Low level crowd analysis using frame-wise normalized feature for people counting

Abstract

People counting is a crucial component in visual surveillance mainly for crowd monitoring and management. Recently, significant progress has been made in this field by using features regression. In this context, perspective distortions have been frequently studied, however, crowded scenes remain particularly challenging and could deeply affect the count because of the partial occlusions that occur between individuals. To address these challenges, we propose a people counting approach that harness the advantage of incorporating an uniform motion model into Gaussian Mixture Model (GMM) background subtraction to obtain high accurate foreground segmentation. The counting is based on foreground measurements, where a perspective normalization and a crowd measure-informed corner density are introduced with foreground pixel counts into a single feature. Afterwards, the correspondence between this frame-wise feature and the number of persons is learned by Gaussian Process regression. Experimental results demonstrate the benefits of integrating GMM with motion cue, and normalizing the proposed feature as well. Also, by means of comparisons to other feature-based methods, our approach has been experimentally validated showing more accurate results.