基于GAN的社会和场景感知行人轨迹预测  

SOCIAL AND SCENE AWARENESS PEDESTRIAN TRAJECTORY PREDICTION BASED ON GAN

作　　者：李兰 张洁 刘杰 胡克勇 Li Lan;Zhang Jie;Liu Jie;Hu Keyong(School of Information and Control Engineering,Qingdao University of Technology,Qingdao 266520,Shandong,China)

机构地区：[1]青岛理工大学信息与控制工程学院,山东青岛266520

出　　处：《计算机应用与软件》2024年第6期72-78,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61902205)。

摘　　要：针对状态精细化长短期记忆网络(SR-LSTM)未考虑周围物理场景对行人轨迹预测的影响,且无法生成多种可能性样本的问题,提出一种基于生成对抗网络(Generative Adversarial Networks,GAN)的社会和场景感知行人轨迹预测模型。此模型引入社会注意力及语义池机制,社会注意力建模邻人当前重要意图,以从相邻行人中选择重要的信息,语义池定义物理场景语义并学习与行人轨迹相关性。由于GAN易发生模式崩溃和下降,采用Info-GAN进行训练生成更真实的样本。在ETH和UYC两个数据集上进行实验,结果表明该方法较于SR-LSTM,ADE降低8.9百分点,FDE降低12.8百分点,且可生成更多合理的样本。In order to solve the problem that state refinement for LSTM(SR-LSTM)does not consider the influence of surrounding physical scenes on pedestrian trajectory prediction,and can not generate a variety of possible samples,a social and scene awareness pedestrian trajectory prediction model based on GAN is proposed.This model introduced social attention and semantic pool mechanism,and social attention mechanism was used to model the current important intention of adjacent pedestrians in order to select important information from adjacent pedestrians.Semantic pools defined the semantics of physical scenes and learn their correlation with pedestrian tracks.Because GAN was prone to mode collapse and decline,Info-GAN was used for antagonistic training to generate more real samples.The experiments on ETH and UYC data sets show that compared with SR-LSTM,ADE of this method is 8.9 percentage points lower and FDE is 12.8 percentage points lower,and more reasonable samples can be generated.

关 键 词：行人轨迹预测 生成对抗网络 注意力机制 语义池机制 长短期记忆网络 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]