Accurate object detection using local shape descriptors

This paper proposes a novel object detection approach based on local shape information. Boundary edge fragments preserve some features like shape and position which properly describe the outline of an object. Extraction of object boundary fragments is a challenging task in object detection. In this paper, a sophisticated system is proposed to achieve this goal. We propose local shape descriptors and present a boundary fragment extraction method using Poisson equation properties, and then, we compute relation between boundary fragments using GMM to obtain exact boundaries and detect the object. To get more accurate detection of the object, we employ a False Positive elimination stage based on local orientation histogram matching. The proposed object detection system is applied on several datasets containing object classes in cluttered images in various forms of scale and translation. We compare our approach with other similar methods that use shape information for object detection. Experimental results show the power of our proposed method in detection and its robustness in face with scale and translation variations.