The Effect of Dataset Size on EMG Gesture Recognition Under Diverse Limb Positions
Hongquan Le, Marc in het Panhuis, Geoffrey M. Spinks, Gürsel Alıcı · 2024
Addressing non-stationarities in surface Electromyography (sEMG) gesture recognition remains a topic of extensive research. A common approach involves collecting sEMG data under diverse conditions, yet the sufficient quantity of data required has not been thoroughly investigated. This study investigates the impact of dataset size on recognition errors across various limb positions. We have found that dataset size significantly affects recognition error in both Linear Discriminant Analysis (LDA) and 1-Dimensional Convolutional Neural Networks (CNN). Notably, CNN only predicted limb-position invariant representations with a minimum of five repetitions for each limb position with balanced dataset. No performance difference was observed between CNN and LDA when recognizing unseen limb positions from two repetitions for each limb position. Finally, larger dataset sizes were observed to increase the recognition errors in classical LDA in case of unbalanced dataset.