Improved Real-Time Monitoring Lightweight Model for UAVs Based on YOLOv8

Chuanlei Zhang, Xingchen Zhao, Di Sun, Xinliang Wang, Guoyi Xu, Runjun Zhao, Ming Gao, Hui Ma

Abstract: This paper proposes a lightweight target detection model based on YOLOv8, which has been improved for real-time detection of circuit system construction site apparatus and equipment. The model addresses the shortcomings of traditional target detection methods, including low accuracy and slow detection speed. In this paper, we replace the C2F module in the YOLOV8 backbone network and the C2F module in the neck network, which is responsible for fusing low-level and intermediate-level features from the backbone network and high-level semantic features from the backbone network as well as features from the detected head, with the lightweight RepNCSPELAN4 structure. This reduces the size of the model and lowers the computational cost. Furthermore, a novel attention mechanism, iRMA, is incorporated into the C2F module to form a novel feature fusion module, C2F_iRMA. This module is employed to replace the original C2F, which is used for further feature fusion of the extracted features from the backbone network in the YOLOv8 neck network. This replacement is intended to enhance the detection accuracy. The experimental results demonstrate that the enhanced model exhibits notable advantages in the domain of real-time detection of construction equipment. It exhibits a 5% increase in detection speed, 99.9% accuracy, and 98.6% recall, which surpass the performance of current mainstream network models.