基于深度机器视觉的香炉山隧洞钻孔多维特征精准定位

Precise multi-dimensional features positioning of Xianglushan tunnel drilling based on deep-machine vision

香炉山隧洞是滇中引水的控制性工程,岩爆是其主要灾害源之一。微震传感器坐标精确定位是岩爆微震监测预警的基础,但是其钻孔位置仍需人工测量,精度低且实时性差,严重阻碍岩爆微震自动监测系统的发展。该文提出一种基于深度机器视觉的隧洞钻孔多维特征精准定位方法YOLO(you only look once)-AT(anchor tracking)。首先,该方法利用改进YOLO V8模型YOLO V8 OBB(oriented bounding boxes)实现香炉山隧洞钻孔轮廓的精准拟合;其次,构建锚点追踪算法AT,在轮廓椭圆上选定锚点,并根据对极线约束和点序约束实现多视角下的锚点追踪;最后,利用视差法求解各锚点的空间坐标,并基于解析几何理论求解香炉山隧洞钻孔的多维特征。利用质量可控的香炉山隧洞合成钻孔轮廓数据库对各种钻孔定位方法进行验证,结果表明:YOLO-AT方法具有良好的钻孔轮廓拟合精度和钻孔多维特征定位精度,对钻孔口中心坐标、钻孔直径、钻孔口平面方向、钻孔方向的定位误差中位数分别为0.835 mm、0.795 mm、0.567°、1.751°,且定位性能受钻孔轮廓质量的影响小,满足香炉山隧洞岩爆微震自动监测的需求,并有向各类地下工程钻孔测量工作推广的应用潜力。

[Objective]Xianglushan tunnel is a key project of the Yunnan Dianzhong Water Diversion Engineering,and rock bursts are one of its main hazards.The precise positioning of microseismic sensor coordinates is the basis for microseismic monitoring and early warning systems for rock bursts,which is controlled by drilling coordinates and angles.However,the drilling location needs to be measured manually,which has low accuracy and poor timeliness,seriously hindering the development of automatic microseismic monitoring systems for rock bursts.Meanwhile,the precise positioning of multi-dimensional features is the key to realizing the full automation of drilling operations,including the coordinates,diameter,plane direction,and busbar direction of drilling.This study presents a method for the precise positioning of multi-dimensional features of tunnel drilling based on deep-machine vision called YOLO-AT.[Methods]The proposed method includes two modules:drilling contour detection and drilling multi-dimensional feature positioning.The drilling contour detection module adds rotation angle prediction to the YOLO V8 model,thereby constructing a YOLO V8 OBB model that can accurately fit the drilling contour.The drilling multi-dimensional feature positioning module establishes an anchor tracking algorithm,selects anchor points on contour ellipses,and achieves anchor tracking from multiple perspectives based on epipolar line constraints and projective invariance of anchor order.The spatial coordinates of each anchor point and multi-dimensional features were solved using the parallax method and analytical geometry theory,respectively.[Results]To verify the robustness of the proposed method to contour quality,a synthetic drilling contour database with controllable quality was established for Xianglushan tunnel,and the positioning results were compared with those of the ALSR,FED,AAMED,YOLO V8,and DLT methods.Results showed that:(1)concerning contour detection accuracy,the proposed method exhibited the best performance among other methods,with the average intersection over union(IoU)ratio,average F1 score,precision,and recall at an IoU threshold corresponding to 0.9 of 0.957,0.948,0.977,and 0.977,respectively.(2)Concerning contour detection stability,the proposed method was less affected by contour quality and always maintained high accuracy.By contrast,YOLO V8 had less fluctuation and low indicators,whereas other methods had poor stability and high accuracy under good contour quality.(3)Concerning the accuracy of multi-dimensional feature positioning,the median errors of the proposed method for the coordinate,diameter,plane direction,and busbar direction of drilling were 0.835 mm,0.795 mm,0.567°,and 1.751°,respectively,which were the best among all methods.(4)Concerning the stability of multi-dimensional feature positioning,the proposed method and YOLO V8-AT had better stability.Conversely,the stability of other methods rapidly decreased as the quality of the drilling contour deteriorated.[Conclusions]The proposed YOLO-AT method can achieve accurate and stable positioning of multi-dimensional features in Xianglushan tunnel drilling,with a performance superior to that of existing methods,thereby having the potential to be extended to various underground engineering drilling positions.Next,the research intends to focus on combining the proposed method and laser ranging technology to measure hole depth and eliminate ambiguity in drilling direction,in addition to analyzing the roles of RANSAC and reprojection methods in improving algorithm accuracy and stability.