Infrared Multi-Object Detection Using Deep Learning

Computer vision has gained great regard in the past few years and one of the most important branches of this field is object detection. It was usual to detect objects using visible images that need good illumination conditions to obtain high performance, but this type of images exhibits unsatisfactory attitude in bad environmental conditions and darkness situations. Other alternatives are emerging, such as infrared images which depend on the infrared radiation emitted or reflected from objects. Infrared imaging systems can be used in complete darkness, so the problem of visible images can be overcome. There are many object detection models which take images as an input and predict location and class type for each element in the image. So, a comparative study using deep learning based infrared object detection models (YOLOv3-4GB, YOLOv5s, YOLOv7) was performed. The models were trained on Forward Looking Infrared (FLIR) dataset that contains 3 classes (person-car-bicycle). The YOLOv7 model attained the best performance and achieved Mean Average Precision (mAP) of 82.4%.