YOLO-Based Object Detection: Architectural Evolution and Applications in Autonomous Driving and Medical Imaging

Object detection is a fundamental task in computer vision, in which both localization and classification of objects are required. In the traditional two-stage method, for example Faster R-CNN, high accuracy is achieved but real-time inference speed is limited, restricting its use in real-time detection applications. With the introduction of the You Only Look Once (YOLO) model, this limitation is addressed by reformulating image detection as a regression problem, thereby achieving faster inference. This paper reviews the structural developments of YOLO models, from the grid-based prediction and anchor box mechanism in early versions, to the recently introduced NMS-free detection method in YOLO26. Based on published papers and research studies from the original YOLO model to the most recent benchmarking works, this paper analyzes the technical mechanism behind each major improvement, and with representative applications in autonomous driving and medical imaging. The result shows that each version of YOLO keeps improving the balance between detection speed and accuracy, and the latest version also achieves better efficiency during real-time applications. In conclusion, the continuous development of YOLO makes it an important model for real-time applications, while achieving accurate small object detection and maintaining high performance under adverse conditions remain challenging.