PART-BODY DETECTION FRAMEWORK FOR PEOPLE DETECTION USING SLICED HOG DESCRIPTORS

We investigate the possibility for using portions of Histograms of Oriented Gradients (HOG) descriptors in a part- based people detection framework. Instead of extracting descriptors from isolated or pre-cropped human parts, we slice the extracted HOG descriptor from whole windows into four, one slice per one human part. Support Vector Machines (SVMs) are used for classifying the slices and the outcome detections are handled by a finite-state machine where three detected parts means that one assumed person is in the window being scanned. Experiments were conducted for our detection framework and another conventional one that uses whole HOG descriptors using images from the INRIA Person Dataset, in which our framework achieved better; detecting 46/50 of occluded people comparing to 36/50 for the conventional framework. Moreover, we achieved less false positive detections of 80 windows comparing to 289 for the conventional framework.