|
LiDAR is a remote sensing method used to produce precise point clouds with millions of geo-spatially located 3D
data points. The challenge comes when trying to accurately and efficiently segment and classify objects, especially
in instances of occlusion and where objects are in close local proximity. The goal of this paper is to propose a more
accurate and efficient way of performing segmentation and extracting features of objects in point clouds. Normal
Octree Region Merging (NORM) is a segmentation technique based on surface normal similarities, and it subdivides
the object points into clusters. The idea behind the technique of surface normal calculation is that for a given
neighborhood around each point, the normal of a plane which best fits that set of points can be considered to be the
surface normal at that particular point. Next, an octree-based segmentation approach is applied by dividing the entire
scene into eight bins, 2 x 2 x 2 in the X, Y, and Z direction. Then for each of these bins, the variance of all the
elevation angles corresponding to the surface normal within that bin is calculated and if the elevation angle falls
below a certain threshold, the bin is divided into eight more bins. This process is repeated until the entire scene
consists of different sized bins, all containing surface normals with elevation variances below a given threshold.
However, the octree-based segmentation process produces obvious over segmentation of most of the objects. In
order to correct for this over segmentation, a region merging approach is applied. This region merging approach
works much like the automatic seeded region growing technique, which is an already well known technique, with
the exception that instead of using height to measure similarity, a histogram signature is used. Each cluster generated
from the previous NORM segmentation technique is then run through a Shape-based Eigen Local Feature (SELF)
algorithm, where the focus is on calculating normalized histograms to describe the local shape and curvature of the
points as well as using Principal Component Analysis (PCA) in order to determine meaningful relationships between
points, primarily using eigenvalues. These extracted features are then applied as the input to a cascade of classifiers,
where an object is classified and results are compared to datasets which have been manually ground-truthed. The
NORM segmentation technique was implemented on two datasets and outperformed other state of the art
algorithms, such as automatic region growing and strip histogram grid methods. The proposed SELF method is
performed on each of the segmented clusters and looks to combine previous research by concentrating on extracting
the global features of each cluster, while simultaneously collecting information about each point on a local level.
The combination of the two novel algorithms, NORM and SELF, prove their effectiveness in classifying five classes
of objects in the scenes. Future work involves improvement of the feature vector to help distinguish between
subclasses such as vehicles of various types, buildings of different roof structures, and vegetations.
|
