PPDet: A novel infrared pedestrian detection network in a per-pixel prediction fashion

Abstract

Infrared images are suitable for pedestrian detection under poor lighting conditions. However, the lack of feature information about pedestrians in infrared images and the diversities of pedestrians make infrared pedestrian detection difficult. To address this issue, this paper proposes a novel infrared pedestrian detection network in a per-pixel prediction fashion. The proposed network consists of three parts, including backbone, neck and head: 1) The backbone expands the receptive field of the network and solves the network degradation problem caused by network deepening; 2) The neck improves feature representation by fusing multi-scale feature information; 3) By designing multiple cascaded heads, pixel-by-pixel prediction of multi-scale pedestrians is achieved. The proposed method is compared with a variety of object detection algorithms on the SCUT dataset, and it achieves better performance in infrared pedestrian detection.