The problem addressed by this project is the inference of 3- D attributes of a moving object from a sequence of noisy images of the object. The 3-D attributes of interest are the translational motion and position, to a global scale factor, the rotational motion and position, and the object structure, again to a global scale factor. This research, high image noise levels are allowed, perhaps as much as 20% of the object image size. Such noise levels can occur even in high resolution imagery, whenever the image of the object is small relative to the sensor field of view. The approach taken in this proposal is to model the 3-D rigid body motion using the principles of kinematics. The kinematic equations propagationg translation and rotation are written in the form of a state space model. The noisy feature points or lines form the measurement model. Given that point or line correspondences over a long sequence of frames are available (or established) our goal to develop recursive and batch techniques for the estimation of 3-D motion and structure parameters. Performance measures such as the theoretically attainable lower bounds for the estimated parameters will be derived. Extensions to the two camera problem will also be attempted and algorithms developed will be tested on both synthetic and real image sequences.
