Landform Features for Infrared Pedestrian Detection

Abstract

Pedestrian detection has long been a research hot spot for its wide range of potential applications, e.g. public transportation, video surveillance. In this paper, we tackle the problem of infrared pedestrian detection as infrared images do not rely on sufficient illumination and have better environmental adaptability. To be specific, we propose a novel landform-based infrared feature extraction scheme, which richly exploits the heat source and temperature distribution information and therefore provides stronger infrared features. Following this feature extraction scheme, we build a LandForm Feature Network (LFF-Net) for infrared pedestrian detection. We show that the LFF-Net achieves a state-of-the-art performance of 70.75% on the widely-used FLIR dataset (improving 3.46% over its baseline), demonstrating the effectiveness of the proposed landform features for infrared images.