Abstract:
To address the difficulty of accurately predicting vehicle trajectories in highway scenarios due to interactive behaviors and driving intentions, a multimodal vehicle trajectory prediction model named Spatio-Temporal Intent-Modal network (STIM), which fuses spatio-temporal features and considers driving intentions, is proposed. An encoder is used to extract the motion features of the target vehicle. An interaction attention module fuses the vehicle state information with the interaction information of surrounding vehicles and calculates their influence weights on the target vehicle. The interactive features and the target vehicle’s state features are then fed into a driving intention fusion module, where feature fusion and Softmax classifiers are used to predict the probability distributions of lateral and longitudinal driving intentions. Based on the intention-aware features, a multimodal trajectory prediction module generates bivariate Gaussian distribution parameters for future trajectories, achieving probabilistic modeling of multimodal trajectories; during inference, the most probable trajectory is selected according to the driving intention classification. Experimental results on the CitySim dataset show that STIM outperforms the compared baselines in RMSE over the 1–5 s prediction horizons, and its 5 s RMSE is 3.02 m, which is 5.3% lower than that of the best baseline HLTP, demonstrating high accuracy and stability in lane-changing trajectory prediction. The proposed model provides an effective method for vehicle trajectory prediction in highway scenarios for autonomous driving.