UAV Vision-Driven Intelligent Measurement for Prefabricated Steel Columns

Prefabricated steel structure construction provides an efficient path for structural construction and also places higher demands on on-site installation accuracy. To address the large number of column components at large-scale construction sites， the low efficiency of traditional manual point-by-point measurement， the wide coverage of orthophotos， and the small proportion of individual columns in the overall image， this study establishes a two-stage rough-to-fine measurement framework for batch inspection of rectangular steel columns at large-scale construction sites. In the first stage， YOLO11-OBB is used for oriented coarse localization of column components， and image rotation and secondary correction of the coarse localization boxes are performed according to the column-group orientation distribution， so as to rapidly obtain unified local measurement regions containing the complete column-top contours. In the second stage， EdgeNAT is used to obtain local edge responses， pixel-level candidate corners are obtained through contour refinement， and two-dimensional Gaussian fitting is applied to estimate subpixel coordinates.This two-stage processing decouples target search in large-scale scenes and local high-precision boundary measurement into separate tasks， and primarily addresses the difficulty of meeting dimensional measurement accuracy requirements when relying only on detection boxes， as well as the strong background interference and large processing scope encountered when directly searching for precise edges in the full orthophoto. After the two stages， column dimensions are calculated from the orthophoto affine parameters， and column-corner engineering coordinates are output through a four-parameter transformation based on control points. Field experiments under a flight altitude of 110 m and an orthophoto ground sample distance （GSD） of approximately 18.5 mm/pixel showed that， for eight length （width） dimensions of four columns， with tape-measured values as reference， the mean absolute error and mean relative error were 18.75 mm and 2.21%， respectively. UAV image acquisition over the 0.074 km² test area required approximately 11 min， and orthophoto reconstruction and algorithm processing required approximately 1.5 h and 0.5 min， respectively； the latter two steps can be performed offline after field data acquisition， demonstrating the framework’s suitability for automated batch inspection of multiple components.