Human activity recognition using an ensemble of support vector machines

Numerous human action recognition algorithms have been developed and evaluated recently. However, the ensemble of classifiers to recognize actions, utilizing diverse feature sets, has remained untouched. Mixing the outputs of several classifiers decreases the risk of a weak choice of a learner or a set of features, and leads to having a more accurate and robust-applicable framework. The weakness of single classifiers becomes more evident when the problem difficulty increases, essentially while having numerous action types or resemblance of actions. In this paper, an ensemble of support vector machines (SVMs) is employed to improve the classification performance by fusing diverse features from different perspectives. The Dempster-Shafer fusion and product rule from the algebraic combiners have been utilized to combine the outputs of single classifiers. The experimental results show that the action recognition performance is improved while employing the ensemble of SVMs and stated fusion techniques.