Abstract
The stereo matching is the most important part of stereo vision. SIFT algorithm is obviously superior to other character descriptions in the feature point scale, rotation and invariant illumination, so it is extended in this paper. Because of traditional Scale-invariant feature transform (SIFT) method’s high dimensions, it is very complex and need much time to calculate. In order to optimize the algorithm, Principal Component Analysis of Traditional Scale-invariant feature transform (PCA-SIFT)+Random Sample Consensus (RANSAC) matching algorithm is proposed. PCA-SIFT uses PCA which can reduce dimension to SIFT in image matching. Although PCA-SIFT can effectively solve the problem of time-consuming, the accuracy of matching is not high. Using PCA-SIFT extract feature points in image and matching them. Firstly, we build Gaussian pyramid, detect extreme value points, and select key points with strong robustness. Using gradient direction distribution features of neighborhood pixels of the key point calculate the direction of key point can be obtained by calculating gradient direction distribution features of neighborhood pixels of the key point. Then using PCA algorithm generates 32d feature descriptors. According to the ratio of nearest neighbor algorithm for image matching, it gets matching points and sort in the size of the ratio of nearest neighbor. RANSAC algorithm is used to carry on the transformation matrix parameters fitting, and the resulting matrix model is used to remove the error matching points. To verify the effectiveness of algorithm, SIFTǃPCA-SIFTǃPCA-SIFT + RANSAC are compared by using many images and data. The experimental results show that the PCA-SIFT + RANSAC algorithm is more stable, more accurate and more rapid.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have