Surveyors, designers, construction teams and engineers all use point clouds. The possibilities point clouds offer only grow as the technology advances. Anyone working with 3D models can benefit from the use of point cloud datasets, and understanding best practice empowers us to go even further.
A detailed, accurate 3D model relies on multiple scans of the same area or object that must be aligned — a process called ‘registration’. Registration is a process that has brought challenges, as well as solutions of varying quality. This article presents an overview of some of the questions posed by point cloud creation and suggests a recent solution to some persistent problems.
Scanners make line-of-sight measurements. The amount of coverage necessary for a detailed 3D model usually requires multiple scans. These scans must be stitched together to create a complete picture of the area. How to align the scans with each other poses the major question of point cloud creation.
In some cases GNSS (Global Navigation Satellite System) technology such as GPS can be used to align scans using positional data of the scanners. But any given number of situations make registration by GPS impossible, for example scanning an area indoors, in tight urban environments, or under thick tree cover.
In these cases, computer software is able to analyse point cloud data to find common features at which to align scans. Two methods of registration have been used in such scenarios, one employing targets and the other solely relying on software.
Targets are point-reducible artificial objects that are easily recognised by the registration software for alignment. At least three targets must be placed for each pair of scans. The location of targets must be planned out, and then the targets must be manually placed in the scanned environment. Different areas present a variety of impediments to target placement, requiring the employment of trained personnel in a time-consuming process.
In order to align adjacent scans without targets, features already present in the scanned environment are used. This can save time in the field, but has added time in the office — often in excess of the time saved.
Traditional targetless registration software simply needs a lot of time to process. On a per scan basis, it takes far longer for the software to examine point cloud data for common features than identifying artificial targets. The sheer amount of data involved leads to more time processing, and each pair of scans must be assessed for three alignments: rotational, vertical and horizontal.
Standard targetless registration programs also require human operators to manually input and cross-check each alignment, because distortions or inaccuracies in a single scan can impact the accuracy of the entire model. The necessity of manual intervention means more labour on top of more time.
Even the amount of time saved in the field can be disappointing. Traditional targetless registration requires up to 60% overlap to ensure accurate alignment, meaning more scans are required to be made than when targets are employed.
Anyone creating point cloud datasets faced the conundrum of whether to use targets or targetless registration software, as both make great demands of time and labour.
The Future of Point Clouds
The solution lies in making targetless point cloud registration viable by reducing processing time and removing the necessity of manual supervision. Fortunately, this is exactly what has been delivered by new registration processes that utilise greater automation and a novel use of vector analysis.
Each scan is converted into a ‘vector sphere’, where the points are represented as vectors and an entire point cloud is collapsed into a single point. The density and directional characteristics of each vector sphere allows it to retain its unique identity, which can be automatically compared to others for common features.
Rotational analysis is performed by placing adjacent scans inside one another. This spherical analysis then allows for separate and rapid 2D point density alignments on the horizontal and vertical axes. This multi-stage automated process drastically increases the speed at which scans are aligned, and enables analysis at each stage to be more thorough. The software's improved ability to detect stationary features results in less required overlap between scans — in some cases the amount of overlap needed is reduced to as little as 30%.
Now that manual operators are no longer required to supervise each step of the registration, scans can be queued to process in the background while other work is tended to, to be returned to when finished. Scans can even be left to process overnight, allowing projects that formerly would have taken a week to be finished in a single day.
The creation of quality models with point cloud technology requires aligning multiple scans of the same space. The question posed by registration was whether to use targets or go targetless to save time in the field. Until recently, targetless point cloud processing required too much time and manual supervision to be preferable to placing targets in most circumstances.
A dramatically faster targetless registration process is now available that requires little to no oversight. Multi-stage vector analysis in combination with increased automation have provided the solution to the problem of targetless registration. This has created a clear best practice for point cloud creation.