基于深度学习的水域实例分割尾矿库干滩长度测量  

Tailings Pond Dry Beach Length Measurement Based on Deep Learning Water Instance Segmentation

作　　者：孙叶青[1] 陈洪飞[2] 童仁园[1] SUN Yeqing;CHEN Hongfei;TONG Renyuan(China Jiliang University,Hangzhou,Zhejiang 310018,China;Zhejiang Institute of Hydraulics and Estuary,Hangzhou,Zhejiang 310020,China)

机构地区：[1]中国计量大学,浙江杭州310018 [2]浙江省水利河口研究院,浙江杭州310020

出　　处：《计量学报》2024年第11期1607-1614,共8页Acta Metrologica Sinica

基　　金：国家重点研发计划课题(2022YFC3003403);浙江省教育厅项目(Y201942369)。

摘　　要：提出了基于YOLOv8的水域实例分割方法,实现了在实时视频流下快速、高效、准确的尾矿库干滩长度测量。首先,完成一份高质量水域实例分割的COCO数据集;其次,分析主流深度学习实例分割算法,选用YOLOv8模型训练出高效识别水线并输出图像坐标;最后,标定相机内外参数,应用相机成像原理,在尾矿库尾部安装监控摄像头,预测出干滩长度。实验证明:此模型不仅能够准确预测出干滩长度,并且对不同尾矿库水域边界分割有较好的稳定性;对实时视频流模式下的野外非接触式测量具有较好的效果,误差控制在2%以内。A method based on YOLOv8 for water instance segmentation has been proposed,achieving rapid,efficient,and accurate measurement of the dry beach length of tailings ponds under real-time video streams.Firstly,a high-quality water instance segmentation COCO dataset is completed.Secondly,mainstream deep learning instance segmentation algorithms are analyzed,and the YOLOv8 model is chosen to efficiently recognize the waterline and output image coordinates.Finally,the internal and external parameters of the camera are calibrated.By applying the principles of camera imaging and installing surveillance cameras at the end of the tailings pond,the dry beach length is predicted.Experiments prove that this model can not only accurately predict the dry beach length but also has good stability in segmenting the water boundaries of different tailings ponds.It has a good effect on non-contact measurement in the field under real-time video stream mode,with an error controlled within 2%.

关 键 词：长度测量 水域实例分割 干滩长度 尾矿库 COCO数据集 YOLOv8算法 实时视频流 

分 类 号：TB92[一般工业技术—计量学]