Learning mid-term human navigation through crowds

Abstract

This paper introduces a method for modeling and forecasting sequences of human decisions in crowd navigation tasks. We focus on mid-term navigation behavior and represent it as a sequence of traversed gaps between individuals. Using data collected from a virtual reality experiment on crowd navigation, the navigable space is discretized through Delaunay triangulation, where the crowd members serve as the vertices. A binary classifier is trained to predict the sequence of triangle edges traversed by each participant, and several geometric feature designs are evaluated to support this prediction. The objective is to determine a minimal yet effective feature configuration that accurately captures human trajectories while maintaining an appropriate balance between predictive performance and model complexity.