A real-time human motion recognition system using topic model and SVM

Human motion recognition is a challenging task, especially when motion capture data is huge. Existing approaches for this task focused mainly on how to extract features from motion capture data to achieve high recognition performance. However, due to the presence of redundant features and the high dimensionality of data, these approaches may not achieve the optimal performance. In order to rapidly and accurately recognize human motion, we present a novel method based on topic model and SVM. It consists of first extracting informative angle features at each frame of motion sequence, which is helpful to eliminate redundant information. Then, significant motion frames are transformed into motion-words to form motion sequences. A motion sequence is a combination of motion-words, which are obtained by extracting key poses and applying hierarchical clustering. Then, topic model are applied to obtain the underlying topic distribution to perform motion recognition. We studied 9 common motions for 13 subjects. The results show the best recognition accuracy achieves 98.41% which outperforms state-of-the-art methods.