Adaptive Metric Learning For Robust Motion Analysis

With the rapid development of science and technology, there is an increasing number of multimedia applications in our daily lives. It seems to be a heavy burden to digest this huge amount of multimedia, such as the images on the Internet, and the videos recorded by the surveillance cameras, which leads to tremendous demanding for high programme computer vision algorithms. Recently, because of the newly proposed feature extraction method and boosting/SVM classifier, motion estimation, pattern recognition and visual analysis have drawn widely attention from the society.motion estimation is one of the key components of other computer vision tasks. When video is regarded as a set of consecutive images, motion is the most remarkable characteris-tic of this video. In order to get a better result from motion estimation, we need to associate the identity objects during the processing. An appropriate matching strategy could largely alleviate the ambiguity.Most existing motion estimation methods employ a fixed pre-specified metric. How-ever, simply using a pre-specified metric is problematic and limited in practice. One phe-nomenon we often observe is that the closest match under a predefined metric in a given feature space may not be the true target of interest. Thus, in this case, it leads to a false positive match. Different from the methods in the literature, our work is targeted on a new learned and adaptive metric for motion estimation. This thesis contains the following three aspects:1. Supervised adaptive distance metric learning. To minimized the K-NN classifi-cation error, supervised adaptive distance metric learning method tries to find the optimal transformation to project the original feature space into another one param-eterized by a transformation kernel. In motion estimation, a problem that we always confronted is to apart the foreground from the background precisely. With the help of supervised adaptive distance metric learning, the distance between foreground and background image patches have been significantly enlarged.2. Distance metric learning under metric preservation. In many cases, the supervised information is not easy to obtain. What we have is merely the raw data extracted from image. To train a better metric based on these raw data, we propose a new dis-tance metric learning method, named metric preservation. It automatically merges the information of two related feature space, such that the matching performance for specific task can be largely improved. 3. Supervised metric preservation. The original metric preservation method can be further improved by imposing supervised information. The objective of supervised metric preservation can be manipulated and imposed class label for learning pro-cessing. By doing so, Supervised metric preservation not only inherits the infor-mation from the source data, but also enhance the discrimination power for motion estimation.The novelties of this thesis:1. Proposed the robust visual tracking method based on adaptive distance metric learn-ing. Distance metric is one of the most important issues of matching. Former researches treat distance metric and motion estimation as two consecutive but sep-arated steps. However, we proposed a new method that integrates them together, and derives a closed-form mathematic result for motion estimation under adaptive metric learning.2. Proposed the metric preservation method. Though lacking of label information, plenty of vision tasks can benefit from metric preservation. For super resolution, the high resolution affinity matrix can be preserved in low resolution feature space. For action recognition, the action pattern of certain people can be preserved for the pattern of the other person. Meanwhile, for motion analysis, metric preservation can be used to explore the relations between different angle of views.3. Proposed the tracking method for low resolution video under metric preservation. Since the limitation of storage space and resubmission, high resolution videos are not easy to obtained. Practically, motion estimation may be performed on low resolution videos. To track the target precisely and efficiently on low resolution videos, we proposed a new tracking method under metric preservation. It firstly proposes and attempts to address the tracking problem in low resolution videos, and ends up with a pretty well results.Based on developing new metric learning methods, we proposed several original visual tracking algorithms. The theoretical analysis and experimental results show that our proposed metric learning methods are adaptive and generally applied to a variety of vision tasks. Furthermore, we treat visual tracking as a solid application, and integrate metric learning methods into its framework. Robust tracker can be always expected under adaptive metric learning.