Semantic-driven spatial fusion for noise-resilient distance measurement in autonomous inspection of insulators

Abstract

Computer vision-based methods have shown great promise in obtaining object distances, significantly improving the efficiency of power distribution system component construction acceptance. However, the complex backgrounds of overhead power lines pose significant challenges to measurement accuracy. To address this, we propose a novel approach that fuses semantic segmentation and spatial reconstruction for noise-resilient distance measurement. The method begins with instance segmentation to generate semantic masks of insulators, followed by binocular vision-based spatial reconstruction. By leveraging depth and density information, DD-Clustereo model is designed to adaptively distinguish valid points from background noise, ensuring precise measurements of the shortest distances between insulators. Experimental results demonstrate that the fusion of semantic and spatial features effectively eliminates background interference, achieving an average error rate of just 2.16%. This work highlights the transformative potential of information fusion in empowering power component inspection through machine vision.