Research On Human Behavior Recognition Method Based On Static And Dynamic History Sequence

Behavior recognition is an important research direction in the field of computer vision,which has important applications in intelligent monitoring,intelligent home,intelligent city and so on.Video based behavior recognition method has always been a research hotspot.In the classification of human behavior research objects,compared with gesture,gait and other single person behavior,interactive behavior and group behavior have certain challenges in spatiotemporal modeling and understanding complexity.Single person behavior and interaction behavior are studied from the key frame selection,feature extraction and network model to construct effective features and improve the recognition effect in this paper.First of all,the existing key frame selection rules will batch extract continuous frame images,which is difficult to describe the motion change process as a whole.Considering the description of the amount of motion in the process of behavior change,active motion trail is constructed based on the dynamic changes of pixel values between adjacent image frames,and the local maximum is extracted to select the frame image.Secondly,the existing motion feature extraction will take into account the different dimensions of the frame image,such as the global and local,the speed of motion change and so on.Considering the continuity of background information and motion information when the action changes,and the behavior information expression of segmented continuous optical flow,a method of constructing and extracting history sequence is proposed,which uses Static History Sequence(SHS)and Dynamic History Sequence(DHS)instead of optical flow,is used to capture each local subtle action as visual information.Finally,the method of grouping features and multi model fusion is proposed to construct PABi-GRU(Progressive Net-Attention-Bidirectional-GRU)network to realize feature extraction and serialization modeling,at the same time solve the modeling problem of complex time relationship in dense annotation video stream.In the process of modeling,the progressive net is proposed to represent the features.The stacked network structure not only extracts deeper information,but also implies the principle of gradual supervision from coarse to fine.Secondly,a pyramid pooling structure is introduced to effectively learn the feature expression between action instances through different levels of features.At the same time,in the construction of serialization model,the attention mechanism and bidirectional Gated Recurrent Unit are used to learn the features of video sequence.Through key frame selection,feature extraction and serialization network,the model is trained and analyzed in single person gesture and interaction behavior data.The experimental results show that the recognition accuracy of PABi-GRU network that realize multi group feature fusion is 88.9% and 83.2% respectively on the single gesture 20BN-JESTER and interactive action UT-interaction test set,which can realize the effective modeling of human behavior based on video sequence.