Dennis W. Strelow

Motion Estimation from Image and Inertial Measurements Degree Type: Ph.D. in Computer Science
Advisor(s): Sanjiv Singh
Graduated: December 2004

Abstract:

Robust motion estimation from image measurements would be an enabling technology for Mars rover, micro air vehicle, and search and rescue robot navigation; modeling complex environments from video; and other applications. While algorithms exist for estimating six degree of freedom motion from image measurements, motion from image measurements suffers from inherent problems. These include sensitivity to incorrect or insufficient image feature tracking; sensitivity to camera modeling and calibration errors; and long-term drift in scenarios with missing observations, i.e., where image features enter and leave the field of view.

The integration of image and inertial measurements is an attractive solution to some of these problems. Among other advantages, adding inertial measurements to image-based motion estimation can reduce the sensitivity to incorrect image feature tracking and camera modeling errors. On the other hand, image measurements can be exploited to reduce the drift that results from integrating noisy inertial measurements, and allows the additional unknowns needed to interpret inertial measurements, such as the gravity direction and magnitude, to be estimated.

This work has developed both batch and recursive algorithms for estimating camera motion, sparse scene structure, and other unknowns from image, gyro, and accelerometer measurements. A large suite of experiments uses these algorithms to investigate the accuracy, convergence, and sensitivity of motion from image and inertial measurements. Among other results, these experiments show that the correct sensor motion can be recovered even in some cases where estimates from image or inertial estimates alone are grossly wrong, and explore the relative advantages of image and inertial measurements and of omnidirectional images for motion estimation.

To eliminate gross errors and reduce drift in motion estimates from real image sequences, this work has also developed a new robust image feature tracker that exploits the rigid scene assumption and eliminates the heuristics required by previous trackers for handling large motions, detecting mistracking, and extracting features. A proof of concept system is also presented that exploits this tracker to estimate six degree of freedom motion from long image sequences, and limits drift in the estimates by recognizing previously visited locations.

Thesis Committee:
Sanjiv Singh, (Chair)
Martial Hebert
Michael Lewicki
Rama Chellappa (University of Maryland)

Jeannette Wing, Head, Computer Science Department
Randy Bryant, Dean, School of Computer Science

Keywords:
Batch shape-from-motion, recursive shape-from-motion, inertial navigation, omnidirectional vision, sensor fusion, long-term motion estimation