Sequential discriminant error minimization: The theory and its application to real-time video object recognition.

摘要

In the context of object detection for intelligent video surveillance, a number of factors have to be taken into account. Two crucial factors are the ability to learn multi-class pattern classifiers with noisy (high Bayes error rate) data and the ability to implement these classifiers on platforms with varying computational capabilities. Recent empirical and theoretical results in discriminative (as opposed to probabilistic) learning algorithms for classification, have proven that directly solving a classification task leads to better generalization and lower complexity requirements than solving an intermediate regression task. The key differences between the various discriminative learning algorithms lie in the precise technique and objective function used to directly maximize generalization performance. Typically, the method of choice has been to minimize the upper bound on the probability of error of a classifier along with some regularization. This thesis suggests an augmented method: minimize the mean squared discriminant error (MSDE) of a classifier. A learning algorithm minimizing MSDE lends itself to estimation-theoretic analysis of its properties. The property of importance is the relative statistical efficiency, i.e., dominance of a learning algorithm in estimating the Bayes-optimal partition of the data. Just as in the case of the bias-variance decomposition of the MSE objective function for regression problems, there exists as associated decomposition of MSDE for classification problems. Minimizing MSDE in a model-constructive framework leads to the formulation of the sequential discriminant error minimization (SDEM) algorithm. This thesis outlines a strategy to derive, implement and analyze SDEM in the context of an automated real-time video surveillance application.