Project Details
Abstract
3D data acquisition and reconstruction is a technique for obtaining geometric digital
models of objects or scenes from sensed data. With the progress of stereo vision algorithms
and availability of low-cost cameras, the recent decade has seen remarkable growth in the use
of binocular stereo vision system for 3D reconstruction. In our previous project, camera
calibration algorithms are developed for extrinsic and intrinsic parameters of a PTZ stereo
vision system. Noticeably, cameras with pan-tilt-zoom maneuverability allow for greater
freedom in data collection. However, commercially available PTZ cameras, such as Sony
D-70, has only a limited pixels of 640 by 480 and their repeatability depends on moving
direction with significant individual difference due to backlash, rendering them unsuitable for
precise applications. We thus plan to build a customized binocular stereo vision system in this
project, which will be equipped with higher resolution cameras of 1920 by 1020 pixels and
driven by servo motors.
In addition to extrinsic and intrinsic parameters of a stereo vision system, corresponding
locations on a pair of images of the same scene are required to triangulate three dimensional
coordinates. When there is significant deviation in view angles or the scene is lack of image
features, the search of reliable corresponding location is vulnerable to failure and is an
important problem pending solutions.
Originally bundled with Xbox 360 to play games without holding a controller, the
Microsoft Kinect is a low-cost range-imaging sensor capable of delivering 3D point clouds
and images in real-time. While spatial and depth resolution may limit its use in high-accuracy
systems, Kinect has attracted interest from developers wishing to exploit the system's 3D
imaging capabilities to industrial and scientific applications.
In this project, we will integrate Kinect with the home-made binocular stereo vision
system by making use of the depth information provided by Kinect to find corresponding
locations on pairs of images of the same scene. The uncertainty model will be used for the
initialization of potential image regions within patches commonly seen by the cameras.
The project will be conducted in two years aiming at enhancing accuracy and reliability
of the stereo vision system by integrating the home-made PTZ binocular system with Kinect.
For the first year, we will build the system, investigate accuracy and resolution of Kinect, and
develop a registration algorithm between the binocular subsystem and Kinect. These efforts
will result in the construction of error distribution model of Kinect and put reconstructed 3D
feature points in the same coordinate system.
In the second year, we will research into the issues of stereo matching for 3D
reconstruction. The project will focus on providing reliable algorithms to search for
corresponding locations on a pair of images from the binocular system. The registration
algorithm and error distribution model of Kinect developed in the first year will be used to
facilitate the integration of data points collected from the maneuver of the PTZ cameras.
Project IDs
Project ID:PB10207-1806
External Project ID:NSC102-2221-E182-073
External Project ID:NSC102-2221-E182-073
Status | Finished |
---|---|
Effective start/end date | 01/08/13 → 31/07/14 |
Keywords
- Kinect range sensor
- Active sensor calibration
- 3D acquisition
- Error
Fingerprint
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.