Researches On Robust Visual Object Tracking Algorithms In Image Sequence

With the popularity of image acquisition equipment, computer vision which studies digital image has undergone an unprecedented development. Object tracking is one core technology of intelligent surveillance, virtual reality and robot vision. It refers to the frontier knowledge of computer science, mathematics and psychology etc. As an inportant part of the computer vision, many researchers from China and abroad work on visual tracking research. But because of complex environments and appearance changes of objects, it is also a challenging study to develop a robust visual tracking algorithm. This paper firstly reviews the classical tracking algorithms, secondly proposes a series of strategies to improve the performance aim at the deficiencies in these classical trackers. Our main research approaches and innovations are listed bellow:1. We present a multi-subspace learning tracking algorithm based on maximum posterior probability. Tracking algorithms based on PCA(Principal Component Analy-sis)linear subspace learning assumes that the object is generated from the same subspace and the error term is subject to Gaussian distribution with a small variance. But these algorithms do not consider the distribution of samples in the subspace, which often re-sults in over-fitting when minimizing the reconstruction error. To solve the over-fitting problem, we first analyze PCA subspace learning algorithms from the maximum like-lihood estimation perspective, then deducing the unbiased estimation of the variance of the subspace and at last proposing a new PCA subspace learning algorithm based on maximum posteriori probability(MAP) method. Tracking algorithms based on PCA only using construction error as the criteria of silimarity but neglecte that the con-struction error can reflect the difference between target appearance and training model. Base on such observation, we propose a local patch model. First, we divide the object image into serval patches. It is assumed that every patch is generated from the inde-pendent subspace. Then, we use the reconstruction error to judge whether a patch is occluded. Finally, computing particles'weight and training new subspace use only the unoccluded patches, which avoids the adverse effect caused by occlusion. Experiments show that the proposed algorithm can effectively solve the influencing factors such as partial occlusions, motion blurs and background interferences.2. We propose a multiple classifiers tracking algorithm based on cluster similarity comparison. A normalized feature space is established according to the variance of the cluster. The target in a new frame could be get by computing the distance between test samples and the center of the cluster. Due to the changes of target and background during tracking, single classifier tracking algorithm based on cluster similarity learns a lot of non-target information which result in the decrease of tracking accuracy. To solve this problem, in this paper, we propose to use tree structure to save former classifiers as a set. In each frame, a subset of classifiers is chosen according to the path in the tree to classify test samples. Experiments demonstrate that such stragety could improve tracking precision of cluster similarity tracker.3. We improve the compressive tracking algorithm via feature selection stragety. Recently, compressive tracking (CT) was proposed for its efficiency, accuracy and ro-bustness on many challenging sequences. Its appearance model employs non-adaptive random projections that preserve the structure of the image feature space. A very sparse measurement matrix is used to extract features by multiplying it with the feature vector of the image patch. An adaptive Bayes classifier is trained using both positive samples and negative samples to separate the target from background. On the CT framework, however, some features used for classification have weak discriminative abilities, which reduce the accuracy of the strong classifier. In this paper, we present an online com-pressive feature selection algorithm(CFS) based on the CT framework. It selects the features which have the largest margin when use them to classify positive samples and negative samples. For features that are not selected, we define a random learning rate to update them slowly. It makes those weak classifiers preserve more target information, which relieves the drift when the appearance of the target changes heavily. There-fore, the classifier trained with those discriminative features couples its score in many challenging sequences, which leads to a more robust tracker.4. This paper presents a cooperative decision-making tracking algorithm which integrates two complementary trackers to track the object with large appearance change and short time occlusion. We also propose a new concept of "reliability" which is used to evaluate tracking results. This concept could extend to other tracking algorithms and improve their tracking performance. Firstly, an improved Bayesian tracker(B-tracker) with adaptive learning rate is presented. The classification score of B-tracker reflects tracking reliability, and a low score usually results from unreliable result. Therefore, if the score is low, we decrease the learning rate to update the classifier fast so that B-tracker can adapt to the variation and vice versa. In this way, B-tracker is more suitable than its traditional version to solve appearance change problem. Secondly, we present an improved incremental subspace learning method tracker (S-tracker). The reconstruction error of S-tracker is used to judge reliability. When the tracking result is unreliable, we present a stop-strategy to suspend update once occlusion occurs. This strategy enables S-tracker to retain more initial target information so that the target can be relocated after it reappears in the scene. Finally, we present a integrated framework(BAST), in which the pair of trackers run in parallel and return two candidate target states separately. For each candidate state, we define a tracking reliability metrics to measure whether the candidate state is reliable or not, and the reliable candidate state will be chosen as the target state at the end of each frame.