Real-Time Multi-Object Detection using Anchor-Free YOLOV8 Architecture

Abstract

Object detection is a foundational task in computer vision, enabling the identification and localization of multiple objects within images or video frames. It plays a crucial role in real-time applications such as smart surveillance, autonomous vehicles, robotics, and urban traffic monitoring. Traditional models, including earlier versions of YOLO (You Only Look Once), rely on anchor-based mechanisms that require predefined bounding box templates. While effective, these systems often suffer from increased computational complexity, limited generalization to varying object scales, and the need for extensive hyperparameter tuning. In this study, we explore and implement multi-object detection using the anchor-free YOLOv8 framework, a modern and reengineered object detection model developed by Ultralytics. YOLOv8 introduces several architectural improvements, including an anchor-free detection head, a C2f backbone for efficient feature extraction, decoupled classification and localization heads, and confidence-weighted non-maximum suppression (NMS) for refining detections. These enhancements collectively improve both detection accuracy and inference speed. To evaluate the system's performance, we conducted experiments on the MOT20 and MS COCO datasets. Our implementation achieved a mean Average Precision (mAP@0.5) of 54.7%, a precision of 75.8%, a recall of 71.2%, and an inference speed of 160 FPS on an NVIDIA RTX 3080 GPU using the YOLOv8-small variant. Compared to anchor-based models such as YOLOv5 and YOLOv7, YOLOv8 demonstrated superior performance, especially in scenes with multiple overlapping objects and varying scales. These results highlight the advantages of anchor-free detection in reducing complexity, improving generalization, and achieving real-time processing speeds. The proposed YOLOv8-based system presents a robust and scalable solution for high-performance multi-object detection in practical, real-world environments.