| 引用本文: | 胡旭晖,宋爱国,李会军.基于表面肌电图像的灵巧假手控制系统[J].控制理论与应用,2018,35(12):1707~1714.[点击复制] |
| HU Xu-hui,SONG Ai-guo,LI-Huijun.A Dexterous Robot Hand Control System Based on Surface Electromyography[J].Control Theory and Technology,2018,35(12):1707~1714.[点击复制] |
|
| 基于表面肌电图像的灵巧假手控制系统 |
| A Dexterous Robot Hand Control System Based on Surface Electromyography |
| 摘要点击 2175 全文点击 832 投稿时间：2018-06-15 修订日期：2018-12-24 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2018.80448 |
| 2018,35(12):1707-1714 |
| 中文关键词 手势识别 表面肌电图像 灵巧假手 |
| 英文关键词 Hand Gesture Recognition sEMG image dexterous hand |
| 基金项目 国家自然科学基金重点项目 |
|
| 中文摘要 |
| 设计了一种通过佩戴阵列型表面肌电传感器, 实时识别受试者的8种手势, 并控制一个自主研发的六自由度灵巧操作假手进行同步动作的人—机协同控制系统. 控制假手的手势识别策略基于神经网络算法, 受试者仅需在首次训练阶段重复完成预先设定的8种手势动作(分别为放松、手腕外翻、手腕内翻、握拳、伸掌、手势二、手势三和竖大拇指）, 之后该系统即能够实时识别受试者随机完成8种手势中的任意一种手势. 本文提出的网络参数随机搜索算法和梯度下降算法, 与目前同规模的神经网络相比提高了网络的训练速度和手势预测精度；该手势识别算法使用Tensorflow机器学习框架学习权值并进行了可视化分析；采用经过优化的手势训练方式既缩短了受试者的手势训练时间, 同时提高了手势训练的熟练度. 本文对一名肌肉无损伤的受试者进行表面肌电信号采集、训练和预测, 对8种手势的综合预测精度达到97%, 且再次佩戴时不再需要进行训练。受试者实际控制假手时, 使用投票算法对实时手势预测结果进行深度优化, 最终假手的动作同步率到达99%. |
| 英文摘要 |
| In this paper, a man-machine coordinated control system was designed by an array sEMG sensor to identify the eight gestures of the subject in real time and operate the self-developed 6 DOFs dexterous hand to perform synchronized actions. The gesture recognition strategy that controls the prosthetic hand is based on a neural network algorithm. The subject only needs to repeatedly complete the preset eight gestures in the first training phase (respectively relax, wrist in, wrist out, fist, palm extensions, gestures two, gestures three and vertical thumbs), then the system is able to recognize in real time any of the eight gestures randomly performed by the subject. The improved network parameter random search algorithm and gradient descent algorithm proposed in this paper increase the network training speed and gesture prediction accuracy compared with the current neural network of the same scale. The gesture recognition algorithm uses the Tensorflow to learn the weights and performs visual analysis. The optimized gesture training method not only shortens the subject''s gesture training time, but also improves the proficiency of gesture training. This article collected, trained, and predicted surface EMG signals for a muscle-non-invasive subject. The overall accuracy of the eight gesture predictions was 97%, and training was no longer required once again. When the subjects actually controlled the prosthesis, the voting algorithm was used to optimize the prediction results of the real-time gestures. The eventual synchronization rate of the prosthetic hand reached 99%. |
|
|
|
|
|