A Comprehensive Review of YOLO Object Detection Models: Evolution, Performance, and Future Directions

Abstract

YOLO, over the past few years, has emerged as the first choice for real-time object detection techniques, affecting real-world applications and participating in the realm of academia for research. The review paper gives a perspective regarding the evolution, which has been witnessed in the YOLO models, starting right from the first YOLOv1, which had a simple yet effective CNN-based architecture, till the latest YOLOv7-EDGE, incorporating transformer-based architectures. The paper also delves into the updations in the architecture, the advancement in the backbones, the measurement of performance parameters such as the Average precision(mAP), the FPS, Latency, FLOPS, and the number of parameters. The paper also includes the training parameters such as the preset setting for the data, GPU utilization, and training time. It also compiles real-world applications across fields such as health care, autonomous vehicles, intelligent surveillance, robotics, and many more, based on implementations involving various versions of YOLO. In this paper, further aspects will be discussed with regards to the remaining difficulties despite the vast progress, which include bias within the dataset, computation cost, and generalization within unseen environments. By compiling these findings, this paper aims to guide researchers and developers in selecting and applying the most suitable YOLO version.