基于气压肌动图和改进神经模糊推理系统的手势识别研究

手势识别是人机交互领域的重要研究内容, 为截肢患者控制智能假肢手提供基础. 当前主流方法之一是利用表面肌电图(Electromyogram, EMG)识别手部运动意图, 但肌电信号存在信号弱和易受噪声、汗液、疲劳影响等缺点. 同时肌电图在识别准确率方面, 尤其是截肢患者手势识别方面仍然具有较大的提升空间. 针对这些问题, 设计了基于气压肌动图(Pressure-based mechanomyogram, pMMG)的穿戴式信号采集装置, 为手势识别提供了优质的信号源. 结合深度神经网络中全连接层结构、典型抽样和标准正则化技术, 提出了一种改进多类神经模糊推理系统(Improved multicalss neural fuzzy inference system, IMNFIS), 与传统自适应神经模糊推理系统(Adaptive neural fuzzy inference system, ANFIS)相比, 泛化能力得到显著提升. 招募了7名健康受试者和1名截肢受试者, 并用8种算法开展离线实验. 所提方法在残疾人手势识别实验中取得了97.25%的最高平均准确率, 在健康人手势识别实验中取得了98.18%的最高平均准确率. 与近年公开报道的多种手势识别研究相比, 所提方法的综合性能更优.

Research on Gesture Recognition Based on Pressure-based Mechanomyogram and Improved Neural Fuzzy Inference System

Gesture recognition, which provides a foundation for amputees to control smart prosthetic hands, is an important research content in the field of human-robot interaction. One of the current mainstream methods is using surface electromyogram (EMG) to identify the intention of hand motion, while EMG signals are weak and show some shortages of being interfered easily by noise, sweat, fatigue, etc. At the same time, there is still large room for the improvement in the recognition accuracy, especially in the gesture recognition of amputees by using EMG. To solve these problems, a wearable signal acquisition device based on pressure-based mechanomyogram (pMMG) is designed in this paper, which provides a signal source of high quality for gesture recognition. We proposed an improved multicalss neural fuzzy inference system (IMNFIS) by combining the full connection layer structure, typical sampling and uniformed regularization techniques in deep neural network, which significantly improved the generalization ability compared with the traditional adaptive neural fuzzy inference system (ANFIS). We recruited seven healthy subjects and an amputee subject, and then conducted an offline experiment in which eight algorithms are used. The proposed method got the highest average accuracy of 97.25% in the experiment of the disabled, and 98.18% in the experiment of the healthy. Compared with many reported gesture recognition researches in recent years, the method proposed in this paper achieves the better comprehensive performance.