Research On Multi View Calibration And 3D Feature Extraction

Multi view calibration and 3D feature extraction are the two important steps in the process of 3D reconstruction. Multi view calibration is a process which collects multi view images of the 3D scene(such as chessboard calibration plates) by a camera, and finds the relationships between the points in the real scene and their corresponding points in the image base on the camera image model, with which to get the internal parameters and external parameters of the camera. The accuracy of the camera parameters directly determine the accuracy of 3D reconstruction results. Feature extraction is a process to extract discernible characteristics from the images on the basis of computer vision algorithms. The characteristics, whose quantity and quality will determine the success of later stage process of 3D reconstruction, we extracted will be used for feature matching and deepness information confirm in later 3D reconstruction stage.The most popular algorithms in multi view calibration are Tsai's Two Steps Calibration Method and Zhengyou Zhang' Calibration. However, these algorithms are complicated in extracting calibration plate corners and corners sorting. Sometimes, corners sorting is artificial.Two most classical algorithms in feature detection are Harris algorithm and SIFT(Scale-invariant feature transform) algorithm. Harris algorithm takes advantage in speed while extracting the Corners and has an improved algorithm which can extract the Corners in uniform distribution, but the key points we got are not suitable for feature matching. Meanwhile, the key points extracted by SIFT algorithm are appropriate for feature matching but not in uniform distribution.The aim of this research is to facilitate the process of extracting calibration plate corners in camera calibration part, promote the speed in calibrating camera and extract key points in a better uniform distribution. The main contributions of this research are as follows:a.) Aiming at the application of camera distortion is little and the calibration board is chessboard, this research proposed an approach for camera calibration on the basis of Zhengyou Zhang' Calibration. The new approach mainly improve the way we extract the calibration plate corners, thus we can pick up the edge of the image through canny algorithm and calculate the intersection of horizontal and vertical lines to locate and sort the corners. The experiments parts prove that the improved camera calibration method we propose has easier compute, angular point automatic sorting. and high reliability and faster run times When compared to the Zhengyou Zhang' Calibration approach, the new approach has less runtime and its results are pretty reliable.b.) An improved feature extraction algorithm is realized by combining the advantages of both improved Harris algorithm and SIFT algorithm. Firstly, we get the locations of the key points with the help of improved Harris algorithm. Then, according to idea of SIFT algorithm we generate feature descriptor for each key points. Finally, we use the improved method to extract key points from two images and match them, the experiment results show that the key points which extracted by our method is more conductive in feature matching when compared with SIFT algorithm.c.) We study the two-view 3D reconstruction. When we have done the steps before, we can acquire the internal parameter matrix of camera and key points as well as the pairs of key points. We deep into solving fundamental matrix, essential matrix, rotation matrix, translation vector and projection matrix, and calculating the three-dimensional coordinates of the objects which are to be reconstructed base on its projection matrix and consequently drawing the 3D point cloud of the objects which realize the prototype of the 3D reconstruction system.