Research On Cross-domain Action Recognition Via Transfer Learning

Action recognition has received much attention as it is an active research topic of computer vision in recent years. It has broad application prospect and high economic value, which can be used for various applications in video-based surveillance, video-based retrieval, human-computer interaction, military, security, medical diagnosis, industrial inspection, traffic management, motion analysis, and biology.To obtain a reliable classification model, traditional classification methods are built upon two assumptions:one is training instances and test instances must have the same distribution, and, the other is training instances should be sufficient. For a new domain, if training instances are insufficient, traditional classification methods are not able to produce a good classification model. This problem can be solved with transfer learning. Transfer learning approaches can improve the performance of the target domain by using some knowledge of the source domain which is relevant to the target domain. However, data from different domains may have different feature distribution. To address this issue, some instances should be selected from the source domain which are similar to the instances of the target domain, or mapping data from different domains into the same abstract space. Then the training phase can be performed,In order to solve the problem of data from different domains may have different feature distribution, we propose a method named collective matrix factorization with graph Laplacian regularization for cross-domain action recognition. Our method employs collective matrix factorization to map data from different domains into the same abstract space. To further enhance the discriminative power of the learned semantic features, we integrate a graph Laplacian regularization term to exploit the label consistency across different domains and the local manifold structure of each domain. Our method jointly learn a common latent space, two linear projection matrices and a linear classifier. For a testing instance, we can obtain the latent semantic representation by the linear projection matrix directly and predict its label by the learned classifier. In order to verify the effectiveness of the algorithm, we selected UCF-101 dataset as the target domain, HMDB51 dataset as the source domain, experimental results verify that this algorithm can achieve good recognition accuracy.