Optimizing You Only Look Once Version 5 (YOLOv5) for Object Detection in Thermal Images

The field of computer vision, has seen exponential growth over the past few years. Within computer vision, object detection is a popular field due to its vast application to real world problems. You Only Look Once (YOLO) is an object detection model trained on the Common Objects in Context (COCO) dataset. YOLO models are widely recognized for their performance in object detection with RGB images; however, their accuracy sees a substantial falloff when dealing with thermal fused images, due to their distinct characteristics. While prior studies have addressed this limitation, none provide open-source code. Our research aims to fine-tune YOLOv5 for thermal fused images and publish our modifications. YOLOv5 comes pretrained with anchors, a backbone, and a head, all optimized for RGB image utilization. Our study tests numerous modifications to these components, to develop a model which excels at detecting the distinct features within thermal images. By sharing our findings and code, we hope to advance the application of YOLOv5 in thermal imaging and facilitate future research into this area.