行人轨迹预测条件端点局部目的地池化网络

Local Destination Pooling Network for Pedestrian Trajectory Prediction of Condition Endpoint

轨迹预测是自动驾驶系统中的核心任务之一。现阶段基于深度学习的轨迹预测算法,涉及目标的信息表示、环境感知和运动推理。针对现有轨迹预测模型在运动推理过程中对行人社交动机考虑不足,无法有效预知场景中行人在不同社交条件下局部目的地的问题,该文提出一种条件端点局部目的地池化网络(CEPNET)。该网络通过条件变分自编码器推理潜在轨迹分布空间来学习历史观测轨迹在特定场景中的概率分布,构建条件端点局部特征推理算法,将条件端点作为局部目的地特征进行相似性特征编码,利用社交池化网络过滤掉场景中的干扰信号,融入自注意力社交掩码来增强行人的自我注意力。为验证算法各模块的可靠性,使用公开的行人鸟瞰数据集(BIWI)和塞浦路斯大学多人轨迹数据集(UCY)对CEPNET进行消融实验,并与平凡长短时记忆网络(Vanilla)、社交池化生成对抗网络(SGAN)和图注意力生成对抗网络(S-BiGAT)等先进轨迹预测算法进行对比分析。在Trajnet++基准上的实验结果表明,CEPNET算法性能优于现有先进算法,并且与基准算法Vanilla相比,平均位移误差(ADE)降低22.52%,最终位移误差(FDE)降低20%,预测碰撞率Col-I降低9.75%,真值碰撞率Col-II降低9.15%。

Trajectory prediction is one of the core tasks in automatic driving system.At present,trajectory prediction algorithms based on deep learning involve information representation,perception and motion reasoning of targets.Considering the problem that the existing trajectory prediction models does not take into account the social motivation of pedestrians and can not effectively predict the local destination of pedestrians in different social conditions in the scene,a Conditional Endpoint local destination Pooling NETwork(CEPNET)is proposed.The network uses conditional variational autoencoder to map out the potential distribution in space,which can study the observation of the history track probability distribution in the specific scene.And then a local feature inference algorithm is built to code the similarity features of conditional endpoint as local destination features.Finally,the interference signals in the scene are filtered out by social pooling network.At the same time,self-attention social mask is used to enhance pedestrian’s self-attention.In order to verify the reliability of each module of the algorithm,the public datasets of Walking pedestrians In busy scenarios from a BIrd eye view(BIWI)and University of CYprus multi-person trajectory(UCY)are used to conduct ablation experiments,and compared with advanced trajectory prediction algorithms such as Vanilla,Socially acceptable trajectories with Generative Adversarial Networks(SGAN)and multimodal Trajectory forecasting using Bicycle-GAN and Graph Attention networks(S-BiGAT).The experimental results on the Trajnet++benchmark show that compared with the benchmark Vanilla algorithm,the Average Displacement Error(ADE)is reduced by 22.52%,the Final Displacement Error(FDE)is reduced by 20%,the predicted collision rate Col-I is reduced by 9.75%,and the true collision rate Col-II is reduced by 9.15%.