ABSTRACT Safety distance monitoring is a critical requirement for live substation operations, where inaccurate distance perception may lead to severe safety hazards. Existing vision‐based approaches commonly estimate distances using single representative points, such as detection box centres, which are sensitive to irregular object geometries, depth noise, and threshold‐induced decision instability in complex operational environments. This paper proposes a vision‐based safety distance monitoring framework that integrates image segmentation and binocular vision for robust spatial distance perception. Operational elements are modelled as three‐dimensional regions rather than isolated points, and a mask‐constrained region‐to‐region distance formulation is introduced by combining candidate point selection with robust statistical aggregation. An enhanced segmentation‐based perception module is adopted to improve planar localisation accuracy under scale variation and deformation, while a lightweight pruning strategy enables real‐time deployment. Experimental results on live substation operation scenarios show that the proposed method reduces the distance estimation error from 2.52 m to 1.04 m in MAE and from 3.48 m to 1.68 m in RMSE while improving the safety decision correctness from 88.6% to 92.4% and reducing both false alarms and missed warnings near safety thresholds.
Wang et al. (Thu,) studied this question.