Human Action Recognition in Unconstrained Trimmed Videos Using Residual Attention Network and Joints Path Signature

Action recognition has been achieved great progress in recent years because of better feature representation learning and classification technology like convolutional neural networks (CNNs). However, most current deep learning approaches treat the action recognition as a black box, ignoring the specific domain knowledge of action itself. In this paper, by analyzing the characteristics of different actions, we proposed a new framework that involves residual-attention module and joint path-signature feature (JPSF) representation framework. The path signature theory was developed recently in the field of rough path and stochastic analysis, which provides a very efficient way to analyze any temporal sequence data. The proposed n-fold joint path signature features entail the Euclidean distances between joints and respective angles. For our experiment, JPSF for three modalities of joints (spatial location, bi-folds and tri-folds) are computed over the temporal length of action sequence. Then all these PSF are concatenated and fed to a CNN to give the recognition result. Experiments on three benchmark datasets, J-HMDB, HMDB-51 and UCF-101, indicate that our proposed method achieves state-of-the-art performance.