Research On Detection Of Cow Estrus Information Based On Video Analysis

With the rapid development of dairy farming in China,it is urgent to use information technology to improve the level of scientific management of dairy farming.It is of great significance for saving the cost of frozen bull semen,improving the rate of pregnancy,shortening the interval of calving,and greatly increasing the production efficiency of the dairy farm if the cows'estrus information is grasped timely and accurately,and in which case the artificial insemination is done at the best time.The traditional manual observation method has problems such as high labor cost and low efficiency.Contact-sensor detection methods have the problem of limited power consumption in hardware,and can also cause problems such as stress response in dairy cows.In consideration of the above problems,this article studies the detection methods of cow estrus information based on video analysis.The study focused on the preprocessing of video frames of daily behavior with a complex background,object detection of cows,extraction of features of targets,and construction of classifiers for behavior of cows.It provides ideas for detecting cows'estrus through video analysis technology.The main work and conclusions are as follows:?1?The method of preprocessing the videos of cows'daily behavior collected in complex environments,combining multiple methods,was studied to reduce the interference of background and improve the contrast between cows and background.The cows in the activity fields are frequently blocked and the lighting environment is complex.In order to effectively detect cows,the mask technology was used to remove irrelevant background in frames so that the interference was reduced and reduce the amount of computation of image processing.In order to reduce the salt and pepper noise and Gaussian noise produced during the collection of videos,a median filter algorithm was used to reduce noise.In order to reduce the influence of light,contrast tests using the gray-scale transformation,histogram equalization,homomorphic filtering,and enhancement methods based on the Retinex theory,was conducted.The results show that the homomorphic filtering method can better improve the contrast between cow target and background.?2?The method of cow object detection using background subtraction combined with the areas'color and texture features was proposed.Based on the analysis of theoretical and experimental results of the inter-frame difference method,optical flow method,background subtraction method,and Gaussian Mixture Model method,combined with the requirements of this study and the characteristics of the cow target,a cow object detection method based on background subtraction with target region color and texture features was proposed.The cow object detection tests of traditional background subtraction method and inter-frame difference method were carried out with the same samples.The results show that the detection accuracy of the method of the article is 98.3%,which is 22.1%higher than the traditional background subtraction method and 6.7%higher than the inter-frame difference method.And the omission rate of the method of the article is 6.4%,which is 4.2%lower than the traditional background subtraction method and 16.6%lower than the inter-frame difference method.The detection algorithm of this article well detected the target areas of the cows from the background.?3?The features of mounting posture and body uplift were proposed and defined as the main features of cows'mounting behavior,and seven of the above features were optimized.Based on the analysis of the spatial characteristics and time characteristics of cows'mounting behavior,this article proposed that mounting posture and physical uplifting characteristics are the main characteristics of cows'mounting behaviors different from other behaviors.According to the mounting posture and physical uplifting characteristics,the geometric characteristics and optical flow characteristics of the cow's target area were selected,and the geometric features were defined as S_mean,V_Xup,V_Yup,V_Xdown,V_Ydown and the optical flow features were defined as F_Ymax,F_Ymin,V_F.Based on the analysis of the feature characteristics of test videos,the seven featureswereproposedanddeterminedtoconstructthefeaturevector Fr=[Smean,VXup,VYup,VXdown,FYmax,FYmin,VF],for behavior recognition.?4?The SVM model for cow behavior identification was constructed.Based on the characteristics of samples,a Support Vector Machine classifier was designed and trained as a target region behavior classifier.In order to compare the results of different classification models,the error Back Propagation Neural Network and K-nearest neighbor classifier were designed and trained,for the test of classifying the target areas including different behavior.The results show that,for the same samples,the average recognition accuracy of the SVM classifier was 90.9%,which is 5.6%and 10.6%higher than that of the BP neural network and the KNN classifier,respectively.The average false positive rate of the SVM classifier was 4.2%,2.7%and 9.0%lower than the BP neural network and the KNN classifier,respectively.This article adopts the SVM classifier as the recognition model for cows'mounting behavior.