Environment Classification Method Using Autoencoder to Select Appropriate Crowd Model for Robot Simulation

Abstract

Various crowd models developed in the field of crowd simulation have been used to learn action policies of reinforcement learning for the navigation of autonomous mobile robots. However, there is no single crowd model that can be applied to all environments because crowd movement is highly dependent on the environment. In this paper, we propose a method for selecting an appropriate crowd model for each environment by classifying the spatio-temporal data to be simulated into multiple categories. For that, we extract features related to crowd movement and geographic shape for use in the robot simulation. Specifically, we generate image data to visualize pedestrian movement trends, perform feature extraction using an autoencoder, and classify the results into several categories. Through evaluation experiments using pedestrian movement trajectory datasets, this paper shows that visually similar feature images are classified into the same categories.