Enhancement of sEMG-based gesture classification using mahanobis distance metric

This paper investigates how to improve the performance of Human-Machine Interface systems using surface electromyogram (sEMG) signals to detect human intention in the forearm muscles. Our goal is to make a robust bionic interface that can interact with computers or robots with human motion. Since recognition performance is highly sensitive to the number and the types of target motions to classify, features of them should be properly handled to generate a total number of user commands. This paper introduces two metric parameters: the Repeatability Index (RI) and the Separability Index (SI), both of which quantify feature distributions in feature space. By evaluating the distributions, we could judge which distribution is desired or not for a training dataset. Furthermore, we could know that which motions should be included or not to improve the performance of the system. Among possible target motion sets, we exploited proposed parameters to find optimal target motions, which lead to achieving precise recognition rates. We confirm the validity of this method through off-line simulation experiments. Using a database of 10 subjects, rather than emphasizing a high level of accuracy, we focused instead on determining the correlation between proposed paratmer and system performances. This research could accelerate the development of wearable sensors, which could then become a familiar and easily applicable part of our daily lives.