Human Action Recognition Using Markov Random Walk Based Semi-supervised Learning

To accommodate the great challenge of obtaining accurate and detailed annotations of training data in human action recognition,in this paper we propose a semi-supervised algorithm based on Markov random walk.We first calculate mesh features of each image in the human action sequence.Then the mesh features are vector quantized through a rival penalized competitive neural network and the actions described by time-sequential images are converted into symbolic sequences.We then construct a Markov chain to reveal the labeled actions,unlabeled actions and action categories in training set according to their edit distance.Furthermore,we predict class label of unlabeled actions through a procedure of Markov random walk.The test sequence is then classified by the maximum posteriori probability criteria.The experiments on Weizmann dataset demonstrate the effectiveness of our method.The average recognition accuracy can exceed 80% even when only very small amount of labeled actions acquired.