Estimation of human motion intentions using high density EMG signals

HDEMG研究中的主要挑战之一是肌肉运动期间和不同实验会话之间的电极移位。为了解决这个问题，我们一直在考虑不同的技术，可以提供足够的信息的肌肉活动的变化，而不管电极的确切位置。因此，我们已经理解了HDEMG信号的空间变化可以提供足够的信息来区分具有与SEMG信号相似的肌肉活动模式测量的手指运动。在这个过程中，我们测量了6种不同类型的手指运动的HDEMG信号。然后使用预处理的HDEMG数据的均方根（EMS）值生成HDEMG信号的激活图。采用Gabor特征提取热图的空间变化，并采用基于纠错输出码的多类支持向量机分类器对6种不同的手指运动进行离线分类，平均准确率为95.8%。Gabor特征成功地从热图中提取了与肌肉活动相关的信息。进一步使用RMS值的时间序列数据来训练深度学习网络，以85%的准确率对6类运动进行分类。

One of the major challenge in HDEMG research is the electrode shift during muscle movement and between different experimental sessions. To address this we have being considering different techniques that can give enough information of variation of the muscle activity irrespective of the electrode's exact location. Thus, we have understood the spatial variation of the HDEMG signals can provide enough information to differentiate finger motions that have similar muscle activity pattern measure with SEMG signals. In the process we measured, HDEMG signals for 6 different types of finger motions. Then activations maps of the HDEMG signals were generated using the root mean square (EMS) values of the preprocessed HDEMG data. Gabor features were used extract spatial variations of the heat maps and error correcting output codes based multi class support vector machine classifier was used to classify 6 different finger motions with an average accuracy of 95.8%, in an offline study. The Gabor features were successful in extracting information related to the muscle activity from the heat maps. Further time series data of RMS values were used to train deep learning network to classify the 6 classes of motion with 85% accuracy.