Graph-Based Action Models for Human Motion Classification

Recognizing human actions is an important ability for service and domestic robots. This paper presents a novel approach for learning and recognizing motion models from human motion capturing data. The key idea is to represent observed motion trajectories as a graph, where the nodes correspond to poses and the edges indicate pose similarities. We optimize this graph using least squares minimization and non-maximum suppression to obtain a generalized model for the respective action. The resulting motion models can then be used to recognize actions in unlabeled motion capturing data. Experiments based on real-world data show that the learned motion models can reliably classify a large set of different motions. Furthermore, we show that the learned models robustly generalize over different people.