Individual Information Perception And Body Condition Score Assessment For Dairy Cow Based On Multi-sensor

With the large-scale development of China's dairy farming,using computer,sensors technologies and other information means to achieve accurate breeding gets more and more attention.Accurate breeding requires real-time acquisition of individual information of dairy cows,in which some identity information,quantity of exercise and body condition score are very important issues.Body condition score(BCS)is a practical tool to evaluate the energy storage,nutrition and health status of dairy cows based on the fatty deposits of cows.At present,the body condition evaluation is generally based on artificial evaluation,which has subjective error and low efficiency.It is difficult to extract and segment the key parts of the body using traditional digital image processing technology,and it is difficult to model the classifier to some extent because of the complex and changeable illumination.In addition,the mapping from cattle BCS to individuals,making a health management system that integrates cow identity information,quantity of exercise and body condition monitoring,are the urgent problems to be solved in the field of dairy accurate breeding.In view of the above problems,this paper comprehensively uses two sensors data,including the image and wearable device,to carry out research around the detection and tracking on the key parts of cows' body,body condition scoring method,individual information perception and so on.These works include following parts:1.Building an automatic data acquisition platform of cows images,and 8972 cows'back top view images were obtained on this platform.The images include two sides of the cow's buttock,ischial tubercle,caudal root,lumbar angle and part of the spine,which are closely related to BCS.The data were manually labeled by veterinary experts to provide data sets for neural network training.2.In view of the original SSD algorithm does not consider the connection between the layers of convolutional neural network well,this paper use the idea of tight connection between the layers in DenseNet,while introduce Inception v4 module to expand the sensory field of neural network.In this way we design an efficient hybrid network model for cow tail detection,improved SSD algorithm.The detection speed of the algorithm for cow tail is 115 fps,which is nearly 2 times faster than the 39 fps of the original SSD algorithm.And the model size is only 23.1MB,which can save the cost of hardware storage.3.A target tracking algorithm based on deep learning and improved Kalman filter is also proposed.On the basis of neural network target detection,Particle filter and Kalman filter are used to estimate the target states.In order to solve the problem of different individuals' allocation and assignment,we introduce Hungary algorithmof to realize the automatic counting function of individual cows.The experimental results show that:(1)compared with the Particle filter and the original Kalman filter,the improved Kalman filter has the smallest average position deviation;(2)by setting a reasonable time threshold T,the counting accuracy can reach 96%and above.4.A multi task deep learning algorithm based on the improved SSD model is proposed to solve the problem of cows body condition scoring.The accuracy of BCS recognition is 98.46%when the error is ±0.5.In order to solve the problem of individual BCS recognition,this paper designs a low-power pedometer based on ZigBee.Through this tool,the accuracy of measurement is 94.6%,which lays a foundation for further detection and diagnosis of cow motion related diseases.In summary,this paper uses two kinds of data,image and pedometer,designs deep learning models to complete the tail detection and tracking of individual cows,to achieve the automatic scoring of cows' body condition.Finally,it provides technical support for the information management of animal husbandry and accurate breeding of cows.