This paper presents a novel stereo system, based on a graphics processing unit (GPU), for pedestrian detection in real images. The process of obtaining a dense disparity map and the edge properties of the scene to extract a region of interest (ROI) is designed on a GPU for real-time applications. After extracting the histograms of the oriented gradients on the ROIs, a support vector machine classifies them as pedestrian and nonpedestrian types. The system employs the recognition-by-components method, which compensates for the pose and articulation changes of pedestrians. In order to effectively track spatial pedestrian estimates over sequences, subwindows at distinctive parts of human beings are used as measurements for the Kalman filter. (C) 2010 Society of Photo-Optical Instrumentation Engineers. [DOI: 10.1117/1.3521254]