摘要
为解决吊装作业数据集获取困难与吊装作业过程中重要对象(吊物与吊钩)监管难题,提出虚实结合的方法构建数据集,基于SketchUp软件建立虚拟吊装作业场景获取虚拟吊装作业图片,同时从网络获取吊装作业图片及现场作业视频截图,将真实作业场景的图片与虚拟作业场景的图片共同组成虚实结合的数据集。引入可改变核卷积(Arbitrary Kernel Convolution,AKConv)和鬼魅空洞可分离卷积(Concentrated-Comprehensive Convolution with GhostBottleneck,C3Ghost)改进目标检测算法模型YOLOv5(You Only Look Once version 5),改进后的模型比原始模型在精确率上高出2.6百分点,在推理速度上高出9.1帧/s,且模型所占存储容量降低1.9 MB。搭建可视化操作界面,与优化好的模型整合成吊装作业实时监测系统,实现对吊物和吊钩的安全状态识别和风险预警,及时进行风险管控。
To address the challenges of acquiring datasets for lifting operations and the supervision of critical objects(such as lifting loads and hooks)during these processes,this paper proposes a virtual-real method for dataset construction.It includes creating a virtual lifting operation scene using SketchUp to generate images that simulate lifting operations.Simultaneously,lifting operation images are gathered from online sources and screenshots are extracted from on-site operation videos.By combining images of real operation scenes with those from the virtual operation scene,we create a comprehensive virtual-real dataset.The issue of limited quantity and variety of lifting operation datasets can be effectively addressed by the method of combining virtual and real datasets.This approach,in particular,allows for a significant expansion in the number of images depicting unsafe states of objects.The YOLOv5s network model was enhanced by incorporating Arbitrary Kernel Convolution(AKConv)into the backbone network.This modification offers flexible sampling shapes while minimizing computational resource consumption,enabling more accurate extraction of target features and improving adaptability to multi-scale targets.Additionally,the Concentrated-Comprehensive Convolution(C3)in the neck network was replaced with Concentrated-Comprehensive Convolution with GhostBottleneck(C3Ghost)to achieve a lightweight model with minimal performance loss.The improved model demonstrates 2.6 percentage point increase in accuracy compared to the original,an inference speed that is 9.1 frames per second faster,and a reduction in model size by 1.9 MB.This not only enhances the accuracy of the algorithm but also results in a lightweight model,making it suitable for deployment on mobile or edge devices with limited computing power.The visual operation interface is developed using the Qt framework and is integrated with the improved YOLOv5s network model to create a real-time monitoring system for lifting operations.When implemented at lifting operation sites,the system can identify the safety status of lifting loads and hooks in real time,providing timely risk alerts.This capability offers valuable technical support for intelligent safety management.
作者
张颖
刘洋
赵鹏程
张珂
吴义蓉
ZHANG Ying;LIU Yang;ZHAO Pengcheng;ZHANG Ke;WU Yirong(School of Safety Science and Engineering,Changzhou University,Changzhou 213164,Jiangsu,China)
出处
《安全与环境学报》
北大核心
2025年第2期508-517,共10页
Journal of Safety and Environment
基金
江苏省研究生科研创新计划项目(KYCX24_3146)。
关键词
安全工程
机器视觉
深度学习
吊装作业
可改变核卷积
safety engineering
machine vision
deep learning
lifting operation
arbitrary kernel convolution
