Study On Identification Of Typical Cow’s Self-Protective Behavior Based On Machine Vision Technology

The large scale breeding of dairy cows has caused a great number of harmful diptera insects due to excessive concentration of fecal urine.They cause cow stress and self-protective behaviorwhen they attack,including tail swishing,head throwing and leg stamping etc.,which has an utmost effects on the health and productivity of cows.Therefore,study on the cow’s self-protective behavioris very meaningful to the effect of double-wing insect invasion on the cow’sphysiological and production,the evaluation of the breeding environment and the cow’s temperament breeding.Current research of cow’s self-protective behavior is mostly based on manual observation,whichhas great intensity work,strong subjectivity and low accuracy.This dissertation studies an intelligent identification method based on machine vision technology for the typicalcow’sself-protective behavior,which is an accurate,non-contact and non-stressmethod.The main contents and conclusions are as follows:(1)In order to solve the difficult problem of the cow’s complex body color and low background discrimination,and the instantaneous and sudden of the cow’s self-protective behavior,this dissertation presents an improved feature point detection algorithm based on Shi-Tomasi corner detection,which can be effective to decrease the number of the target feature points,accurately detect the feature pointsof self-protective behaviour,and greatly reduce the computational complexity.In order to meet the requirement of accuracy of the system,an improved the protective behavior tracking algorithm based on pyramid Lucas-Kanade optical flow methodwas presented,by using the bounding box to effectively track tail swishing,head throwing and leg stampingof the self-protective behaviors.(2)In this dissertation,the characteristic vectors of cow’s self-protective behaviors,such as tail swishing,head throwing and leg stamping,are analyzed deeplyand extracted in 9-dimensional feature vectors such as rotation angle,moving distance,vertical displacement,horizontal displacement,area of bounding box,aspect ratio,etc.The correlation and specificity of each feature vector were analyzed.A recurrent neural network model,which was established using random sample data of three typical self-protective behaviorscan achieve accurate recognition results.(3)The intelligent recognition system of cow self-protective behavior developed in this dissertation takes the detail parts of the cow as the tracking target.The improved corner detection and optical flow tracking algorithm are used to track the self-protective behavior of dairy cows.Based on machine learning theory,the recurrent neural network model is developed to extract the typical characteristic vectors of cow’s self-protecting behavior and realize the non-interactive automatic identification for the three typical body protection behaviors of dairy cows.Using a variety of machine learning methods to compare the recognition effects,the recurrent neural network model can achieve an accuracy of 98.4%.By comparing with the artificial experience,the accuracy is in[0.88,1],the recall rate is in[0.87,1],and the detection result is obviously better than the background modeling method such as the frame difference method and the Gaussian Mixture Model.This method which has good robustness can avoid false detectionof the cow body shaking,background changes and the camera jitter,and can detect the blocked behavior,provided a theoretical and technical support for realizing the digital and intelligent analysis of the ruminants’self-protective behavior.