Research On Abnormal Behavior Recognition Technology For Public Safety Based On 3D Convolutional Neural Network

In order to ensure public safety and social stability,surveillance cameras are widely deployed in densely populated public places such as squares,airports and stations.However,massive acquired data poses a challenge to rapidly detect abnormal human behavior in surveillance videos.Therefore,within the field of intelligent surveillance,the detection and recognition of abnormal behavior in surveillance video based on deep learning technology has become a research hotspot.Due to the complex background in the video and the variety of targets,it is difficult to accurately identify abnormal behavior in surveillance video.The existing abnormal behavior recognition models have some shortcomings such as complex structure,big calculation cost and low recognition rate.In order to tackle these problems mentioned before,this thesis proposes an improved three-dimensional convolutional neural network model based on channel attention,and introduces a long short term memory network to achieve accurate recognition for abnormal behavior in surveillance videos.Besides,an abnormal behavior recognition system is also built.The contributions of this thesis are listed as follows:(1)Optimize the C3D network structure to improve the accuracy of abnormal behavior recognition.Consider the low recognition rate of the C3 D network model,a convolutional layer and a Re Lu activation function are added to the C3 D structure so as to improve the performance of the overall network.Besides,the SE module is integrated into the optimized C3 D structure,and the effective information is improved by assigning weights of attention.Beyond that,the fully connected layer of C3 D model is replaced by the global average pooling layer to complete dimensionality reduction and feature extraction.Therefore,this proposed method achieves efficient dimensionality reduction and significant feature compression.Finally,the long short term memory network is introduced for timing-modeling of the high-level features.It is verified by experimental results that the recognition accuracy of the proposed model for abnormal behavior in surveillance video is higher than that of the traditional C3 D network model.(2)Enhance the R3D network structure to reduce the amount of model parameters and improve the recognition accuracy for abnormal behavior.Given that the deep R3 D network has a large amount of parameters and low recognition accuracy,this thesis introduces global average pooling layer to the R3 D model so as to reduce the amount of parameters of the model.Besides,the SE module and the 3D residual block are merged to realize the adaptive calibration of the channel weights by the network.Finally,the effectiveness of this optimization scheme is verified by experimental results.(3)Develop an abnormal behavior recognition system.By integrating the channel attention mechanism,three-dimensional convolutional neural network and long short term memory network in deep learning technology,this thesis develops an abnormal behavior recognition system.This system mainly contains the functions of selecting models,choosing videos and displaying recognition results in real time.Besides,the UCF Crime dataset is used to test the model.In the end,the test results show that the system is competitive in responding speed and practicality.