Semi-Supervised Learning Using Incremental Support Vector Machine and Extreme Value Theory in Gesture Data

A variety of problems are related to real-world gesture recognition, such as continuous data streams, concept drift, novel and outlier samples, noise, scarcity of manually labeled data, on-line classification and the fact that the same gesture may implement in different way. Two important features should be included in the classifier to overcome these problems, which are the ability of detecting the outlier and the ability to update itself incrementally. Since outliers affect the performance of the classifier, they should be excluded from subsequent classifier updates in semi-supervised scenarios. The updating should be done incrementally, i.e. the old data should not be required for the updating process. As the SVM classifier known to be accurate for both linearly and non-linearly separable data, we extend it here to work within a semi-supervised scenario. We present an SVM classifier that has efficient ability to detect outliers in a multi-class system using the extreme value theory. The experiments show superior of the proposed algorithm by accuracy, detecting the outliers and the computation time of incremental processing. The comparison of the experimental results on an unbalanced multi-class gesture database between the proposed algorithm and the SVDD classifier clearly shows the advantages of the proposed approach.