A novel genetic algorithm for projective invariant 2D object shape recognition

The aims of model-based recognition are to identify unknown object(s) in a scene and to estimate their locations with reference to the corresponding models. Early research works were mainly related to affine invariant recognition. These approaches had shown that an object shape can generally be represented by its outermost boundary, and in particular at locations having high curvature known as dominant points. However, the number of dominant points detected from the same object are generally not unique and their localisation is variant for various capturing environments. Consequently, the affine invariants computed from these points are often unreliable. Furthermore, an approximation to perspective projection is assumed in order to employ the affine transformation. Therefore the camera must be placed sufficiently far away so that the thickness variations of the scene objects are relatively small compared to their distances from the camera. This thesis investigates the use of genetic algorithm for projective invariant recognition of near planar objects. The approach does not require dominant points detection, the computation of projective invariants and the approximation to perspective projection. The problem considered is to find the projective transformation of a reference contour which can be best matched against a scene contour. A novel genetic algorithm has been developed to find the optimal solution in order to improve search efficiency and to prevent the process of trapping in local optima. Two variations of the genetic algorithm has been evaluated and presented in this thesis. The approach has been tested on a reasonable number of industrial handtools. Experimental results show that, in general, high matching score is obtained from two shapes belonging to the same object viewed at different positions; and low matching score is obtained from the matching of two non-identical objects. These demonstrate the feasibility of this approach in projective invariant recognition. Possible future research based on the proposed algorithm has also been included in the thesis.