Research On Human Action Recognition Method Based On Infrared Video For Chemical Enterprise Video Surveillance

With the development of modern social science and technology,many chemical enterprises are also rapidly developing towards the direction of intelligent mode.In such companies,there are often many situations where manual operations are carried out under dangerous conditions.At the same time,several chemical accidents reported in recent years have exposed significant safety hazards,mainly due to unsafe behavior of on-site personnel,and most of them occur during nighttime operations.Therefore,people attach great importance to them.Human action recognition is widely used in many fields such as video classification,human-computer interaction,medical diagnosis,security monitoring,and so on.Currently,people have a certain understanding of motion recognition under visible light,but there is still little research on infrared motion recognition.However,infrared video is not only suitable for allweather monitoring,but also provides higher privacy,so it is of great research significance.For this reason,this article conducts research on the difficult problem of human action recognition arising from intelligent video surveillance in nighttime scenes in chemical enterprises.The specific research work is as follows:(1)This paper constructs an infrared human action recognition dataset to compensate for the lack of existing infrared human behavior datasets.This dataset includes different behaviors of different volunteers in real nighttime scenes to simulate the real environment of chemical scenes.At the same time,two other infrared image and infrared video datasets are introduced to facilitate the related research in this article.(2)This paper proposes a cross domain infrared human action recognition method based on depth transfer learning to solve the problem of identifying unmarked images in infrared video.Firstly,select the data fields needed for the experiment,where the source field uses well labeled gray scale images,and the target field uses unmarked infrared images.Because the two fields have different data distributions,a domain adaptation is proposed to reduce the distribution differences between the two fields;Secondly,input source and target domain data into a deep neural network,and extract their respective image features through convolution pooling and other operations.At this time,in order to calculate the loss of source and target domain features,a distance loss called lightweight density divergence is proposed,which not only minimizes inter domain differences,but also maximizes intra class density,better achieving the effect of reducing loss;Finally,defining three FC layers as domain discriminators and applying lightweight density divergence to the adversarial domain adaptation framework not only confuses the domain discriminator,but also ensures good alignment of the two data distributions,thereby reducing error losses and achieving better experimental results.(3)Aiming at the shortcomings of low image resolution and poor signal-to-noise ratio in infrared video,this paper proposes an infrared human action recognition method based on deep neural network and attention mechanism.In order to implement the fusion strategy more efficiently,first,we extract each infrared image from the infrared video,preprocess the image information,and obtain the spatial and temporal information of the actions in the video;Secondly,respectively input VGG16 network model for spatial feature extraction operations such as convolution and pooling,and input Bi-LSTM network model incorporating attention mechanism for temporal feature extraction;Finally,the classification results of the two networks are obtained through a decision level score fusion strategy.Experiments have shown that the algorithm proposed in this paper has a good effect on action recognition.