Evaluation of an Environment Classification Method for Optimal Crowd Model Selection in Autonomous Mobile Robot Simulations

Incorporating crowd models into robot simulators is a common practice in the field of autonomous mobile robot navigation research. Although crowd models developed in the field of crowd simulation are frequently used, there is no universal model that can be used for all scenarios. Therefore, it is essential to select an appropriate crowd model for the specific simulation environment. From this perspective, until now, we have been studying a method for selecting an appropriate crowd model for each category. This method observes the movement trajectories of pedestrians in the environment to be simulated, and classifies the environments into multiple categories based on the observation results. In this process, feature images are generated by superimposing the observation results on a time axis. The latent variables in the feature images are then extracted using an autoencoder, and the extraction results are clustered. However, the impact of overlapping time intervals (temporal granularity) and different extraction methods of latent variables, which are important in generating feature images, has not yet been clarified. In this paper, we assess their influence on category classification accuracy. Based on the obtained classification results, we also develop scenarios for selecting a crowd model for each category.