Robust Edge Detection and GPU-Based Smoothing for Extracting Surface Primitives from Range Images

Abstract

It is important to construct 3D virtual models of man-made fields in which people work and live. Recent mid-range and long-range laser scanners can be used to acquire 3D shapes of cities, buildings, factories, heavy goods, transportation infrastructure, and so on. However, they tend to produce outliers and very noisy points near silhouettes and sharp edges of objects. This problem makes it difficult to reconstruct bounded faces. In addition, since enormous volumes of point-clouds are captured from a broad range of scenes, efficient processing methods are required. In this paper, we propose a robust edge detection method and an efficient GPU-based smoothing method for reconstructing primitive surfaces. We first calculate straight edge lines and silhouette lines from raw scanned data, and then eliminate noises and outliers by our GPU-based smoothing method for calculating surface equations. Then primitive surfaces are extracted using sharp edges, silhouette lines and surface equations. Our method is useful to robustly extract surface primitives from practical noisy pointclouds.

Detection of Surfaces and Edges in Large Point-Clouds Using Region-Growing on Spherical Space

“…Mixels on a silhouette line (b) Reflection from an inclined face Fig. 3: Outliers at silhouettes and tangent surfaces [2] Processing on Spherical Space. While the directions of the laser beams are relatively precise, the distance values are very noisy on feature points.…”

“…Therefore, we discard the distance values of feature points. In our previous method [2], points were projected on 2D perspective screens, but a wide range of points could not be covered by a single screen. Therefore, in this paper, we project all points onto a single 2D spherical screen.…”

“…Planar faces in engineering facilities also have parallel lines at their boundaries in many cases. Therefore, we search for planes and cylinders in regions between parallel lines [1,2]. We applied Delaunay triangulation to feature points to detect adjacent segments.…”

“…In our previous work [1,2], we proposed methods for localizing search space to efficiently detect cylinder and planar faces. Many components in facilities consist of primitive surfaces.…”

“…We searched for edges in 3D space in the first paper [1], but this method was not robust because naive thresholds are required. In the next work [2], we discarded distance values and projected points on 2D perspective screens. This method realized robust edge detection, but perspective projection covers only limited view angles.…”