CASIA OpenIR  > 复杂系统认知与决策实验室  > 先进机器人
面向中医机器人的手部穴位识别与定位方法研究
孙玲瑶
2021-05-25
Pages93
Subtype硕士
Abstract

     中医作为我国具有原创优势的医疗卫生资源,在我国的经济社会发展、民生保障与改善问题上发挥着重大作用。其中,包含针灸、艾灸及穴位按摩在内的中医穴位疗法,因不良反应与副作用较小,常常作为国内患者优先选择的治疗方案。然而,在风湿性疾病多发的秋冬季节,需要接受穴位疗法的患者增多,在超负荷工作的情况下,医师的劳累程度会直接影响到患者的疗效。此外,穴位疗法治疗效果的好坏,除了与辩证、治法、选穴、手法等因素有关之外,还与所找穴位位置的正确与否有着极大的关系。因此,研究能够自动进行穴位识别及定位的机器人,成为了实现智能化穴位治疗的前提,具有十分重要的现实意义。本文将计算机视觉和深度学习技术应用到穴位识别任务领域当中,研究端到端的穴位识别算法,并基于ROS机器人平台完成了结合机械臂的穴位三维定位,为中医机器人的发展提供了一定的理论基础和实验验证。本文的主要工作如下:

    (1)适用于穴位识别的基础卷积神经网络模型的分析研究

      本文针对穴位识别基础网络模型选择的问题做出了分析研究。首先,参照人体关键点检测领域COCO数据集的格式,结合传统中医腧穴定位方法,标注并构建了一个手部穴位数据集,为后续模型的训练与实验验证提供了基础。然后,参照人体关键点检测领域中PCK指标的评价方式,设计了穴位识别量化评价标准,为本文的对比实验提供了评价依据。随后,在所建立的穴位识别数据集上,对CPM、SHN和HRNet三个深度卷积神经网络进行了对比实验,并使用穴位识别量化标准进行评估,最终根据实验结果选择HRNet作为进行穴位识别任务的基础网络模型,为后续穴位识别算法的研究提供基础。

    (2)基于改进HRNet的手部穴位识别算法研究

      本文在HRNet网络模型的基础上,面向穴位识别任务需求,从实际场景出发,做出了针对性改进。分别针对网络、数据和损失函数三个方面做出改进。首先,在网络方面,提出了基于级联方式的的网络框架改进,通过级联手部检测框来提高输入图像的分辨率以提高检测精度。然后,在数据方面,在穴位点热力图生成部分引入了基于offset的无偏编码解码方法,以消除穴位点坐标编解码过程中的统计误差。随后,在损失函数方面,引入了Soft-OHKM方法,对不同穴位点回传的损失函数进行加权,对损失函数值更大的点给予更大的梯度回传,以更好地对模型拟合进行修正。最后,在所建立的穴位识别数据集上进行了实验,实验结果验证了穴位识别算法的有效性。

    (3)基于ROS的中医穴位识别与定位机器人系统

      本文将基于改进HRNet的穴位识别算法在ROS机器人平台上进行了系统性实现,完成了中医穴位识别与定位机器人系统的软件设计,搭建了系统化实验平台,结合Kinect深度相机和Kinova 7自由度机械臂,在系统中实现了穴位的识别与三维定位,并控制机械臂到达穴位点处,对穴位点的位置作出指示。该系统模拟了中医机器人进行针灸、艾灸或穴位按摩之前的自动找穴过程,为中医智能机器人的后续研发提供了有力的研究基础。

