A Risk-Aware Lightweight Vehicle Trajectory Prediction Framework for Right-Turn–VRU Interactions at Urban Intersections

At urban intersections, trajectory prediction supports downstream planning, particularly for right-turning vehicles interacting with vulnerable road users (VRUs). Conventional proximity-based interaction modeling may fail to distinguish interactions relevant to future vehicle motion from weakly related neighbors. To address this issue, this study proposes a lightweight Risk Interaction Spatio-Temporal Graph Convolutional Neural Network (RI-STGCNN), which extends Social-STGCNN through risk-informed selective information propagation. An Interaction Risk Index (IRI) integrating relative distance, relative velocity, Time to Closest Approach (TCA), and participant type is used to characterize interaction relevance, while risk-aware filtering and feature fusion selectively propagate interaction information. The model generates a single deterministic 12-step trajectory from eight historical frames. Experiments on the SinD dataset show that RI-STGCNN achieves an ADE of 0.68 m and an FDE of 1.31 m, reducing the errors of Social-STGCNN by 16.0% and 13.8%, respectively. In right-turn–VRU conflict scenarios, the reductions reach 27.9% and 24.1%. Additional comparisons show lower errors than Multiclass-SGCN, whereas HiVT-64 achieves lower errors with a substantially larger parameter count. An exploratory analysis of 720 training-set interactions finds an association between IRI and future vehicle–VRU separation (Spearman’s ρ = −0.401), without establishing calibrated collision risk. With 19.6 k parameters and a single-sample inference time of 1.75 ms on a GTX 1650, the proposed model provides a compact approach to single-trajectory vehicle prediction.