Skip to Main Content
An algorithm for real-time recursive estimation of the two-dimensional shift between successive images in an image sequence occurring, e.g., in a down-looking airborne TV sensor, is presented. The shift, which is linearly related to ground velocity, is adaptively tracked by means of the gradient of a similarity function relating the two successive images. An implementation is presented in which substantial memory capacity and computational complexity are saved by using only a single line in the image frame and by binary quantization of the video signal. A detailed analysis of the algorithm is presented using stochastic mathematical models for the terrain texture, image noise, and velocity variations. Under some simplifying assumptions, closed-form solutions for the error statistics, including temporal power spectral density, are derived. The probability of loss of lock in tracking and the expected time for its reacquisition are evaluated and are found to be low at typical operating conditions. The analysis also indicates considerable robustness of the algorithm to widely different image statistics. Numerical examples indicate very good performance in autonomous navigation applications.