基于级联卷积神经网络的山羊姿态估计

Goat pose estimation based on deep cascaded convolutional neural networks

为了实现单只和多只山羊的姿态估计,试验采集陕西关中地区某山羊养殖场的山羊视频和图片,标注山羊的轮廓、关键部位和连接,构建山羊姿态数据集;将山羊骨架分解为18个关键部位和17个连接,并通过深度卷积神经网络预测、匹配、提取山羊骨架,即采用单分支级联卷积神经网络预测单只山羊的关键部位,并按顺序连接提取山羊骨架,采用双分支级联卷积神经网络同时预测多只山羊的关键部位和连接,再用匈牙利算法进行匹配来提取每只山羊的骨架。结果表明:单分支级联卷积神经网络模型的关键部位正确预测率达到了82.50%;双分支级联卷积神经网络模型关键部位和连接的正确预测率分别是85.32%和83.76%,骨架提取的正确率为78.26%;平均对象关键点相似度为0.85时两个模型的平均召回率和平均准确率均达到80%。说明本试验构建的神经网络模型预测山羊骨架的准确性和效率优良。

In order to realize the pose estimation of single and multiple goats,videos and pictures of goats from breeding farms in Guanzhong area,Shaanxi Province were collected,and the outlines,key parts and connections of goats were marked to construct data sets.The goat skeleton was decomposed into 18 key parts and 17 connections,and the goat skeleton was extracted by deep convolutional neural network prediction and matching,that is,the single branch cascade convolutional neural network was used to predict thekey parts of a single goat,and the goat skeleton was extracted by sequential connections.The two-branch cascade convolutional neural network were used to predict the key parts and connections of multiple goats,and then the Hungarian algorithm was used to match and extract the skeleton of each goat.The results showed that the accuracy rate of key parts by single branch cascade convolutional network model reached 82.50%.The accuracy of key parts and connections by two-branch cascades convolutional network model were 85.32%and 83.76%,respectively,and the correct extraction rate of skeleton was 78.26%.When the average object key point similarity was 0.85,the average recall rate and average accuracy of the two models both reached 80%.The results indicated that the neural network model constructed by this experiment was accurate and efficient in predicting goat skeleton.