Online action recognition from RGB-D cameras based on reduced basis decomposition

• Published in: Journal of real-time image processing Vol. 17; no. 2; pp. 341 - 356
• Main Authors: Arunraj, Muniandi, Srinivasan, Andy, Vimala Juliet, A.
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.04.2020; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Approximation; Cameras; Classification; Classifiers; Collaboration; Computer Graphics; Computer Science; Datasets; Decomposition; Gesture recognition; Human activity recognition; Image Processing and Computer Vision; Methods; Multimedia Information Systems; Original Research Paper; Partial differential equations; Pattern Recognition; Reduction; Regularization; Sensors; Signal,Image and Speech Processing; Statistical analysis; Reduced basis decomposition; Greedy algorithm; Human action recognition; Collaborative representation classifier; Depth motion map
• ISSN: 1861-8200; 1861-8219
• DOI: 10.1007/s11554-018-0778-8

Human action recognition from RGB-D cameras has recently become one of the major fields of research. While accuracy improvement was given more importance in previous action/gesture recognition methods, there are opportunities to work on improving the computational efficiency too. This paper introduces an efficient dimensionality reduction technique and classification mechanism to recognize actions from depth motion map features. For our proposed work, a recently introduced technique called reduced basis decomposition (RBD) is employed, which manages faster dimensional reduction with its unique mechanism of generating compressed basis vectors. The RBD has an offline error-determination and an online approximation mechanism, and it is faster than PCA/SVD. For classification, this paper employs a Probabilistic Collaborative Representation Classifier (Pro-CRC). The recommended classifier works based on probability in connection with l 2 -regularization. The combined effect of the methods above helps in achieving the state-of-the-art efficiency. In the standard protocol tests carried out in the MSR-Action3D dataset, our proposed method achieved a considerable accuracy of 91.7% which is better than the currently efficient method. Further, our proposed method also proved its effectiveness in the challenging, subject-generic test with a reported accuracy of 89.64% and an average accuracy of 85.70% in the cross fixed tests which included 252 combinations of all the subjects without repetition.