A Robust Method for Airborne Video Registration Using Prediction Model

Abstract

In order to generate high precision registration for airborne video with or without GPS data, this paper presents a new robust method to remove matching failures fast. Rather than registering each frame of the video sequence individually, which is popular in existing applications, our predicted progressive sample consensus (P-PROSAC) algorithm firstly predicts a movement model of camera from previous registering results and/or GPS data, and then use it to construct a set of ranked candidate correspondences, from which our P-PROSAC algorithm draws samples. Previous robust estimator-PROSAC has improved the efficiency of RANSAC estimator. However, it is under the assumption that the similarity measure predicts correctness of a match, which is strongly challenged by moving targets, repetitive patterns and noises. Our P-PROSAC uses a prediction model of camera movement which can best describe the character of inliers so as to find solutions much earlier. Experiments on real-world aerial video demonstrate that our approach can significantly reduce the calculation time.