Dempster-Shafer Fusion of Context Sources for Pedestrian Recognition

This contribution presents the design of an image-based contextual pedestrian classifier for an automotive application. Our previous work shows that local classifiers working with image cutouts are in many cases not sufficient to achieve satisfactory results in complex scenarios. As a solution the work proposed incorporating contextual knowledge into the classification task, significantly improving the classification results. Contextual knowledge is described by a set of different and independent context sources. This paper discusses the fusion of these sources on the basis of the Dempster-Shafer theory. It presents and compares different possibilities to model the frame of discernment and the mass function to achieve optimal results. Furthermore, it provides an elegant way to take uncertainties of the context sources into account. The methods are evaluated on simulated and on real data.