Real-Time Human Detection with Thermal Camera Feed using YOLOv3

Human detection is needed for various applications such as advanced driver assistance systems and autonomous driving, security and surveillance etc. Thermal imaging is often adopted for night time because of its capability of capturing the energy emitted from human body where visible light camera fails. In this paper, we employ YOLOv3 for an accurate real-time human detection using thermal images. We modified the network parameters according to the characteristics of the human, making this method more suitable for detecting human. Subset of Korea Advanced Institute of Science and Technology (KAIST) multispectral dataset consisting of 47650 thermal images is used for training and testing of YOLOv3. During experimentation, it is observed that humans are detected at 17 millisecond which is much faster than a local machine detection. Our test result also shows improved performance of the detector with thermal image with average precision of 95.5% and miss rate of 4.7%.