Human Body Action Recognition Based On Joint Data And Extreme Learning Machine

With the development of technology, the human motion capture data is widely applied in a lot of fields such as human-machine interaction, interactive entertainment as well as education, health care and others, due to its ability to maintain the details and the strong sense of reality. And human action recognition, as a problem in the field of computer vision, becomes the key technology of somatosensory games, security and the retrieval of multimedia information, so it’s critical to increase the performance of human action recognition.In this thesis, we improve the feature and classifier based on the existing feature descriptor of human action. And we experimented both on Microsoft MSR-Action3D dataset and the HDM05 Mo-cap dataset of Bonn University, we got good classification results. The main work and the contributions of this thesis is shown as follows:(1)Based on the Covariance Descriptor and the Histogram of Oriented Displacement (HOD) Descriptor, we created a new descriptor containing both the static information of the joints position in every frame and the dynamic information which shows the position change of the same joint in different frame, by combining the two descriptors linearly. And we got better classification results with the Extreme Leaning Machine, ELM.(2)We improved the ELM to classify the action feature by introducing the linear regression classifier to the ELM. We experimented on the two data sets, and got better recognition results compared with the original ELM algorithm.(3)We introduced the idea of Sparse Representation to the ELM to classify the human action samples, and used the Sparse Representation Classifier to classify the samples which is randomly projected to the high-dimensional space. We did the comparative experiments on the MSR-Action3D data set, and got better recognition performance when the number of hidden layer units is big.(4)In order to improve the anti-over-fitting capability of the ELM, in which the number of hidden nodes is large, we applied the Dropout strategy to the ELM. We did the comparative experiments on the MSR-Action3D data set, and the results showed that the algorithm is better than the original ELM when the number of hidden layer units is large.