A Radar-Nearest-Neighbor based data-driven approach for crowd simulation

In this work, a learnable data-driven motion model namely Multi-Feature Fusion Recursive Neural Network (MFF-RNN) is proposed. The model yields pedestrians' velocities by learning from the designed motion states consisting of the relative distances and velocities with neighbors, as well as individuals' previous velocity sequences. A novel Radar-Nearest-Neighbor (Radar-NN) method is developed to determine the nearest neighbors of a pedestrian by treating him/her as a radar and detecting the surrounding environment within a limited circular receptive field. Bidirectional flow scenarios are adopted to evaluate the performance of the proposed model and the lane formation phenomenon can be successfully reproduced. The simulation results coincide with that of experiments and are superior to the work of Ma et al. in pedestrian trajectories, distributions, as well as fundamental diagrams. By calculating five evaluation metrics, it shows that the errors of our model are reduced by 34.1-79.0% compared with their work.