Human Detection in Thermal Imaging Using YOLO

In this paper, we consider the problem of automatic detection of humans in thermal videos and images. The thermal videos are recorded on a meadow with a small forest with up to three persons present on the scene at different positions and ranges from the camera. To simulate realistic conditions that can happen during surveillance and monitoring of protected areas, all videos are recorded at night but different weather conditions--clear weather, rain, and fog. We present the results of human detection on a custom dataset of thermal videos using the out-of-the-box YOLO convolutional neural network and the YOLO network trained on a subset of our dataset. YOLO is an object detector pretrained on the COCO image dataset of RGB images of various object classes. Test experimental results have shown significantly improved performance of human detection in thermal imaging in terms of average precision for trained YOLO model over the original model.