| Precision livestock farming(PLF)is a refined breeding methodology that can improve the efficiency of livestock and poultry production,promote animal welfare,optimize the breeding environment and feed supply.This concept has great application value in the economy,society and environment.It is also a technology with great potential to optimize intensive farming and to solve the problem of “large but not strong,many but not superior” in the livestock industry of China.This research uses computer vision technology,machine learning technology and multi-sensor fusion technology to solve the problem of current computer vision monitoring methods in PLF.In accordance with demand analysis,algorithm proposal,quantitative evaluation,algorithm verification and practical application,the following topics were conducted: development of object detection algorithm in dynamic background environment,implementation of multiple object tracking algorithm in grouphoused animals,extraction of the region-of-interest in high-density feeding conditions.These results provided the technical backup for recursively monitoring the phenotype and behavior of livestock animals in different feeding environments.Based on this,approaches were proposed to model animal behaviors in group level and interactions among individuals,aiming to provide new methods to perceive the phenotypic information and behavioral patterns of livestock animals in ways of online,real-time and non-destruction.The main work and conclusions of this study are as follows:(1)A moving object extraction algorithm based on the improved Gaussian Mixture Model(GMM)was proposed.The complexity and dynamics of the breeding environment of livestock animals brought challenges to the current object detection models,resulting in the decline of robustness and real-time performance.To solve this problem,based on the idea of recursive background modeling,the classical Gaussian Mixture Model was improved through the penalty factor,locate updating strategy,and a binary classification algorithm based on chromaticity and brightness deviations.Compared with the classical GMM in a complex environmental conditions,owing to the penalty factor,the model complexity was reduced by 50.85% in the sense of average,the foreground false detection rate and the background false detection rate were reduced by 19.50% and 13.37% respectively,the single-frame running time was reduced by 29.25%.The locate updating strategy solved the problem of foreground melting and the proposed binary classification algorithm solved the problem of the shadow mis-classification.In conclusion,the improved GMM met the requirement of real-time detection of moving object under dynamic background.Then,the proposed algorithm was applied into the application of automatic body condition score of dairy cows in commercial farms:(1)Improved GMM was capable to precisely extract the contour of cows and establish a 3D coordinate system.(2)The image features corresponding to the manual scoring indicators were designed and the correlation analysis is carried to verify these image features.(3)The ensemble learning algorithm was proposed to solve the problem of dataset imbalance.Crossvalidation results showed that the average accuracy of the improved GMM was 56% within a 0.125 deviation,76% within a 0.25,94% within a 0.5 deviation range.Compared with other existing technologies,the system has stronger robustness under dynamic background environment,and it also can make more accurate prediction of cows with extreme body conditions.This means that the technology meets practical requirements.(2)A novel action recognition model based on ALR-GMM was proposed.the classical activity index model was heavily relied on the background subtraction which was not sensitive to the dynamic background environment.To solve this problem,a novel activity index calculation method was proposed based on the idea of recursive background modeling.An adaptive learning rate adjustment mechanism was applied to the GMM by introducing the hyperbolic tangent function,named ALR-GMM.With this improvement,ALR-GMM could differentially update its learning rate on different image regions in order to model the background and perceive animals’ motion pattern synchronously.Testing on the manually labeled images,the results showed that when the initial learning rate is 0.2,the average absolute error was reduced from 0.0265 to 0.0233,and the average relative error was reduced from 18.08% to 14.34% comparing to the background subtraction.It showed that ALR-GMM meets the requirements of the calculation accuracy index.In order to verify the feasibility of ALR-GMM,studies on the behavioral response of laying hens to mite stress and on the recognition of aggressive behaviors caused by environmental stress in group-housed pigs were carried out.The results showed that there was a positive correlation between the activity index of laying hens and the number of red mite infestations,which provided a reliable realization for accurate insecticidal.The accuracy of the aggression behavior recognition was 97.6%.,which brought new method to improve animal welfare.(3)An action recognition model based on the Convolutional Neural Network(CNN)and Long Short-term Memory Network(LSTM)was proposed.Because the existing animal action recognition models cannot monitor fine-scale actions,and also cannot accurately obtain the spatial-temporal features of individual animal behaviors in group-housed environment.A tracking-by-detection algorithm and a CNN+LSTM based spatial-temporal feature description method were integrated to solve this problem.In order to verify the effectiveness of the algorithms,first,the performance of the tracking algorithm was evaluated.Second,the sow posture classification was used to verify the ability of CNN to express spatial features.Finally,the tail biting classification is used to verify the ability of LSTM to express temporal features.The results showed that the accuracy of the tracking algorithm is 78.35%,which was capable to locate 92.97% pigs.CNN was able to classify four basic postures(sows lying,lying down,sitting back,and standing)with an average accuracy of 88.1% indicating that the convolution feature could be used for describing spatial features.LSTM achieved an accuracy rate of 96.25% for video sequence of tail biting behavior,indicating that LSTM could be used for describing temporal features of animal behaviors.(4)An object extraction algorithm based on color and depth information was proposed.Since the self-similarity and occlusion of animals,it was a big challenge to locate individuals by the current object detection methods.To solve this problem,a novel method that combined color and depth information was proposed to extract the region of interest.Based on the twostream calibration,the proposed method located the region of interest from the depth image,then extracted the color,texture,shape and spatial relationship features from the color image.To verify the feasibility,a Kinect-based automatic weighing system for broiler breeders was developed to obtain individual weight and gender information under commercial farms.The results showed that the locating accuracy was 77.3%,the accuracy,sensitivity,precision and specificity of the gender classification model were 99.7%,98.8%,100%,and 100%,respectively,which met the requirements of practical usage.This study provided an effective tool for optimizing the breeding process by maximizing egg production and fertility. |
PDF Full Text Request