基于行人轨迹的大规模人群模拟快速自主导航  

Fast Agent Navigation for Large Scale Virtual Crowd Based on Pedestrian Trajectory

作　　者：刘景昊 王雷[1] 盛捷 LIU Jing-hao;WANG Lei;SHENG Jie(Department of Automation,University of Science and Technology of China,Hefei Anhui 230026,China)

机构地区：[1]中国科学技术大学信息学院,安徽合肥230026

出　　处：《计算机仿真》2020年第10期195-199,共5页Computer Simulation

基　　金：中科院创新基金(CXJJ-17-M139);中科院重大专项课题(KGFZD-135-18-027)。

摘　　要：多出入口环境下虚拟行人的逼真动态行走是大规模虚拟人群仿真研究的难点问题,现有研究方法的训练成本与仿真性能无法满足高逼真度大规模人群仿真的需要。提出一种从视频中快速学习行人轨迹的智能体快速自主导航仿真框架,将行人运动规则提取转化为符号回归问题的求解。得到的行人运动规则为一个非线性函数,虚拟行人可以据此实现快速导航。实验结果证明,上述方法能够高效生成逼真的行人运动,在Unity3D仿真环境中性能表现优异。The realistic dynamic walking of virtual pedestrians in the multi-entry environment is a difficult problem in large-scale virtual crowd simulation research. The training cost and simulation performance of the existing research methods cannot meet the needs of high-fidelity and large-scale crowd simulation. This paper proposes a fast autonomous navigation simulation framework for fast learning pedestrian trajectories from videos. The pedestrian motion rule distillation is transformed into a symbolic regression problem. The rule is a nonlinear function, which can be used for virtual pedestrian navigation.The experimental results show that the method can efficiently generate simulation scenarios for large-scale crowd dynamic walking.

关 键 词：虚拟人群 多智能体 行人轨迹 符号回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]