关键词: 动作识别, 局域性约束线性编码, 词典学习, 时间金字塔匹配, 深度图像

Abstract: Large intra-class variations of action features lead to low classification accuracy of action recognition, on the other hand, current algorithms exist drawbacks in computational complexity and extension of recognizable action classes. A method based on locality-constrained linear coding (LLC) for action recognition from depth images was proposed. In order to reduce the intra-class variations and increase classification accuracy, joints' positions, velocities and acceleration features were concatenated to form local action features, then LLC was used to calculate sparse representations of local action features. Analytical solution of LLC ensures computational speed of our method is up to 760 frames per second. Dictionary is composed by sub-dictionaries learned by K-means from features of each class separately, so global optimization is avoided during extending recognizable action classes. Moreover, to avoid classifier to be over-fitting, a dimensionality reduction method based on labels of dictionary items was proposed. The proposed method was evaluated on MSR-Action3D dataset captured by depth cameras. The experimental results show that the proposed approach achieves classification accuracy of 85.7%.

Key words: action recognition, locality-constrained linear coding, dictionary learning, temporal pyramid matching, depth images