Research On The Identification Method Of Suspect Dangerous Behavior In Public Places Based On Deep Learning

At present,deep learning theory has many practical applications in computer vision,speech recognition,natural language processing and man-machine game.The development of deep learning has greatly improved the effect of behavior recognition based on video information.However,as the parameters of the deep neural network model based on video data are more than those of other models,the requirements for hardware computing resources are higher,so the practical application in the field of intelligent robot is a great problem.To this end,this paper improves the 3D convolutional neural network model,reduces the parameter setting,and thus improves the training and testing speed of the deep neural network.At the same time,for the application scenario of this topic,the improved 3D convolutional neural network was applied to the data set of suspect dangerous behavior recognition with migration learning method,achieving a high accuracy rate.Firstly,according to the mechanism of 3D convolutional neural network model,this paper proposes an improved 3D convolutional neural network model by dividing 3D convolutional kernel and introducing the idea of residual network.Through experimental verification,the model is compared with the improved 3D convolutional neural network model,and the results show that it effectively reduces the training and testing time,and reduces the requirement of the neural network model on hardware computing resources.Secondly,aiming at the specific content of behavior recognition research of this topic,this paper selects the public place of railway station square as the video shooting site to collect the data set of suspect’s dangerous behavior recognition.This data set simulates some dangerous behaviors of suspects that may occur in public places and is recorded into video data.Finally,five categories of dangerous behaviors were established: "normal behavior","armed behavior","fighting","robbery" and "holding".This data set effectively restores dangerous behaviors in real scenes and has certain practicability.Finally,aiming at the problem that the number of data set samples constructed in this paper is small,the migration learning method is adopted to input the dangerous behavior data set into the improved 3D convolutional neural network for training.After constructing the migration learning model,the feasibility of the migration learning method is analyzed and verified by visual analysis,which proves the effectiveness of migration learning.According to the change trend of loss rate curve,the stopping times of training iteration were selected.Finally,the final accuracy of the test of the model in the data set constructed in this paper was obtained through experiments,and each predicted classification category and actual classification category were described in detail by means of confusion matrix,and it was concluded that the method provided in this paper had good recognition effect and practical value in the suspect dangerous behavior identification data set.