DeepThermal Outdoor: A first-person thermal imaging dataset

Abstract

Recently, thermal imaging modules equipped for infantry soldiers have been a trend to improve the combat ability of soldiers. Soldiers have to perform many different tasks at the same time, so it is necessary to equip them with the tools of automatic target detection, especially human objects detection, in practice. Hence, there is a need to intelligently optimize the effectiveness of thermal imaging equipment. New artificial intelligence and deep learning(DL) approaches are applicable methods that show superior accuracy compared to previous methods. However, state-of-the-art DL methods depend on the generality and diversity of the training data set. To address this issue, our paper presents the DeepThermal Outdoor thermal imaging data set, which is collected from equipment mounted on the body of infantry at various terrain locations. The labeled dataset focuses on human objects with different locomotion postures, and it contains 10,190 images and 22,464 labeled human-objects. Finally, the experiment is conducted with several DL methods using the proposed dataset, and the results show its contribution to the improvement of the performance of DL methods to detect humans on thermal images as well as to evaluate the practical applicability of a DL.