关键词: 轨迹预测, 深度学习, 编码-解码框架, 人工势场, 注意力机制

Abstract: Accurate obstacle trajectory prediction is a key to correct decision making and precise control of autonomous vehicle. Considering the influence of lane information and surrounding obstacles on vehicle motion in complex environment, a vehicle trajectory prediction model aggregated vehicle-lane information is proposed based on encoding-decoding framework. Firstly, the directed graph is adopted to describe the lane nodes. Then the target vehicle and surrounding obstacles are encoded via GRU. Simultaneously, the artificial potential field is introduced to represent the vehicle-vehicle interaction. By concatenating the lane node vector and repulsive force vector, the attention mechanism is utilized to explore the spatial-temporal coupling mechanism. Finally, the multi-modal trajectory prediction of obstacles is achieved by scoring and clustering the lane nodes through the strategy network. The proposed prediction model is trained on the nuScenes motion prediction dataset and evaluated with the state-of-the-art baseline models’ performance. The results demonstrated that the proposed prediction model has lower prediction error and better robustness. On the other hand, it has better interpretability by introducing the potential field into attention mechanism.

Key words: trajectory prediction, deep learning, encoding-decoding framework, potential field, attention mechanism