MPDFF: Multi-source Pedestrian detection based on Feature Fusion

Abstract

Pedestrian detection from UAV images is vital for many fields. Given that visible images are susceptible to bad illumination, thermal images with the ability to characterize the temperature of an object can provide auxiliary information. It is useful to fuse the visible and thermal images to improve the pedestrian detection performance. Unfortunately, studies on pedestrian detection with UAV visible-thermal image pairs are still rare. Therefore, we propose a method for Multi-source Pedestrian Detection based on Feature Fusion (MPDFF). With the registered visible and thermal image pairs as the input, MPDFF can achieve better characterization of pedestrians by concatenating the features from the two images. MPDFF performs much better than the methods that use only single-source images, which demonstrates that visible and thermal images are complementary in pedestrian detection.