Trajectory prediction with hierarchical spatio-temporal feature extraction and historical prediction embedding

Abstract

The autonomous decision-making and planning capabilities of intelligent vehicles are key to ensuring safe and efficient driving in complex road scenarios. Accurate prediction of the future behavior of surrounding vehicles is a necessary condition for realizing autonomous decision-making and planning in intelligent vehicles. In this research, a trajectory prediction model (HSTH) based on historical prediction embedding and hierarchical spatiotemporal feature extraction is proposed. aiming to address the key challenges in autonomous vehicle trajectory prediction. Traditional methods perform inadequately in handling non-Euclidean spatial relationships and capturing dynamic associations between time steps, resulting in wasted computational resources and unstable predictions. The HSTH model extracts temporal and spatial features through causal temporal attention modules and spatial graph convolution modules, respectively, and achieves adaptive fusion of both through a gating mechanism. Furthermore, the model uses historical prediction attention embedding to model dynamic correlations between consecutive predictions, improving the stability and consistency of the predictions. To reduce computational complexity, HSTH decomposes the trajectory prediction task into two parts: local context extraction and global interaction modeling, and uses a Gated Recurrent Unit (GRU) network to construct a decoder that further generates the probability distribution of future vehicle trajectories. The proposed model is tested on the Argoverse dataset, and experimental results show that, compared to traditional models, the HSTH model achieves optimal performance. Specifically, under the multi-modal setting (k = 6), HSTH obtains a minADE of 0.7613, minFDE of 1.0981, and MR of 0.1066, while under the stricter single-prediction setting (k = 1), it achieves minADE of 1.5218, minFDE of 3.3851, and MR of 0.5382. These results confirm that HSTH consistently outperforms recent baselines such as LaneGCN, SceneTransformer, and Hivt, providing strong support for autonomous driving decision-making.