Abstract
In this paper, we propose an online recognition method for daily actions, such as walking and standing. The proposed method has following characteristics: (1) simultaneous recognition that is able to output multiple action names when human act more than one action, such a situation ashuman is waving hand on standing, (2) modeling action classifiers with kernel methods, (3) effective optimization for the parameters of the recognition system with margin-based query learning. The characteristic (2) unifies the process for modeling and learning the classifiers, and makes us easy to incorporate prior knowledge about action. The characteristic (3) reduces the burden of process for annotating action, which is an inevitable task for supervised learning. The experimental results using real motion capture data show that the proposed margin-based query learning is very effective to achieve high performance of the recognition system with very small sized query and annotation process.
