Research On Human Behavior Recognition Method Of Moving Target In Tarmac Under Low Light Conditions

With the development of the civil aviation,traveling by air has become the first choice of young people and middle class.However,the rapid growth of the passenger population has also brought tremendous burden to the airport security.It is urgent to establish an all-weather intelligent video surveillance system on the tarmac which is an extremely important part of the airport.Human behavior recognition,the core technology of intelligent video surveillance,has developed rapidly in recent years,most of the research focuses on visible light data but few achievements in infrared data.Relevant technologies based on infrared video data are still unable to achieve real-time and effective human behavior recognition due to the lack of color texture information and blurred edge of the target in infrared image.Therefore,this paper focuses on human behavior recognition in infrared video.The main problem of traditional behavior recognition technology is that the feature representation ability is weak,the robustness is poor in complex and varied scenarios,and the feature dimension is huge.It is often necessary to use a special coding method to reduce the dimension,which undoubtedly increases the computational cost.The convolutional neural network method emerged in recent years has strong feature learning ability.It can adaptively capture the information through data distribution,and extract abstract semantic features.Its end-to-end model structure also greatly reduces the cost of preprocessing.Therefore,this paper uses the convolutional neural network method to model the human motion information in infrared video.The main contributions are as follows:First of all,two three-dimensional convolutional neural networks suitable for RGB data and optical flow data were proposed.This two networks made structural adjustments according to the data types,and rationally utilized the advantages and disadvantages of three-dimensional convolution and pooling.They can respectively represent the motion information from the original RGB data and optical flow data without complicated preprocessing and subsequent feature coding.Using the powerful adaptive learning ability of the three-dimensional convolutional neural network,they can extracted discriminative features separately from RGB and optical flow data.Secondly,in order to recognize the human behavior in the video surveillance of the apron scene under low light conditions,this paper migrated the training results of the two networks from visible light data to infrared data respectively,realizing the relationship between the two data modalities.By using the training model of visible light as the initial parameter,the problem of over-fitting during training from scratch was avoided.Finally,in order to further improve the effect of the model,this paper proposed a method of combining the three-dimensional convolutional neural network structure with the two-stream network.By combining the output features of the last convolutional layer of the two networks with the maximum pooling,average pooling and direct addition,the two networks are combined in the decision-making layer,which further improves the discrimination of the feature.