Abstract:

A crossdomain dictionary learning algorithm was applied to the recognition of human actions. By introducing the auxiliary domain data set, the algorithm learned from the original training set (target domain) together with the dictionary to obtain the dictionary pair, and then obtain the sparse coding of the action class, and effectively expanded diversity of training set. The algorithm is a combination of learning framework of dictionary learing and training classification, which can be used to classify the reconstruction errors in the learning process. The experiment was carried out under the condition of MATLAB simulation. The UCF YouTube data sets were regarded as the original training set, and the HMDB51 data sets were regarded as an auxiliary domain data set. 7 coherent human actions of the two data sets were selected to do the work of recognition according to the flow chart of the proposed algorithm. The crossdomain dictionary learning algorithm was compared with another two kinds of human actions recognition algorithms. The results show that the recognition rate of the proposed algorithm is significantly improved. It is proved that the crossdomain dictionary learning algorithm is effective in human actions recognition.

Key words: human action recognition, dense point trajectory, crossdomain dictionary learning, sparse representation