Surface electromyography (sEMG) is one of the most prevalent methods of upper-limb prosthesis control nowadays, especially when combined with machine learning. However, this approach requires a significant amount of annotated data, including gesture classes and joint angles, the recording of which often requires expensive equipment and a major time expenditure. In this paper, we present an sEMG dataset intended for use in prosthesis control system development, and a novel method for the accurate reconstruction of finger movements using depth cameras. We recruited 6 able-bodied volunteers and recorded a set of 7 hand gestures, which we designed to be easily classifiable and similar to those found in other, major publicly available datasets. For data capture, we used an 8-channel device with semi-dry electrodes and a sampling rate of 500 samples-per-second, while also recording finger movements using a pair of commercially available depth cameras. Additionally, we developed an algorithm for the accurate localization of the joints of the hands using 3D point clouds.
Paper
Full text
A Surface Electromyography Dataset for Hand Gesture Recognition
Semantic Scholar · Engineering · 2022
Abstract
Surface electromyography (sEMG) is one of the most prevalent methods of upper-limb prosthesis control nowadays, especially when combined with machine learning. However, this approach requires a significant amount of annotated data, including gesture classes and joint angles, the recording of which often requires expensive equipment and a major time expenditure. In this paper, we present an sEMG dataset intended for use in prosthesis control system development, and a novel method for the accurate reconstruction of finger movements using depth cameras. We recruited 6 able-bodied volunteers and recorded a set of 7 hand gestures, which we designed to be easily classifiable and similar to those found in other, major publicly available datasets. For data capture, we used an 8-channel device with semi-dry electrodes and a sampling rate of 500 samples-per-second, while also recording finger movements using a pair of commercially available depth cameras. Additionally, we developed an algorithm for the accurate localization of the joints of the hands using 3D point clouds.