Human detection based on CENTRIST and scale of edge selection

In this paper we propose human detection algorithm based on CENTRIST visual descriptor and scale of edge selection. Contour is the most feasible method to detect human object, and the sign of neighbor pixel is the key of contours. CENTRIST feature is particularly suitable for human detection since it obtains the sign information and contours could be represented precisely by feature vectors. However, CENTRIST descriptor is based on Sobel edge detection and is always challenged by object with local texture inside and complex background, such as a man in a tweed suit or beside tree and leaves. To address this issue, we introduce an improved RS-edge based CENTRIST, which keeps the contours most salient to people visual system and removes local texture. A linear SVM classifier is used to detect human object based on descriptors above. Experiments on INRIA datasets and real video surveillance are conducted to validate the performance of detector. Moreover, the human detecting algorithm has been ported onto float-point DSP of embedded DM8168 multimedia processing platform in real classroom recording system to detecting teacher.