Conventional vision-based monitoring systems on construction sites typically output object classes, bounding boxes, or masks but not the metric parameters needed to calculate machinery safety distances. This study proposes a monocular scale-recovery method for extracting machinery geometry and delineating dynamic safety radii. Instance segmentation and monocular depth estimation generate instance-level relative point clouds, which are scaled to metric units using on-site semantic anchors of known size, such as safety helmets. The point clouds are refined through depth gating, photometric consistency filtering, connected-component analysis, and outlier removal to extract machinery dimensions and structural parameters. These measurements and operating-envelope parameters defined by machinery class or model are incorporated into a dynamic safety-radius model. Field experiments show that the mean absolute relative error (AbsRel) decreases from 11.2% to 6.3%, the proportion of valid points exceeds 95%, and the end-to-end processing latency is approximately 120.6 ms. The method therefore offers a low-cost approach to dynamic geofencing using existing monocular surveillance infrastructure.