Study On The Machine Learning Based Swine Herd State Monitoring And Individual Behaviour Recognition

Swine industry is very important related to national economy and people's livelihood.Pig industry is of great significance to ensure the safe and reliable supply of pork food.With the development of large-scale swine industry,the demand for artificial intelligence automatic monitoring of pig behavior is also growing.The study on an automatic intelligent pig behavior monitoring system can effectively improve the quality of scale farming and the operating efficiency and production capacity.This paper proposes an intelligent study on the behavior recognition of multi-target pigs based on machine learning.The main research work is as follows:(1)The characteristics of pigs' video image are analyzed,and the basic principle of multi-target tracking method based on foreground target segmentation algorithm is described in detail.This paper introduces the implementation principle and technical algorithm of the pig monitoring system,and lays a foundation for the development of efficient and practical pig behavior identification system.By histogram equalization and bilateral filtering,an adaptive threshold segmentation method is used to automatically calculate the local threshold and complete initial extraction of foreground target.The multi-target pigs are tracked automatically by using the Kalman filtering algorithm.The image is simplified by morphological processing method and filtered out the background interference pixels such as grille and wall seams.The GrabCut algorithm is used to divide the foreground target pixels and accurately extract the pigs' pixel area.It can effectively improve the accuracy of the foreground target.The mean range of Structure similarity SSIM is in [0.88,1].The method proposed in this paper greatly reduces the calculation amount of GrabCut algorithm,which is significantly more efficient than the traditional algorithm.The average computing time was 23% less than the traditional GrabCut algorithm.The segmentation method proposed in this paper not only realizes the effective segmentation of non-interactive multiforeground target,but also satisfies the real-time requirement of video monitoring system,and provide reliable foreground segmentation data for the intelligent warning of video monitoring system.(2)We propose a method to identify the extruding state of pigs based on machine learning.In the top view because the adhesive region cannot be separated by the foreground segmentation image,the characteristic vector is extracted by fitting the ellipse such as the ellipse area,rotation angle,elliptic eccentricity,ellipse's long axis and short axis.The correlation and specificity of the eigenvectors are compared.Support vector machines,adaboost,random forest and other machine learning methods were used to train and classify the pig multidimensional feature vector samples whether multi-target pigs are extruded.Intelligent early warning is given to prevent piglets from squeezing for a long time,thus increasing the survival rate of piglets.(3)The individual behavior of swine is identified based on machine learning method.The body of a pig is marked by a rotated rectangular box in the side view,by extracting the feature vectors which include the center of gravity,rotation angle,target area of the foreground,the distance of the center of gravity to the abdomen,the height of the rotating box,the width-height ratio and so on.Support vector machine is used for sample training to establish classifier model to identify the pig multidimensional feature vector samples.It could realize the recognition of walking and lying posture of pigs.The accuracy of the classifier based on the support vector machine can reach above 95%.However,the accuracy of traditional machine learning methods depends on the extraction accuracy of multidimensional feature vectors,and can not effectively meet the requirements of pig behavior identification in different complex environments.(4)Study on the individual behavior identification of pigs is based on deep learning technology.The deep learning model of GoogLeNet convolution artificial neural network was trained and established in piggery environment.The model does not need to perform target segmentation,tracking,feature extraction,etc.,directly and effectively detect pig individual posture.This method effectively overcomes the constraints of complex environments and can achieve an accuracy of over 96% in recognition for standing and lying.The convolution artificial neural network based on deep learning overcomes the shortcomings of traditional image processing,video tracking method and traditional machine learning method,and this paper expands the application field of convolution artificial neural network based on deep learning.It provides the necessary theoretical basis and technical support for further research and development of deep learning technology and pig behavior monitoring system.