Dimensionality reduction for whole-body human motion recognition

We address the problem of feature space dimensionality reduction for the recognition of whole-body human action based on Hidden Markov Models. First, we describe how different features are derived from marker-based human motion capture and define a total number of 29 features with a total of 702 dimensions to describe human motion. We then propose a strategy for a systematic exploration of the space of possible subsets of these features and the identification of meaningful low-dimensional feature vectors for motion recognition. We evaluate our approach using a data set consisting of 353 motions grouped into 23 different types of whole-body actions. Our results show that a lower-dimensional feature space is sufficient to achieve a high motion recognition performance and that, using just four dimensions, we can achieve an accuracy of 94.76% on our data set, which is comparable to feature vectors that consider a much higher-dimensional feature like the joint angles.