Real-time people counting for indoor scenes

Abstract

People counting in indoor environment is a challenging task due to the coexistence of moving crowds with stationary crowds, recurrent occlusions and complex background information. The performance of existing crowd counting methods drops significantly for indoor scene since the stationary people are missed due to moving foreground segmentation and the counting results are often disturbed by occlusions. To address the above problems, in this paper we propose a counting approach for indoor scenes, which can count not only moving crowds but also stationary crowds efficiently. Firstly, a foreground extraction assisted by detection is introduced for crowd segmentation and noise removal with a feedback update scheme. Then we build a multi-view head-shoulder model for people matching in the foreground and estimate the number of people with an improved K-mean clustering approach. Finally, to reduce the disturbance of occlusions, we present a temporal filter with frame-difference to further refine the counting results. To evaluate the performance of the proposed approach, a new indoor counting dataset including about 570,000 frames was collected from four different scenarios. Experiments and comparisons show the superiority of the proposed approach.