An Improved Method of Crowd Counting Based on Regression

Recently intelligent crowd counting has attracted researchers ’ attention in computer vision and related fields [1-9]. The existing predominant techniques for crowd counting fall into two categories: 1) object detection and tracking based crowd counting; 2) crowd density estimation based on features and regression analysis. In the first category work always involves pedestrian detection and tracking. Since the counting results depend heavily upon detection response, applying a state-of-the-art detector has becoming an essential key to improve system performance. In [1-4] generic detectors including HOG based head-shoulder and JRoG based part detectors have been validated fairly effective in detecting and counting people. This kind of methods shows excellent performance in scenes with sparse crowd due to its accurate localization. However, as the crowd becomes large and heavy occlusion arises, individual detection and tracking both become almost impossible. Methods in the second category, such as [5-7], estimate the crowd density by extracting a set of holistic or local features in regions of interest (ROI) and then modeling the number of people based on features. This kind of methods does not consider individual localization and has been proved effective in dense crowd counting, especially in public environments where heavy occlusions happen frequently. In [8], the feature-regression based system i