Human detection with a multi-sensors stereovision system

Abstract

In this paper, we propose a human detection process using Far-Infrared (FIR) and daylight cameras mounted on a stereovision setup. Although daylight or FIR cameras have long been used to detect pedestrians, they nonetheless suffer from known limitations. In this paper, we present how both can collaborate inside a stereovision setup to reduce the false positive rate inherent to their individual use. Our detection method is based on two distinctive steps. First, human positions are detected in both FIR and daylight images using a cascade of boosted classifiers. Then, both results are fused based on the geometric information of the sterovision system. In this paper, we present how human positions are localized in images, and how the decisions taken by each camera are fused together. In order to gauge performances, a quantitative evaluation based on an annotated dataset is presented. © Springer-Verlag Berlin Heidelberg 2010.