Algorithm Studies Of The Online Semi-supervised Infrared Tracking

This dissertation is based on a pre-research project of the** infrared tracking plat-form. On-line infrared tracking is one of the most challenge computer vision problems. It has a wide range of realistic applications:precision guidance, flight control, unmanned aerial vehicle reconnaissance, intelligent video surveillance, video-based human-computer interaction, intelligent vision navigation, and so on. Unfortunately, Because the contrast ratio of infrared images is usually lower compared to visible images, it is ineffective for existing methods to accurately detect the infrared target and track robustly in practice. The tracking-by-detection method, which is new in recent decades, is adopted as the basic in-frared tracking system. For the sampling subprocess, this dissertation focuses on how to reduce the error of labels to abate the drift problem in the process of infrared tracking. The contributions of the thesis are as follows:Firstly, based on the sufficient dimension reduction and semi-supervised learning, the theory of sufficient semi-supervised feature extraction is proposed in this dissertation for the first time. Then two sufficient semi-supervised feature extraction algorithms (the fusion re-finement (FR) method and the semi-supervised kernel fusion refinement (SemKFR) method) to extract the infrared image feature are designed. The former is a linear method, and the latter is a nonlinear method. Comparing with other feature extracting methods, these two methods proposed here extract features not only utilizing a small amount of labeled samples and many unlabeled samples but also aiming at minimizing the loss of information in the feature extracting process based on the Central Space (Section 2.5). So the features extract-ed can be more distinctive. The experiments in this thesis use the infrared images from the VOT-TIR2015 database to testify the effectiveness of the FR and SemKFR methods.Secondly, this dissertation studies a novel online semi-supervised feature extraction method, which is based on the semi-supervised rough common vector (SRCV) method, which is less vulnerable to the "small sample size problem". It has been observed that the sufficient semi-supervised methods proposed in last chapter are time consuming for the online feature extraction. Meanwhile, there exists a "small sample size problem" [1], which means the training samples are scanty, while the dimension of the image patches are large, in the initial phase of the online infrared tracking. In our experiment, although the training sample is small, the SRCV method can find the important features. Furthermore, in order to adapt for the online infrared tracking setting, we put forward the incremental SRCV (ISRCV) method to learn the features online. The ISRCV method utilizes the the random projection tree (RPTree) to approximate the manifold structure of the incremental sample, and an objective function is constructed from the RPTree, which can iteratively updated in the tracking process. The experiments on the VOT-TIR2015 database validate the effectiveness of the ISRCV method to track the infrared target.Thirdly, a side information based online semi-supervised feature extraction method -the incremental semi-supervised generalized common vector analysis (ISSGCVA) is pro-posed in this dissertation. In the tracking-by detection framework, even partly labeling a little samples may still introduce label mistakes. In the proposed ISSGCVA method, the sample need none labeled information available, instead some similar link sample pairs and dissimilar link sample pairs are leverage to extract features. In this way, the ISSGCVA method can reduce the label information's impact on the drift problem. Moreover, The ISS-GCVA method broadens the restrictions that the transformation matrix must lie in the null space of the similar scatter matrix and can be used regardless the similar scatter matrix is singular or not. Then based on the ISSGCVA method, an online infrared tracking system is constituted via combing the naive bayes classifier. The ISSGCVA based infrared tracking sysytem outperforms its competitors, demonstrating its effectiveness in the experiments.After that, a novel online semi-supervised method - the semi-supervised incremen-tal flexible manifold embedding (ISemFME) method, is presented as the classifier in the framework (Fig.1-2). The ISemFME classifying method not only inherits the advantage of the semi-supervised learning, which can exploit the labeled and unlabeled sample to learn the classification model, but can updates its parameters learned before to accommodate the changes of target appearance during the infrared tracking. In addition, a buffering strate-gy is introduced to reduce the time and space complexity of the ISemFME method. In the VOT-TIR2015 database experiments, the ISemFME based tracking system demonstrates outstanding performances.Last, in view of the infrared small target doesn't hold a certain shape, the performance of lots of feature extraction methods on the small target is disappointing. In this dissertation, a combined method - Enhanced one-bit transform (En1BT) is suggested for the infrared small target extraction. The simulation demonstrates the effectiveness of the EnlBT method to extract the infrared small target.Around the tracking-by-detection system, this dissertation studies two important problems in the infrared tracking:how to extract robust features of the target and how to construct the classification function in the environment with observation noises. The semi-supervised learning method based online infrared tracking system is proposed, and verified in the simulation experiments. The proposed online semi-supervised infrared tracking theory, model and algorithms have guiding significance for other computer vision and machine learning studies.