Automated Harnessing of Complex Distributed Attributes for Robust Object Recognition
Robert K. McConnell, PhD
Olin College, 1000 Olin Way, Needham, MA 02492, Milas Hall Auditorium.
Classification, verification and anomaly detection are often key components of robotics and automation projects.
Most of us are comfortable automating such recognition decisions based on properties that can be adequately represented by a simple mean value and Gaussian distribution. Dimensions of simple shapes, areas and weights usually fall into this category.
Unfortunately, sometimes the distinguishing information provided is more complex, and the simple models are unjustified. Yet, all too often, the methods suitable for those simple models are still applied. This commonly results in less than satisfactory outcomes and complexity tends to be regarded as an obstacle to reliable results. Such should not be the case.
Attribute complexity offers the advantages of increased potential to distinguish classes, redundancy to increase reliability, and robustness. This complexity appears in many forms and in most properties that can be sensed including, but not limited to, touch, sound and visual attributes. Color-based recognition provides a good example because it's easy to visualize and, in the field of machine vision, institutionally misunderstood.
The presentation will demonstrate the advantages of complex attribute distributions for recognition and segmentation. Using examples from color, multispectral and hyperspectral imagery it will first show the basics of a general integrated approach to classification, verification and anomaly detection based on such attributes and then describe a new, closely related, method for selection of an optimum subset of the available attributes.
The meeting is free and open to the public.
Reservations are not required.
Other WAY-2C Presentations
Last Updated 2/21/15