Human Action Recognition Based On The Representation And Measurement Of 3D Skeleton Snippet Using Manifold Division

Human action recognition has been one of the most popular research in computer vision field for its wide applications including video surveillance,human-computer en-tertainment,health care and social assistance,and so on.The ability of machines to anal-yse and recognize actions in videos is the main objective of human action recognition.However,the inherent attributes of a human action video such as background clutter,partial occlusion,changes in scale,viewpoint,lighting,and appearance make this task more challenge.Previous work focused on extracting local space-time features from 2D frame images,recently,the emergence of depth camera such as Microsoft Kinect and its corresponding real-time skeleton extraction method promote the study from both aspects of depth maps-based methods and skeleton-based methods.Compared with 2D frame images,depth images reflect pure geometric or shape clues and are insensitive to lighting conditions,which are more robust in practical application.Different from approaches based on features extracted from RGB or depth im-ages,3D skeleton information is used in this paper to represent human action.On the one hand,3D skeleton information outperform other low-level appearance features for human action recognition.On the other hand,pose-based features is a natural and in-trinsic representation as it conforms to the study of how humans understand actions.In this paper,the locality of a human action is represented as postures according to relative 3D skeleton information,and sequence of postures are regarded as a human action man-ifold.Two human action recognition methods named human action recognition based on 3D skeleton snippet representation using manifold hierarchical division and human action recognition based on 3D skeleton snippet distance measurement using manifold serialized division are proposed.Human action recognition based on video with its ex-tracted human body 3D skeleton position information is the main problem of this paper.The solution is divided into three steps including the video division,action-snippet rep-resentation and the process of global temporal relationship.The main contributions of this paper includes the following aspects:(1)Human body 3D skeleton information is used to represent human action,and human action(sequence of human postures)is regarded as manifold.(2)According to 3D skeleton information,major posture feature and main dynamic tendency feature are proposed to describe a human action(snippet),besides,their distance measurement method is also defined.(3)The local linearity of human action manifold is computed for posture sequence division,which benefits the purpose of recognition and classification.(4)Two human action recognition approaches are proposed,and several popular hu-man action datasets are used to validate the effectiveness and efficiency.