Unsupervised Learning and Reverse Optical Flow in Mobile Robotics

Book excerpt: As sensor resolution increases and costs decrease, the amount of data available on mobile robotics platforms is exploding. Unsupervised machine learning algorithms, and their ability to produce useful information without large labeled training sets, are an important tool for benefiting from this abundance. In this thesis the application of unsupervised learning to three subfields of mobile robotics is discussed. Tracking multiple moving objects from an unmanned aerial vehicle, road following in loosely-structured environments, and autonomous offroad navigation. The thesis focuses on building dynamic activity-based ground models for multi-object tracking, the combination of optical flow techniques and dynamic programming to estimate the location of a road, and the use of optical flow techniques to improve the quality of an autonomous robot's obstacle classification.