Infrared pedestrian detection with converted temperature map

Infrared pedestrian detection aims to detect persons in outdoor thermal images. It shows a unique advantage in dark environment or bad weather compared to daytime visible images (the RGB image). Most current methods treat infrared detection the same way as with visible images, e.g. regarding the infrared image as a special gray-scale visible image. In this paper, we tackle this problem with more emphasis on the underlying temperature information in infrared images. We build an image-temperature transformation formula based upon infrared image formation theory, which can convert infrared image into temperature map with the prior of pedestrian pixel-temperature value. The whole detection process follows a two-stage manner. In the first stage, we use a common detector which treats the infrared image as the gray-scale visible image to provide primary detection results and a pedestrian position prior (the highest-confidence pedestrian detection box in each image). In the second stage, we convert infrared images into corresponding temperature maps and train a temperature net for detection. The final results consist of both the primary detection and the temperature net outputs, detecting pedestrians with characteristics in both image and temperature domain. We show that the converted temperature image is less affected by environmental factors, and that its detector shows amazing complementary ability with the primary detector. We carry out extensive experiments and analysis on two public infrared datasets, the OTCBVS dataset and the FLIR dataset; and demonstrate the effectiveness of incorporating temperature maps.