Object detection from two dimensional intensity images as well as three dimensional range images is considered. The emphasis is on the robust detection of shapes such as cylinders, spheres, cones, and planar surfaces, typically found in mechanical and manufacturing engineering applications. Based on the analyses of different HT methods, a novel method, called the Fast Randomized Hough Transform (FRHT) is proposed. The key idea of FRHT is to divide the original image into multiple regions and apply random sampling method to map data points in the image space into the parameter space or feature space, then obtain the parameters of true clusters. This results in the following characteristics, which are highly desirable in any method: high computation speed, low memory requirement, high result resolution and infinite parameter space. This project also considers use of fuzzy clustering techniques, such as Fuzzy C Quadric Shells (FCQS) clustering algorithm but combines the concept of "noise prototype" to form the Noise FCQS clustering algorithm that is robust against noise. Then a novel integrated clustering algorithm combining the advantages of FRHT and NFCQS methods is proposed. It is shown to be a robust clustering algorithm having the distinct advantages such as: the number of clusters need not be known in advance, the results are initialization independent, the detection accuracy is greatly improved, and the computation speed is very fast. Recent concepts from robust statistics, such as least trimmed squares estimation (LTS), minimum volume ellipsoid estimator (MVE) and the generalized MVE are also utilized to form a new robust algorithm called the generalized LTS for Quadric Surfaces (GLTS-QS) algorithm is developed. The experimental results indicate that the clustering method combining the FRHT and the GLTS-QS can improve clustering performance. Moreover, a new cluster validity method for circular clusters is proposed by considering the distribution of the points on the circular edge. Different methods for the computation of distance of a point from a cluster boundary, a common issue in all the range image clustering algorithms, are also discussed. The performance of all these algorithms is tested using various real and synthetic range and intensity images. The application of the robust clustering methods to the experimental granular flow research is also included.
If you have any questions please contact the ETD Team, libetd@njit.edu.
![[NJIT logo]](http://archives.njit.edu/vhlib/images/njit_logo.gif){kind=logo}
![[Van Houten Library logo]](images/vhl-top-image.jpg)
