Self-Taught Learning Features for Human Action Recognition

Abstract

Human action recognition has been an import subject in computer vision with its application in robotics, video surveillance, human-computer interaction, user interface design and multimedia video retrieval. But it is also a challenge work for the complex of feature extracting and limited labeled data. Previous feature extracting approaches are almost manual features, such as HOG, SIFT, HOF and so on. These manual features receive significant results in some specific applications. But algorithm's generality is a problem. Moreover, they need mass of labeled data to do training for recognition. In this paper, we propose self-taught learning features and unsupervised learning pre-processing. According to this method, we extract feature through unsupervised self-taught with large group of unlabeled data, then fine-turning with small number of labeled data. And the Soft max regression can classify and recognition human action. Experiments demonstrate that our method can receive outperform results.