Toward Realistic Human Crowd Simulations with Data-Driven Parameter Space Exploration

Understanding human crowd motion is crucial for realistic crowd simulations and content creation. Over the multiple decades, several computational algorithms have been developed to devise a more realistic simulation that will vastly improve the user experiences. However, the gap between real and simulated human motion and behavior is still large. One of the promising thrusts of the crowd simulation community to reduce the gap is to utilize a data-driven trajectory prediction model, using the access to a large amount of data. However, building a crowd simulation model based on the learned microscopic trajectory model is still a challenging task. In addition, unlike individual or group-level human trajectory data, large-scale real human crowd motion data is not readily available. To overcome these, we investigate the utility of synthetic simulation data. We propose a novel human crowd motion estimation framework that can predict simulator parameters from trajectory data. Our initial findings show promising results that our method can robustly estimate simulation parameters.