OPTIMIZATION METHOD FOR GESTURE CLASSIFIER

Abstract

The work is devoted to the study of the possibility of optimizing the process of synthesis of
gesture classifiers by selecting the most significant channels of electromyographic (EMG) activity
of the muscles of the forearm. The first part of the study is devoted to the development and analysis
of the performance of gesture classifiers with a different number of EMG channels, ranked by
significance based on the Pearson criterion. The solution of the problem of classification of gestures
by EMG signals was first implemented on the basis of ensembles of decision trees trained by
the gradient boosting method. For this, software was developed that allows automatic synthesis
and training of gesture classifiers. Next, a series of studies was carried out to find the optimal
number of EMG channels based on three criteria: the classifier learning rate, the performance of
the trained model, and the area under the ROC AUC error curve. To do this, a cycle of training
and testing of the classifier was carried out for data sets recorded at different positions of the electrodes
on the forearm. Then, range diagrams of the studied criteria were constructed for various
numbers of EMG channels involved in the work from 1 to 8, ranked by significance in each of the
samples. It was found that the optimal number of EMG channels involved under the experimental
conditions was 3-6, since a further increase did not lead to a decrease in the classification error,
while significantly degrading the performance. The proposed method allows you to automatically
select the channels, the electrodes of which are located above the most informative areas of the
forearm in case of an accidental change in the position of the sensors. The second part of the work
contains the results of a full-scale experiment to demonstrate the possibility of controlling a
wheeled robot through EMG analysis.