Doctoral Dissertations

Date of Award

5-1996

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Electrical Engineering

Major Professor

Mohan M. Trivedi

Committee Members

Donald Bouldin, Marshall Pace, Steve Serbin, Reid Kress

Abstract

There are three main points of focus in this dissertation, all of which address important aspects of the object recognition problem. These are: adaptive range segmentation, graph matching and calibrated 3-D ranging. Computationally efficient, robust and accurate means to accomplish these types of problems are critical to improving the state of the art in object recognition.

In addition to the above three main areas, a fourth objective was also set. It was desired to create components using a systemic perspective, allowing them to be put into an integrated framework. Their ability to be integrated into a system is considered to be an important aspect of their real-world utility. Integrating the three main contributions into a complete system permitted validation of the performance of individual components in a system context.

The novelty of the research in 3-D ranging has produced new approaches to calibration and to acquisition. These techniques make for a system that acquires data at frame rate, is accurate and is easy to calibrate. The calibration technique that has been developed permits modeling of all elements of the sensor pack- age in a single step via a least norm solution. This is simpler and can provide greater accuracy than more traditional methods that model the camera and laser separately. The range sensor has an accuracy of better than 0.005 inches.

An adaptive approach for robust, accurate and computationally efficient seg- mentation of range images has also been developed. It is well matched to the range acquisition process. Adaptivity is a new trend in range image processing. Adaptive schemes are not bound by rigid geometries when accessing data, for example. This improves performance by better handling outliers. The new approach uses a combination of multiple adaptive mechanisms. Tests demonstrated that segmented regions can be computed in less than 3 seconds, under a wide variety of test conditions. Stable operation was achieved above 90% of the time in the majority of testing trials.

A new approach to graph matching has also been developed as a part of this effort. It provides a tremendous computational advantage over established methods that are either based on search, or on iterative relaxation techniques, or on alternate exhaustive formulations. The new method requires a computational effort that is of polynomial order. It is able to maintain a performance level for subgraph isomorphisms at or above 95% under a wide variety of conditions and with varying levels of noise. The performance on the full size comparisons associated with graph isomorphisms has been found to be 100/100, also under a variety of conditions.

Careful thought went into test conditions for individual components and for integrated tests. Types of surface curvature and discontinuities were important considerations, as were the complexity of objects and the number of database models. A suite of discontinuities and surface types were used in testing. Testing was performed using both synthetic scenes (4000+) and real scenes (500+). Results are reported for individual processing modules and for the integrated system performance.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS