Regional crowd flow estimation from aerial view

In recent years, crowd analysis has been widely studied due to its realistic applications in many areas. In this paper, we pose a novel challenge for monitoring large-scale crowd scenes from an aerial view via estimating specific crowd flow for each partitioned region. However, existing methods are difficult to estimate the specific crowd flow for each region flexibly, simply, and accurately, especially lacking clear crowd appearance features from a top-down view. To accomplish this, we present a crowd flow estimation model whose goal is to estimate the flow into and out-of a certain region over any given time span. Specifically, we set up a two-stream network that jointly regresses crowd density and individual velocities, so as to directly approximate the instantaneous flow at each location. To enhance the flow estimation, we utilize the local relationships between crowd distribution and individual velocities via a proposed locality-confined attention module. Furthermore, we incorporate the additional spatio-temporal regularization for the top-down view by reversing future frames via the proposed inverse-temporal loss. In experiments, we apply drone-based overhead crowd videos to evaluate our approach in the task of crowd flow estimation, and we show that our approach surpasses the performance of prior methods and can also be applied in a variety of crowd analysis applications for understanding social scenes.