| Summary: | 碩士 === 國立中正大學 === 電機工程研究所 === 104 === To the best of increasing robotic vision in 3D conceptual for recognizing this living world, this thesis proposed a 3D recognition system by combining the local feature and global verification technique. To approach this, we modified the state-ofart methods and organized it as a robust hybrid flow. Another contribution to this thesis, we release the finest parameters to the Kinect sensor as well as the dataset. In the proposed framework, we expect the pre-process can deal with range filtering, noise reduction, and point cloud refinement. After this, the captured point cloud is more reliable and better to describe the object surface. The Second part is focused on recognition and pose estimation. We here refer two robust methods, SHOT descriptor and Hough Voting, one for the local feature generation and the other contributes to the object alignment. Finally, through the ICP to refine the pose matrix, we remove the false positive while verifying the good instance. Moreover, we design a keypoint selective mechanism after the hypothesis verification stage back into local conception.
|
|---|
