Image matching is a very important technique in image processing. It has been widely used for object recognition and
tracking, image retrieval, three-dimensional vision, change detection, aircraft position estimation, and multi-image
registration. Based on the requirements of matching algorithm for craft navigation, such as speed, accuracy and
adaptability, a fast key point image matching method is investigated and developed. The main research tasks includes:
(1) Developing an improved celerity key point detection approach using self-adapting threshold of Features from
Accelerated Segment Test (FAST). A method of calculating self-adapting threshold was introduced for images with
different contrast. Hessian matrix was adopted to eliminate insecure edge points in order to obtain key points with higher
stability. This approach in detecting key points has characteristics of small amount of computation, high positioning
accuracy and strong anti-noise ability; (2) PCA-SIFT is utilized to describe key point. 128 dimensional vector are formed
based on the SIFT method for the key points extracted. A low dimensional feature space was established by eigenvectors
of all the key points, and each eigenvector was projected onto the feature space to form a low dimensional eigenvector.
These key points were re-described by dimension-reduced eigenvectors. After reducing the dimension by the PCA, the
descriptor was reduced to 20 dimensions from the original 128. This method can reduce dimensions of searching
approximately near neighbors thereby increasing overall speed; (3) Distance ratio between the nearest neighbour and
second nearest neighbour searching is regarded as the measurement criterion for initial matching points from which the
original point pairs matched are obtained. Based on the analysis of the common methods (e.g. RANSAC (random sample
consensus) and Hough transform cluster) used for elimination false matching point pairs, a heuristic local geometric
restriction strategy is adopted to discard false matched point pairs further; and (4) Affine transformation model is
introduced to correct coordinate difference between real-time image and reference image. This resulted in the matching
of the two images. SPOT5 Remote sensing images captured at different date and airborne images captured with different
flight attitude were used to test the performance of the method from matching accuracy, operation time and ability to
overcome rotation. Results show the effectiveness of the approach.
|