Behavior Recognition Based On Deep Learning And Its Application On Infrastructure Sites

In order to further ensure the safety and stability in public places,human behavior recognition technology based on video surveillance has received widespread attention from the whole society.Existing human behavior recognition methods based on deep learning have high computational complexity,poor model generalization,and are difficult to accurately recognize.This paper analyzes the advantages and disadvantages of the existing models.From the perspective of the model's practicability and recognition performance,two improved models are proposed based on the original two-stream convolutional neural network.Aiming at the problem that the convolutional neural network(CNN)can only extract the static features and local motion features in the video,this paper proposes to integrate Long Short-Term Memory(LSTM)in the spatial stream network,and replace the original two-dimensional convolution network with a three-dimensional convolution(C3D)network in the time stream network,forming an improved two-stream CNN-LSTM and C3D network structure,further strengthen the learning of timing information in the video,and then the weighted fusion method is used to weight the decision scores output by the Softmax layer of the two-stream network to obtain behavior classification results,but the accuracy rate is not high enough.In connection with the problem that the two-stream network model is only fused at the decision score stage,which leads to a low accuracy of behavior recognition,this paper proposes to fuse the spatial stream network and the time stream network at the feature level,in this way,the two stream network structure must be the same.In view of the superiority of the three-dimensional convolution,on the basis of the above improvements,continue to replace the spatial stream network with the C3D network to form an improved two-stream C3D network structure,and the fusion methods include Conv5b fusion,Fc7 fusion,and hybrid fusion.The network achieves a better recognition effect on the behavior recognition task.For behavior recognition,most of them use public data sets,and rarely study the problem of abnormal behavior recognition in specific scenarios.In addition to using the public UCF-101 data set,this paper also established an infrastructure site abnormal behavior detection data set,using transfer learning to identify seven behaviors:walking,running,working,spaning the cordon,triping,fighting,and throwing down.The experimental results show that the improved two-stream CNN-LSTM and C3D network structure achieves an accuracy rate of 90.24%in the UCF-101 data set,and an accuracy rate of 95.97%in the infrastructure site abnormal behavior detection data set.In the two datasets,the accuracy of the improved two-stream C3D network structure is 92.47%and 97.51%respectively.The improved network model not only has a good recognition effect for 101 types of behaviors in the UCF-101 data set,but also shows superiority in the identification of 7 types of behaviors in specific infrastructure site data sets.This paper combines theory with actual application scenarios to make behavior recognition research makes more sense.