Capturing Human Movements for Simulation Environment

Chengxin Wang, Muhammad Shalihin Bin Othman, Gary Tan

Abstract: In this paper, we proposed a novel data-driven framework to translate human movements from real-life video feeds into a virtual simulator in Unity 3D. In the proposed framework, YOLOv3 is used for pedestrian detection. Thereafter, a modified offline tracking algorithm with the min-cost flow was built to associate detected pedestrians from frame to frame. Finally, 2D trajectories are produced where a script would translate them into the Unity 3D platform. The proposed framework has the ability to display realistic behavior patterns where we would be able to introduce threats and analyze different strategies for improving evacuation and rescue in disaster situations.