Open Access

ARTICLE

Robust Analog Gauge Reading via Virtual Point-Based Geometric Rectification and P2-YOLO-Pose

Jaekyung Lee^1,2, Youngjun Kim², Byungsung Ko², Taewon Kim², Jaeheon Park², Jiwon Lee², Wonhee Kim^1,*

1 School of Energy Systems Engineering, Chung-Ang University, Seoul, Republic of Korea
2 KEPCO Research Institute, Daejeon, Republic of Korea

* Corresponding Author: Wonhee Kim. Email:

(This article belongs to the Special Issue: Data-Driven and Physics-Informed Machine Learning for Digital Twin, Surrogate Modeling, and Model Discovery, with An Emphasis on Industrial Applications)

Computer Modeling in Engineering & Sciences 2026, 147(1), 35 https://doi.org/10.32604/cmes.2026.080624

Received 13 February 2026; Accepted 30 March 2026; Issue published 27 April 2026

Abstract

Automated reading of analog gauges in industrial environments is essential for predictive maintenance and safety monitoring. However, conventional computer vision approaches encounter two fundamental bottlenecks: polar unwrapping techniques induce severe nonlinear scaling distortions under oblique viewing angles and axis-aligned bounding boxes (AABBs) are geometrically inefficient for encapsulating high-aspect-ratio rotating needles. To overcome these limitations, this paper proposes a novel end-to-end framework that innovatively redefines gauge reading as a structural pose estimation task. We model each gauge as a topological five-keypoint skeleton (kstart,kmid,kcenter,kend,ktip), and localize these landmarks using a customized P2-YOLO-Pose architecture. By integrating a high-resolution P2 feature layer (stride 4) while excising the macro-scale P5 layer, the network yields a 40% enhancement in small-gauge detection recall with a negligible (<1%) frame-rate degradation. Furthermore, to address the intrinsic lack of salient vertices in circular dials, we introduce a Virtual Point (VP) generation algorithm. This algorithm exploits the point symmetry of the detected keypoints to autonomously synthesize four spatial correspondences, thereby enabling markerless, homography-based perspective rectification for corner-free objects. An adaptive control mechanism based on aspect ratio analysis (AR≤1.5) dynamically regulates the geometric warping to prevent algorithmic over-correction. Extensive evaluations on an 11,000-image field dataset acquired from an operational power data center demonstrate a Pose mAP50 of 99.45% and an mAP50-95 of 99.37%. Under severe vertical tilt conditions, the VP-based rectification curtails the absolute reading error from 3.5% to 0.6% compared to the uncorrected baseline, attaining measurement precision commensurate with physical ArUco marker-based ground truths. Operating in real-time at 25.9 FPS, the proposed system is currently deployed within an integrated inspection platform coupled with an autonomous quadruped robot (Boston Dynamics SPOT), facilitating reliable, perspective-invariant visual inspections across 10 distinct classes of analog gauges in an active industrial facility.

---

Keywords

---