New applications of Bayesian recursive estimation in the areas of image processing and computer vision.

摘要

This is a study of applying Bayesian recursive estimation approach to practical estimation problems. Estimation problems exist in many important areas, such as surveillance, navigation, traffic management, and medical application. All these problems can be summed up to state estimation of non-linear and non-Gaussian systems. Bayesian recursive estimation approach can be used to combat these problems effectively. The main contributions from this thesis work include some fit models developed for these practical problems and some systems implemented effectively to address these problems.;As an active research area, video surveillance attempts to detect and track moving objects in video, and analyzes the object behaviors automatically. In this thesis, a novel object-tracking model to address these difficulties is described, in which multi-clue information and a mask are used to develop the tracking approach based on Bayesian recursive estimation method. Experiment results demonstrate that this model can be used to tracking the moving object in the cluttered background effectively.;In radiation therapy treatment of tumors on lung or abdomen, breathing-induced tumor motion can significantly limit treatment efficacy. Based on Bayesian recursive estimation formula, a tracking algorithm is developed to track tumor in real time, which enable the beam to move with tumor. A quasi-periodic model is used and no internal markers are required. As demonstrated by the experiment results, our periodic model provides more accurate tracking results than sinusoidal model and linear model.;For vision-based Unmanned Aerial Vehicle (UAV) navigation, a novel approach to position and orientation estimation is described in this thesis. In this approach, position and orientation estimation problem is formulated as a tracking problem and solved by using Bayesian recursive estimation method. The state and observation models are established based on analysis of imaging geometry of the UAV's video camera in connection with a digital elevation map (DEM) of the flight area, which helps to control estimation error accumulation. The efficacy of this approach is demonstrated by simulated experiment results.