Deep Learning–Based Footfall Detection and Anthropogenic Activity Characterization Using Fiber-Optic Distributed Acoustic Sensing

Fiber-optic distributed acoustic sensing (DAS) has been found effective for vibration-based monitoring of human activities with high resolution in space and time. Here, we use deep learning–based computer vision algorithms for footfall signal extraction from the 2D array like DAS data. Unlike previous studies that primarily focused on detecting linear walker movement features along the fiber, we treat each individual footstep as an independent event target for picking. Using the effective algorithm, we detect a massive amount of footstep signals from a month-long DAS recording. We also correct for the polarity reversal characteristics of DAS and use backprojection techniques to determine each footstep location and reconstruct anonymous walking trajectories of walkers along the coastline, revealing unique walking behaviors under different weather conditions. This seismic experiment demonstrates the potential of combining DAS with advanced deep learning algorithms to achieve precise signal recognition and real-time monitoring, thereby opening new possibilities for characterizing anthropogenic activities.