|
|
|
“Motion Analysis and Passive Navigation Using Long Image Sequences”
by Chandra Shekhar
May 1992
This dissertation deals with the analysis of visual motion from image sequences, with emphasis on the use of long image sequences, recursive estimation techniques and the integration of feature matching and motion estimation. The objective is to estimate motion parameters (pose, velocities, etc.) and 3-D structure parameters relating the camera(s) to the scene. The basic idea is to extract and match salient points from the image sequence, and to relate their image plane trajectories to the motion and structure parameters. This is accomplished by using simple models for translational and rotational motion, and, for monocular sequences, a central projection model for imaging. A Kalman filter, or one of its variations, is used to estimate the unknown model parameters. Various methods of initialization are developed, including least-square batch algorithms, linear methods, and iterated filter-smoothers. The recursive estimation of motion and structure parameters is interleaved with the extraction and matching of feature points. This method is then applied to the following problems in motion analysis:
1. Passive navigation: This deals with the case of a single moving camera in a stationary environment.
2. Target tracking : Here the camera is stationary, and the goal is to track the motion of a rigid object within the camera's field of view.
3. Obstacle avoidance: In this application, the situation is more general; a (stereo) camera is moving in a traffic environment containing several moving obstacles, some of them possibly nonrigid.
The algorithms developed are tested on real and synthetic data, demonstrating their robustness in the presence of measurement and modeling errors.