Research On Feature-based Automated Registration Method For Aerial Image Sequence Stitching

Automated Aerial Image Stitching technology has many perspective applications, such as early warning, disaster control, digital map generation, military reconnaissance, target tracking and UAV-aided navigation etc. With the development of Digital Image Stitching technology, the applications will be more extensive.Automated image registration is a key part of aerial image sequence stitching. Its role is to automatically implement global registration of entire aerial image sequence, preparing for image sequence stitching.Feature-based registration method is chosen as the basis of our research and to handle the complex geometric transform between adjacent frames. We mainly focus on the characteristic of aerial image sequence to implement registration rapidly. More detailedly, this paper makes the following contributions:1. Efficient feature detection method for aerial image sequenceWe select SIFT as the feature of aerial image sequence for the reason that SIFT detector can be easily adaptive to the complex scene (OR content) and the existence of complex geometric transformation between adjacent frames of aerial image sequence. At the same time, we simplify the work flow of SIFT detector according to the fact that the change of image scale between adjacent frames is tiny. Experiments show that the simplified SIFT detector can maintain the stability of feature distribution and accelerate the SIFT feature detection.2. Stable feature matching method for aerial image sequenceWe briefly introduce a few metrics for feature similarity measure and feature matching principles briefly at the beginning. Then an algorithm named as B-NNDR (Bi-directional Nearest-Neighbor Distance Ratio) is proposed and applied in SIFT feature matching. It is shown that the B-NNDR algorithm can detect matching pair of SIFT feature efficiently and reduce the probability of error matching effectively.3. Fast Global motion model estimation algorithm for aerial image sequenceBecause the B-NNDR algorithm mentioned above cannot eliminate error matching completely, RANSAC method is applied here to expel those matching pairs (called outliers here) which do not agree with the Global Motion Model. At the same time, in view of high credibility of those feature matching pairs, we improve the procedure of global motion estimation between adjacent frames. Experiments show that the improved method can effectively reduce iterations of global motion parameter estimation and accelerate the estimation process.