Crowd Simulation Using Velocity Field Map and LSTM Neural Network

In this paper, we propose a novel method that uses a velocity field map and long short-term memory (LSTM) neural network to control agents to avoid collisions with nearby agents by considering their previous trajectories. Our method uses a series of continuous velocity field maps that represent the surrounding conditions of the controlling agent. A velocity field map comprises 2-channel images containing the positions and velocities of the controlling and nearby agents. We employed an LSTM neural network to predict the current velocity field map from the velocity field maps of the previous frames to determine the desired velocity of the controlling agent. The agent is controlled using a social force method based on the desired velocity. The novelty of our method lies in the use of the velocity field map and LSTM neural network. By using the velocity field map, the information of the controlling agent and its surrounding salutation is efficiently represented. In addition, by using a series of continuous velocity field maps and LSTM, the agent is efficiently controlled considering the previous frames. We present experimental results and discuss the advantages of the proposed approach.