A multi-resolution approach for infrared vision-based pedestrian detection

Abstract

This paper presents the improvements of a system for pedestrian detection in infrared images. The system is based on a multi-resolution localization of warm symmetrical objects with specific size and aspect ratio; the multi-resolution approach allows to detect both close and far pedestrians. A match against a set of 3D models encoding human shape's morphological and thermal characteristics is used as a validation process to remove false positives. No temporal correlation, nor motion cues are used for the processing that is based on the analysis of single frames only.