ICP registration with SHOT descriptor for arresters point clouds

Abstract

Abstract Arresters are one of the critical components of the power system.
However, due to the arrester’s regular and uniform umbrella skirt, both traditional
manual detection methods and existing computer vision approaches exhibit limitations
in accuracy and efficiency. This paper proposes an automatic, robust, efficient arrester
point cloud registration method to address this problem. First, a robotic arm
maneuvers a depth camera to capture point cloud data from various perspectives.
Then, the fast global registration (FGR) point cloud coarse registration method
based on the signature of histograms of orientations (SHOT) descriptor to produce
preliminary registration results. This result is ultimately used as the initial value of
the improved iterative closest point (ICP) algorithm to refine the registration further.
Experimental results on various data sets collected from arrester and public data sets
show that the algorithm’s root mean square error(RMSE) is less than 0.1mm, meeting
the requirements of the engineering application of arrester detection.