Reliable Location and Regression Estimates with Application to Range Image Segmentation

  • Authors:

  • M. Baccar;L. A. Gee;M. A. Abidi

  • Affiliations:

  • University of Tennessee at Knoxville, Department of Electrical & Computer Engineering, Knoxville, TN, USA;University of Tennessee at Knoxville, Department of Electrical & Computer Engineering, Knoxville, TN, USA;University of Tennessee at Knoxville, Department of Electrical & Computer Engineering, Knoxville, TN, USA

  • Venue:
  • Journal of Mathematical Imaging and Vision
  • Year:
  • 1999

Quantified Score

Hi-index 0.00

Visualization

Abstract

Range images provide important sources of information in manythree-dimensional robot vision problems such as navigation andobject recognition. Many physical factors, however, introduce noiseto the discrete measurements in range images, identifying the need toreassess the error distribution in samples taken from real rangeimages. This paper suggests the use of the L norms to yieldreliable estimates of location and regression coefficients. Thisparticular approach is compared against two commonly used approaches:Equally Weighted Least Squares, which minimizes the L_2 norm; andthe Chebychev approximation, which minimizes the L_1 norm. Theproblem is a weighted least squares case where the weights arederived from the chosen parameter, p, and its ability to yield avariety of location estimates spanning from the sample mean to thesample median. These two estimates have a wide application in imageprocessing that includes noise removal. This paper will show theproblems associated with these two techniques, and suggestexperimental solutions to minimize them. A specific operating rangeis determined in which the L norms perform well and a regressionmodule is used in conjunction with a region-growing segmentationalgorithm to provide a reliable partition of range images.