Data for: Towards an End-To-End System for Real-Time Gesture Recognition from Surface Vibrations

The dataset contains vibration signals recorded for gesture recognition on a standard office desk. Four piezoelectric sensor units were mounted on the underside of the tabletop in an orthogonal cross configuration and connected to the custom acquisition hardware. The signals were recorded at 1 kHz per channel and stored as .xdf files.

The study includes 15 participants performing six gesture classes:
- swipe-left
- swipe-right
- swipe-up
- swipe-down
- tap
- knock

For each participant, ten sessions were recorded. Each session contains ten repetitions of each gesture class. This results in:
- 60 gesture events per session
- 600 gesture events per participant
- 9,000 gesture events in total

Gestures were performed near the geometric center of the tabletop, aligned with the center of the four-sensor layout. Swipe and tap gestures were performed with one finger, while knock gestures were performed with the knuckles. Event timestamps were generated with a semi-automatic annotation workflow and manually refined where needed.

Raw Data
The raw data are published as .xdf recordings. They contain the acquisition stream recorded during data acquisition.

Raw data characteristics:
- Acquisition format: .xdf
- Acquisition stream: vibration-sensor signal stream in each .xdf recording
- Recorded acquisition channels: 0-8
- Vibration sensor channels used for analysis: 1, 4, 6, 8
- Channel indexing: zero-based

Preprocessed Data
The preprocessed dataset contains event-centered samples exported from the raw .xdf recordings.

Preprocessed data characteristics:
- Export format: one .npz file per event
- Stored array: x
- Data type: float32
- Shape: [4, T]
- Channel order: [1, 4, 6, 8]
- Window length: 1000 ms

The exported .npz samples have already been preprocessed and sliced before publication.

Applied preprocessing and slicing settings:

- filter_band = 225-375
- norm_method = zscore
- frame_ms = 1000

Recommended Use
Use the raw .xdf data when the original recordings or alternative processing steps are needed.

Use the preprocessed .npz dataset when working directly with ready-to-load event windows for model training or evaluation.