Interaction-Aware Prediction for Cut-In Trajectories With Limited Observable Neighboring Vehicles

Abstract

Predicting the future trajectories of risky maneuvers such as cut-in is of great significance for intelligent vehicles to alarm potential crashes in advance. A more critical issue is the use of limited observable neighboring vehicles to achieve reliable trajectory prediction. Motivated by it, this paper studies the trajectory prediction for cut-in maneuvers from ego-vehicle's-view. An interaction-aware model is developed for long-term prediction based on the attention mechanism, where both the encoder and the social interactor are implemented through multi-head attentions and feed forward layers. First, an attention-based trajectory encoder module with temporal embedding is presented to extract hidden features from historical observations. Then another attention-based module is constructed to characterize social interactions between the target vehicle and its neighbors. In the interactor, a novel social mask layer is designed to improve prediction performance in different scenarios, especially for discretionary cut-in maneuvers. Experimental comparison with the state-of-the-art models on public dataset demonstrates the excellent performance of our proposed model on cut-in trajectory prediction. In addition, ablation studies are conducted to highlight the strengths of the main modules in our model.