EMG-EPN-107

The dataset, named EMG-EPN-107, contains electromyography (EMG) and orientation data collected from 107 participants for the study and benchmarking of hand gesture recognition systems. It was developed by the Artificial Intelligence and Computer Vision Research Laboratory “Alan Turing” of Escuela Politécnica Nacional, Quito, Ecuador. Data acquisition was carried out using two devices: the G-ForcePro armband, which records EMG signals at 500 Hz and inertial measurement unit (IMU) data at 50 Hz, and the Myo armband, which records EMG signals at 200 Hz and IMU data at 50 Hz, allowing the capture of both muscular activity and motion-related information from the forearm.

The dataset comprises 11 hand gestures grouped into 12 classes, including five gestures (wave-in, wave-out, fist, open, and pinch), six directional gestures (up, down, left, right, forward, and backward), and a relax or no gesture class. Each participant performed 30 repetitions per gesture, resulting in 360 samples per user, with each sample lasting 5 seconds. In addition, each user performed five repetitions of a synchronisation gesture, each lasting 10 seconds, which are intended to support signal alignment and calibration.

The dataset is organised into two levels of division. The 107 participants are split into a training group comprising 54 users and a testing group comprising 53 users. For each participant, the recorded samples are further divided into training and testing repetitions, each containing 180 samples. For benchmarking purposes, labels are partially hidden: only the samples corresponding to the testing repetitions of users in the testing group have hidden labels, enabling unbiased evaluation of classification models.

Each user folder contains a photograph of the participant’s forearm showing the placement of the acquisition device, where the face is not visible in order to preserve anonymity. Additionally, each folder includes MATLAB (.mat) files that store both the recorded signals and associated metadata within a structured variable named userData. The most relevant information includes non-identifiable participant data such as age, gender, handedness, occupation, and forearm-related measurements; acquisition metadata such as repetition counts, duration per repetition, synchronisation settings, and recording date; and device information including the acquisition system used and EMG sampling rate. The recorded data are organised into three main subsets corresponding to synchronisation samples, training samples, and testing samples.