Research On Bovine Body Pose Estimation Based On Convolution Heatmap Regression

Bovine pose estimation refers to the process of detecting the position of each part of the bovine body from the image and calculating its orientation information and scale information.It describes the projection of each joint of the bovine body in the two-dimensional plane of the image through a line segment or rectangular frame.The distribution of the length of the segment or the size of the rectangular box represents the posture of the cow.Based on the estimation results of the body posture,the data about the physical indicators of the cattle can be further obtained,for example,the speed of movement,the time of stay,the amount of activity,etc.According to the physiological characteristics of cattle,these data can be used to assess the health status of cattle,which is of great significance for the intelligent breeding of cattle.In order to solve this problem,this paper takes the images of dairy cows taken from farms as research objects and proposes a posture estimation method based on concatenation of convolutional neural network modules.From the training data,an end-to-end prediction model is established.This model can directly output the line segment pose estimation map of the cattle body.The main research contents and conclusions are as follows:(1)Related technical principles and models.In this paper,a deep learning model based on convolutional neural network cascading is used to solve the bovine pose estimation problem.In order to ensure the feasibility of the model,this paper studies the theoretical knowledge of convolutional neural networks and decides to use VGGNets and FCN as the theoretical basis for the pose estimation model of bovine body.(2)Design and analysis of bovine body pose estimation model.In view of the research problem,using the idea of estimation of human pose estimation,we refer to the selection scheme of human pose features to select the key features for the bovine pose,and propose a network model based on cascading of two convolutional neural network modules.The first module is responsible for roughly detecting the corresponding heat maps of each characteristic part of the bovine body.The second module is responsible for feature regression of the stacked images of the original image and feature detection heatmaps.The cascade of models and the stacking of data enables an accurate estimation of the characteristics of the bovine body.(3)Research and production of model data sets.Taking into account the selection of the posture of the bovine body in the form of line segment,in order to obtain the coordinate data corresponding to the feature points in the image more conveniently,the development of the feature point calibration software is completed according to selection scheme of the feature points.For the data set of the bovine pose estimation model,this paper completed the production of standard input data by framing,cropping,and scaling the original video data.Then according to the characteristics of the bovine pose estimation cascade model,different types of standards output data are created for the two sub-models using the calibrated feature points.(4)Experiments and comparisons of paper models.In the experimental analysis stage,two main aspects of the work are done.The first aspect mainly analyzes the changes in the accuracy rate of the model in the training set and the verification set under different iterations in the training process and select the model with the best performance in the verification to carry out the test evaluation.The second aspect mainly compare the proposed algorithm model with the non-cascaded pose estimation model and the cascaded pose estimation model based on quadratic regression.Experimental results show that the accuracy of the bovine pose estimation model based on convolution heatmap regression can reach up to 85.46% on the training set,and the highest accuracy on the validation set can reach 83.91%.The accuracy of the optimal model on the test set is 80.24%.Compared with the non-cascaded model,the prediction accuracy rate of the model used in the test set is increased by up to 22.31%.Compared with the cascadebased feature quadratic regression model,the accuracy rate is improved by 2.37%.The experimental results demonstrate the feasibility and robustness of the bovine pose estimation method proposed in this paper.Provide a technical reference for the intelligent breeding technology of cattle.