Search

Constructing geometric digital twins of highways at present still demands substantial human effort. Unlike most previous work that uses deep learning models to segment point clouds of highways into class level or object instance level, this paper further segments pavements into a more detailed level (lanes, hard shoulders, central reserves). The central curves of each lane marking are fitted in a two-step method, approximated by a polynomial and then converted into the Frenet coordinated system. The fitted curves with smoothly changing curvature are used to separate points of road surfaces into lanes, hard shoulders, and central reserves, resulting in a mean Intersection over Union (mIoU) at around 90%. This automatic approach extracts geometric and object category information from point clouds and stores the information in a graph, showing the hierarchical relationships among various components and offering the potential for expansion into more comprehensive digital twins encompassing the entire highway network. • Segment point clouds of pavements into a more detailed level. • Fit curves to lane markings in point clouds and use them to segment pavement points. • Generate graph representation of highway geometric digital twins from point clouds. [ABSTRACT FROM AUTHOR]

To prevent safety accidents caused by scaffolds, safety managers on-site need to check a list of regulations whenever scaffolds are assembled, used, and disassembled. However, numerous errors can result from conventional manual inspection methods, leading to potential safety accidents at construction sites. This paper presents a three-step methodology to automate the inspection process of scaffolds with minimum human intervention: 1) acquisition of point cloud data from a construction site using a Terrestrial Laser Scanner (TLS), 2) classification of each point into seven different elements using a deep learning-based 3D segmentation model, RandLA-Net, and 3) inspection of the required regulations using a robust regulation checking algorithm. The efficacy of this methodology was proven by validating two construction sites that were different from the training dataset, showing 100% and 76.5% regulation checking F2 scores, respectively. • This paper presents a methodology to automate the inspection process of scaffolds using TLS-acquired point clouds. • Point clouds were segmented into 7 classes using a deep learning algorithm. • A robust regulation checking algorithm is constructed to find violations in the data. • Experiments showed regulation checking F2 scores of 100% and 76.5% on two sites, respectively. [ABSTRACT FROM AUTHOR]

This paper proposes an efficient and accurate process to find instances of a specific object in a point cloud scene where the object is unique, a member of a trained class, and the unique object's class is known. Object finding is essential for multiple applications like object detection, pose estimation, and asset tracking. Current template matching and object detection methods are slow or approximate feature matching to match objects, which generalize classes of objects and do not differentiate between specific objects. This paper proposes a specific-object finding methodology based on existing point cloud segmentation, fully convolutional geometric features, and a color-based iterative closest point algorithm. A point cloud template and scene are used, the latter is segmented, the resulting points matching the template's label are isolated, generating candidates. The candidate's geometric features are matched and compared with the point cloud template. The methodology results generate high accuracy for specific-object matching. • Methodology for finding specific objects in point clouds using algorithmic and machine learning for industrial applications. • An automatic resampling framework to specify similarity thresholds. • Network-based candidate generation using point cloud segmentation and convolutional geometric features. • Identification of matches based on double color-based registration. • A pipeline for finding specific objects that handles different point cloud densities. [ABSTRACT FROM AUTHOR]

Abstract Since the mid-19th century, the railway network has occupied a crucial place at the heart of the world's transport systems. Its infrastructure is often situated in harsh environments where an extreme event, or even daily use, could lead to a catastrophic accident. This is one of the main reasons why inspecting these constructions is so important. Despite the advances in this field, the human component continues to be part of the final inspection process. In order to improve on this, this paper shows the use of laser scanning as a leading technology in automating the inspection of railway infrastructures. The proposed methodologies provide the essential processed and classified data needed for the structural health monitoring of the various assets related to railways. It is divided into three main parts, which pre-process the point cloud, divide the cloud into ground and non-ground points, and detect the elements present in each of these clouds. The methods are validated in three case studies, each containing different railway tunnels. The results demonstrate that laser scanning technology, together with customized processing tools, can provide data for further structural operations with no requirement for either training in geomatics or high-performance computers for the data processing. Significant results are obtained for the developed classification methods: the classification of the tunnel elements returns a global F-Score of between 71 and 99% in a point-by-point comparison. With regard to the labelled rails classification, a global F-Score of 100% is achieved for the analyzed datasets, and between 56 and 73% for the point-by-point classification. Highlights • Automatic classification of large point clouds in tunnels' environments • Extraction of geometrical characteristics • Classification of lining, power lines, ground cantilever arms and rails • Point-by-point classification: global F-Score between 56 and 99% • Possible rails' classification (labels): global F-Score of 100% [ABSTRACT FROM AUTHOR]

