|Ph.D Student||Jacob Barhak|
|Subject||Reconstruction of Freeform Objects with Arbitrary Topology|
from Multi Range Images
|Department||Department of Mechanical Engineering||Supervisor||Full Professor Fischer Anath|
Reverse Engineering (RE) is the process of reconstructing a computerized model from a digitized 3D object. Laser scanners are commonly used since they can sample 3D range images fast and very accurately relative to other technologies. There are several open problems in the literature that are viewed as a bottleneck in the RE reconstruction process: (1) The data size is enormous and it includes noise. (2) The topology is unknown, and therefore point connectivity relations are undefined .
This research proposes a new approach which utilizes neural network techniques in order to overcome these problems. Neural networks generalize information by learning from data examples, making them suitable for RE tasks. This research utilizes and extends neural networks that employ competitive learning techniques, such as self organizing maps (SOM) and the neural gas neural network technique .
This work proposes two new approaches for reconstructing objects from multiple range images, both relying mainly on neural network algorithms. In the first approach, a parametric surface is constructed from each range image. Then, surfaces are merged together in order to reconstruct the boundary surface of the volumetric object. The second approach reconstructs a triangular mesh, approximating the geometry and deriving the topology from the cloud of points. This approach solves the difficult topology detection problem by extending the neural gas concept beyond the original algorithm .
For feasibility of the reconstruction process, examples of several freeform objects with arbitrary topology will be presented.