An HOG-CT human detector with histogram-based search

This paper addresses the problem of human detection in still images. We first describe a novel descriptor concatenating the local normalized histogram of oriented gradients (HOG) and the global normalized histogram of census transform (CT) of images for human detection. The detector is trained by using cascade learning method based on AdaBoost. In addition, we propose an easy histogram-based search method, termed the block histogram, which can reduce the computational cost and speed up the process of detection when sliding in the test image. Experimental results on the INRIA person dataset show that the proposed method can achieve competitive results both in discriminating power and detection speed as compared to the state-of-the-art.