Human Action Recognition Based On 3D Skeletons

Human action recognition has become more and more popular research in recent years,which belongs to the field of computer vision and behavior pattern recognition.Human action recognition is widely used in human-computer interaction,security intelligent monitoring,video retrieval based on content and other related fields and make a great practical value.During the construction of information social,human action recognition has been more and more important.At the same time,human action recognition can also promote the other fields of computer vision,such as face recognition,gait analysis and so on.In the past,most of human action recognition method is based on 2D video and images without ideal effect due to the limitation of technology.With the development and popularization of 3D technology such as depth camera,human action recognition has gradually developed from plane to facade.3D human action recognition has become the main research method to recognize human's action.The study of this paper is made in this case.In the process of collecting motion characteristics,different movement speed will lead to a large difference of the data in the time series.Therefore,in the process of human action recognition,the alignment of data is an important factor affecting the accuracy of identification.The research of this paper mainly focuses on the alignment and correction of action characteristics.In this paper,we use a method to describe the behavior by Lie algebra.In order to make this method get better results,this paper attempts to apply the Deep Canonical Time Warping to the Lie Algebra relative pairs representation.The Deep Canonical Time Warping uses deep learning to align the data,and can help to get a better effect when applying to the alignment of Lie Algebra relative pairs representation.At the same time,in order to reduce the time of calculation and to improves the experimental results,this paper makes some changes to the network structure of Deep Canonical Time Warping and adjusts the test samples by correlation.