|M.Sc Student||Shmukler Alexander|
|Subject||Verification of 3D Freeform Parts by Feature-Based|
|Department||Department of Mechanical Engineering||Supervisor||Professor Anath Fischer|
|Full Thesis text|
Precision inspection of freeform parts is one of the most important quality control tasks in the manufacturing industry. The aim of this inspection is to verify that the geometric dimensions and part tolerances of the produced part meet quality requirements. This is achieved by fitting the scanned data to the CAD model. This verification is complicated since the produced part includes defects and distortions. Currently, industry uses semi-manual verification, which is expensive, often inaccurate and very time-consuming. The goal of fast and automatic alignment of the scanned data to the CAD model can be achieved by developing geometry-based registration methods to verifying the overlapping shapes.
The most common method for registration of two 3D shapes is the Iterated Closest Point (ICP) algorithm. The quality of the alignment obtained by this algorithm depends heavily on the initial position and on choosing good corresponding points. As a result the ICP can converge slowly or find the wrong pose, especially in the presence of noise.
This research presents a method for automatic registration and alignment of two 3D freeform shapes, one from the scanned data and the other from the CAD model. The method makes no assumptions about their initial positions. Instead, the proposed algorithm uses a multi-scale shape descriptor to select features on the scanned data and identify their corresponding features on the CAD model. The proposed shape descriptor is invariant with respect to local shapes and is robust to noise. A coarse alignment is computed by finding and registering the best matching triplet of features. The ICP algorithm uses resulting coarse alignment to achieve a tuned alignment. The proposed method is automatic, efficient and straightforward to implement. The algorithm can also be effective in the case of partial scanned inspected shapes.
The feasibility of the proposed method is demonstrated on a number of freeform engineering and medical models.