Finding Structure in Data

Martial Hebert

Abstract

In many areas of robotics (most notably mobile robots), one is faced with the problem of extracting "structure" (objects, surfaces, etc.) from a large, unstructured set of data points from a sensor. The situation is complicated by many factors: the vast majority of the data comes from spurious sources in the environment that we don't care about (e.g., vegetation); the data is accumulated incrementally over time so that fixed data structures won't work; the resolution of the data may vary widely (so that anything that is tuned to a fixed resolution won't work); and the characteristic of the sensor and the data acquisition conditions may vary, thus requiring adaptation to different sensor modalities. I'll discuss some approaches found in the literature, including statistical models of natural environments and geometric approaches. This talk will be an informal discussion of some of the issues and solutions to this class of problems. I'll show some example applications from systems that we are currently working on (all in the context of robots operating in natural environments). I'll conclude with a few tough (for us, at least!) challenges with which I hope you might be able to help.

Although I will attempt to give this talk, the material was provided from the work of: Daniel Huber, Ranjith Unnikrishnan, and Nicolas Vandapel.

 

Speaker Appointments

For appointments, please contact Martial Hebert (hebert@cs.cmu.edu).