Pedestrian detection in infrared images using cross-modal region-of-interest memory

Abstract

Multispectral pedestrian detection using both visible and thermal images shows good detection performance. However, there are several problems in applying it to real-world applications. In this paper, we propose a novel thermal image based pedestrian detection method that utilizes visible images appearance information during the inference time. In the proposed method, a cross modal Region of Interest memory(RoI) is proposed to store visible and thermal image region of interest information during the training phase. Then, at the inference time, the visible image information is read from the cross modal RoI memory using the thermal image. Therefore, even if there is no input visible image, high-performance thermal image based pedestrian detection is performed by utilizing the complementary information of thermal-visible image. The proposed method achieves state-of-the-art pedestrian detection performance in pedestrian detection using only thermal images during the inference time. In addition, the proposed method shows good performance comparable to multispectral detection methods.