Data driven simulation for pedestrian avoiding a fixed obstacle

MARTIN, RAFAEL F.; PARISI DANIEL R.

Buenos Aires

Conferencia; StatPhys 27; 2019

The problem of navigation and collision avoidance is a central issue when developing realistic pedestrian simulation models.Traditionally, models are first proposed and, then calibrated with experimental data. But recently a new approach of data driven models have arose. The idea is, instead of proposing an explicit model, to use the available experimental data to decide the movement of simulated particles. Natural tools for doing this are artificial intelligence methods, in particular the use of neural networks.Up to now, multi-layer neural networks were presented [1]. Instead, we propose to work with a radial basis neural network (General Regression NN: GRNN) which only has one intrinsic parameter: the spread. [2]This approach can be fitted into the general framework proposed in [3] and thus the data is used to dynamically adjust the desired velocity of the simulated pedestrian moving by a first order model.In this work, we study the scenario where a pedestrian had to avoid a fixed obstacle.To adjust the spread of the GRNN, a leave-one-out cross-validation process was carried out with two different error functions. One based on the minimum distance to the objective and the other is the mean of the difference between positions of simulated and experimental data points along the whole trajectory.Finally, using the optimal parameters, we performed several simulations to observe the performance of the proposed method in the scenario studied. It showed up to be a robust approach to reproduce the one pedestrian-one obstacle avoidance problem. Also, this is a promising methodology to extend to the complete pedestrian navigation navigation problem.References[1] Wei, X., Lu, W., Zhu, L., & Xing, W. (2018). Learning motion rules from real data: Neural network for crowd simulation. Neurocomputing,310,125-134[2] Specht, D. F. (1991). A general regression neural network. IEEE transactions on neural networks, 2(6), 568-576.[3] Martin, Rafael F., Parisi, Daniel R. (2018). Pedestrian collision avoidance with a local dynamic goal. Proceedings of the PED2018 Conference. In Progress.