Research And Design Of Deep Learning Based Behavior Identification Algorithm For Monitoring Video Of Financial Institutions

The application of behavior recognition is very extensive,covering human-computer interaction,intelligent monitoring and intelligent robots.While at present,behavior recognition algorithms are still far from largescale practical applications.Considering the complexity of the scene,the variability of the action,and the camera's movement,behavior recognition is a very challenging task.In this research context,this article conducts some research on behavior recognition algorithms based on deep learning and its application in monitoring videos of financial institutions.Many current behavior recognition algorithms mainly rely on the scene and image features of the video,and pay insufficient attention to the key human pose features used to distinguish behaviors.In addition,self-built datasets and some public datasets are often too small in scale,have the problem of insufficient data,which will lead to the training effect is not good enough,and the model is easy to overfit.In view of the shortcomings of the above existing studies,this thesis proposed a prototype framework for abnormal behavior detection by combining pose estimation algorithm and behavior recognition algorithm,which uses Open Pose to extract human body posture and can improve the accuracy of behavior recognition by about 7% on the self-built dataset.And the thesis made further research on the combination of pose estimation and behavior recognition,established a pose feature classification network and integrated other network branches,made the overall network accuracy higher than the basic network.For the problem of insufficient data volume,the thesis uses transfer learning to pre-train the network on a large-scale dataset and then fine-tune it on a small dataset,which can greatly improve the model results and shorten the training time,and this can effectively alleviate the problem of too small dataset.The experimental results show that by combining pose estimation and behavior recognition,and using pose features to improve the behavior recognition effect,the accuracy of the overall network is improved by 1.5% to 6.9% compared to the basic network,and the model effect is improved by 10% to 30% through transfer learning.