Study On 3D Model Retrieval Technology Based On 3D Reconstruction

The 3D model has more information and better straightforward than 2D image. With the development of the 3D model retrieval technology, people find that it is possible to solve some problem can't be solved in plane use of 3D model retrieval technology combine with 3D reconstruction, the 3D model retrieval technology based on 3D reconstruction has been emphasis and development. How provide efficient the 3D model retrieval technology based on 3D reconstruction algorithm become a forefront and challenging task.The 3D model retrieval technology based on 3D reconstruction is the computer technology which extracting 3D model of target objects from a video scene and then classifying and retrieving in accordance with the features of the models. Work process is divided into three sections, the first step, automatically selecting feature points, the second step, reconstructing points cloud model from feature points, the third step, classifying and retrieving in accordance with the different features of the models. In extracting feature points section, normal methods are obvious defect that these have to need user's participation in the processing of feature points selection, so it can't achieve the effect of feature automatic selection. In 3D reconstruction section, usually use Structure From Motion algorithm, but traditional SFM algorithm strictly rely on accurately correspondence between feature points, however, getting the accurately correspondence is difficult. In 3D model retrieval section, the reconstruct result is 3D cloud model, compute similarity of 3D cloud is very difficult, for the above problem, mainly including:(1) In selecting feature points section, I use the marking feature points algorithm based on Horn-Schunck optical flow algorithm and the canny operator to automatically get feature points. First, extract target object edge from image sequence, Then use Horn-Schunck optical flow algorithm compute optical flow field of all images. Finally select the points that flow vector is maximum on edge as feature points.(2) In 3D reconstruction section, I use SFM algorithm based on EM algorithm, the characteristic of the algorithm use a "virtual measurement" instead of accuracy correspondence. The "virtual measurement" integrate all of the measurements of 3D points and feature points and then use priori knowledge combined with Expectation Maximization Algorithm to improve "virtual measurement". Finally convergence "virtual measurement" to measurement and then get the 3D cloud model of target object.(3) In 3D model retrieval section, I use Crust algorithm make surface grid, then I use a new compute Extended Gaussian Image method with kernel density estimation to complete 3D model match and retrieval. Experiments show that the implementation of the algorithm is high efficiency and classification is accuracy.