Accumulator-based three dimensional building boundary reconstruction framework for satellite images

Mapurisa, Willard

Accumulator-based three dimensional building boundary reconstruction framework for satellite images

Thesis / Dissertation

2024

Publisher

University of Cape Town

Department

School of Architecture, Planning and Geomatics

Faculty

Faculty of Engineering and the Built Environment

Abstract

Three-Dimensional (3D) building boundary detection and reconstruction from satellite images has various applications in computer vision such as 3D city modelling, radio network planning, urban planning, mapping, and navigation. Traditionally, 3D building models have been produced from multiview satellite images using dense pixel matching. The point clouds or Digital Surface Model (DSM)s produced from the matching are then used to estimate 3D building boundaries. From these point clouds 3D wireframe models can be generated by recovering the surface boundaries of objects in the point cloud. Ideally, it is desirable for recovered edges to follow the object edges without generalisation and with good localisation. However, the accuracy and fidelity of most of the reconstructed boundaries are diminished because of point mismatches, occlusions, and the failure to recover sharp object discontinuities during image matching. An alternative 3D boundary detection and reconstruction approach that has been less widely explored is the use of edge features exclusively for 3D reconstruction. The advantage presented by using edge features in 3D reconstruction is the ability to preserve the shape of the building boundary during reconstruction. However, the density of edges recovered using existing detectors has not reached a level sufficient to recover full closed boundaries of objects. This limits the exclusive use of edge features for 3D boundary reconstruction. 3D boundary reconstruction from edge features involves three processes namely edge detection, boundary detection and reconstruction, and edge matching to recover 3D building geometry. With edge detection, fragmentation and missing edges hinder the successful recovery of complete object boundaries. Furthermore, the need for thresholding parameters for edge detectors contributes to a reduced number of detected edge pixels, resulting in piecewise open boundaries. A 3D reconstruction from recovered fragmented lines results in a sparse 3D model. Based on the challenges presented above, the research presents a novel edge detection and building boundary reconstruction framework from satellite images. By recovering dense edge features, detailed 3D wireframe models are produced. The proposed solution works on the idea that edge information extracted from different detectors with parameter variation can be aggregated to obtain better edges. The recovered edges ultimately result in improved boundary detection and reconstruction. The solution was achieved through two objectives. The first objective was to develop a iv sub-framework for edge detection that reduces fragmentation and missing edges through aggregation. The second objective was to develop a boundary detection and reconstruction sub-framework from the detected edges. From the first objective, an Accumulator Based Edge Detector (ABED) framework was developed that aggregates edges by automated parameter tuning. Good edge localisation is maintained by a localisation filter during edge aggregation while keeping salient edges. ABED introduces genericity since gradient operator agnostic. From the dense edge map produced by the edge detection sub-framework, a building boundary shape reconstruction sub-framework was developed termed Accumulator Based Line Detector (ABEDL). The boundary recovery method works by recovering longer unbroken lines through a Breath-First Search (BFS) line search algorithm with parameter relaxation. The boundaries are then matched to produce detailed 3D building wireframes. To evaluate the frameworks, ten different building sites were selected from five different locations around the world. The sites were chosen because of different building architectural styles, differing imaging resolution, and varying view angles. For benchmarking, ABED was compared against Parameter Free Canny Edge Detection (CannyPF) [1] and Parameter Free Edge Drawing (EDPF) [2] for each site. ABED recovered more edge pixels with improved recall for all ten sites. Furthermore, the Root Mean Square Error (RMSE) for edge position errors was less than 1.4 pixels when compared with ground truth images. ABEDL was benchmarked against eight state-of-the-art line detectors namely Edge Drawing Lines (EDLines) [3], Line Segment Detector (LSD) [4], Canny Lines (CannyLines) [1], Progressive Probabilistic Hough Transform (PPHT) [5], Active grouping and geometry-gradient combined validation Line Detector (AG3Lines) [6], Fully Convolutional Line Parsing (F-Clip) [7], L-CNN End-to-End Wireframe Parsing (L-CNN) [8] and Hough-Transform Holistically-Attracted Wireframe Parsing (HT-HWAP) [9]. ABEDL showed good F1-Score on all ten sites. Furthermore, ABEDL recovered more boundaries for all sites. The major finding of the developed frameworks is the possibility of using satellite images to reconstruct 2D and 3D building boundaries using dense edge features as opposed to dense pixel matching. The edge detection sub-framework introduces genericity in edge detection and does not require predefined parameters. Furthermore, ABEDL presents the ability to reconstruct object form or outline and detail such as windows, allowing the reconstruction of LOD 2 and 3 detail.

Keywords

Engineering

Reference:

Collections

PhD / Doctoral

Full item page