Research On Video Action Recognition Based On Transfer Learning

Human action recognition is a hotspot and one of the most difficult research areas in computer vision.It has broad application prospect in video surveillance,video retrieval,human-computer interaction,movement analysis and so on.Transfer learning has proven to be an effective solution for cross-domain action recognition,which can diminish the difference of data distribution in different domains,and solve the problem about the shortage of training samples in target action videos.Based on this,this paper propose two effective transfer learning algorithms separately to fulfil the cross domain action recognition task.The first transfer learning algorithm for action recognition is based on domain similarity,which utilizes relevant data from other domains to enhance the original learning system.First,by using canonical correlation analysis to add additional constraints on the target classifiers,we can leverage the useful knowledge from the related domains.Then,we learn a reconstructive,discriminative and domain-adaptive cross-domain dictionary pair to map data from different domains into a same abstract space.Finally,the human behavior is classified according the mapping features and classification model.We manually leverage images from Web pages,and evaluate the proposed method for human action recognition on the UCF Sports Action dataset,achieving effective results.The second transfer learning algorithm for action recognition is combined with semi-supervised learning,which uses multiple base kernels to map different domains and utilizes semi-supervised learning to exploit label information in the target domain.First,feature extraction is performed on motion images and videos.We extract the static features of images based on the spatial pyramid method,and extract the motion features of videos based on the dense trajectory method.The features is reduced by Multidimensional Scaling,which removes the redundant information.Then,we combine the Maximum Mean Discrepancy with multiple kernel learning as a process of transfer learning to reduce the mismatch of data distributions between different domains.Finally,a semi-supervised learning process combined with k-Nearest Neighbor is incorporated,making full use of labeled and un-labeled samples in all domains to achieve human action recognition.Experimental results on multi-group data sets show the stability and effectiveness of the proposed algorithm.