Mode-Driven Volume Analysis Based on Correlation of Time Series

Tensor analysis is widely used for face recognition and action recognition. In this paper, a mode-driven discriminant analysis (MDA) in tensor subspace is proposed for visual recognition. For training, we treat each sample as an N-order tensor, of which the first N-1 modes capture the spatial information of images while the N-th mode captures the sequential patterns of images. We employ Fisher criteria on the first N-1 modes to extract discriminative features of the visual information. After that, considering the correlation of adjacent frames in the sequence, i.e., the current frame and its former and latter ones, we update the sequence by calculating the correlation of triple adjacent frames, then perform discriminant analysis on the N-th mode. The alternating projection procedure of MDA converges and is convex with different initial values of the transformation matrices. Such hybrid tensor subspace learning scheme may sufficiently preserve both discrete and continuous distributions information of action videos in lower dimensional spaces to boost discriminant power. Experiments on the MSR action 3D database, KTH database and ETH database showed that our algorithm MDA outperformed other tensor-based methods in terms of accuracy and is competitive considering the time efficiency. Besides, it is robust to deal with the damaged and self-occluded action silhouettes and RGB object images in various viewing angles.