Transfer learning of human poses for action recognition

Abstract

In order to increase the success rate of a human action recognition system trained with limited labelled video sequences, we propose an approach which combines an efficient use of the scarce data and a transfer learning improvement. The efficient use of data is implemented using the Fuzzy Observation Hidden Markov Model so as to outperform the constraints of the classical approaches when training with small datasets. Additionally, we use a transfer learning procedure that takes advantage of the fact that some human body poses are shared among actions and then key poses can be trained from external sources. Thanks to this method we have improved the recognition performance in new action classes introduced in the target domain. In order to confirm the usefulness of the approach we have tested the performance using the IXMAS dataset as target domain and the ViHASi dataset as source domain. © 2013 Springer International Publishing.