Research On Visual Tracking Algorithm Based On SIFT

With the development of science and technology, visual tracking has attracted the attention of many researchers. Visual tracking is one of the key technologies used in video monitoring, navigation and many other applications. So it is widely used in these applications and also brings great prestige to the society. Although researchers have come up with a lot theories and methods, visual tracking still faces huge challenges since the actual scene is complex. Based on the existed methods, visual tracking still needs to be studied and improved.Modeling is one key technology for visual tracking and it can affect the tracking results directly. Recently, Scale invariant feature transform (SIFT) as a robust feature for scale, rotate, blur and illumination has drawn many researchers'attentions. Some of them have used SIFT to improve the target's model. The registration based on SIFT can reach sub-pixel accuracy, so we can improve the accuracy of tracking by using SIFT.This paper introduces some common features, especially SIFT and also summarize the tracking algorithms. Based on these theories, tracking algorithm is further studied and applied into two programs in this paper. By combining the target tracking and target recognition, a SIFT based target tracking and identification method is proposed. Tracking algorithm models the target with its appearance model and SIFT feature. Then it uses the probabilistic voting method and mean-shift method to estimate the object's optimal center. The recognition part adopts the SIFT feature to construct bag of words model. After that, using support vector machine algorithm to get a recognizer for each target by training the sample. The whole algorithm's first step is extracting foreground by using Gaussian Mixture Model. Then recognize the foreground and track each target to get the optimal position.A point tracking method based on SIFT and HOG is proposed when doing a tracking project. After matching the front and back images'SIFT feature, a homography matrix is obtained. Map the tracking point to the second image to get the candidate point. In order to improve the accuracy, random sample consensus algorithm is introduced to get a more accurate matrix. HOG feature is applied to precisely locate the candidate point. At last, update the point's SIFT and HOG feature to eliminate the offset caused by the changed environment.By introducing the algorithm of point tracking based on SIFT into the program of detecting object's speed, this paper presents a speed measurement algorithm based on SIFT. The algorithm gets the target's moving distance by matching two images.