Crowd behavior analysis is one of the significant issues in video surveillance. In this paper while having a video sequence of a public place, the goal is to extract the regions of interest in the input video that can be the regions where more people commute. To identify these regions, first a non-parametric kernel density estimation method is applied to create a crowd density map. Then, we propose to use a kernel density estimation based on the Bayesian risk. By defining a new loss function for the Bayesian risk, we can decrease the effects of outliers in order to extract the distribution’s peaks of the resulted map more accurately. To evaluate the proposed method, we utilize 9 different real-world datasets on various scenarios where among them PETS2009 and Mall dataset are popularly used. We also make use of multiple synthetic datasets as our evaluation benchmark. Experimental results present that the proposed method has higher accuracy to find different centers of attention in compared with other similar state-of-the-art methods on both real-world and synthetic datasets.

Keywords