Abstract:
Pedestrian detection is a key problem in various computer vision applications, such as driver-assistance or surveillance systems. A Hough forest framework is based on the idea of a discriminative codebook learning of the appearance of patches from an object. It uses a Hough voting process with regard to the centroid of the object class. This thesis compares the Hough forest detector with a more general random forest detector which uses other classes alone with “pedestrian” and “background” classes, but does not use a centroid voting process. Regarding the general random forest detector, the thesis proposes a new framework which takes advantage of stereo vision to estimate the height of pedestrians by using disparity maps. A random forest classifier considers class probabilities of a candidate from a region of interest with respect to nine classes; eight classes define pedestrians, and one class defines the background. Both approaches (Hough forest or random forest) are tested on a few datasets for pedestrian detection. The detailed results are used for a comparative evaluation of both approaches. Keywords: Hough Forests, Random Forests, Pedestrian Detection