Object Detection on Thermal Images: Performance of Yolov4 vs Yolov4 Tiny trained on Custom Datasets

The process of identifying and pinpointing the location of objects within an image or video is a crucial task in computer vision, known as object detection. While there has been significant progress in object detection using conventional optical RGB images, there has been comparatively less research done on object detection using thermal images. Thermal imaging has the advantage of being able to capture images in low light or even complete darkness, making it an attractive technology for surveillance applications. However, due to the scarcity of publicly available thermal image datasets, the development of object detectors specifically for thermal images has been hindered. In the proposed work a dataset of 2000 thermal images of three classes namely Humans, Dog, and Cats is collected using FLIR thermal camera. The YOLO (You Only Look Once) algorithm, specifically YOLOv4 and YOLOv4 tiny versions, are assessed for their performance to classify thermal images into three classes namely Human, Dog, and Cat. The findings demonstrate the potential of YOLOv4 for object detection on thermal images, especially when trained on large custom datasets. The results of this study may lead to the design of effective and efficient low-light vision systems that can be utilized in thermal imaging applications.