Crowd Density Forecasting by Modeling Patch-Based Dynamics

Forecasting human activities observed in videos is a long-standing challenge computer vision and robotics also beneficial for various real-world applications such as mobile robot navigation drone landing. In this work, we present new forecasting task called crowd density forecasting. Given video of captured by surveillance camera, our goal to predict how the will change unseen future frames. To address task, developed patch-based networks (PDFNs), which directly forecasts maps frames instead trajectories each moving person crowd. The PDFNs represent based on spatially or spatiotemporally overlapping patches learn simple dynamics fewer people patch. Doing so allows us efficiently deal with diverse complex when input involve variable number crowds independently. Experimental results several public datasets demonstrate effectiveness approaches compared state-of-the-art methods.