Research On Object Recognition And Tracking Based On The Improved SIFT

With the progress of society and the rapid development of information technology,images have become a necessary means that human access to information of outside, andobtaining interesting information using computer has become a hot issue of research.Recently, local descriptor technology has been applied widely in image processing. In thispaper, aiming at the limitation of target recognition and tracking, the scale invariant featuretransform algorithm is optimized. The main research contents are summarized as follows:1. In remote sensing images, the geometrical characteristics of building area isdistinctive. In the discussing, considering the large field and complex scene of remotesensing images, the extrema detection method based on significant edge is proposed. Weget the binary image through adaptive threshold segmentation of the pyramid image and forimproving the effectiveness of extrema, the detection only applies on building area withsignificant edge. To solve the problem that the traditional RANSAC algorithm is not robustto purification with a small amount of matching points，at the same time, in view of topsignificant structure of building area, a matching method based on local similarity is putforward which has resulted perfect experimental in this paper.2. In the target recognition and tracking survey, by analyzing the result of framedifference, the registration based on SIFT is applied. Considering the high computationalcost of and timeliness requirement of system, the registration algorithm based on improvedSIFT is put forward. We implement the fast corner detection FAST in the gauss pyramid ofone octave, then assign orientation and describe the feature points by the traditional SIFT.By ensuring high representative of feature points and the accuracy of descriptors, thealgorithm in this paper has improved the speed and precision of the registration. Theregional consolidation is given to eliminate the influence of holes resulting by framedifference. Then the determination criterion based on the local features of tracking target isput forward to judge the identity of candidate target. And the target is tracked by check outconfidence value. In order to improve the practicability of the algorithm, the target isclassified by SVM which is trained by the shape and local feature. Experimental resultsshow that the method of this paper has achieved the accuracy detection and stable tracking. This paper completes the recognition of building region and tracking of moving targetswhich lay a foundation for practical application of the algorithm.