For the real-time and robustness of far-infrared pedestrian detection remains a very challenging problem, the paper presents a vertical gradient projection method based on the pixels from ROIs (Regions of Interest) extracted angle. This method utilizes the image gradient information to conduct early positioning for the vertical banded image area which may contain pedestrians, then forms relatively reliable ROIs by partial dual-threshold image segmentation algorithm inside obtained image. In the pedestrian recognition process, it combines global information and gradient histogram method based on pyramid weighted entropy to describe the far-infrared pedestrian, at the same time assist in SVM algorithm to complete the identification process. Experiments show that the method can improve the search efficiency of ROIs extraction stage, and suppress portion of the candidate region included in the background target.
