Research On Sparse Appearance Model And Efficient Classifier

Sparse appearance model and efficient classifier are regarded as objects of study inthis paper, aiming at improving the robust of tracking algorithm which may be damagedby the change of object appearance in complex tracking scenes. Visual object trackingbelongs to the basic research category of computer vision, has broad applications in safemonitoring, intelligent traffic, military application, video compression, human computerinteraction and so on, improving the robustness of visual object tracking algorithm hasvery important senses.Compressive tracking algorithm, basing on compressive sensing theory, using arandom measurement matrix, extracts locally sparse generalized Haar-like features toconstruct sparse and adaptive appearance model, and improves the robustness ofalgorithm in complex tracking environments. In this algorithm, a rectangle is used todenote object region, while this region includes some background information, so thatsome sparse generalized Haar-like features from background would be extracted. Thesefeatures from background would take part in the construction of appearance model andreduce the accuracy of appearance model, and then affect the robustness of algorithm.The algorithm in this paper firstly gets the classifier score of detected samples, secondlyuses global intensity histogram feature to get the similarity between the object and thedetected object, and finally uses the similarity to amend the classifier score.Experiments show that the disadvantage taken by sparse generalized Haar-like featurefrom background part in object region can be eliminated to some extent after importingglobal intensity histogram feature, improving the stable and accurate of appearancemodel, so as to improving the robustness of tracking algorithm in complex environment.Compressive tracking algorithm uses a naive Bayesian classifier to classify detectedsamples and gives the same importance to the positive samples in positive sample setwhen training naive Bayesian classifier. The algorithm in this paper gives different weight to the positive sample according the distance between the positive sample andthe object when training naive Bayesian classifier. The distance between the positivesample and the object is smaller; the similarity between the positive sample and theobject is larger, so that the weight of the positive sample is bigger. Conversely theweight of the positive sample is smaller. Experiments show that this weighted methodon positive samples according to distance can improve the performance of the naiveBayesian classifier and get the sample whose similarity to object is biggest, improvingthe robustness of algorithm in complex scenes.