Appearance methods of solving recognition and estimation problems in robotics

This dissertation focuses on utilizing appearance methods to solve the recognition and estimation problems in robotics. As an intelligent robot scouts in an environment, it continuously captures images of various objects and scenes. Suppose the models of some objects and scenes are already encoded in the brain of the robot, then one fundamental task for the robot is to process this continuously absorbed visual information, recognize the observed objects and scenes. Geometrically, a continuous test curve is to be detected among perturbed model hypercurves or hypersurfaces after the KL decomposition is applied. Mathematically, the detection theory is used to solve this problem of detection in colored noises having nonidentical distributions. The result of scene recogntion is also extended to detect an intruder trepassing a modeled scene.;In addition, a new concept appearance flow is also described by revealing the relationship among the articulated motion dynamics, optical flow and appearance flow. As an example of using the derived appearance flow equation, an Extended Kalman Filter problem is formulated to estimate the joint angle parameters of a three-link planar arm.