Defining quality metrics for photogrammetric construction site monitoring

Point clouds are becoming a quasi-standard as a representation for capturing the existing context, for construction progress monitoring and quality control. While it is possible to create and track sites with a reasonable amount of effort using photogrammetry, different recording strategies and computational power lead to different properties of the point cloud. While the needed specifications are based on the concrete type of analysis and will vary from recording to recording, the overall properties of the reconstruction toolchain are immanent to assess the performance of further processing. Within this paper we will present different criteria for the quality and evaluation of point clouds in respect to construction sites. These indicators together form a benchmark and can be used to evaluate a given toolchain and estimate the properties of a resulting point cloud. While all these publications tend to apply slightly different use cases, the first hurdle for a successful identification is registration of the captured point cloud to the geometric representation. For the relative registration of the geometric model, two base approaches exist in literature: feature-based recognition and the reduction of distances between the alignments (optimization). An Iterative Closes Point (ICP) (Besl and McKay, 1992) showed good results (Bosché, 2010; Masuda et al., 1996) after a successful manual initial alignment. It also worked well in combinational approaches (Huang and You, 2013). A generalized approach for surface to pointcloud was deducted by Segal (Holz et al., 2015; Segal et al., 2009). All these approaches need initial alignment or filtering since the ICP is a non-convex method. This problem leads to a global alignment when dealing with noisy construction site captures (Braun et al., 2016; Tuttas et al., 2017). The SfM/MVS based approach was identified as less accurate but much cheaper (Golparvar-Fard et al., 2011) compared to laser scanning and lead to a discussion of quality (Toth et al., 2013). To predict the quality of site captures, recent developments showed two different approaches: deducing the quality of the recording based on the result of the identification process (Rebolj et al., 2017) or using pure point cloud related properties and toolchains (Angel Alfredo Martell, 2017; Dyer, 2001; Haala et al., 2013). Following up on these publications, we present simple metrics of quality and emphasize on developing robust independent criteria. 3 PROCESS OF RECONSTRUCTION 3.1 Structure from Motion The structure from motion pipeline provides us with the first step to create a 3D model from taken images. It eliminates the need of a calibrated camera where the extrinsic and intrinsic parameters are fixed and known. In all taken images, points of interest are calculated and their correspondence is determined. Rejecting flawed correspondences, each camera is registered relative to the initial match. With the help of the bundle adjustment, the overall error is reduced significantly as multiple images will be refitted to the current model. The main goal of the SfM is to generate the initial camera configurations of a scene captured by one or more cameras with multiple images. Since on construction sites, the images are quasi random, it is the first step in retrieving correct 3D information. The correspondences exist as a point cloud deduced for the camera alignment. They already have partial geometric information of the dense reconstruction. The camera alignment is the basis for all common reconstruction pipelines used in construction site monitoring. 3.2 Multi View Stereo Multi-view stereo algorithms vary significantly in their principles. Seitz categorized the existing methods by six major properties (S. M. Seitz et al., 2006). 1. Scene representation Voxels, volumes or levelset methods represent the approximate surface, polygon meshes as facets and depth maps as 2D representation. 2. Photo consistency Two main competitors: Determined by the discretization and projection in scene (reconstruction grid) or image space (pixels of the image). 3. Visibility model The model verifies, if the view needs to be considered during calculation. This is especially important with larger scenes. 4. Shape prior During the reconstruction, assumptions for the shapes are imposed e.g. approaches that minimize surfaces. 5. Reconstruction algorithm Different types are used: calculating the cost of voxels, evolving a surface iteratively, enforced consistency in depth maps and merging them into a 3D scene, fitting a surface to an extracted set. 6. Initialization requirements Needed initialization may be bounding boxes, fore-/background separation. Image-space algorithms restrict the disparity or depth values. These properties will later define the quality of different algorithms and in case of 1.) if a point cloud is a suited output. In the next section we will look at current applications available and group them regarding to these fundamental properties.