基于Transformer的自动驾驶交互感知轨迹预测

针对自动驾驶运动规划中预测周围交通参与者(如车辆、自行车、行人)未来轨迹的问题,提出了一个基于Transformer的轨迹预测模型(Trajectory Prediction Transformer,TPT)来帮助自动驾驶车辆预测周围交通参与者的未来运动轨迹。首先,为了有效地考虑交通参与者和交通环境之间的交互信息,将交通参与者建模为交通智能体。并将交通智能体的历史运动轨迹和周围交通环境信息编码为多通道图,作为模型的输入。然后,利用改进的Transformer对交通环境进行建模,并捕捉交通智能体与交通环境之间值得关注的交互信息,预测其未来运动轨迹。最后,在大规模自动驾驶数据集Lyft进行的实验表明,TPT模型能够在不同预测时长下取得优于其他对比模型的预测结果,且用时更短。

Transformer-based interaction-aware trajectory prediction for autonomous driving

Aiming at the problem of predicting the future trajectories of surrounding traffic participants (such as vehicles, bicycles, and pedestrians) in autonomous driving motion planning, a Transformer-based trajectory prediction model (Trajectory Prediction Transformer, TPT) is proposed to help self-driving vehicles predict the future motion trajectories of surrounding traffic agents. First, to effectively consider the interaction information between traffic participants and the traffic environment, traffic participants are modeled as traffic agents. The historical motion trajectories of traffic agents and the surrounding traffic environment information are encoded as a bird"s-eye view as the input of the model. Then, the traffic environment is modeled using the improved Transformer, and the noteworthy interaction information between traffic agents and the traffic environment is captured to predict their future motion trajectories. Finally, experimental validation is conducted on the large-scale autonomous driving dataset Lyft, experimental results show that the TPT model can achieve better prediction results than other comparative models with reduced time.