Summary: | 碩士 === 國立臺北科技大學 === 自動化科技研究所 === 103 === 3D modeling of real object is an expend topic. This topic have a lot of applications in medical imaging and robot vision. Therefore, the model dataset became bigger, so we need to remove the redundant points or simplification the model. Kyo ̈stila ̈etc. proposed to use Mahalanobis distance as a criterion to remove the redundant points, but the selection of the covariance of the camera model and the Mahalanobis distance threshold effects the correctness of the model. Thus we propose to use radial basis function network (Radial Basis Function Networks, RBFN) to learn the relationship between the two input variables (covariance and threshold) and an output variable (the sum of the point cloud volume and the three-dimensional image model error). We adopt a RGB-D camera to create a three-dimensional model and use the SURF (Speeded Up Robust Features) algorithm to obtain the image feature points, followed by the ICP (Iterative Closest Point) algorithm to calculate the translation and rotation matrix between the two images. Then, by the proposed method, the sub-optimal covariance and threshold are found. In the experiment of reconstructing the 3D model of an apple, 361037 cloud points were created and 82% of them were removed by the proposed method and the volume error was only 3.83%. Compared to the previous work by Kyo ̈stila ̈, there is a great improvement in removing redundant points of a 3D model and maintaining the map quality.
|