A hierarchical method for pedestrian detection with random forests

Abstract

Due to many uncontrolled factors, pedestrian detection is one of the most challenging problems in computer vision. In this paper, a fast and accurate hierarchical method for pedestrian detection with random forests is proposed, which can combine holistic information and local information based on image pyramid model. Image pyramid can effectively realize multi-layer information fusion and hierarchical detection. At the first low spatial resolution layer, a holistic random forests classifier is trained with dominant orientation templates (DOT), which is used for detecting candidate pedestrian area. At the second high spatial resolution layer, local image patches and their offset vectors relative to object center are extracted which are used for learning visual words and geometric constraints by parts-based hough forests. Accurate pedestrian detection is implemented in corresponding candidate area at the second layer by means of hough voting. We test the proposed method with two challenging pedestrian databases: INRIA and TUD-pedestrian. According to the theory analysis and experimental results, our method obtains lower computation complexity and higher precisions than previous works.