Head-shoulder Detection Using Joint HOG features for People Counting and Video Surveillance in Library

Pedestrian detection is an important problem in video surveillance. While pedestrians often have diverse postures and mutual occlusion which make the detection quite difficult, their head-shoulder portions are relatively stable. Thus we choose to use head-shoulder outline features of a pedestrian for detecting. First, we apply a hierarchical classification method using Haar features and HOG features to head-shoulder location detection. Second, we define a combined feature named Joint HOG based on the symmetry of head-shoulder portion. Third, we filter out most negative samples by using the Haar classifier. Finally, we execute an elaborate HOG verification and thus obtain the head-shoulder target box expected. Experimental results show that our method achieved a real-time processing accuracy rate of nearly 90%, arguing that it is applicable to people counting and video surveillance in library.