A Study Of Robust Object Tracking Under Complex Background

The technology of robust object tracking under complex background is the basis to accomplish reconnaissance and strike for the UAV(Unmanned Aerial Vehicle) system, and it is also the key for various types of precision guided weapons to capture object in real-time at the stage of the terminal guidance. However, it is still very difficult to track a single moving object robustly due to these challenging factors, such as illumination variation, occlusion, drastic change of object’s appearance due to scaling rotation and nonrigid deformation, image fuzzy caused by abrupt motion, cluttered background and similar disturbance etc. To overcome the influence of those factors, the research is launched on four major components of a typical tracking system, and some corresponding tracking algorithms based on semi-supervised online learning theory are proposed in this dissertation. The main contents are summarized as follows:1. The basic theory of semi-supervised online learning and two types of typical tracking methods called Online MIL tracker and TLD tracker are introduced. Both of them have strong reference to guide the research of robust object tracking under complex conditions. To solve the sampling fuzzy problem and weaken the supervision effect, the sampling label of the Online MIL tracker is fuzzied by packaging the instances with similar label into a bag and the object appearance is indicated with the label of bag instead of instance, the essence of which is the improvement of the object appearance representation.A novel object tracking framework which explicitly decomposes the tracking task into tracking, learning and detection is proposed by the TLD tracker. An online learning mechanism which makes the tracker possesses of the "memory" ability is used to fuse the result of tracking and detection. As a result, the lost object can be recaptured as long as it reappears in the filed of view.The method expands and improves the traditional object tracking theory based on object detection, makes up for the shortage of unstable tracking performance obtained by pure detection or pure tracking method, whose essence is the innovation of the tracking method.2. An adaptive scale feature compressed tracking algorithm is presented to reduce the dimension of sample features, and enhance the adaptability of the discriminative methods to the scale variations. The compress sensing theory is introduced into the object tracking. Firstly, a gaussian random matrix satisfied the Restricted Isometry Property(RIP) criterion is utilized to compress the feature and achieve dimensionality reduction. Secondly, the compressed feature is used for classifying. It not only benefits to reduce the amount of computation and improve processing performance, but also can characterize the target characteristics and ensure the tracking accuracy better due to the compressed features retaining most information of the original object features. Meanwhile, in order to adapt to the scale variations, a sample sets which reflect the location and scale variation as much as possible are obtained by structural constraints of sampling, so as to find the sample which is optimally match of the current target state while tracking.3. An online weighted feature selection object tracking method via maximizing the response difference of positive and negetive samples is proposed to solve the problem of feature redundancy while representing the object appearance model.It is not a linear relationship between the input feature amount and the output performance of classifier. More worse, not only huge computation will be consumed, but also the classifier’s performance would be aggravated as long as the amount of features exceeds a certain number. In this dissertation, multiple feature selectors(i.e. the weak classifier) are selected to compose of a strong classifier by a defined response difference function, the strong classifier is used to classify samples and the sample with the highest classifying score is regarded as the current tracking result. Different weight is endowed according to the distance and overlapping degree between the sample and object while selecting feature selectors to compose the strong classifier, so as to enhance the positive sample and suppress the negative sample and enhance the discriminating ability of the classifier on positive and negative samples response, which is able to capture the optimal positive sample to describe the current target state.4. An object tracking method based on the sparse representation of compressed feature is proposed to solve the problem of PCA subspace representation of compressed feature. The PCA subspace representation is able to enhance the descriptive ability of target appearance greatly and overcome the influence of noise and illumination variation via representing individual candidate object with the principal components of target template sets.The object tracking is considered as a sparse approximation problem of the compressed feature via sparse representation of the target appearance model with compressed feature subspace and trivial templates by a generative representation strategy and incremental subspace learning update mode.To solve the updating problem of target appearance model under occlusion condition, an inverse indicating strategy is proposed, which searches the image patch with maximal observation likelihood in the original image space according to the corresponding value in the compressed subspace. Compared with the appearance representation based on template sets or PCA subspace, it also needs to solve a sequence of regularized least squares problem, but the computational complexity is greatly decreaed due to the less dimensional superiority of the compressed feature. The proposed method performs with strong robustness against the influence of illumination variation, partial occlusion scale and pose change against several state-of-the-art methods.5. Two object tracking methods based on context auxiliary are proposed due to the superiority of context information in dealing with the problem of anti-occlusion and similar appearance disturbance.The first one is called anti-occlusion object tracking method based on two-level implicit shape model, which purpose is to solve the problem of locating object accurately under heavy occlusion.A two-level voting model is constructed by the two-level codebook, which come from the object itself and the surrounding target respectively. The codebook is endowed with different voting weights according to corresponding occlusion degree, so as to improve the locating precision under occlusion condition, whose essence is to take advantage of the sparse context to assist object trackingThe second one is the object tracking method based on Scale and Orientation Adaptive Spatio Temporal Context(SOASTC),which treats the object tracking as a problem of searching the extremum confidence map of the location likelihood under a bayesian framework, the prominent advantage is that it makes use of the Fast Fourier Transform(FFT) algorithm to accelerate computation while updating the spatio-temporal context model and obtaining the target location likelihood estimation, resulting in a high running speed. In order to estimate the scale and rotation angle and adapt the variation of target scale and orientation, the principal component analysis method is used to calculate the covariance matrix of the target region weighted image.The proposed method has strong resistibility to occlusion and illumination variation, and has a certain adaptability to such challenging factors as image fuzzy caused by abrupt motion, similar disturbance, nonrigid deformation and cluttered background. The essence is to make use of the dense context to assist object tracking.