Robust Methods for Visual Tracking and Model

Title of dissertation Robust Methods for Visual Tracking and Model Alignment Hao Wu, Doctor of Philosophy, 2009 Directed by Professor Rama Chellappa Department of Electrical and Computer Engineering The ubiquitous presence of cameras and camera networks needs the development of robust visual analytics algorithms. As the building block of many video surveillance tasks, a robust visual tracking algorithm plays an important role in achieving the goal of automatic and robust surveillance. In order to maintain a persistent tracking of objects, it is critical to know when and where the tracking algorithm fails so that remedial measures can be taken to resume tracking. However, most present evaluation methods are ogg-line and based on manually labeled ground truth data. Online evaluation methods in the absence of ground truth are of urgent need. We propose a novel performance evaluation strategy for tracking systems based on particle filter using a time-reversed Markov chain. The posterior density of the time-reversed chain is computed and the distance between the prior density used to initialize the tracking algorithm and the time-reversed posterior density function forms the decision statistic for evaluation. This backward tracking-based performance evaluation strategy is also general enough to be applied to many other tracking algorithms. In this dissertation, we also present a new bidirectional tracking strategy to achieve better performance. Instead of looking only forward in the time domain, we incorporate both forward and backward processing of video frames using a time-reversibility constraint. This leads to a new minimization criterion that combines the forward and backward similarity functions and the distances of the state vectors between the forward and backward states of the tracker. The bidirectional track strategy significantly improves tracking robustness and accuracy. We illustrate the improvements due to the proposed approach for the popular KLT tracker and a search-based tracker. Some objects of interest in surveillance applications like faces have relatively stable structures, which allows us to build parameterized shape models to localize the objects more precisely. There are some widely used algorithms for model alignment; however, most of them suffer from the problem of converging to local extrema when used in practice. In this dissertation, we present a machine learning method to learn a scoring function without local extrema to guide the gradient descent/accent algorithm and find the optimal parameters of the shape model. The method is called Boosted Ranking Model (BRM). By arranging the training samples in some special structure, we feed them pairwise into the rank training algorithm and learn a strong ranking function from a pool of weak features. Theoretically, this method can learn a function with arbitrarily few local extrema as long as the training samples are dense and the representation ability of the features are good enough. The extensive experimental results show that our proposed algorithm, BRM, outperforms existing algorithms. Robust Methods for Visual Tracking and Model Alignment