LiDAR (Light Detection and Ranging), a laser scanning method in the area of remote sensing technology, is the most accurate tool in creating point cloud which maps discrete changes at a very high resolution and provides precise information on the objects surveyed by sequentially measuring with high point density as well as the ability to resample areas in a quick and efficient way. LiDAR has wide range of applications in 3D modeling for various fields including 3D Object detection in autonomous driving and creating detailed tomographic maps, 3D printing, 3D gaming and various virtual reality (VR). It, however, remains a challenging area as the raw point cloud is not the final product and additional processing like classification, filtering, and modelling is required in order to retrieve information from the point cloud. Although the conventional tools like LAS can be used to assign a classifier to each LiDAR point, they are subject to errors as one LiDAR point can be classified as a 2D and 3D object at the same time.

In this research project, the main objective is to come up with an algorithm to reduce the Classification Rrrors (Misclassification Rate) in 3D Classification of 9-million LiDAR Point Cloud. Considering the local geometric characteristics of each LiDAR point, a classifier is assigned to each LiDAR point to determine its shape based on the laser plus reflection. To address the misclassification, we utilized a Dynamical Principal Component Analysis which approximates the spatial distribution of points in the neighborhood by an ellipsoid. We then performed a K-d Tree Data Structure to find the K-Nearest Neighbor (KNN) classifiers in the data. To be able to update the radius of search (dynamic feature), we defined three-dimensionality descriptors: linear, planar, and volumetric to introduce Local Point Density indicators. Using parallel computing with 12 cores via MatLab (took six days to run the analysis for 9-million LiDAR point), we illustrated that performing the PCA with a dynamic radius search results in a more consistent 3D point cloud recognition compared to a PCA with a fixed radius search or analysis with conventional LAS tool.