Object recognition and visual tracking, whose main purpose is how to automatically recognize and track the objects instead of human, are currently the most actively researched areas of pattern recognition and computer vision. During many practical applications, object recognition and visual tracking are the challenging jobs owing to the difficulties arising from diverse appearance of objects, geometrical transformations, illumination changes, occlusions, low-image resolution and noises etc, and many theory and technical problems remain unsolved. Recently, statistical machine learning algorithms, such as Gaussian process model and kernel machines, have been successfully applied in object recognition and visual tracking areas. It is fully believed that novel statistical learning technologies have strong potential to further contribute to the development of object recognition and tracking. In this dissertation, we explore object recognition and visual tracking based on Gaussian processes model from three aspects: Discriminative methods for object category recognition are typically non-probabilistic, and predict class labels without directly providing an estimate of uncertainty, then do not have good generalization capabilities with limited training data. As a statistical machine learning method with considerable attention, Gaussian processes provides a principled, practical and probabilistic model conveniently modeling the uncertainty, and a well-founded framework for model selection, model learning and model prediction. The excellent classi¯cation performance of Gaussian process model can be widely used in object categories recognition. In the supervised learning framework, the posterior distribution of Gaussian processes can not be affected by unlabeled data, which makes the location of decision boundary not be in°uenced. Based on the various characteristics of Gaussian processes, we consider how to effectively expand this supervised learning method into the semi-supervised framework through incorporating unlabeled data, and build the better performance of classifier for object recognition. Tracking algorithms can be roughly classified into two categories: deterministic methods and stochastic methods. Deterministic methods are fast and efficient, but can not recover from the temporary tracking failures because of being sensitive to occlusion and clutter. By estimating the probability distribution and maintain multiple hypotheses in the state s...
修改评论