Study On Recognition Method Of Unsafe Behavior Of Miners Based On Improved Two-Stream Algorithm

As the main source of national energy supply,it is of great significance to ensure the safety and intelligence of coal mining.People,as an important link in the mining process,their non-standardized operations will directly lead to accidents.Therefore,it is necessary to identify and monitor the behavior of miners in coal mine scenarios.This paper combines behavior recognition technology and scene classification algorithm to study the unsafe behavior recognition of miners in complex coal mine scenes.The research contents of this paper are as follows:(1)First,this paper uses the key frame extraction technology to filter the redundant frames of the self-built miner behavior video data-set,and then identify the behavior of the target based on two-stream network.In the aspect of behavior recognition,a recognition algorithm based on dual complementary optical flow features is proposed.The sparse optical flow and the dense optical flow are combined to generate dual optical flow features,and form the final characterization of behavior joint with RGB features.In order to retain the semantic information,two methods are used to perform feature fusion in the middle/late stages of the network,and the superiority of the Support Vector Machines fusion method is proved.After feature fusion,the recognition accuracy of Two-Stream,TSN and I3 D networks increased by1.1%,0.7% and 0.5%,respectively,and the recognition accuracy of I3 D models based on migration learning reached 81.5%.(2)The image enhancement method is used to preprocess the dark scene image,and the scene classification model is used to realize the coal mine scene recognition.A high-level scene feature fusion algorithm based on the dual-stream mode is proposed,which achieves higher-level learning of Gist features as well as semantic abstraction of the original image.The two-stream features are spliced to realize the fusion of global and local features after the global pooling,and finally the abstracted features are used to visualize the changes of attention points using class activation mapping.This article is experimentally verified on the Alex Net,Google Net,VGG and Res Net networks.The algorithm has improved scene recognition accuracy by 1.0%,1.2%,0.5% and 0.7%,respectively.The scene classification accuracy of the fine-tuned Res Net network has reached 94.8%.(3)Aiming at the problems of model over-fitting and local optimization,this paper proposes a hybrid model optimization algorithm based on quantum genetics.By quantum genetic evolution for the encoded convolution kernel and training with(4)gradient descent algorithm,the optimization of the parameter of the network convolution kernel is realized.This paper also uses quantum genetic algorithm to search for the global optimal of specific network hyperparameters.Experiments show that in terms of parameter optimization,the algorithm has improved the accuracy of Alex Net and Le Net networks on the MINIST by 1.05% and 0.32%,respectively;the recognition accuracy of seven classic networks have been improved on the CIFAR10.In terms of hyperparameter optimization,the recognition accuracy of all individuals in the population has improved.After 20 generations of evolution,the recognition accuracy of Alex Net and Google Net has increased by 0.84% and 0.7%,respectively.The algorithm also optimizes the behavior recognition network and scene classification model,and the accuracy is increased by 3.9% and 1.8%,respectively.The article has 47 figures,25 tables and 83 references.