Range Image Feature Extraction with Varying Degrees of Data Irregularity

Abstract

The use of range images has become prominent in the field of computer vision. Due to the irregular nature of range image data that occurs with a number of sensors, edge detection techniques for range images are often based on scan line data approximations and hence do not employ exact data locations. We present a finite element based approach to the development of gradient operators that can be applied to both regularly and irregularly distributed range images. We have created synthetic irregularly distributed range images for each edge type, and the gradient operators developed are evaluated with respect to their performance in edge detection across varying levels of data irregularity.

Combining Gradient Operators and Dihedral Angle for 2D and 3D Feature Extraction

“…The feature maps illustrated in Figure 4 demonstrate that the proposed dihedral angle approach provides better feature maps than the other techniques, and this is most evident in Figure 4(l), the face image. [13] (e) Significant gradient change [13] (f) Significant gradient change [13] (g) Scan line approach [5] (h) Scan line approach [5] (i)Scan line approach [5] (j) Dihedral angle (k) Dihedral angle (l) Dihedral angle [17] and compare our proposed feature finding approach, the dihedral angle technique (denoted as DA), with that of the significant gradient change thresholding technique [13] (denoted as SGC) and the well-known scan-line approximation algorithm [5] (further denoted as SA). Pratt [17] here A I is the actual number of edge pixels detected, I I is the ideal number of edge pixels, d is the separation distance of a detected edge point normal to a line of ideal edge points, and α is a scaling factor, most commonly chosen to be 9 1 , although this value may be adjusted to penalise edges that are localised but offset from the true edge position.…”

“…However, as previously discussed, when gradient operators are applied to range images, standard thresholding yields object surfaces rather than edges: hence one approach to determining features is to look for significant changes in the gradient responses of the range image; this approach is presented in [13] and used for comparative evaluation in Section 4. In this paper, we present an approach in which we extract features via the computation of the dihedral angle between two surfaces via application of the irregular gradient operators.…”

“…In addition, the results are computed on five randomly generated meshes, comprising five of each range edge types: convex roof, concave roof, convex crease, concave crease, jump, and the Figure of [13] and scan-line methods [5] are similar within the various cases of the edges types; the directional derivative approaches has slightly higher Figure of Merit values than the scan-line methods. However Figure 6 clearly demonstrates that the proposed dihedral angle technique becomes superior for higher degrees of irregularity than other methods.…”

“…In [13,14], we addressed the data distribution problem by generating directional derivate operators that are shape adaptive, and hence can be applied directly to irregularly distributed data. In [13,14] we noted that the standard thresholding technique for finding edges, typically applied to intensity images, is not appropriate for use on range image data, as it does not identify edges, but surfaces; to overcome this drawback, we identified a process of finding significant changes in the gradient response as a means of determining edges in range images.…”

An Adaptive Technique for Accurate Feature Extraction from Regular and Irregular Image Data

“…In [15,16] we addressed the data distribution problem by generating directional derivate operators that are shape adaptive, and hence can be applied directly to both irregularly and regularly distributed data, thus providing the facility for the operators to be applied directly to both range and intensity images. In [15,16] we noted that the standard thresholding technique for finding edges, typically applied to intensity images, is not appropriate for use on range image data, as it does not identify edges, but surfaces in range images; to overcome this drawback, we identified a process of finding significant changes in the gradient response as a means of determining edges in range images.…”