People-Flow Estimation from Footstep Sounds: A Feasibility Study via Simulation Using Large-Scale Pedestrian Trajectories

Yu Kitano, Nobutaka Ono

Abstract: In this paper, we investigate the feasibility of estimating people-flow from footstep sound signals by modeling footstep sound events from large-scale pedestrian trajectory data. Using about $\text{1 3}$ million trajectory records collected on an outdoor sidewalk, we simulate footstep sound signals around a single microphone as a controlled, noise-free observation and train models to estimate the average number of people within a fixed radius. A one-dimensional convolutional neural network (1D-CNN) achieves strong correlation with the ground truth within approximately three meters, while accuracy decreases at larger distances as acoustic cues attenuate. These results demonstrate the feasibility of privacy-preserving and cost-effective crowd sensing based on acoustics.

External IDs:dblp:conf/bigdataconf/KitanoO25