Online Boosting Learning For Object Tracking

Video object tracking is the important research branch in computer vision, and has applications in many fields. Recently, visual tracking based on learning has caused many scholars'attention, since it can achieve good tracking performance. The representative method treats tracking as a classification problem between object and background. Instead of building complex model to describe the visual object, this method intends to find a decision boundary between object and background. When the appearance of object changes, it only needs to update decision boundary, rather than the object appearance model. Currently, the representative tracking algorithms based on learning include Ensemble tracking and Adaptive linear weak classifiers boosting for online learning which proposed by Avidan and Tougiq Parag respectively. The basic principle behind the two methods is to train a set of linear weak classifiers for visual tracking in an online manner by use of simple image feature like color, intensity etc, and may fail to track the visual object in the complex scence. By incorporating better image features like tensor or gabor features and replacing the linear weak classifier with a nonlinear weak classifier, this thesis has done substantial research work on boosting based on visual object tracking. Specificially, this theis has made the following contributions:1) Gabor filter has been used to get better image feature for visual tracking by online boosting. Compared with intensity and color, Gabor filter has good spatial locality and orientation selectivity , and can extract multidirectional spatial frequency feature and local structure feature. As a result, it has a higher discriminative power between background and foreground. However, a trival application of high dimensional Gabor feature to tracking will affect tracking speed. So, we turn to two schemes to reduce dimension and select the most discriminative feature. (a).Using local gabor filter bank to extract the Gabor feature vectors; (b).Adjusting filter bank parameters adaptively.2) Tensor feature has been explored for visual tracking by online boosting. As complement to typical vector patterns, tensor feature can capture gradient direction information, and has good distinguishing ability for the object which has strong texture property. By combining tensor feature with online boosting algorithm, we have achieved good tracking result for textured visual object.3) Online boosting method using the recursive least-squares (RLS) algorithm has been proposed for visual tracking. Linear classfier cann't acquire good discrimination power. So,we employ a nonlinear version of the recursive least square algorithm(RLS) here. It performs linear regression in a high-dimensional feature space induced by a Mercer kernel,and can therefore be used to recursively construct minimum mean-squared-error solutions to nonlinear least-squares problems. In order to regularize solutions and keep the complexity of the algorithm bounded, we use a sequenctial sparsification process that admits into the kernel representation a new input sample only if its feature space cannot be sufficiently well approximated by combining the previously admitted samples. So, using this sparsification proceduce, we can update weak classifiers online.Classifier and feature selection are two important fields in pattern recognition. In this paper, we focus on feature and classifier. First, using Gabor filter and tensor feature to extract some discriminative features. Then, we choose the classifier based on kernel rather than the linear classifier. When the tracking scenario is very complex, this method can still achieve good results. Experimental results verify the effectiveness of the algorithm.