A Comparative Survey of YOLO Variants in Object Detectors

Abstract

As a single-stage object detection framework, YOLO (You Only Look Once) has gained prominence for its remarkable balance between speed and accuracy across diverse detection tasks. This research article provides a comprehensive survey of the YOLO family of algorithms, tracing their evolution from the earliest version to the most recent advancements. The review offers an in-depth analysis of the performance and defining characteristics of each iteration, with particular emphasis on YOLO’s applications in real-time detection, especially on embedded systems. Special attention is given to recent developments in model compression and optimization techniques aimed at addressing the challenges posed by the large size of YOLO models, thereby enabling deployment on resource-constrained devices. Practical implementation examples are also discussed to illustrate its applicability in real-world scenarios. Finally, the study highlights potential research directions, including novel architectural innovations and advanced training strategies, that may further enhance the YOLO framework. Overall, this review serves as a valuable reference for researchers and practitioners seeking insights into the evolving landscape of YOLO-based object detection.