基于多信息融合的驾驶视角下行人轨迹预测

Pedestrian trajectory prediction from driving perspective based on multiinformation fusion

行人轨迹预测是实现在城市内完全自动驾驶的重要支撑,并且广泛应用于机器人路径规划、自主巡航等领域.驾驶视角下交通场景复杂多变、行人未来位置不确定性大,只考虑观测轨迹信息预测行人轨迹会有较大位移误差.针对这个问题,提出一种多信息融合网络(multi-information fusion network,MIFNet)来预测驾驶视角下未来行人轨迹的多种可能.MIFNet在观测轨迹信息的基础上引入姿态信息和光流信息,分别采用骨架序列重组和划分局部光流的方法避免遮挡造成的信息失真.为了更有效地融合这些信息,提出一种基于信息评价的跨信息融合注意力机制,综合考虑了预测过程中不同信息间的重要程度和同一信息间不同特征的重要程度.MIFNet在PIE数据集上预测1.5 s的平均位移误差取得了最佳成绩,在JAAD数据集1.5 s的长时轨迹预测任务中预测误差最小,并且模型参数量、推理时间较最新模型大幅度下降.

Pedestrian trajectory prediction is an important support for fully automatic driving in the city,and is widely used in robot path planning,autonomous cruise and other fields.The traffic scene from the driving perspective is complex and changeable,and the future position of pedestrians is uncertain.In this paper,a multi-information fusion network(MIFNet)is proposed to predict multiple possibilities of future pedestrian trajectories.Pedestrian posture information and optical flow information are used in the MIFNet on the basis of observed trajectory information,and the ways of reconstructing skeleton sequences and dividing local optical flow are used to avoid information distortion caused by pedestrian occlusion.In order to fuse these information more effectively,this paper proposes a crossinformation fusion attention mechanism based on information evaluation.The importance of different information in the prediction process and the importance of different features between the same information are comprehensively considered.The MIFNet achieves the best results in predicting the average displacement error of 1.5 seconds on the PIE dataset,and the long-term trajectory prediction task of 1.5 seconds on the JAAD dataset.The prediction error is the smallest,and the number of model parameters and inference time are greatly reduced compared with the latest model.