【摘要】 为了实现畜牧业集约化、规模化、精准化管理中及时分析动物的运动姿态，针对不同环境中动物关节点识别准确率较低的问题，提出了一种双路径网络(Dual path net, Dpnet)特征提取方法。首先，在HRnet(High resolution net)网络输出分支后面添加混合空洞卷积层(Hybrid dilated convolution, HDC)和并行注意力模块(Parallel attention module, PAM)以识别关键点坐标。然后，利用模型生成的热力图与本研究设定的精确圈进一步精准定位关键点位置，通过斜率法区分并匹配动物四肢的关节点归属，从而获取正确的骨架结构。最后，模型通过骨架结构提取并分析了不同速度下动物的同侧步距、步频、关节角度与步态占空比4种运动参数。试验结果表明，该方法中模型识别分类动物5类关键点的精确率为91.1%,召回率为91.0%,5类关键点的平均相似度为87.0%,可有效监测农场或牧场犬类、牛群、羊群等个体的健康状况。更多还原

【Abstract】 In order to achieve more intensive, large-scale and precise management of the animal husbandry, timely analysis of animal movement posture is a key link in animal breeding. Aiming at the problem of low recognition accuracy of animal joint points in different environments, a feature extraction method based on dual path network(Dpnet) was proposed. Firstly, the hybrid dilated convolution(HDC) and the parallel attention module(PAM) were added after the output branches of the high-resolution network(HRnet) to identify the coordinates of key points. Secondly, the heat map generated by the model and the precise circle set in this study were used to further accurately locate the positions of key points, the slope method was used to distinguish and match the joint points of the animal limbs, so as to obtain the correct skeleton structure. Finally, four motion parameters of animals(ipsilateral stride, stride frequency, joint angle and gait duty cycle) were extracted and analyzed by the model at different speeds through the skeleton structure. The experimental results showed that the accuracy rate of the model to identify and classify the key points of the animal was 91.1%, the recall rate was 91.0%, and the average similarity of the five key points was 87.0%, indicating that this method could effectively monitor the health status of dogs, cattle, sheep and other individuals on farms or ranches.更多还原