Use of structured light for three-dimensional reconstruction
暂无分享,去创建一个
The adoption of structured light illumination has been proven an effective and accurate visual means for 3D reconstruction. The system consists of a projector that illuminates controlled pattern or patterns to the target object, and a camera grabbing image or images of the illuminated object. Once correspondences between positions on the projector's pattern panel and positions on the camera's image plane are established, simple triangulation over light rays from the projector and the corresponding light rays to the camera would recover 3D information about the target object. Key issues involved in the approach include (1) Calibration: how the projector and camera can be calibrated so that metric measures about the object can be extracted from the image data; (2) Image Feature Extraction: what image features to use and how to extract them accurately from the image data; and (3) Coding : how the illuminated pattern can be designed so that each position of it embeds in the pattern a unique code which is preserved on the image side, so that position correspondences between the projector's pattern panel and the camera's image plane can be easily established. Each of the issues can affect the accuracy of the system. This thesis aims at providing improved solutions to each of these issues.
An accurate and convenient system for calibrating projector-camera system is presented. A consumer-grade LCD panel is used in place of the traditional printed pattern as the calibration plane. While patterns shown on the panel are used for camera calibration, when we turn the panel off (with its pose kept in space), patterns illuminated by the projector and reflected from it and captured by the camera can be used for the calibration of the projector. This way, patterns for calibrating the camera will not overlap with patterns for calibrating the projector, avoiding confusion in the image data. In addition, even a household-quality LCD panel has industrial-grade planarity. Experiments show that a setup as affordable as this can still have the system parameters calibrated in far less images with much higher accuracy.
On the image features for establishing correspondences between the projector's pattern panel and the camera's image plane, we propose to use rhombic pattern with binary (i.e., black and white) or colored elements as the projected pattern, and the grid-points between neighboring rhombic elements but not the centroids of the pattern elements themselves as the feature points. We show that, grid-point in the pattern owns the two-fold rotation symmetry, or so-called cmm symmetry, which is largely preserved on the image side after perspective projection and imaging. We propose a grid-point detector that exploits such a symmetry. By avoiding the direct use of raw image intensity, the detector is less sensitive to image noise and surface texture. Comparison with traditional operators shows its promising robustness and accuracy.
On the coding issue, we investigate a number of options. Coding can be established over time, and one widely used scheme in this direction is the adoption of Gray code in a series of binary patterns that are projected at different instants. We describe how the traditional Gray code patterns, if augmented by the use of strip shirting, can have the resolution of 3D reconstruction enhanced. The whole system setup of such a system is affordable, nonetheless experiments show that high accuracy can be achieved based on it. The disadvantage of such a system is mainly that multiple image captures are necessary for the operation.
Finally, we explore how coding in structured light mechanism can be made even unnecessary. We adopt the above concept of recovering surface orientation from grid-lines, and show that by the use of a regular pattern, like a binary pattern with rhombic pattern elements, an orientation map about the imaged object can be recovered. Specifically, we show that here the correspondences over grid-lines between the projector's pattern panel and the camera's image plane can be approximated with a linear mapping and in turn boost the accuracy of surface normals calculation. We go on and show that, as long as no less than one reference point on the imaged object is available where absolute 3D is known, the above orientation map can even be converted to an absolute depth map by a simple integration process. (Abstract shortened by UMI.)