Vehicle trajectory prediction with temporal–channel attention and GRU

Accurate vehicle trajectory prediction is crucial for autonomous driving systems, as it provides essential inputs for downstream tasks such as behavior planning and motion control. In complex traffic environments, a vehicle’s future motion depends not only on its own historical dynamics but also on interactions with surrounding vehicles, resulting in strong spatiotemporal dependencies. Although recent deep learning approaches have achieved promising performance, many existing methods focus primarily on modeling temporal sequences or spatial interactions, while insufficiently considering the heterogeneous importance of different feature channels in vehicle motion data. To address this limitation, this paper proposes a vehicle trajectory prediction framework that jointly models temporal dependencies and feature-channel relationships through a dual-attention architecture. Specifically, a temporal attention module is designed to capture long-range temporal dependencies in motion sequences, while a channel attention module learns adaptive importance weights for heterogeneous feature channels. A gated fusion mechanism is further introduced to integrate the complementary representations from the two branches. Based on the fused representation, a GRU-based decoder generates multi-step future trajectory predictions. The proposed model takes as input the motion history of a target vehicle and its six surrounding vehicle classes, and predicts trajectories over a 5 s prediction horizon using 3 s historical observations. Extensive experiments on the HighD and NGSIM datasets demonstrate that the proposed framework consistently outperforms several representative baseline methods across different prediction horizons, achieving average RMSE values of 0.78 m and 1.36 m, respectively. Ablation studies further confirm the effectiveness of the temporal attention module, channel attention mechanism, and gated fusion strategy. Overall, the proposed approach provides an effective solution for capturing both temporal dynamics and feature interactions in vehicle motion data, enabling more accurate and robust trajectory prediction in complex traffic environments.