Research On Key Techniques For Non-contact Automatic Extraction Of Cow Linear Appraisal Indicators

In the field of animal husbandry in China,dairy farming is an important part of animal husbandry production.Cow linear appraisal has become an important part of domestic and foreign dairy cow breeding research,and also a core link in dairy production.The cow linear appraisal can predict the milk production performance of dairy cows,estimate breeding values,and organize selection schemes,which can then guide the selection,breeding and breed improvement of dairy cows,and ultimately improve the production efficiency and economic benefits of dairy farms.The traditional manual appraisal method is time-consuming and labor-intensive,and both efficiency and automation need to be improved.Moreover,contact measurement will also cause stress response to cows.The cow linear appraisal based on two-dimensional images mostly selects appraisal key points,and the loss of spatial three-dimensional information will introduce errors.Therefore,with the current research progress,the need for three-dimensional data for cow linear appraisal has become increasingly urgent.In this paper,aiming at the problems of low efficiency,high difficulty and low degree of automation of traditional linear appraisal methods,we research the key techniques for non-contact automatic extraction of cow linear appraisal indicators based on deep learning algorithms and computer vision technology.First,the depth cameras were used to capture the cow rump and side image sequences,and then the object detection and individual identification were performed on the cow objects in the image.Then,the key point detection model was performed to detect the appraisal key points of the cow rump.Finally,the automatic extraction of cow linear appraisal indicators was completed by combining three-dimensional reconstruction of the cow rump and the detected appraisal key points.The main work contents of this paper are as follows:(1)Cow image acquisition and processing.Selected structured light and TOF depth camera to capture the visible light and depth image sequences of the cow's rump and side respectively,and post-processed the sequences to obtain experimental data,which were used to build a cow rump and side image database to prepare for the subsequent experiments.(2)Cow object detection.Used the YOLO object detection model to extract the overall cow object in each frame of the image sequence,and the cow rump and side object images were taken as the images to be identified.(3)Individual identification of cows.The individual identification of cows is the basis of cow linear appraisal,two individual identification methods based on convolutional neural networks were proposed to identify cow's rump and side.For the side images with more discriminative contour and texture features,a deep part feature fusion model was proposed for cow identification.Before the model is trained,the overall cow object of the side images to be identified was divided into three parts,head,trunk and legs,by frame differencing and segmentation span analysis.In this identification model,three independent convolutional neural networks were trained firstly to extract deep features from these three parts,then a feature fusion strategy was designed to fuse the features,i.e.,deep parts feature fusion,and finally a support vector machine classifier trained by the fused features was used to identify each individual cow.(4)Cow appraisal key point detection.Convolutional neural networks were used to detect key points in the cow rump images.Used annotation software to label the cow appraisal key points of a part of the rump images under the guidance of cow linear appraisal experts,and the convolutional neural network was fine-tuned by using the labeled image of the rump images as the input to achieve regression prediction of cow appraisal key points of the cow rump images.(5)Automatic extraction of cow linear appraisal indicators.Used three-dimensional reconstruction technology to construct a three-dimensional coordinate system from the depth image and restore the spatial three-dimensional information of the image.The computer vision available appraisal indicator of the cow rump,the rump width,was automatically extracted by converting its measurement to the distance between the two cow appraisal key points of the rump images in the three-dimensional coordinate system.This paper achieved non-contact individual identification of the cow's rump and side,and automatic extraction of the rump width based on computer vision and deep learning algorithms.It can provide a reference for cow identification,cow linear appraisal based on other views,which is also conducive to the promotion and application of non-contact cow behavior detection and intelligent analysis.