Many important applications of computer vision are found in manufacturing and defense industries. Such applications include inspection, measurements, robotic assembly and autonomous vehicle guidance. The area of three-dimensional vision presents great challenges and complex problems. It has attracted considerable research efforts in recent years. Although significant progress has been made, the transfer of fundamental ideas to concrete applications has not been successful at the practical implementation stage. This is due to the complexities involved in the process of emulating the human vision capability in a computer system.
In this dissertation, a stereo-based multi-camera system for complete 3-D information extraction and object surface reconstruction in a robot workspace is developed. The system consists of N number of cameras arranged in an N/2 number of periodic stereo pair structure. The cameras sense the working area of a robot in the form of N images which are processed to obtain the 3-D data in the robot's environment. The extracted information is provided to a surface reconstruction algorithm for object description. The image reconstruction phase is performed in the scene domain on the combined data of adjacent camera pairs.
The system advances the 3-D vision capability of industrial robots. Specific contributions include a camera calibration procedure that determines the system's parameters directly from the output digital image using only three known world points. This procedure uses a pinhole camera model and assumes a linear image transformation process between the image plane of the camera and the output digital image. The parameters of the system are computed by solving a set of linear equations.
The number of cameras for entire coverage of the robot workspace is determined by defining the common area of a camera. This is essential because the 3-D information of any portion of the object that does not appear in both cameras can not be recovered. The working environment of the robot is described based on the knowledge of the common area and the number of camera pairs used.
The shifting property of the Fourier transform is utilized for disparity estimation. The result reduces the cost and increases the accuracy of the matching procedure. This is true because the matching process is directly proportional to the search limits in the other image. In addition, the matching is only performed in the common area of a camera pair which is determined from the geometry of the set-up.
Matching of stereo image pairs is also addressed. An image matching technique that makes use of the estimated disparity is developed. The method combines the advantages of both the area-based and feature-based approaches. The feature-based matching results guide a local window operation that identifies correct matches within a neighborhood. The local matching measure is based on the smoothness in disparity values in neighboring pixels on the surface of the object.
Surface reconstruction for complete object representation is performed in the scene domain. In this regard, the Lagrangian polynomial is employed locally to approximate the object points based on the available data. Here, the known depth points retain their original values. This initialization process improves the convergence rate and the performance of the quadratic variations technique.
Finally, an algorithm for corner detection on digital curves was developed and employed for object representation to test the efficiency and reliability of the developed system for 3-D measurements. First, the thinned image is scanned to assign candidate corners. Then, false corners are eliminated by operating locally at the initial assignments. The final result is a list of corners, each identified by its position, and the number and direction of the edges intersecting at its center. The algorithm is capable of finding two and multiple-side corners and is suitable for parallel implementation.