Determining Posterior Probabilities on the Basis of Cascaded Classifiers as used in Pedestrian Detection Systems

Abstract

Cascaded classifiers are widely spread in automotive pedestrian detection systems. Since there has been no research on probabilistic information derivable on the basis of a cascade, these systems are limited in the sense that they only exploit the binary classification results. In contrast to that, this paper presents a mathematically founded model regarding the computation of posterior probabilities on the basis of such classifiers. This is highly relevant in respect of the further development of robust and reliable detection systems.