In the past few years, a great interest for the classification of hand gestures with Deep Learning methods based on surface electromyography (sEMG) signals has been developed in the scientific community. In line with latest works in the field, the objective of our work is to design a novel Convolutional Neural Network architecture, for the classification of hand-gestures. Our model, while avoiding overfitting, did not perform significantly better compared to a much shallower network. The results suggest that the lack of diversity in the sEMG recordings between certain hand-gestures limits the performance of ML models.

However, the classification accuracy on a dataset we developed using a commercial device (Myo Armband) was substantially higher (approximately 24%) than a similar benchmark dataset recorded with the same device.

In order to contribute to the acquisition of sEMG data, particularly from devices that do not require professional calibration, we developed a sizeable sEMG dataset. Our dataset, MyoUP, was inspired by the Ninapro dataset and all of the recorded hand-gestures, are identical to some of the Ninapro (http://ninaweb.hevs.ch). The recording device we used was the Myo Armband, by Thalmic labs. The Myo Armband is a relatively cheap and easy-to-wear device, with a sampling frequency of 200Hz and 8 dry sEMG channels that has been widely adopted in scientific research.

The MyoUP dataset contains recordings from 8 intact subjects (3 females, 5 males; 1 left handed, 7 right handed; age 22.38 ± 1.06 years). The acquisition process was divided into three parts: 5 basic finger movements, 12 isotonic and isometric hand configurations and 5 grasping hand-gestures. Volunteers became accustomed with the procedure before performing each set of exercises. Subjects were instructed to repeat each gesture 5 times, for a 5sec period, interleaved with 5sec interruptions to avoid muscle fatigue. A supervisor assisted the subjects in wearing the Myo Armband to their dominant hand so that the device would be placed in a comfortable position for the subject and the device would detect the sEMG signals accurately. The sEMG was visible to the subject on a screen along with a picture of the hand-gesture that had to be performed.

https://github.com/tsagkas/MyoUP_dataset

By training our CNN with sEMG recordings from the MyoUP dataset, we managed to develop a real-time hand gesture recognition software.

N. Tsagkas, P. Tsinganos and A. Skodras, "On the Use of Deeper CNNs in Hand Gesture Recognition Based on sEMG Signals," 2019 10th International Conference on Information, Intelligence, Systems and Applications (IISA), PATRAS, Greece, 2019, pp. 1-4. doi: 10.1109/IISA.2019.8900709

URL: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8900709&isnumber=8900660