Research On Human Behavior Recognition Technology Based On Deep Learning

With the rapid development of electronic science and information technology,human behavior recognition in video has gradually become one of the more popular research directions in industry and academia.Human behavior recognition has extensive application prospects and practical value in real life such as intelligent video surveillance,intelligent medical treatment,security monitoring,smart home,and human intelligent interaction.However,due to the intricate background of human behavior in the video,the sudden movement of a certain behavior of the human behavior and the change of the viewpoint of a camera,how to efficiently and accurately extract the characteristics of human behavior remains a huge challenge.In response to these problems,this topic is based on the two techniques of deep learning and target recognition.The purpose of the human body motion recognition in video is to complete the algorithm research of human behavior recognition.And training and testing on the UCF101 dataset,realizing the recognition of human behaviors in videos based on deep learning.This theme improves the C3 D network structure,proposes a behavior recognition algorithm based on RC3 D convolutional neural network,extracts temporal and spatial features at the same time,and embeds the 3D residual network structure into the 3D convolutional neural network.While deepening the network width,the network performance is improved,and the problem of gradient disappearance is avoided.In order to make the RC3 D network learn more robust features and effectively prevent overfitting,the Dropout method is introduced in the fully connected layer.The algorithm is trained and tested on the UCF101 dataset.The experimental results show that the average accuracy of human behavior recognition based on RC3 D convolutional neural network is 54.5%,which is 14.33% higher than C3 D convolutional neural network.RC3D convolutional neural network still has problems such as single extraction time information,insufficient model generalization ability,and low recognition rate.Therefore,this theme adds a BN layer to the 3D residual network structure of the RC3 D network to form a BRC3 D convolutional neural network,optimizes the network structure,and guarantees the stability of the input data distribution at each layer.Combining the idea of dual stream,a behavior recognition algorithm based on dual stream BRC3 D convolutional neural network is constructed,and a weighted fusion strategy is adopted to achieve the fusion of spatial and temporal features.The algorithm is trained and tested on the UCF101 data set.The experimental results show that the average accuracy of human behavior recognition based on the dual-stream BRC3 D convolutional neural network is 91.34%.The algorithm effectively accelerates the training speed of the model and also improves Generalization and recognition capabilities of the model.