The Research Of Key Technology On Image-based Reconstruction Of Workpiece Surface

Spray painting is an important process in the manufacture of many durable products,such as automobiles, furniture and appliances. Many works focus on this area when CADmodel of work piece is kown. In this paper, we focus on the surface reconstruction of workpiece based on images. The key technology of surface reconstruction is camera calibrationand stereo matching algorithm. Although, many global stereo matching methods, such asgraph cuts, belief propagation and object stereo, have achieved excellent performance, theyare troubled by computational complexity or varied parameters. Belief propagation stereomatching requires many iterations to ensure convergence of the message values. Thematching result is affected by parameters directly. However, proper parameters vary with datawhich is commonly intensity differences. Many excellent stereo matching methods are limitedto vision applications due to their computational complexity. Some algorithms take a longtime (even over20minutes) to obtain a disparity map on a pair of reasonable size images.In order to achive the good performance in the area of engineering, this work address thekey problem of computer vision, such as camera calibration and stereo matching. Takingaccuracy and effient of algorithm into consideration, we present a corner detection methodbased on surport vector machine and two stereo matching algorithms as follows:(1). Camera calibration is key process in stereo vision. There exist many stereocalibration methods at present. However, a common problem is the corner detection. Wedetect corner of calibration pattern with suport vector classification.(2). Stereo matching is critical technology in vision measurement. MRF models areestablished to do with stereo problem. A parallel multi-scale belief propagation algorithm isused for MRF energy minimization and generating disparity map. Parallel algorithm isimplemented based on traditional sequential algorithm with CUDA technology. In energyfunction, data term is conjugated with Gradient and intensity of images, smooth term ismeasured with the absolute difference of disparities between two adjacent pixels. Withstandard Middlebury stereo data sets as input, experiments show that the proposed algorithmhas good real-time performance；Running time is much less than the traditional sequentialalgorithm and the generated disparity map is excellent.(3). We present an efficient stereo matching algorithm. Given two grayscale stereoimages, each pixel of them is encoded as a multidimensional point for stereo matching problem. These multidimensional points in the right image are used to build kd-trees. Thenearest neighbor searching is performed on the multidimensional space and an initial disparitymap is generated. Furthermore, a segment-based refinement approach is applied to generate amore accurate disparity map. Experimental results show that our algorithm is efficient andquality.We reconstruct a workpiece surface from a pair of rectified images with proposedalgorithms. The experiments illustrate that the method is feasible.