Open AccessHuman action recognition and analysis algorithm for fixed and moving cameras
Author(s) -
Abdelwahab M.A.,
Abdelwahab M.M.
Publication year - 2015
Publication title -
electronics letters
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.375
H-Index - 146
eISSN - 1350-911X
pISSN - 0013-5194
DOI - 10.1049/el.2015.1037
Subject(s) - optical flow , principal component analysis , artificial intelligence , computer science , feature (linguistics) , pixel , action recognition , pattern recognition (psychology) , translation (biology) , relation (database) , action (physics) , computer vision , algorithm , fourier transform , eigenvalues and eigenvectors , discrete fourier transform (general) , image (mathematics) , mathematics , fourier analysis , data mining , short time fourier transform , class (philosophy) , mathematical analysis , philosophy , linguistics , biochemistry , chemistry , physics , quantum mechanics , messenger rna , gene
A new algorithm for human action recognition is presented. The use of both front and side views of the optical flow (OF) in multiple layers representing different angles is proposed. The side view of the OF, created from the frontal view, is introduced as a new feature. It improves recognition accuracy and provides more information about the action such as the number of repetitions. Two‐dimensional (2D) discrete Fourier transform is applied to the obtained OF features that makes the algorithm not sensitive to translation and alignment. 2D principal component analysis is used to extract features from the eigenspace maintaining the spatial relation between pixels and increases the recognition accuracy. Results of experiments performed on four diverse datasets, Weizmann, IXMAS, KTH, and UCF sports, representing fixed and moving cameras, confirm these excellent properties compared with recent reported methods.