Detection Methods Of Cow Mounting Behavior Based On Video Analysis And Deep Learning

With the rapid development of our country's dairy farming industry,there is an urgent need to use information technology to improve the scientific management level of breeding.Timely and accurately grasping the estrus information of dairy cows and timely artificial insemination are of great significance for saving feeding and artificial insemination costs,increasing the pregnancy rate of dairy cows,shortening the calving interval,and maximizing the production efficiency of dairy farms.Aiming at the high labor cost and low efficiency of traditional manual observation methods;hardware power consumption of the contact sensor detection method is limited,and it will cause the cow to have a stress response;the method based on video analysis relies on the results of threshold segmentation,and it is not easy to extract features of cow estrus behavior,and so on,this paper studies the detection method of cow estrus information based on machine vision,focusing on the preprocessing of daily behavior video of cows in the complex field environment,the target detection algorithm of moving cows,the classifier model of cow behavior recognition based on Alex Net,and the cow estrus based on improved YOLO?v3 Behavior recognition methods provide ideas for detecting estrus of cows through machine vision.The main work content and conclusions of the paper are as follows:(1)A comparative study of the combination of multiple methods,the method of preprocessing the daily behavior video of dairy cows in complex environments,reduces background interference and improves the contrast between the cow's target and the background.In order to reduce the uniform noise,impulse noise and Gaussian noise generated in the process of image acquisition and transmission,the method of bilateral filtering is used for denoising;in order to reduce the influence of light transformation,weather,etc.,the piecewise linear transformation is used to enhance the contrast of low-brightness images,the Gamma transformation(?=0.5)is used to enhance the contrast of the image at noon on a sunny day,and the histogram equalization is used enhance the contrast of the image in rainy and foggy weather.The test results show that the above three methods can improve the contrast between the cow's target and the background.(2)A target detection algorithm for moving cows based on background subtraction is proposed.Considering that the location of the camera is fixed when the cow video is collected in this article,the video shooting scene is fixed,and the background of the collected video is not much different,and the dynamic background modeling of the video is selected to accurately detect the moving target in the scene.On the basis of the frame difference method,multi-frame averaging method,GMM and Vi Be algorithm theory and experimental results analysis,combined with the requirements of this research and the characteristics of the cow's target,a moving cow target detection method based on the improved Vi Be algorithm is proposed.In the same cow video test set,the improved Vi Be algorithm,the classic Vi Be algorithm and the mixed Gaussian model of the moving cow target detection experiment are carried out,the average precision of the detection algorithm in this paper is 82.66%,which is 4.94% and 2.58% higher than the classic Vi Be algorithm and the Gaussian mixture model method,respectively.The precision of the detection algorithm in this paper is 86.80%,which is 4.93% higher than the classic Vi Be algorithm.And 2.37%,in terms of target detection speed,the real-time index of the method in this paper is 0.78(?1 is considered to meet the real-time performance),the real-time index of the classic Vi Be algorithm is 0.91,and the real-time index of the Gaussian mixture model method is 1.30.The method in this paper can quickly and accurately detect the moving cow target from the background.(3)Constructed a cow behavior recognition classifier model based on Alex Net,and performed pattern recognition on the detected target area of moving cows.Recognizing the estrus behavior of dairy cows can also be regarded as the classification of dairy cow behavior images,that is,regional distribution of estrus behavior images and non-estrus behavior images.This paper builds a deep convolutional neural network model based on Alex Net.After the model is trained,it can accurately identify and classify cows.Behavior,the recognition accuracy rate is 97.6%.Using the moving cow target detection algorithm and the cow behavior recognition classifier model to test,the method has an accuracy rate of 100%for cow estrus behavior detection and a recall rate of 88.24%.(4)A recognition method of cow estrus behavior based on improved YOLO?v3 is proposed.Aiming at the problem of the method of first detecting the cow's target and then identifying the behavior of the detected target area,there are many interference factors and slow recognition speed.In order to further improve the efficiency of cow's estrus behavior recognition,combined with the characteristics of dairy cow video data,from the anchor point The YOLO?v3 model is improved in three aspects: frame size set,feature extraction network and bounding box loss function,and a one-stage end-to-end cow estrus behavior recognition method is constructed.The test results show that for the same test sample,the average recognition accuracy of the model in this paper is 99.15%,and the recall rate is97.62%,which is 2.63% higher than YOLO?v3 accuracy and 7.28% higher recall rate.Compared with Faster RCNN,Although the recognition accuracy rate is lower by 0.21%,the recall rate is increased by 7.28%.The recognition speed of the model in this paper is 31 f/s,which can meet the real-time demand for recognition of estrus behavior of dairy cows.