local multi-head spatiotemporal attention based long short-term memory model (NL-MHA-LSTM) is introduced which employs an attention mechanism to assign context weights to relevant neighbor vehicles. It extends beyond pairwise effects to model long-range dependencies. The model emphasizes the most influential vehicles and formulate position-aware interaction representations. A comprehensive set of experiments is conducted on the pubicly available HighD dataset. The results demonstrate that the proposed model outperforms all state-of-the-art methods, achieving a 53.4% reduction in RMSE at the 5s prediction horizon relative to the next-best model. In addition, a detailed ablation study is conducted to systematically evaluate the impact of different attention mechanisms and varying numbers of attention heads on prediction accuracy across multiple horizons.