IAHA: Intention-Aware Hybrid Attention Transformer for Trajectory Prediction

Trajectory prediction is an essential support for realizing high-quality and detailed path planning for autonomous driving. To address the challenges of diminished prediction accuracy and limited long-term trajectory forecasting capabilities in highly interactive scenarios, such as on-ramps and intersections, an Intention-Aware Hybrid Attention Transformer network (IAHA) is proposed in this paper, which is employed to achieve reliable and robust long-term trajectory prediction. Within the constructed dynamic vehicle interaction maps, an intention-aware interaction feature extraction module is developed, which consists of social pooling layers and a graph neural network. Moreover, to capture complex spatiotemporal correlations in highly interacting environments, by introducing a hybrid multi-head attention mechanism, an enhanced graph attention network is embedded into the Transformer. Experiments on the public data set exiD demonstrate that the IAHA outperforms the state-of-the-art by 30%, exhibiting superior long-term trajectory prediction performance.