Location:
CGVG >
CVG >
Research >
Projects
Research - Projects
[
overview |
projects |
software & demos |
publications |
seminars
]
Outdoor 3D Scanner Using Off-the-shelf Digital Camera
Principle Investigator: Dr. K.-K.Y.Wong
The generation of realistic 3D models of real world objects is of great interest
in many fields and has many practical applications. For instance, such 3D models
can be used in reverse engineering, model-based tracking system, video games,
virtual reality, movie making and internet showroom. Traditionally, in computer
graphics, such 3D models are constructed using specialized design software in a
polygon-by-polygon fashion. Such an approach is very time consuming and the quality
of the output model depends very much on the skill of the operator. The introduction
of laser scan systems allows 3D objects to be "scanned" into the computer directly,
and makes the 3D modeling process faster and easier. In spite of that, such systems
are very expensive and require careful calibration before use. Besides, they cannot
cope with specular surfaces nor surfaces with low reflectance, and can only handle
objects of limited size. This project aims at providing a cost-efficient solution to
the above problem through the use of computer vision techniques. Vision-based systems
do not require direct mechanical measurements on the objects, and can also handle
outdoor objects with various size and reflectance. In this project, the following
2 problems will be addressed:
1. Motion Estimation - The estimation of the relative motion between the object
and the camera using features extracted from the images. Most of the existing
vision-based techniques use information from the viewpoint independent corner features
only, whereas algorithms exploiting the viewpoint dependent silhouettes have only
been developed in special cases of constrained motion. The key issue here is to
combine the information provided by those 2 complementary sources (i.e. corner
features and silhouettes) to provide a robust solution to the motion estimation
problem with less constraints placed on the motion itself; and
2. Model Reconstruction - The construction of a 3D model from the image sequence
using the estimated/known camera motion and the extracted image features. Existing
systems usually produce 3D models composed of extremely large number of triangles.
Such models suffer from the geometric aliasing problem and cannot be rendered
efficiently. The key issue here is to achieve a smooth, compact and deformable
representation for the object model, which can be further refined using other image
features available.
The results of this research project will provide a robust solution to the structure
and motion problem, and thus represent significant original contributions to the
study of 3D modeling from 2D images. In addition to original research output, this
project also provides a complete framework upon which a practical software for 3D
model generation from 2D images can be built. With the decrease in price of digital
cameras, such a software system would provide a cheap solution for home users, as well
as professionals, to create their own digital archive of 3D models.
