Pedestrian detection at night time in FIR domain: Comprehensive study about temperature and brightness and new benchmark

This paper presents a novel method to detect pedestrians in the far infrared (FIR) domain at night. In existing infrared data, the brightness is distorted by the contrast of the scene, causing the performance to degrade. The radiometric temperature was introduced to cope with this issue. The temperature was unaffected by the contrast and was more stable than the brightness because of the limitation to the specific thermal range with thermoregulation. The dataset was constructed across each season and four versions of thermal infrared radiometry aggregated channel feature (TIR-ACF) are presented with normalization using the maximum temperature of humans. In these experiments, the proposed method outperformed the brightness-based baseline method by a maximum 11% on the log-averaged miss rate for seasonal variations. As a result, the physical temperature enhanced the performance and helped detect pedestrians that cannot be found using the brightness with a reduction of false positives.