A novel hierarchical framework for human head-shoulder detection

This paper proposes a novel framework for human head-shoulder detection. It can be used to detect persons in sports videos or family albums, where poses of persons are various and bodies are often partially occluded. In our method, human head-shoulder is decomposed recursively along the omega-shape contour. A new feature named Oriented Integration of Gradients (OIG) is introduced to describe subparts of human head-shoulder. Part detectors are trained based on OIG, and an improved Hough voting system is built to combine results from these detectors. The main contribution of our work has three aspects. Firstly, we propose a novel contour based decomposing method which can be used to decompose objects that have salient contours. Secondly, a simple but effective feature named OIG is introduced. The descriptive ability of OIG is comparable with that of HOG, and it is much easier to calculate. Thirdly, we build an improved Hough voting system to solve problems caused by deformable models. We test our method on images selected from PASCAL.