Title:  A Concurrent Learning Approach to Monocular, Vision-Based Regulation of Leader/Follower Systems

Committee: 

Dr. Vela, Advisor

Dr. Verriest, Chair

Dr. F. Zhang

Abstract:

The objective of this proposal is to enable a vision-based leader-follower to regulate range and bearing angle without communication, persistency of excitation or known geometry. The challenge is that computer vision does not directly provide a range measurement. In this proposal, a concurrent learning approach is used to estimate target size in a history stack. The history stack is then used to calculate a range pseudomeasurement which is used as a measurement in the existing Kalman filter. Concurrent learning has its origins in parameter identification identification which can be broken into two methods: recursive system identification and batch processing. Recursive system identification is on-line and requires persistency of excitation or can suffer from parameter drift. The parameter can also be included in the Kalman filter, in which case it can also experience parameter drift and requires an initial condition. Batch processing is off-line and requires specific motion to ensure a non-singular solution. The methods of recursive system identification and batch processing are combined without their limitations in concurrent learning. Concurrent learning is a time-switched method of building a stack of parameter estimates over real time, thus reducing the need for persistency of excitation at all times to simply a need for some excitation at any time. It doesn't require an initial condition or suffer from parameter drift.