Dynamic Finger Task Identification Using Electromyography.

As high-fidelity wearable sensing technologies become available for control of physical devices and virtual environments, human-machine interaction (HMI) is becoming ubiquitous in our daily lives. Electromyography (EMG) is a wearable sensing modality that directly observes wearer intention by recording muscle activations. In this paper, we propose to use EMG signals to robustly identify dynamic task performance. EMG-based methods have been used with high success for static gesture recognition, but do not fare well with dynamic tasks. The methods that tackle the dynamic case rely on large amounts of data and fail to maintain robustness against unexpected movements. We present a method based on a single feature extracted from a set of time-series EMG signals that robustly captures a given dynamic task. This feature is used to train an algorithm using a single subject's data with the goal of dynamic task identification. The trained algorithm has high success rate (96%) when used for the same subject, and maintains this accuracy for new subjects with minimal calibration. This consistency demonstrates that our task signature identification method is agnostic to subject-specific variability. We further validate our method's robustness with unexpected non-target tasks and observe a low rate of false positive identification. Thus, the proposed method can be used to robustly detect the given task under the prescribed task condition for any new subject given the initially trained algorithm and as few as two repetitions of the new subject's EMG signals. The method is robust to unexpected dynamic behavior, requires only a small training dataset, and is among the first of its kind for dynamic task identification problems.