InGesture Dataset

Data are organized in a clear, hierarchical folder structure under a single root directory named InGesture. At the top level, each participant has a dedicated folder (e.g., Subject01, Subject02, …), and within those, one or more session subfolders contain the raw recordings and segmented gesture files. Two complementary formats are provided:

Segmented gesture files

Each CSV file corresponds to a single occurrence of a specific gesture.

Filenames encode the gesture label, repetition index, and sample count (for example, fluid_intake_1_900.csv).

Columns include:

timestamp (sensor capture time)

x, y, z (accelerometer readings in m/s²)

gx, gy, gz (gyroscope readings in °/s)

Because each file contains only one gesture class, no label column is needed—classification can be inferred from the filename or parent folder.

Continuous session files

Each file captures an uninterrupted ~10-minute recording at 200 Hz, yielding about 120,000 samples per session.

Columns include:

timestamp

accX, accY, accZ (linear acceleration)

gyrX, gyrY, gyrZ (angular velocity)

label (integer code 0–7 denoting the current gesture)

Gesture codes map as follows:
0 – free activity
1 – fluid intake
2 – answering phone
3 – scratching head
4 – passing hand over face
5 – adjusting glasses/touching temples
6 – holding chin
7 – stretching hands behind neck

Across 65 sessions (most participants contributed one session, some two or more), there are roughly 2,080 segmented files (8 gestures × 4 repetitions × 65 sessions) and 65 continuous recordings. All data were acquired with a WT901BLECL5 inertial measurement unit secured on the dominant wrist and sampled at 200 Hz. Real-time gesture boundaries were annotated via a synchronized mobile app, ensuring that each sensor sample aligns precisely with its gesture label.