Scale-invariant Feature Points In The Three-dimensional Measurement

Many applications of 3D object recognition, such as augmented reality or robotic manipulation, require an accurate solution for the 3D pose of the recognized objects. This is best accomplished by building a metrically accurate 3D model of the object and all its feature locations. In this thesis, we describe a system for constructing 3D metric models from multiple images taken with an uncalibrated handheld camera. In order to reconstruct more accurate 3D metric models, we introduce the SIFT (Scale Invariant Feature Transform) feature instead of the traditional Harris corner. This application of SIFT feature in 3D modeling not only simplifies the approach of reconstruction, but also gets better 3D models and camera parameters.The works completed in this paper are as follows:1 We Use SIFT feature to reconstruct 3D models instead of Harris corner, and construct a new framework of 3D reconstruction.2 All kinds of matching algorithm of SIFT feature are researched and compared, so that the most accurate and efficient matching algorithm is used for matching the image features.3 The minimum spanning tree of the net optimization is introduced and two new algorithm are advanced to compute minimum spanning tree. Then the models could be reconstructed even if the input image set is unordered.4 Weighed reprojection error is used for making cost function when optimize the model and camera parameters using bundle adjustment. And the incremental model construction and bilinear alternation strategy are provided to improve the efficiency of bundle adjustment.5 The principle of image 3D measurement is analyzed. The reconstructed model is used for image 3D measurement. The result of measurement and augmented reality are drawn by OpenGL.Experimental results show that 3D reconstruction based on SIFT feature is feasible. And the reconstructed models could be used for image 3D measurement or model based tracking.