| In the past years,many single-modality learning methods for RGB human action recognition have been proposed.However,the human action recognition of RGB was easily influenced by some factors which restrict the development of researching work further,such as invariant to lighting and background variations.In order to reduce the interference of these factors,many researchers have paid attention to human motion recognition based on multi-view,multi-modal and cross-domain.Therefore,the paper is composed of three parts as follow: 1)proposing a collaborative sparse representation leaning model(CSR)for RGBD action recognition;2)analyzing and evaluating that different cross-domain learning algorithms exert an effect on open-set human action recognition;3)a real-time action recognition system based on video is designed and developed.Specific work as follows:1)Collaborative sparse representation leaning model for RGBD action recognition.In order to make use of the RGB features and the depth features of the videos simultaneously in human action recognition,therefore,the collaborative sparse representation leaning model(CSR)for RGBD action recognition which adaptively fuses different modal features is proposed.Firstly,the dense trajectory features from the each RGB video is extracted,and the bag-of-words approach is utilized to RGB features;for each depth video,the CNN features was extracted,the temporal model that fetch timing information is utilized,and the principal component analysis(PCA)method is adopted for decreasing the feature dimension.Then,the fusion and reconstruction of multi-modality features are unified into an objective function,meanwhile,the difference of reconstruction sparse coefficients between RGB features and depth features is controlled,so as to fuse the features of different modality adaptively.Large scale experimental results on challenging and public DHA,M2 I and Northwestern-UCLA action datasets show that the performances of our model on two modalities are much better than traditional sole modality.2)Analyzing and evaluating that different cross-domain learning algorithms exert an effect on open-set human action recognition.Cross-domain learning problem aims to leverage the small-scale data from target domain together with a large-scale data from an auxiliary domain to augment the generalization ability for model learning.Therefore,in this chapter,cross domain learning algorithm is applied to improve the performance of action recognition in open-set human action recognition,meanwhile,the performance of six different cross domain learning algorithms is evaluated under different auxiliary domain and target domain.Finally,we compare the performance of cross-domain learning to the non cross-domain learning on open-set human action recognition,it can be observe that regularized multi-task leaning algorithm have superiority and robustness on open-set human action recognition,however,the performance of cross-domain learning on open-set human action recognition will be difference when the dataset has different auxiliary domain and the target domain.3)A real-time action recognition system based on video is designed and developed.In order to achieve practical application of the human action recognition algorithm,in the chapter,an real-time action recognition system based on video is designed and developed for accurate and real-time action recognition,which base on the theoretical basis of computer vision and pattern recognition,and the object-oriented programming thought is adopted.In the system design and development,the human action recognition algorithm such ad dense track feature extraction,feature coding and SVM classification are utilized,meanwhile,in order to improve the recognition velocity,the concept of video window and multithreading technology are introduced.The system test results on the KTH dataset show that the system can detect the human action in the video in real time and accurately. |
PDF Full Text Request