Using three-view analysis, this paper proposes a new approach for single polyhedron recognition. Both the 3D surface data and 2D image projection patterns are used to find the exact corresponding model of an unknown input object. The recognition process first applies a hypothesis-generation-verification process to match the 3D partial object of the unknown input, which is reconstructed by applying the three-view matching strategy (Chiou et al., Proc. ICS'88, pp. 1516-1521 (1988)) and the polyhedron reconstruction process (Hung et al., Pattern Recognition 22, 231-246 (1989)), against the data-based models for selecting the possible corresponding models (candidates). It then backprojects each candidate onto the three image planes and matches the outer boundaries of the three backprojections by a simple method for judging the exact corresponding model. Measurement error is also analyzed theoretically to give a bound of the threshold value so that it can be automatically chosen in each decision making process. The approach is implemented on a PC-level vision system. Experimental results show that the system is feasible and reliable even in a noisy environment.
All Science Journal Classification (ASJC) codes
- Signal Processing
- Computer Vision and Pattern Recognition
- Artificial Intelligence