ACTA THERIOLOGICA SINICA ›› 2022, Vol. 42 ›› Issue (4): 451-460.DOI: 10.16829/j.slxb.150639

• METHOD AND TECHNOLOGY • Previous Articles     Next Articles

Giant panda pose estimation method based on high resolution net

Yu QI1(), Han SU1, Rong HOU2,3, Peng LIU2,3, Peng CHEN2,3(), Hangxing ZANG1, Zhihe ZHANG3   

  1. 1.School of Computer Science, Sichuan Normal University, Chengdu 610101, China
    2.Chengdu Research Base of Giant Panda Breeding, Sichuan Key Laboratory of Conservation Biology for Endangered Wildlife, Chengdu 610086, China
    3.Sichuan Academy of Giant Panda, Chengdu 610081, China
  • Received:2021-11-17 Accepted:2022-03-14 Online:2022-07-30 Published:2022-07-22
  • Contact: Peng CHEN


漆愚1(), 苏菡1, 侯蓉2,3, 刘鹏2,3, 陈鹏2,3(), 臧航行1, 张志和3   

  1. 1.四川师范大学计算机科学学院,成都 610101
    2.成都大熊猫繁育研究基地,四川省濒危野生动物保护生物学 重点实验室,成都 610086
    3.四川省大熊猫科学研究院,成都 610081
  • 通讯作者: 陈鹏
  • 作者简介:漆愚 (1997- ),男,硕士研究生,主要从事计算机视觉和姿态估计研究. E-mail:
  • 基金资助:


Long-term behavioral monitoring of captive giant pandas (Ailuropoda melanoleuca) can help animal managers better understand the panda’s physiological cycle and health status in a timely manner, and help breeding facilities quickly take corresponding husbandry actions to improve breeding management. At present, neither animal managers nor scientists can monitor giant pandas 24 hours a day and obtain corresponding behavioral information on time. Accurate animal pose estimation is an important factor in animal behavior research and is also the basis for many downstream applications. Understanding the pose of giant pandas can greatly promote the research of panda behavior and improve its conservation and management. In order to improve the accuracy of giant panda pose estimation in complex environments, this paper proposed a pose estimation method based on the high-resolution network HRNet-32. To solve the problem of large-scale differences in different parts of the giant pandas, an atrous spatial pyramid pooling module was introduced in HRNet-32, which used dilated convolution with different dilated rates to form a similar pyramid form, so as to capture multi-scale information while enhancing the feature’s receptive field. Meanwhile, the giant panda pose estimation was regarded as a homogeneous multi-task learning problem, the joint points of the giant panda were grouped, and the part-based multi-branch structure was introduced to learn the representations specific to each part group. The results of several comparison experiments show that the model proposed in this paper, PCK@0.05, had a high detection accuracy (81.51%). The method proposed in this paper can provide technical support for the behavioral analysis and health assessment of giant pandas.

Key words: Giant panda, Posture estimation, Image analysis, Deep learning


对圈养大熊猫 (Ailuropoda melanoleuca) 开展长期行为监测能及时了解其所处生理周期和健康状况,有助于繁殖饲养机构迅速采取相应繁育保护措施提高饲养管理水平,但目前无法对大熊猫进行24 h监控并及时地获得相应的行为信息。准确的动物姿态估计是动物行为研究的关键,也是诸多下游应用的基础。了解大熊猫的姿态可以促进大熊猫行为研究并提升保护管理水平。为了提高复杂环境下大熊猫姿态估计的准确率,本文以高分辨率网络 (High resolution net, HRNet) 为基础网络架构提出了一种大熊猫姿态估计方法:针对大熊猫不同部位尺度差异较大的问题,在HRNet-32中引入了空洞空间金字塔池化 (Atrous spatial pyramid pooling, ASPP) 模块,在提升特征感受野的同时捕获多尺度信息;同时对大熊猫身体关键点进行分组,引入基于部位的多分支结构来学习特定于每个部位组的表征。多次对比实验结果表明本文所用模型具有较高的检测精度:在PCK@0.05中所用模型精度达到了81.51%。本文提出的方法可为大熊猫的行为分析和健康评估提供技术支撑。

关键词: 大熊猫, 姿态估计, 图像分析, 深度学习

CLC Number: