Planar Surface Extraction For Complex Facades From Unstructured TLS Point Clouds

Presently more and more applications-such as 3D modelling, as-built surveys, documentation, restoration and reconstruction-need to create 3D models of man-made objects. That is in order to describe the complete details for the object features with best fit reality. Therefore 3D modelling is considered as a complicated and complex research problem in computer vision, photogrammetry and remote sensing, since it is consisting of many different processes such as extraction, detection, reconstruction of the object’s elements and also it is require automatic processing of massive 3D data acquisitions, for complex object, to extract elements of the recorded objects.In the context of acquires massive 3D data from complex object, on one hand, in fact the new variables offers of Terrestrial Laser Scanner (TLS) instruments are very efficient tools for the acquisition of massive quantities of data (3D point clouds) due to their speed, which are giving them considerable potential for data acquisition for 3D modelling. On other hand, in commonly case of complex object, TLS acquires massive (up to some millions) unstructured (randomly distributed) 3D point clouds with different local densities and in presence of random noisy points, that is due to many different factors such as; instrument resolution, registration step, elements’colours, elements’roughness and errors of measured points. Therefore the extraction of object features for find 3D modeling of building facade from massive unstructured point clouds with different local densities, especially in the presence of random noisy points, is not a trivial task even if that feature is a planar surface and/or interprets linear (edges). Thus this process is considered as one of the complicated cutting edge issues. The selected topic is crucial important for 3D documentation of man-made buildings.Segmentation is the most important step in the feature extraction process. In practice, most segmentation approaches use geometrical information to segment the 3D point cloud. The features generally include the position of each point (X, Y and Z), locally estimated surface normals and residuals of best fitting surfaces; however, these features could be affected by noisy points and in consequence directly affect the segmentation results to be have bad-segmentation results (i.e. over-segmentation, under-segmentation or no segmentation). This finally can be leads to a failure in the extraction process and consequently a failure for creating the 3D model.While the RANSAC (random sample consensus) algorithm is effective in the presence of noise and outliers, it has significant disadvantages, but the plane is detected by RANSAC may not necessarily belong to the same object surface, i.e. spurious surfaces may appear, especially in the case of parallel-gradual planar surfaces such as stairs.The innovative idea in this study, to avoid the spurious surfaces of parallel-gradual planar surfaces, is a modification for the RANSAC algorithm called Sequential Normal Vector RANSAC "Seq-NV-RANSAC". This algorithm checks the normal vector (NV) between each existing point clouds and the hypothesised RANSAC plane, which is created by three random points, under an intuitive threshold value. After extracting the first plane, this process is repeated sequentially (Seq) and automatically, until no planar surfaces can be extracted from the remaining points under the existing threshold value.Since the bad-segmentation results (i.e. under-and over-segmentation) are still standing as a big obstacle to extract planar surfaces with best fit reality. Therefore an extension of "SEQ-NV-RANSAC" approach, to avoid the bad-segmentation problems, is proposed using topology information and intuitive threshold value. First, in order to avoid the under-segmentation problem, each one group-which is resulted previously by proposed "SEQ-NV-RANSAC" approach-is checked to get all neighbours points which have Euclidean distance less than the threshold value as a one surface group. This process will be repeated until no more points can be adding to that surface group. Then a new surface group will be created to check the remaining points. Second, in order to solve the over-segmentation, three checks are proposed; the similarity of normal vectors (NV), the perpendicular distance and the intersection zone using bounding box test. Then we are presenting the details steps to interpret linear features, for all extracted planar surfaces previously, using Sequential RANSAC for Extract Edges "Seq-RANSAC-Edges" as a new proposed approach. Firstly, Intersection Edges algorithm (IntEdges) is used to find these point clouds which belong to the intersection line between two neighbours planar surfaces by creating the topology relationship for all existing planar surfaces. Secondly, Free Boundary Points algorithm (FreeBoundaryPoints) is used to find the outlines points which are envelop each existing planar surface as a one ring. Thirdly, in order to avoid the duplicated edges at intersection parts, Merge the Edges points algorithm (MergeEdPos) is used to merge the results from the both algorithms IntEdges and FreeBoundaryPoints. Fourthly, Seq-RANSAC-Edges algorithm is used to segment each edge points individually and also extracts edges line by fitting the segmented edges points for all existing planar surfaces automatically and in sequence even in presence of more than one edge in each group, as respect to the common data.For all approaches, are involving in this study, the general workflows are showed and the methodology for the proposed approaches are explained demonstrating by flowcharts and equations. Finally, the proposed approaches are applied on the practical data and the final obtained results are demonstrating by figures and evaluated.Finally, these approaches are proposed prevent the extraction of spurious surfaces and edges, bring an improvement in quality to the computed attributes and increase the degree of automation of surface extraction. Also since the best fit is achieved for the real existing features (planar surfaces and linear edges), therefore these results can be the considerable and potential for creating 3D modelling directly from massive unstructured 3D point clouds.