Dynamic Vision in the Dynamic Scene: An Algebraic Approach

In this paper, we address the issue of dynamically recovering 3D position vectors of a moving target based on its images in three views under the following scenario: (1) dynamic vision in dynamic scene, meaning that both camera and target are in motion; (2) The distance between target and camera is extremely large, so that we have only one image observation of the moving target for each view, the recovery of relative motion of target with respect to camera is impossible. By imposing the kinematical and dynamic constraints on the motion of target, its future position and velocity, given previous ones can be described by what we call f and g function. By combining with motion of the platform, the slant range for the middle time (or middle view) can be estimated by solving an eight-order polynomial. Since the three position vectors corresponding to the three image points on the three views can be assumed to lie in a single plane (approximate true if interval of the consecutive frames is small, and absolutely true if we consider the target to be a satellite moving on its orbit), the other two position vectors can be further derived. Initial experimental results demonstrated the correctness of the developed algorithm