Research And Application Of SIFT Feature Matching Technology

In this information age,image has become an essential way for human beings to acquire information.Therefore,how to exploit image processing techniques to obtain outside information has gained the special attention of domestic and foreign researchers.Scale-invariant feature transform(SIFT)algorithm is widely used in feature-based matching due to its robustness and uniqueness under the circumstances of image scale variation and rotation.However,SIFT algorithm also has its limitations in the timeliness of feature generation and matching accuracy.Focusing on two main research directions in image processing field,object recognition and motion tracking,this thesis takes advantage of the SIFT idea,and designs enhanced object recognition and motion tracking algorithms by improving SIFT method.Specifically,the research tasks of this work include:(1)Relying on the information entropy of image for threshold determination,this work proposes an adaptive method for adjusting the key point threshold;(2)This work introduces histogram distance to calculate EMD distance.Based on the features extracted by SIFT algorithm,we combine the improved EMD method with multi-gradient SIFT features to perform distance comparison and computation pruning;(3)Aiming at multiple object recognition,this work proposes an improved algorithm that employs bi-directional matching of SIFT features to achieve efficient target detection;(4)This work proposes a new method that integrates SIFT vectors and DBSCAN clustering algorithm,for the purpose of replacing the tracking module in original TLD algorithm.In addition,we also make adjustment to the detection module in TLD algorithm.Following the design ideas described above,in this thesis we have implemented the improved SIFT based object recognition and target tracking algorithms.The proposed algorithms have been verified on various test data sets.Experimental results demonstrate that,(1)the proposed method solves the problem of massive meaningless matching of feature points during the procedure of complex image matching;(2)the proposed method addresses the limitation that Euclidean distance is not applicable in a number of matching scenarios;(3)the proposed method provides an implementation of recognition and detection in multi-object scenarios;(4)the proposed method solves the problem in TLD algorithm that the tracking module is difficult to maintain tracking robustness during long-term tracking of motion objects.The algorithms proposed in this work are designed specifically for correcting the shortcoming of traditional SIFT method.Not only the matching accuracy of image object has been enhanced,but also the computational efficiency has been improved compared with traditional SIFT method.The proposed algorithms are expected to show better applicability in object detection and target tracking applications.