Misalignment-Robust Pedestrian Detection Framework for Visible and Far-Infrared Image Pairs

Inspired by recent progresses of hardware technologies of consumer electronics, joint use of visible and far infrared (FIR) images becomes more common for various applications such as pedestrian detection. In actual situation, there is misalignment between the visible and FIR images, because the axis of these two cameras are different. Although various existing pedestrian detection algorithms have been proposed, the existing methods are sensitive to the misalignment because these methods assume that the input visible and FIR image pair is strictly aligned. This paper presents a novel misalignment robust pedestrian detection framework for the visible and FIR image pairs. The keys of the proposed framework are the following two points: 1) a set of perturbed FIR images is used to contain the aligned visible and FIR image pair, and 2) our proposed framework employs a merged criterion that combines a learning-based detection approach and a image-processing based similarity measure. Experimental results on the real image dataset show that the proposed framework outperforms existing methods in terms of log-average miss rate.