Learning Cross-domain Dictionary Pairs for Human Action Recognition

This paper present a cross domain dictionary learning way, via the introduction of auxiliary domain, as the extra knowledge, the intra class diversity of the original training set (also known as the targetdomain) is effectively enhanced. Firstly, use local motion pattern feature as a low-level feature descriptor, and then through a cross domain reconstructive dictionary pairlearning, which brings the original target data and the auxiliary domain data into the same feature space to get correspondingsparse codes of each human action categories. Finally, classification and recognition is carried on thehuman action representation. Usingthe UCF YouTubedataset as theoriginal training setand the HMDB51 data set asthe auxiliary data set, the recognition rate of the proposed framework is significantly improvedon the UCF YouTube dataset.