Emerging vision-based frameworks have demonstrated the great potential to robustly perform volumetric measurements on point cloud models, which has several applications for site material management (e.g., during earthworks). However, prevalent vision-based frameworks to date involve human interventions to manually trim objects of interest from point cloud models, which would be time-consuming and labor-intensive. In addition, point cloud models for volumetric measurements are often incomplete and noisy. To address such challenges, we automatically detect and segment target objects in point cloud models via a deep learning-based approach and then map the semantic values onto point cloud models for 3D semantic segmentation. Once target objects are segmented, the associated volumes are quantified through the proposed vision-based computational process. For evaluation, case studies were performed on material piles in the real-world. The proposed method has the potential to enhance vision-based volumetric measurements, which supports systematic decision-making for material management in jobsites. • New vision-based volumetric measurements to reduce human intervention. • 3D point cloud segmentation based on 2D semantic segmentation. • The integration of volumetric measurements with vision-based object recognition. • Results show the potential for enhanced vision-based volumetric measurements. [ABSTRACT FROM AUTHOR]

Multi-planar building facades, pose significant challenges to existing point cloud segmentation techniques, as most have an underlying assumption that all structural elements appear within only a single plane, which is often untrue. For architecturally complex buildings, this prevents the automatic extraction of the load bearing structure for use in structural engineering applications. To address these deficits, the multi-planar algorithm is introduced as a means to differentiate between structural and non-structural elements for multi-planar facades. The algorithm considers statistically the position of groups of points with respect to their neighbours to identify and segregate the principal facade of architecturally ornate, multi-planar buildings. The viability and robustness of the algorithm is demonstrated through its application to 3 buildings in Dublin, Ireland, which when compared to independent survey data, resulted in 98% accuracy for single complex openings and an overall average accuracy of at least 91%. The superiority of the technique is demonstrated against four prominent segmentation techniques and through the application of the output files into an ANSYS finite element model. • Created distinctive 3D feature segmentation method for multi planar building façades • Practicable and scalable with respect to building types and architectural complexities • Developed evaluation technique to measure the accuracy of segmentation method • Utilized targeted triangulations on classified point cloud to calculate the opening ratio on the façade • Generated geometry input for computational solid model without intermediary software [ABSTRACT FROM AUTHOR]

This paper introduces the Gridded-RANSAC approach for automated point cloud façade segmentation applicable to highly ornamental masonry buildings with non-rectilinear openings. The aim is to segregate the principal façade. The paper also introduces a novel means to quantify the resulting openings through repetitive application of Delaunay triangulation. The Gridded-RANSAC method begins with a grid concept to organize the inherently unorganized data set followed by a non-traditional, local deployment of the Random Sample Consensus (RANSAC) method, then a more traditional global application of RANSAC, and finally a local neighbourhood similarity check. When applied to historic, 19th and 20th century façades in Dublin Ireland, three classes were defined: principal façade, recessed sections (normally openings) and protruded parts (mostly balconies, cornices and capitols). When compared to drawings from an independent licensed surveyor, the Gridded-RANSAC approach has a typical accuracy of better than 90% overall and in excess of 98% accuracy for complicated individual features. The superiority of the technique is demonstrated against four prominent segmentation techniques and through the application of the output files as geometry input for computational solid models without intermediary software. • Created distinctive 3D feature segmentation method for highly ornate building façades • No supplementary imagery or intensity information is utilized. • Developed evaluation method to measure the accuracy of segmentation technique • Generated geometry input for computational solid model without intermediary software [ABSTRACT FROM AUTHOR]