3D Reconstruction Of Free-form Surfaces Based On Binocular Vision

The 3D reconstruction of free-form surfaces has been widely used in the fields of reverse engineering and virtual reality. The stereo vision is one of the important methods to reconstruct 3D models of free-form surfaces. The camera calibration and the correspondences between images of stereo system have been the difficulty and the focus for decades in the research domain of stereo vision. In this paper, the experimental system is built to reconstruct the model of free-form surfaces, and the camera calibration and image matching of stereo system are deeply researched. Genetic algorithms (GA) are adopted to calibrate the cameras of stereo vision system, and the new method of encoding with the adaptive adjudging of camera parameter search interval is presented to improve the standard GA. The new approach not only provides enough loose bounds for camera parameters, but also ensures the search resolution with the length of the chromosome bit string unchanged. Therefore, the improved GA is effective to solve the nonlinear and high dimensional function optimization problem and provides new clues to enhance the computational performance of standard GA on solving the nonlinear and intricate function optimization. The implicit camera system model is made with BP neural network simulating the relationship between the 3D objects and its images in the stereo system, which avoiding the system errors caused by unperfect mathematic relation. With the accurate NC movable workbench to acquire the dense calibration data and the regularization method to be applied during the training process of BP neural network, more precise calibration is got and the neural network's generalization capacity is enhanced. The digital dots and stripes are projected respectively onto the surface of the object in order to acquire discriminable features from the surface, which is lack of image features. By this way, the accuracy of image matching is increased effectively. The area-partitioned matching approach based on projective transformation is presented. On the left and right images respectively, after the shapes of areas formed by big dots are transformed into standard shapes, all the small dots within the standard area are matched automatically with their corresponding partners within the other standard area. The problem of matching dots is solved successfully. The new algorithm to rectify stereo rigs is presented after investigating the epipolar geometry of stereo vision system in this paper. The algorithm calculates epipole positions and adjusting coefficients for epipolar lines, and sets up the formulation of projective rectification. After the recification transformations on the left and right image planes are done, the pairs of conjugate epipolar lines become collinear and parallel to the horizontal image axis. Therefore, when computing stereo correspondences between the left and right images is being conducted, the search is done only along the horizontal lines of the rectified images, which is highly advantageous to image matching. After the 3D model of the object is established, the textures from the real images are mapped rapidly to the surface of the 3D model according to the relationship simulated by BP neural network between the object and its images. Then the 3D models have better visual effect. The results of experiments on 3D reconstruction indicate that the process of reconstructing 3D models of free-form surfaces is implemented effectively using two images acquired from the calibrated stereo vision system by means of projecting auxiliary patterns and presenting corresponding algorithms of image matching.