Other Abstract

      Traditional Chinese Medicine(TCM), which is a unique and originals resource in China, plays an important role in economic and social development, people's livelihood security and improvement. Among them, TCM acupoint therapy, including acupuncture, moxibustion, and acupoint massage, is often the preferred treatment option for domestic patients because of its low side effects and side effects. However, in the autumn and winter seasons when rheumatic diseases are frequent, more patients need acupuncture therapy which will increase workload. And in the case of overloaded work, the degree of the doctor's fatigue will directly affect the patient's curative effect. In addition, the effectiveness of acupoint therapy is not only related to factors such as dialectics, treatment, acupoint, selection, manipulation, etc., but also has a great relationship with whether the location of the acupoint is correct or not. Therefore, to invent a robot, which can automatically identify and position acupoints, has become the basic prerequisite for the realization of intelligent Chinese medicine acupoint therapy. And it also has very important practical significance. This paper applies computer vision and deep learning technology to the field of acupoint recognition tasks, studies end-to-end acupoint recognition algorithms, and completes the three-dimensional positioning of acupoints combined with a robotic arm based on the ROS robot platform, which provides a certain theoretical basis for the development of Chinese medicine robots and experimental verification.

      The main work of this paper is as follows:

      (1) Analysis and research of basic convolutional neural network model suitable for acupoint recognition.

      This article analyzes and studies the selection of basic network models suitable for acupoint recognition. First of all, this article refers to the COCO data set format in the field of human key point detection, combined with the traditional Chinese medicine acupoint positioning method, annotated and constructed a hand acupoint recognition data set, which provides a practical basis for subsequent model training and experimental verification. Then, referring to the evaluation method of PCK indicators in the field of human key point detection, this paper designs a quantitative evaluation standard for acupoint recognition, which provides an evaluation basis for the comparison experiment. Subsequently, based on the established acupoint recognition data set, this paper conducted a comparative experiment on the three deep convolutional neural networks of CPM, SHN, and HRNet, and used the quantitative criteria for acupoint recognition to evaluate then selecting HRNet as the most suitable one based on the experimental results. The HRnet, the most suitable model for the task of aciwoint recognition, is used as the research foundation for the subsequent network model of acupoint recognition.

      (2)Research on hand acupoint recognition algorithm based on improved HRNet.

      Based on the HRNet basic network model, this article aims at the needs of acupoint recognition tasks. starting from the actual scene of the experiment, and making targeted improvements. This paper makes improvements in three aspects: network, data and loss function. First, in terms of network, a network framework improvement based on cascading method is proposed. In order to get better detection accuracy, the resolution of the input image is improved by cascading hand detection frames. Then, in terms of data processing, an offset-based unbiased encoding and decoding method is introduced in the acupoint point heat map generation part to eliminate the statistical error caused by the process of encoding and decoding the acupoint point coordinates; And in terms of loss function. The Soft-OHKM method is introduced to weight the loss function returned by different acupoint points, and give a larger gradient return to the point with a larger loss function value to better correct the model fitting. Finally, experiments are carried out on the established acupoint recognition data set, and the experimental results can verify the effectiveness of the acupoint recognition algorithm.,

      (3)TCM acupoint recognition and positioning robot system based on ROS.

      In this paper, the improved acupoipt recognition algorithm proposed in this paper is systematically implemented on the ROS robot platform, and the software design of the TCM acupoint recognition and positioning robot system is completed, and a systematic experimental platform is built, combining the Kinect camera and Kinova arm. This system could carry out acupoint recognition and three-dimensional positioning of acupoints, and control the Kinova arm to reach the acupoint, and give instructions to the position of the acupoint. This system simulates the automatic point-finding process before acupuncture, moxibustion or massage operation performed by a Chinese medicine robot, and provides a strong research foundation for the follow-up research and development of Chinese medicine intelligent robots.

Keyword中医机器人 穴位识别 穴位定位 卷积神经网络 关键点检测
Language中文
Sub direction classification机器人感知与决策
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/45010
Collection复杂系统认知与决策实验室_先进机器人
Recommended Citation
GB/T 7714
孙玲瑶. 面向中医机器人的手部穴位识别与定位方法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2021.
Files in This Item:
File Name/Size DocType Version Access License
孙玲瑶毕业论文.pdf(7456KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[孙玲瑶]'s Articles
Baidu academic
Similar articles in Baidu academic
[孙玲瑶]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[孙玲瑶]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.