Research On Target Extraction And Tracking Method Of Dairy Cows Based On Video Processing

In modern dairy farming,manual monitoring is time-consuming,labor-intensive,and low in accuracy,so the use of video perception and monitoring technology is of great significance.However,there are unfavorable factors such as large amount of data,serious environmental impact,and many noise interference in the video of dairy cows.There are still great difficulties in intelligent video processing technology.The specific technologies such as automatic positioning and extraction of video cow targets,and tracking of small targets on cow legs need to be broken through.This paper uses image and video processing,machine learning,deep learning and other technical means to conduct targeted research with Holstein cows in large-scale dairy farms as the main research object.The main research contents and conclusions are as follows:(1)In order to realize the automatic extraction of dairy targets in large-scale farms,the correlation filtering algorithm was integrated into the basic framework of target extraction,and a target extraction algorithm based correlation filtering-edge detection(CFED)was proposed.Firstly,the correlation filter designed by color names(CN)and histogram of oriented gradient(HOG)is used to obtain the target range of cows,and then 13 edge filter templates in different directions are used to convolve the target range image to obtain the image edge.Finally the cow target is extracted by fusing the edge information and color features.Target extraction experiments were carried out on 9 videos in different environments of dairy farms.The results show that the average overlap rate between the targets extracted by the algorithm and the real results reaches 92.93% The average overlap rate of the algorithm in this paper is 35.63,32.84,20.28,and 14.35 percentage points higher than OTSU,k-means clustering,frame difference method and gaussian mixture model(GMM)method respectively.The false positive rate and false negative rate were 5.07% and5.08%,respectively.The average processing time per frame is 0.70 s.The results show that the proposed CFED algorithm has better target detection ability,and can provide an effective method for accurate and rapid extraction of cow targets.(2)In order to realize automatic tracking and monitoring of cow legs in large-scale farms,this paper proposes a siamese tracking algorithm with attention mechanism(Siam-AM).First,the features of the cow’s legs are extracted in the first frame of the video,and the search areas in the subsequent frames are also input into this network to extract features.Then the features are weighted by the attention module,and the similarity between the different regions and the first frame image features is compared,and the one with the largest similarity is taken as the predicted leg position of the cow.In the experiment,60 videos of cows walking on the side were collected,and Labelimg software was used to create a tracking data set of cow legs.The data set and transfer learning algorithm were used for network training,so that it could automatically learn the appearance and motion characteristics of cow legs.The experimental results on 10 test videos show that the average tracking accuracy of the Siam-AM algorithm is 93.80%.The average tracking accuracy of the Siam-AM algorithm is 4.63 and 25.27 percentage points higher than the fully-convolutional siamese network(Siam FC)and the kernelized correlation fifilter algorithm(KCF),respectively.The average frame rate reached 57 fps,which achieved the purpose of end-to-end high-precision real-time tracking of cow legs,and provided an effective method for accurate tracking of cow legs and gait detection.(3)Aiming at the problems of low accuracy of cow lameness detection and low automation attributes,the leg coordinates obtained by cow leg tracking are used as the feature input of the lameness detection classifier,and then the lameness detection is performed on the test video by the SVM method.The results on the test video show that the accuracy of the SVM classifier reaches 96.21%,which is significantly higher than 94.67% of the BP classifier,93.28% of the KNN classifier and 91.35% of the decision tree classifier.The true rate of SVM classifier is 95.14%,which is 0.63%,1.91% and 4.45% higher than BP neural network algorithm,KNN algorithm and decision tree algorithm respectively.The false positive rate of SVM classifier is 6.28%,which is higher than BP neural network algorithm.The network algorithm,KNN algorithm and decision tree algorithm are 0.86%,3.39% and4.25% lower,respectively.Therefore,the leg features of dairy cows can be used for accurate detection of lameness behavior of dairy cows.