Multi-feature fusion based region of interest generation method for far-infrared pedestrian detection system

Given the uncontrolled outdoor environments and different physical properties of clothing, the appearance of pedestrians in far-infrared (FIR) images changes dramatically. Finding a robust region of interest (ROI) generation method for pedestrian detection remains challenging. Previous researches can obtain reliable results in some conditions. But they always got inappropriate results in warmer conditions. This study presents a Multi-feature fusion based ROI generation method for FIR pedestrian detection system to solve this problem. We extract two kinds of salient feature regions, namely, highlighting feature areas and vertical feature areas. A reasonable threshold is set to derive the highlighting feature areas and Scharr operator is used to find vertical edges. These areas are not necessarily connected to each other in the image. We think an upright pedestrian is a highly structured target consisting of highlighting feature areas and vertical feature areas. The distribution of feature areas is demonstrated using a skeleton model. So we apply the dilating morphological operation to ensure that these adjacent feature areas within pedestrians' regions will connect together. The size of the structuring element is set adaptively according to the size of feature areas. Finally, the experimental results show the robust performance of our method in different ambient conditions.