|M.Sc Student||Arbel Joseph|
|Subject||Segmentation and Recognition of Meshes for Scanned Parts|
Based on Curvature Histograms
|Department||Department of Mechanical Engineering||Supervisor||Professor Anath Fischer|
Recognition of 3D geometric objects plays an important role in computer vision and inspection applications in manufacturing and medicine. The most common methods used today for geometric 3D object recognition are semi-automatic and require intensive human intervention. The motivation of this research is to improve the infrastructure required for making these inspection processes more automatic.
This work focuses on recognition of object segments and features. 3D geometric objects are defined in terms of a triangulated mesh reconstructed from scanned data. This work focuses on inspection applications for manufacturing modern mechanical parts. The main problem in recognizing and classifying objects from scanned data is that the data is noisy, thus leading to errors in recognition. Moreover, modern mechanical parts have complex shapes due to their multiple functionalities.
This thesis proposes an automatic method for recognizing functional segments in the part. The proposed method is based on discrete curvature analysis. The thesis deals with the noise problem by applying smoothing and filtering and by implementing a segment growing algorithm. The advantages of this new approach are that it involves insight into part functionality. Moreover, the method can recognize basic shapes and freeform surfaces, and is fast. The method's disadvantages include low recognition accuracy and low robustness. The method’s proposed phases are: (a) segmentation based on curvature sign classification, with planes recognized in this phase; (b) segmentation based on histogram clustering by using a curvature histogram, with freeform surfaces classified in this phase; (c) segmentation and recognition based on metric classification of the following basic shapes: spheres, cylinders, cones and tori.
All the segmentation processes use a segment growing algorithm, which increases the reliability of the results and reduces noise related to recognition and verification errors. The segmentation yields classification shape parameters that can be used as a basis for object verification. The method’s recognition, dimensional parameter extraction and verification performance are demonstrated on synthetic and scanned models.