Hybrid Attention with Memory-Based Conditional Refinement for Unified Intent Trajectory Prediction

Abstract

Current mainstream trajectory prediction models have achieved good results in predicting long-term trajectories. However, there is still room for improvement with reference to lane-change prediction. This is because comprehensive evaluation metrics for trajectory prediction may overlook data distribution, affecting the accuracy of lane-change trajectory judgment. This paper proposes a new paradigm based on conditional refinement memory network (HAMR-LSTM) to address these issues. We propose a trajectory classification labeling method based on lane-changing windows for model learning. A hybrid attention mechanism is incorporated in the spatio-temporal feature learning process, with the memory networks are used to refine the temporal attention through conditional searching. The proposed method is evaluated on the publicly available next-generation simulation (NGSIM) dataset. The accuracy of lane-change intention recognition 0.5 s before the lane-change point reaches 96.96%. The improvement in lane-change intention recognition accuracy has also led to enhanced performance in trajectory prediction.