28 April 2026 in IEEE Internet of Things Journal
The ability to sense vehicles in the surrounding environment and predict the trajectories of these vehicles is critical to the safe planning and control for autonomous driving. There are two problems with the current mainstream Graph Neural Networks (GNN)-based interaction-aware models: first, the weight matrices of GNNs are either predefined or model-learned, and there is no method of combining the two types of weight matrices; second, the deep learning-based model lacks interpretability. In response to the above problems, an interpretable DCM-GAT-GRU model based on the weighted Graph Attention Networks (GAT) and the Learning Multinomial Logit model (L-MNL) sampler is proposed. The weighted GAT combines both predefined and model-learned weight matrices to fully extract environmental interactions and improve prediction accuracy. The L-MNL sampler utilizes a stochastic feature sampling strategy to optimize feature fusion, thereby increasing both the interpretability and prediction accuracy of the model. Experiments on the autonomous driving open dataset nuScenes show that the multimodal trajectories predicted by the proposed model is competitive among leading vehicle trajectory prediction models, where minADE5, minADE5 and minFDE1 improve by 0.88, 0.83 and 2.28 compared to the baseline model, while providing interpretable estimation of explanatory variables. This paper contributes to the exploration of interpretable mechanisms for trajectory prediction and promote the safety of autonomous driving.