Bilkent University
Department of Computer Engineering




Yalım Doğan
PhD Student
(Supervisor: Prof. Dr. Ugur Gudukbay)
Computer Engineering Department
Bilkent University

Crowd simulations imitate the group dynamics of individuals in different environments. The realism of such simulations is assessed in comparison to real-life scenarios. Many applications in entertainment, military, security, and education require augmenting simulated crowds into videos of real people. In such cases, virtual agents should realistically interact with the environment and the people in the video. An important component of this augmentation task is determining the navigable areas in the video. We propose a method that automatically locates and reconstructs the navigable areas on the ground plane of the video, using semantic segmentation and pedestrian detection. We then place the 2D mesh of the reconstructed navigable regions into a 3D crowd simulation environment and project the input video over it. We report the performance of our approach using real-life surveillance videos, based on the accuracy of the determined navigable regions and camera configuration. The results indicate that our method generates accurate navigable areas for realistic augmented crowd simulations. We also investigate its limitations in the case of realism and interactability. Various single and multi-view reconstruction for static backgrounds and deep learning-based object/person reconstruction techniques are to be discussed.


DATE: 14 October 2020, Wednesday @ 21:00