An Evaluation of Artificial Neural Networks as an alternative to Example-Based Crowd Simulation

Abstract: In the ever-developing field of crowd simulation several different algorithms have been proposed to enable data-driven simulations: methods using real-life data to generate realistic pedestrian movements. With different use-cases and trade-offs it is vital to understand the particulars of each approach. This paper investigates two alternative methods for crowd-simulation: The ‘Example-Based’ (EB) method introduced by Lerner, Chrysanthou, and Lischinski [1] and an Artificial Neural Network (ANN) approach based on the work of Ma, Lee, and Yuen [2]. These methods were applied to simulate agents in two scenarios from the UCY Dataset [1] then compared against the real-life scenario and each other in terms microscopic- and macroscopic navigation along with collision avoidance qualities. Both methods showed realistic seeming microscopic qualities with pedestrian speeds and acceleration similar to the real-life scenario. The ANN method showed much better capability to avoid collisions, although the EB method performed better compared to the ANN when not aided by ‘position-modification schemes’ for collision avoidance. Macroscopic qualities varied such that the ANN method more closely aligned the real pedestrian behaviours in the sparser Zara Scenario, while neither method performed well in the dense scenario. In all, the ANN method shows the most promise, especially in terms of performance, but further investigations into more complex neural networks and greater, more representative, amounts of data should be made to create more robust methods.