Meta-classifier for pedestrian attitude recognition

This paper presents an innovative pedestrian detection algorithm with applications to driving assistance systems. The novelty of the approach resides in the construction of an expert module for pedestrian recognition based on a meta-classification scheme applied on different pedestrian attitudes. The designed module is part of a real-time stereo based driving assistance system. The proposed meta-classification scheme learns the discriminant features of a partitioned pedestrian space. The complex pedestrian object is decomposed into different attitudes like pedestrian standing, running and for each attitude a classifier is trained. Our experiments show that the proposed meta-classification scheme outperforms a single classifier trained on the whole un-partitioned object space. For classification a probabilistic approach based on Bayesian networks was used. Two types of features extracted from the image have beed involved in the training process: magnitude of first order partial derivatives computed in four directions and histograms of gradient orientations (HOG).