Research On Human Behavior Detection Methods Based On Deep Learning With The Application In Public Buildings

In the safety maintenance of public buildings,video monitor plays an important role.Traditionally,video surveillance is mainly operated by people.Thus,there is no way to deal with the emergency events promptly.Lately,the artificial intelligence has been rolling up and become popular again.In video monitor,it is broadly applied.The research on detection methods for human abnormal behavior now is a hotspot as well as the key to the transformation and upgrading of video intelligent surveillance in building security.Traditional methods for pattern detection depend on manual work when designing features,consuming much time and the intelligence is low.However,the deep learning methods for behavior detection improve the deficiency that is equipped with high computational efficiency and excellent ability of generalization.In this study,research on the methods related is done and carried out in public buildings for human abnormal behavior detection.The following is the research content:Firstly,investigating the background,the basic theory as well as the research condition concerned for video monitor in the security surveillance of public buildings,and then summarize them.Secondly,one method for abnormal behavior detection based on the improved two-stream CNN is proposed.In view of the two-stream CNN network can not effectively use the time sequence information in videos,LSTM network is introduced into the two-stream network to realize the learning of the time sequence information by using its memory function.In the meantime,the output of the two-stream networks with weight assigned is fused for the improvement of performance.Finally,experiments concerned are done to analyze the model performance.Thirdly,to solve the problem that the two-stream CNN network can not effectively deal with the long-time segments,the structure of the two-stream network is unchanged.For the detection of human abnormal behavior,one method based on the improved TSN(Temporal Segments Networks)method is proposed.Considered that the distribution of behavior data in videos is unbalanced,in which the GAN network is introduced.The game learning between Generator and Discriminator is used to process the data.Moreover,for the further development of the detection effect of the method,the attention mechanism is lead in the network.Next,the features of the two networks above respectively are fused together.Then,the new feature obtained is put into the later network for further training.At last,experiments concerned are done to analyze the model performance.Finally,one behavior detection system is designed for the actual scene to detect the behavior of human in buildings.It can help prevent the occurrence of dangerous abnormal behavior.Thus,the matter can be handled in time.After test,it can be found that the system can detect the behavior of human in buildings effectively.