3D Reconstruction Using Image Sequences

Computer vision mainly studies how to use computers to implement the humanvisual function, namely, by using the two-dimensional projective images to achieve theperception, recognition and understanding of the three-dimensional objective worldseene. With the rapid development of computer technologies，people has more and moredemand for3D models.3D reconstruction based on image sequences has become one ofthe hot research topics in the field of computer vision because it has the advantages ofbeing low cost, simple and realistic. This thesis studied the correlative algorithms of3Dreconstruction using image sequences and systematically summarized the studies andachievements of the author in3D reconstruction based on two images,3Dreconstruction based on image sequences, dense reconstruction and the practicalapplication of3D reconstruction.(1) A3D reconstruction algorithm based on SIFT and corner detector is presentedin this paper. The basic idea of the proposed algorithm is: SIFT feature points areaccurate, but many of them are not what needed in reconstruction. Corners can betterexpress the basic shape of objects. By combining the SIFT feature points with the Harriscorners it is possible to obtain more vivid and detailed3D models. Compared with themethod without corners, reconstruction results of the proposed method is much closer tothe real object.(2) A novel3D reconstruction method using image sequences based on projectivedepth and the simplified Iterative Closest Point (ICP) is proposed. This paper firstpresents a corollary about the relation between the projective depth and3Dreconstruction points based on the ambiguity of reconstruction and then gives thedetailed proof. It proposes a new algorithm for3D construction from a sequence ofimages. In order to avoid accumulation of errors, this algorithm modifies thereconstruction results based on a simplified ICP algorithm. Compared with existingsequential algorithms, our algorithm can reduce the impact of accumulated errorsbecause the reconstruction process is equal to the pileup of the reconstruction resultsbased on two images. Moreover, compared with methods using Measurement MatrixFactorization, our algorithm does not require the feature points to be visible in all theimages. Therefore it is possible to reconstruct the whole shape of the object withsufficient corresponding points by the proposed algorithm.(3) For dense points reconstruction, we developed a new method that optimizes the3D points by using the average girth of the triangles obtained from triangulation. It iscomputationally expensive to eliminate false matches with the fundamental matrix whenthere are a large number of points. Our algorithm can effectively accelerate theoptimization speed of the reconstruction result without affecting the reconstruction quality. We also studied the dense matching algorithm based on region growing andused the sparse points obtained from the SIFT feature matching algorithm andoptimized by the fundamental matrix as the seed points of the dense matching algorithm.As a result, the accuracy of the dense matching has been greatly improved. Wecalculated the projective matrices of the image sequence relative to the same referencecoordinate system according to this paper’s algorithm, the projective matrices were usedas the input of the up to date multi-view stereo algorithm, and the influence of photoconsistency had been avoided.(4) To cope with the installation error of the camera, we developed an algorithm tocalculate the angle between the moving orientation of the vehicle and the drivewaybased on monocular vision using geometrical method. In order to facilitate comparison,we also achieved the angle detection between the moving orientation of the vehicle andthe driveway based on binocular vision according to the theory of3D reconstruction.The angle between the moving orientation of the vehicle and the driveway wascalculated with a hypothesis that there was an angle between the optical axis of thecamera and the ground, the actual deviation angle was also computed. Images capturedfrom the real scene validated the accuracy of our method. The prominent advantage ofthe proposed algorithm is that we do not need to measure the camera height above theground and the pitch angle and also the deflection angle, all the value can be obtainedfrom calculation, which can further decrease the measure error.