This research explores several novel approaches to improve visualization and segmentation
of point clouds acquired with 3D laser scanning. 3D laser scanning is used
in a wide variety of applications including surveying and mapping, transportation asset
management, facilities management, building information modeling, crime scene investigations,
cultural heritage and geologic instigations. Although most 3D laser scanners can
co-acquire photographs during the scanning process, this information is mostly being used
to color the point cloud for visualization purposes, rather than utilizing it as a quantitative
source of information to enrich point cloud processing.
Two key limitations of most digital cameras are their restrictions in capturing the
full luminance range of the scene and a full field of view. These limitations result in overexposed/
underexposed pixels and lead to inconsistencies between adjacent photographs
when linking them with scans. High Dynamic Range (HDR) photography is a promising
solution to these problems.
HDR combines multiple images with varying exposure levels to cover the dynamic
range in the scene. However, HDR presents unique challenges of its own. Inconsistencies
in object positions between images can result in ghosting effects in the composite image. Additionally, it can still be difficult to properly map the color tones to ensure details are
properly displayed in regions of extreme lighting.
This research presents novel methods to address these challenges so that point cloud
data can be segmented more effectively with the aid of the HDR photographs. First,
a new method is developed to detect these moving objects in digital images without
requiring a camera response function, supervision of objects in the scene, or selection
of a reference image. The proposed approach yields effective results with a Matthews
correlation coefficient of 0.82 on HDR photographs from both indoor and outdoor scenes.
Second, a new tone mapping operator is invented with the intent of preserving
detail in both dark and bright areas of the images. It utilizes geostatistical variography to
reproduce optimal color values. The approach proved to be highly versatile and worked
consistently well in producing realistic images for a variety of scene types.
Third, this research contains two novel approaches for point cloud segmentation
utilizing computer vision algorithms. The first approach applies a superpixel algorithm
for clustering the colorimetric data. Then, each superpixel is classified as either "ground",
"wall", or "vegetation" based on its corresponding normal vector. The proposed method
effectively generates accurate segmentation results when compared to ground truth. The
second approach uses both colorimetric and geometric data to segment lidar point clouds.
The fixed angular acquisition pattern of the scan permits creation of 2D structured
panoramic image maps (PIMPs) representing various subsets of the data including normal,
intensity, range, and RGB color. An image segmentation algorithm is applied to
each PIMP, and then the union of the segmented PIMPs is mapped back to the 3D point
cloud. Compared with RANSAC, the proposed approach has significantly higher perceptual
segmentation and efficiency. The use of HDR photographs and corrected laser
intensity significantly improves the segmentation by the average MCC value of 0.87.