| In the era of social media, social security is an important topic such as public surveillance. Even tons of web cameras are widely used in public places, e.g., bank, square and airport, it still needs lots of time and labors to keep attention. Therefore, how to recognize human action in a specific circumstance with a complex background is a basic question for social security problem.;In this thesis, we focus on the social security problem, in particular human action recognition, and give the analystics in two lines, (1) machine learning algorithms for action recognition, (2) applying algorithms for novel problems in action recognition, e.g., missing-modality problem, dimensionality reduction. These two lines are detailed in following.;For machine learning algorithm, extracting features from high-dimensional action data is crucial in human action recognition. The usual approach is finding a subspace, i.e., projecting high-dimensional data into a low-dimensional subspace containing main pattern of original data and fewer variables, for classification. First of all, data representation is crucial for action video which contains spatiotemporal information. To this end, we propose high-order tensor to represent the action videos, and employ tensor decomposition methods for dimensionality reduction. Second, different problems in action recognition tasks are solved by machine learning algorithms, such as transfer learning, low-rank learning, manifold learning.;Low-rank tensor learning is appropriate for some problems in action recognition task such as missing-modality problem, intra-class diversity problem and manual dimensional setting problem. In missing-modality problem, training ans testing data are different modalities, which may not obtain good performance in the test phase. Intra-class diversity problem contains three situations, such as multi-subject, multi-modality and sub-action challenges, which will decrease the performance of recognition. Sub-action problems indicates some body poses are different even for the same action, e.g., drinking action by standing person and drinking by sitting person. |
