【摘要】 行人轨迹预测对智慧城市建设、公共危机管理具有重要意义.复杂场景中的行人轨迹不仅包含行人个体运动时序性特征,还包含行人与周围其他运动实体之间的交互性特征.如何根据场景变化,对这种时序性和交互性特征进行深度刻画并进行轨迹预测,是复杂场景行人轨迹预测的关键问题.本文采用多头注意力机制和对抗生成方法,提出一种基于多头注意力机制的生成对抗网络模型(Multi-head Attention Generative Adversarial Model,MAGAM),对复杂场景下多行人轨迹进行建模.论文首先通过多头注意力机制融合行人的相对位移信息,从不同方面学习轨迹特征空间中各子空间特征的权重信息,实现对行人之间相互影响的交互性轨迹特征刻画;然后采用对抗生成机制和多轨迹生成策略,实现对复杂场景下不同个体移动轨迹的生成与预测.最后,本文在两个公开的数据集(ETH和UCY)进行了实验验证.实验结果表明,在ADE、FDE和AnlDE三个指标上,本文提出的MAGAM模型比基准模型误差平均降低了26.90%、21.02%和24.06%.本文对模型的预测结果进行可视化分析,直观展示了本论文模型的合理性.更多还原

【Abstract】 Pedestrian trajectory prediction plays a vital role in intelligent city construction and public crisis management.Distinct from the single trajectory prediction which rely on strong temporal correlation,in the complex scenes,the trajectory reflects not only the temporal characteristics of a single person,but the interactive features between human and other moving objects nearby.Therefore,how to deeply describe such temporality and interactivity,and then to generate accurate trajectory prediction results according to the change of the scene has become a major problem in the field of trajectory prediction today.In recent years,deep learning has attracted great attention and achieved success in the trajectory prediction tasks.However,most of these methods capture the influence between pedestrians from a single view,and they fail to consider the multiple factors which have an effect on the decision of pedestrians,such as going straight or turning.To this end,in this paper,we propose a multi-head attention generative adversarial model(MAGAM) which combines the multi-head attention mechanism and the generative adversarial network to model the pedestrian trajectory in the complex scenes.Specifically,the MAGAM model employs multi-head attention mechanism with relative displacement information to learn the attentive weight of subspace features in the whole trajectory feature space on different aspects,to realize the characterization of the interactive trajectory features that resulting from mutual influence between pedestrians.Moreover,the adversarial generation strategy and multitrajectory generation strategy are used to achieve the reasonable generation of individual moving trajectory in the complex scenes.During the training process,the generator firstly extracts the personalized temporal features of pedestrians from historical observation sequences with long short-term memory(LSTM) based encoders.Secondly,the locations of pedestrians and temporal features are integrated into the multi-head attention model to learn the different weights and output the interactive state of the pedestrians.Thirdly,the interactive states and the Gaussian noise are fed into the LSTM-based decoders to generate multiple prediction trajectories.Then the discriminators are employed to judge whether the input trajectory belongs to the truth trajectory or generated trajectory as much as possible.By training with the adversarial mechanism,we could obtain the approximate truth results when modeling convergences.Finally,in order to estimate the performance of the proposed model,we conduct the experiments on two public datasets(ETH and UCY) which are widely used in the trajectory prediction tasks.We evaluate the prediction results based on three evaluation indicators:the average displacement error,the final displacement error,and the average no-linear displacement error.Compared with the existing trajectory prediction methods,the three metrics of the MAGAM model on all the datasets reduced by 26.90%,21.02 % and 24.06 % on average.And the prediction results and the interactive scenes among pedestrians are visualized and analyzed which demonstrates the rationality of the results.Additionally,the performance of the MAGAM model including the average convergence accuracy,the average convergence time and the average prediction time is verified through related experiments,compared with the baselines,the MAGAM model gets the longest convergence time and prediction time.更多还原