Pedestrian Detection Based on Motion Compensation and HOG/SVM Classifier

Pedestrian detection algorithm for videos taken by cameras amounted on a moving platform is studied. Firstly, Harris operator is used to extract corner points in each frame. Then Normalized Cross-Correlation (NCC) method is used to match those corner points between two adjacent frames. Based on a set of selected matching points, the global motion parameters are calculated using the projection model of a camera. The video frames are then corrected with the obtained motion parameters. Moving targets are detected and segmented by calculating the difference between two adjacent frames in the corrected video. The histogram of gradient (HOG) feature is computed for each detected target and is used as the input to a SVM-HOG classifier, which is obtained via off-line training and machine learning, to check the validness of existence of a pedestrian. Experimental results on PC show the effectiveness of this method.