Max-Margin Methods And Their Applications In Activity Analysis

Margin is often used to measure confidence of classifier in machine learning. And Max-Margin algorithm in support vector machine becomes a hotspot in the field of machine learning, due to its strong theory guarantee and excellent performace in practice. The most common activity in daily life is human interaction. So researching the method of human interaction recognition has very important practical significance. This paper proposes two kinds of methods to recognise human interaction based on max-margin algorithm.The first method is human interaction recognition based on max-margin conditional random field. The conditional random field parameters can be obtained by maximum likelihood method traditionally. But this method has poor generalization ability and can not support kernel function, so it can not be used in high dimensional feature space. In response, this paper proposes max-margin conditional random field. Specific optimization implementation uses block-coordinate Frank-Wolfe algorithm. This algorithm is a variant of the classic Frank-Wolfe algorithm, and it is online. This algorithm need not to choose stepsize, which can be computed by a closed form. Meanwhile this algorithm has a dual margin guarantee, which can be used as iteration stopping criterion. In addition, the convergence rate remains stable even though it employs approximate max-inference function. The method that employs max-margin algorithm to learn parameters retains the structured advantage of conditional random field and takes advantage of max-margin algorithm.The second method is human interaction recognition based on max-margin markov network. A new hierarchical recognition method is put forward. Employing markov network models the high-layer semantics. Relative to single layer recognition of the first method, there are four layers in this progressing. The first is single tracking, which is obtained by combining the local sparse structured model and variable template updating strategy algorithm. The second is feature extraction. Silhouette, optical flow and context feature are extracted respectively. Then final feature is acquired by combining the three features. The third is atomic activity recognition by Large Margin Nearest Neighbors. Finally, training max-margin markov network and template can get the model of interaction, which can model the atomic semantics to obtain final interaction activity.The experimental results on UT datasets demonstrate that the two methods are both effective.