Document Type
Thesis
Date of Award
5-31-1988
Degree Name
Master of Science in Mechanical Engineering - (M.S.)
Department
Mechanical Engineering
First Advisor
Rajesh N. Dave
Second Advisor
Harry Herman
Third Advisor
Bernard Koplik
Abstract
The detection of straight lines in 2-dimensional images leading to polygonal approximation to the boundaries of the objects is one of the important stages in scene analysis as it leads to the description of the images. The comparison of three diverse approaches for this problem in terms of the efficiency and accuracy is made, and their performance for different types of data and different sets of parameters is studied in this thesis.
The data points are generated using either an edge detection algorithm using chain encoding on real images or a synthetic data generator, which can create a variety of clusters. Both the above methods generate ordered data points, however nonordered data points are generated by random shuffling of the ordered data points. The vision system used is IRI DX vision system with video camera, digitizer, four frame buffers, and high and low resolution options.
There are three main line de-yption techniques considered in this study. They are: Hough transform, fuzzy clustering, and eigenvector line fitting schemes. The Hough transform is carried out by calculating the transformation curves for each point and incrementing the corresponding cell in the accumulator with the curve parameters. At the end, cells with high counts are checked and lines are drawn with these parameters. The fuzzy clustering technique, hereafter called the boundary fitter and the eigenvector fitting methods start with finding the corners in the data set using a corner detection algorithm which uses the differential tangential deflection as its basis. The data between successive corners form a cluster. Eigenvector line fit fits a line for the data points in each set and the boundary fitter algorithm uses fuzzy membership of each point in different clusters for calculating the center of each cluster, mixing coefficient, and the error function. The output is one line per cluster. The performance of the algorithms is studied by varying the values of the parameters after each process.
The boundary fitter gives the best approximation to the boundary though very slow. The other two methods are much faster but cannot give the best approximation to the boundary. Hough transform method is found to be unaffected by the ordering of the data points unlike the boundary fitter and eigenvector fit. The corner detector scheme is found to be reliable and fast however, the need for further investigation is realized. All the algorithms are coded into an extremely user friendly software which allows the user to compare various schemes to find lines in his own data sets.
Recommended Citation
Bhamidipati, Somayajulu, "Comparison of various techniques for the detection of lines in 2-dimensional images" (1988). Theses. 3012.
https://digitalcommons.njit.edu/theses/3012
