MLRMV: Multi-layer representation for multi-view action recognition

Daily action recognition has gained much interest in computer vision. However, viewpoint changes will lead to sizable intra-class differences in the same action. To deal with this problem, we propose a novel multi-view daily action recognition approach based on the multi-layer representation. In use of motion atoms and motion phrases, we construct the middle-level feature representations in multi-view daily actions. A multi-view unsupervised discriminative clustering method is proposed for constructing motion atoms, and the classification accuracy of motion atoms is improved by jointly learning atom dictionaries and the classifier. Moreover, we present discontinuous temporal scale motion phrases and a grading mechanism of motion phrases to strengthen the representative ability of motion phrases and the final recognition accuracy. Finally, the experimental results based on the WVU dataset, the NTU RGB-D dataset, and N-UCLA dataset show that the proposed methods have the state-of-the-art performance, compared with the classic methods such as IDT, MoFAP, JLMF, and so on. (c) 2021 Elsevier B.V. All rights reserved.