We are inviting IDEALS users, both people looking for materials in IDEALS and those who want to deposit their work, to give us feedback on improving this service through an interview. Participants will receive a $20 VISA gift card. Please sign up via https://forms.illinois.edu/sec/4069811
Files in this item
|(no description provided)|
|Title:||Fast Waveform Pattern Matching With Significant-Point Frames|
|Author(s):||Mayse, William Clark|
|Doctoral Committee Chair(s):||Ray, S.R.,|
|Department / Program:||Computer Science|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||The basic technique of scale-space filtering has been modified to render it suitable for detecting instances of short-duration "patterns" in one-dimensional waveform data. Scale-space filtering involves convolving a signal with a parametric kernel, such as the Gaussian function. The contour plot and interval tree record migrations and disappearances of the level-crossings of linear differential operators across multiple convolutions. These structures can be viewed as compressed, hierarchical representations of the signal, and have served as bases of successful systems for pattern matching The interval tree proves unsuitable, however, for representing signals of short duration, for which migrations of the level-crossings across data-interval boundaries cause loss of contour-plot information.
The significant-point (SP) frame is introduced as a simpler interpretation of the contour plot. Treating the multi-convolution data in an uncoupled, non-hierarchical manner avoids the requirement of preserving contour identities across convolutions. This simplifies computation and eliminates the aforementioned information loss. It also allows a two-stage matching approach which further reduces the necessary computation. Subsets of the pattern and search-data SP frames are compared; computing the remainder of the pattern and data frames is required only where this preliminary search indicates a possible match.
A parametric computational-cost model has been developed for the five major non-convolution tasks in the foregoing paradigm. The method's complexity varies with the parameter values, but at worst compares to those of typical noise-tolerant matching methods, and is generally two orders of magnitude lower. A system based on the paradigm has been implemented and tested on 80- to 160-point patterns and random-waveform data. The observed performance corroborates the cost model, and indicates the capability (with convolution-hardware support) of real-time matching on workstation-class computers at data rates of several kHz. Highly parallel implementation and extension to higher-dimension domains appear feasible.
Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1992.
|Date Available in IDEALS:||2014-12-17|