Classification of gross upper limb movements using upper arm electromyographic features

This research paper explores the possibility of using Electromyogram (EMG) signals for classifying point to point upper limb movements during dynamic muscle contraction in the context of human-robot interactions. Previous studies have mostly focused on classifiers for gesture recognition using steady state EMG. Only few studies have used non-steady-state EMG classifier when gross upper arm muscles are in motion. To investigate it further, our study was designed to take EMG measurements from 4 upper limb muscles of 10 participants while interacting with HapticMaster robot in assisting mode. The participants were asked to move the robotic arm in a rectangular path consisting of 4 segments named S1 to S4. The EMG signals were analyzed by splitting them into non-overlapping windows of 100 milliseconds width. The initial windows within the initial 1 seconds of each segment iteration were considered to train and test a Support Vector Machine classifier. Various EMG features were calculated for different number of windows and used for classifying different segments. For the different combinations of features and muscles, it was noticed that the near-the-body segments S1 and S4 displayed the highest median accuracy for the feature combination (Waveform Length + Mean Average Value + Zero Crossing Count + Signal Slope Change) which were 100% each. For the same feature combination, it was also noticed that the segments S2 and S3 had the least accuracy, 76.2% and 73.8% respectively, possibly due to the away-from-body movements. In general, the accuracy was found to be more stable and higher for S1 and S4 segments. Considering 700 milliseconds (so 7 windows) for classification provided the best accuracy and the best muscle combination was Trapezius + Deltoid + Biceps Brachii + Triceps Brachii.