Agent navigation using deep learning with agent space heat map for crowd simulation

We propose a novel method of crowd simulation that employs a deep learning technique for controlling individual agents. We use a convolutional neural network to learn agent space heat maps that are generated from example crowd animations. A heat map contains the positions and speeds of nearby agents and the temporary target position that indicates an appropriate heading direction for the agent to reach the final target position efficiently. When controlling an agent, an agent space heat map that contains possible temporary target positions is estimated from the trained model and the agent space heat map that contains only the positions and speeds of nearby agents. In addition, evaluation functions are used to choose a temporary target position from the estimated heat map. Individual agents are controlled using a force-based model so that they move toward the estimated temporary target position. Our approach realizes human-like navigation by combining the intuitive and logical aspects of decision-making.