Acquiring 3D models of objects via a robotic stereo head David Virasinghe Department of Computer Science University of Adelaide Supervisors: Mike Brooks and Anton van den Hengel
Stereo Vision Stereo is an important concept of human vision. Yorick 8-11R cameras are mounted to a movable platform, which mimics degrees of freedom of a human head. Each camera can be moved along four axes.
The 3D reconstruction process Comprises of three main stages: Camera calibration Image matching Reconstruction
Camera Calibration Camera calibration involves computation of internal and external properties of the camera. It requires an image of an object with some known 3D measurements. We use a calibration grid.
Extracting Image Coordinates To extract 2D image coordinates of the corners of the squares in the calibration grid we use the following process: 1. Apply edge detection to the image. 2. Perform line fitting. 3. Find where lines intersect.
Edge detection
Line fitting
Junction detection
Tsai Camera Calibration The model has 11 parameters. Five internal parameters: f – focal length of the camera, – radial distortion coefficient, C X, C Y – the principle point, S – scale factor. and six external parameters: R X, R Y, R Z – rotational angles, T X, T Y, T Z – translation components.
The Projection Matrix Encapsulates the orientation and properties for the camera. A projection matrix can be decomposed as A is a matrix describing the camera’s internal properties.
Image Matching Involves finding corresponding points in left and right images that depict same points in the scene. A program called Image-Matching was used to perform matching. Implements a robust technique for image matching by exploiting the only available geometric constraint, the epipolar constraint.
Image Matching The algorithm consists of three steps: 1. Establish initial correspondences 2. Estimate robustly the epipolar geometry 3. Stereo matching
Image Matching 1. Establish initial correspondences A corner detector is first applied to each image to extract high curvature points. Then a classical correlation technique is then used to establish matching candidates between the two images. Matching ambiguities are then resolved using a relaxation technique.
After Corner Detection
After Correlation
After Relaxation
Image Matching 2. Estimate robustly the epipolar geometry The fundamental matrix is recovered. 3. Stereo matching Establish a new set of correspondences using a correlation based approach that takes into account the recovered epipolar geometry.
After stereo matching
Image-Matching is unpredictable
Reconstruction Determine depth of points by using triangulation. Triangulation results in 3D cloud of points being determined; however to visualize the structure of the 3D object easily points need to be connected. We use Delaunay triangulation.
Point Clouds
Delaunay Triangulation
YorickIn3D In this project a GUI has been created that enables a user to perform the 3D reconstruction process.
Conclusion We have successfully created an easy-to-use program that allows the 3D reconstruction process to be performed and creates accurate reconstructions. We have discovered a process that accurately extracts image coordinates used in calibration.