A Comprehensive Comparison Between YOLOv5 and YOLOv8 for Embedded Systems Applications

Object detection has a vast potential in computer vision, which can be used in any application starting from autonomous vehicles to real-time monitoring systems. Among these methodologies, the YOLO (You Only Look Once) series has become one of the most successful ones in terms of speed and accuracy. This paper seeks to discuss and compare YOLOv5 to YOLOv8 to highlight the improvement made in the present model and their ability to be implemented in embedded systems. Hence, YOLOv5 adds and refines lightweight architecture learning and bounding box anchors for improved deployment. YOLOv6 enhances the primary parameter of real-time applications, namely inference speed, while YOLOv7 further introduces the techniques of model scaling and removing the concept of anchors for precision. YOLOv8 was subsequently improved with transformer-based modules and model scalability to work on both high-end and low-end devices. Specifically, the results obtained through the experiments are discussed, focusing on the so-called key performance indicators such as mean average precision mAP, frames per second FPS, and computational performance, and their possible intersection for real-time object detection applications. Further, a literature review highlights their reallife application in various business contexts. This paper provides the reader with a roadmap for choosing the best implementation for YOLO at the system level to maintain a high degree of accuracy, speed, and resource use.