A Study on Estimation of Pedestrian Dynamics in Video and its Application to Crowd Simulation

Abstract

We propose a method for generating crowd simulations from video. The parameters necessary for crowd simulation, i.e., number of people, actions, and speed, were extracted using image recognition technology. The number of people and their actions were estimated using a dedicated algorithm with high accuracy even in low-light conditions or in situations where people overlap. Speed is estimated by using the moving distance between frames. By providing the above parameters to the crowd simulation, we were able to generate a simulation that is close to the actual video.