Pedestrian detection from traffic scenes based on probabilistic models of the contour fragments

Abstract

Driving assistance systems usually have a pedestrian detection module for alerting the driver in case of a dangerous situation. In this paper we describe such a module that is used for obstacles classification in pedestrians and non-pedestrians. The obstacles are defined by their region of interest (ROI) in the grayscale scene image. Random size and location of pedestrians' contour-edge fragments are extracted and filtered. They are used for building a very large codebook of pedestrians' contour fragments. A novel multi-level clustering is introduced in order to sequentially group these fragments first on location, then on size and finally on the content. A new method is proposed for computing a set of probabilistic contour fragments models inside each individual cluster. It is used for characterizing the entire codebook in just few models, one for each cluster. These models are used in a fast matching process against the obstacles ROIs that should be classified. A SVM classifier is trained on the matching scores vector and applied for detecting the pedestrians.