Human detection in aerial thermal imaging using a fully convolutional regression network

Abstract

Thermal imaging can play a critical role in surveillance by promising higher robustness to the bad weather and night vision. Human detection and localization are important surveillance tasks for security purposes and maintaining law and order. This paper proposes a novel regression-based method for human detection from thermal infrared images. A fully convolutional regression network is designed to map the human heat signature in the input thermal image to the spatial density maps. The regressed density map is then post-processed for human detection and localization in the image. 25% data holdout validation scheme is used to train and test the proposed regression model using two benchmark thermal image datasets, the autonomous system lab thermal infrared dataset and the Ohio state university thermal pedestrian database. The proposed regression-based method can detect humans with 99.16% precision and 98.69% recall, outperforming the state-of-the-art conventional hand-crafted and CNN-based techniques for human detection from thermal images. Further, the designed fully convolutional regression network has much reduced computational complexity; yet, the detection performance is on par with the state-of-the-art fully convolutional architectures for predicting the density maps for human detection in thermal images.