Soemantoro_Manchester.pdf (1.05 MB)
Download file

An AI automated self-organising, feature imitating approach for point cloud data reduction

Download (1.05 MB)
conference contribution
posted on 14.05.2021, 10:53 by Raska Soemantoro, Hao Yue, Muhammad Omer, Lee Margetts
This paper presents a novel AI based self-organising data reduction technique which combines feature detection and topological learning to reduce the memory footprint of point cloud data in an adaptive way, reducing point density in featureless parts of the point cloud whilst maintaining sufficient points to preserve details of interest to the engineer. As a case study, this is applied to a 3D LiDAR scan of a masonry bridge with localised details such as anchors, cracks and patches of vegetation. The process comprises 4 stages: For each point, the distance to its neighbouring point is calculated using a nearestneighbours search. Afterwards, 3D point cloud data is projected onto a 2D plane using principal component analysis. Next, the variables are mapped into the feature space where they are grouped using a clustering algorithm. Clusters correspond to specific physical features within the point cloud. Points within these clusters are assigned a feature weight that dictates their level of bias compared to points that do not belong to an engineering feature. Finally, a weighted self-organising map (SOM) algorithm is applied to learn the topology of the original point cloud, with a specific focus on feature points. This algorithm is a type of artificial neural network that uses competitive learning to generate an output map iteratively adjusted to represent the input. An advantage of the SOM is that the output map can be used to define the position of vertices in a finite element mesh. The results show that the standard SOM was able to reduce the data size by over 90%. However, such a reduction in data size risks loss of smaller engineering features. Upon application of a feature weight, the method was able to produce the same data reduction with up to 35% decrease in quantization error compared to a standard SOM. The research is of timely importance. It reduces the computational cost of point cloud visualisation and improves the reverse engineering process, enabling conversion of point clouds into CFD and FE meshes. The work will benefit academics, researchers and engineers in developing their models.

History

School

  • Science

Department

  • Mathematical Sciences