Off-campus UMass Amherst users: To download dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.
Non-UMass Amherst users, please click the view more button below to purchase a copy of this dissertation from Proquest.
(Some titles may also be available free of charge in our Open Access Dissertation Collection, so please check there first.)
Reducing noise in three-dimensional models recovered from a sequence of two-dimensional images
Abstract
The goal of this dissertation is to develop a technique for constructing models of a scene using camera images obtained by a moving robot. Such models are useful for a navigating robot, especially for positioning itself in the world, following paths and avoiding obstacles. In order to construct a 3D model from images, the information about the camera's motion between viewpoints is needed. However, estimating the camera's motion turns out to be a major source of error in the constructed 3D model; in all previous work on motion this motion error has been neglected. The main contribution of this dissertation is isolating the motion error, estimating its effect, and correcting for it using an incremental algorithm. Motion error manifests itself in cross-correlations of errors between points in the 3D model. The algorithm developed in this dissertation weights an individual 3D model by the inverse of its covariance matrix (which contains the cross-correlations), reflecting the accuracy of the model. Such weighted 3D models--obtained as the robot moves--are then combined. The performance of the algorithm was compared against three algorithms which neglect the motion error: Horn's two-frame algorithm, a multi-frame blind averaging algorithm, and a standard multi-frame Kalman Filtering algorithm. In three experiments considered (involving a robot workcell sequence, an indoor lobby sequence, and an outdoor rocket-field sequence), the algorithm consistently outperformed (by a factor of 2-3) the other three algorithms. In further experimentation, the constructed 3D model was used to determine the position of a robot with a accuracy of 2-3%. The computational complexity of the algorithm is $O(n\sp3)$ (for n points in the model). In preliminary experiments, it was determined that reducing computational time by ignoring parts of the covariance matrix does not appear promising, whereas dividing larger 3D models into smaller subsets of points (while maintaining the full covariance matrix for each subset) may turn out to speed up the algorithm without sacrificing accuracy. Furthermore, it is estimated that constructing and updating a model made up of 22 points takes only 1.8 seconds on a fast Silicon Graphics machine (SGI) every time the robot moves.
Subject Area
Computer science
Recommended Citation
Thomas, Joe Inigo, "Reducing noise in three-dimensional models recovered from a sequence of two-dimensional images" (1993). Doctoral Dissertations Available from Proquest. AAI9408355.
https://scholarworks.umass.edu/dissertations/AAI9408355