Robust Identification of Dense or Sparse Crowd Based on Classifier Fusion

摘要

For a video surveillance system, crowd behavior analysis and crowd managing are important tasks. Along with the event in which crowd participates, its volume and density are also important in managing the crowd. Hence, characterizing the crowd as dense or sparse is an essential component of a crowd handling system. In this context, most of the existing methods try to estimate the headcount. Unlike those, the proposed method exploits the domain-knowledge based low-level features to classify the crowd image as dense or sparse. We present three simple systems working with three different feature sets. These are all free from the burden of background estimation. Experiments are carried on a dataset formed by taking the images from UOF-CC50 and SanghaiTech. Performance of all three feature sets are satisfactory, and Corner-Point based methodology provides the best result.