Research And Application Of Hu Sheep Behavior Recognition Based On Deep Learning

Hu sheep behavior recognition plays an important role in the study of Hu sheep’s health status,emotional expression and scientific breeding.In the intelligent breeding and precision breeding of Hu sheep,its behavior recognition is an indispensable link.With the wide application of video capture devices,behavior recognition technology based on video understanding has become the focus of research in the field of computational vision.The traditional behavior of Hu sheep mainly relies on manual identification or data collected by sensors for analysis and identification.The former is time-consuming and labor-intensive and will cause data loss,while the latter is too expensive.Therefore,this paper proposes a behavior recognition method of Hu sheep based on deep learning,which uses deep learning to extract the spatiotemporal information and important features of Hu sheep movements from the video,and then understands and recognizes Hu sheep movements.This paper uses the deep learning method and only uses RGB data for training to achieve accurate identification of the three behaviors of Hu sheep:rumination,eating,and running.The research content and work are as follows:1.Make a Hu sheep action dataset.In the Hu sheep farm,video data of three movements of Hu sheep rumination,eating and running are captured by video capture equipment.The recorded videos are classified according to categories,and the classified Hu sheep behavior videos are edited into 10s short video clips as required,and simple preprocessing is performed to form the training set and test set required in this paper.2.In view of the gradient explosion of 3D convolution when building a deep network and in order to reduce information loss and enhance global interaction,a 3D residual convolutional neural network and a Quadruplet Attention Module(QAM)are proposed.The 3D convolution module of the residual structure is designed to form a network model,and a quadruple attention module composed of four branches is added to enhance the ability to extract features.Recognition accuracy of 96.6%and 68.5%is achieved in the human behavior datasets UCF101 and HMDB-51,respectively.A recognition accuracy of 99.2%is achieved on the Hu sheep behavior dataset.3.Aiming at the disadvantage that 2D convolution cannot extract temporal features,a TSM(Temporal Shift Module)Hu sheep behavior recognition network based on attention mechanism is designed.Use ResNet-50 as the backbone network for feature extraction,extract the spatial feature information of video frames,introduce a time-shift module into the network,and achieve efficient temporal modeling by moving the extracted feature maps along the temporal dimension.The dimension fuses the different information of the previous frame and the next frame with the current information,and achieves the function of making up for the lack of time information.By adding an attention module to suppress background interference,the perception of the key action areas of Hu sheep is enhanced.Use cosine restart learning rate to speed up model training during training.The accuracy rate reached 97.57%on the Hu sheep behavior dataset.