Range data recognition using three-dimensional invariant features
Hon-Son Don, Chong-Huah Lo · 1990
The dissertation is concerned with computer recognition of range data using 3-D invariant features. We have derived new 3-D invariant features and developed pattern recognition algorithms using these features. The recognition systems based on these invariant features are robust and reliable. Specifically, we have done the following: (1) An innovative pattern recognition algorithm using 3-D moment invariants and multi-layer neural classifier--it can do multiview 3-D object recognition and is suitable for real time implementation. (2) An efficient invariant complex waveform representation for 3-D curves--this representation can reduce the complexity of the matching and segmentation of 3-D curves to that of 1-D waveforms. Moreover, the point correspondence problem can be efficiently solved by waveform matchings. (3) A range image analysis algorithm which is capable of analyzing a complex scene--the range image is first segmented using the split-and-merge technique and the functional approximation method. The segmented image is represented by a region adjacent graph with invariant attributes. The distance between attributed graphs is computed. This scalar distance measure is useful in model-based object inference and object classification.