An Improved Local Invariant Feature Matching Algorithm And Its Application

In recent years,image matching technology has been widely used in various fields of life,such as medical detection and identification of diseases,cloud image recognition in weather,ID face recognition in railway traffic,fingerprint identification in security monitoring,military field Risk target detection,etc.With the rise of machine learning,deep learning and artificial intelligence technology in recent years,traditional image processing technology has been continuously affected.However,deep learning is gradually manifested on the basis of continuous maturity,such as lack of data sets,limitation of device resources,adaptation of application scenarios and consumption of time costs.Traditional image processing related technologies are becoming more and more mature,and image matching technology still plays an important foundation as a basic branch.The image matching algorithm mainly includes two categories: gray-based matching algorithm and feature-based image matching algorithm.The gray-scale correlation algorithm needs to combine the gray information of all the pixels in the image to process,resulting in a large amount of calculation,slow running speed,and sensitive to grayscale changes of the image.The feature-based algorithm has weak dependence on the image itself and low computational complexity.Therefore,feature-based image matching algorithms have always been one of the hotspots of research.In this thesis,the problem of image matching algorithm based on local invariant feature is analyzed and studied.At the same time,the feature matching feature is used in target tracking.The main work is as follows:Firstly,in-depth study of the advantages and disadvantages of several feature matching algorithms such as SIFT,SURF,ORB and Harris.The advantages of SIFT algorithm are scale invariance,rotation invariance and robustness to illumination.The disadvantage is that the real-time performance is poor.The SURF algorithm is improved on the basis of SIFT algorithm,and the performance of the algorithm is improved on the basis of the original.The Harris algorithm mainly detects corner points or straight lines in the image,and the feature points are densely distributed and do not have scale invariance.The most obvious advantage of the ORB algorithm is good real-time performance.At the same time,the disadvantage is that it does not have scale invariance and feature points are dense,which is not conducive to post-processing.Combining the advantages of SURF and ORB algorithms for scale variation and good real-time performance,a feature matching method based on SURF and ORB is proposed.Secondly,the algorithm proposed according to the requirements needs to be combined with the feature detection of SURF and ORB algorithms,and the feature description of the ORB is used for feature description.The feature points obtained by the ORB algorithm using FAST feature detection are relatively dense.Therefore,SURF replaces FAST to acquire feature points during feature detection.The spatial structure design of ORB algorithm does not adapt to image scale changes,and the design ratio of spatial structure is improved,so that the improved algorithm satisfies Scale invariance requirements.In the process of feature matching,the mismatch matching algorithm is improved,and the matching feature point pairs are deleted by mistake,which improves the matching accuracy.The experimental results are verified by the images taken in the surrounding environment and the images in the standard dataset(Mikolajczyk).The improved S-ORB algorithm finds the rotation angle and scale change of the image.The overall matching performance is better than the ORB algorithm and the SURF algorithm.Improved,the algorithm performs well in terms of matching accuracy.Thirdly,combining the improved feature matching algorithm with target tracking,the implementation principle and process of each module of TLD target tracking algorithm are studied in detail,and the algorithm is improved.Several commonly used target detection and tracking algorithms are studied in detail,and the implementation process of TLD algorithm is mainly studied.The TLD algorithm decomposes the tracking task into three parts: detection,learning and tracking.In the research,it is found that the feature points of the tracking module are selected by LK optical flow method.The feature matching and mismatch matching algorithm are combined to replace the optical flow method to achieve image feature matching.The experimental results show that the algorithm can significantly improve the running speed of the whole algorithm,and the accuracy of the algorithm tracking is improved compared with the original algorithm,and the image rotation and scale changes are improved.