Crowd density estimation using sparse texture features

This paper presents a technique for crowd density estimation in surveillance images, which needs neither individual detection and tracking nor a complex training. This is done by building a set of feature templates for different crowd density scenes, and calculating the similarity between templates and features that are extracted from surveillance video frames. These templates can be selected by staff according to the situation of surveillance scenes. Thus our approach can be deployed with minimal setup for a new site. In order to get sparse features, a generative model of sparse texture representation is improved for crowd scene description: firstly, multi-scale local image patch is generated to deal with perspective projection; secondly, a novel statistic descriptor, Gray-Gradient Dependence Matrix, is introduced to extract features; thirdly, an adaptive clustering is utilized to identify clusters. By computing the weighted average of these clusters, a more compact representation of the image can be obtained. Three aspects of experimental results show that the proposed approach is efficient and accurate in crowd density estimation.