# BGT60 Gesture Recognition Dataset
This dataset was originally acquired as part of Andrea Ronco's Master's Thesis.
This dataset was used in the paper "TinyssimoRadar: In-Ear Hand Gesture Recognition with Ultra-Low Power mmWave Radars", published at the 9th ACM/IEEE Conference on Internet of Things Design and Implementation (IoTDI) in 2024.

## Dataset Structure
The dataset is organized into individual folders, each corresponding to a participant who contributed data. Within each participant folder, there are subfolders dedicated to different gestures performed by that individual. These gesture-specific subfolders contain a single .npy binary file encapsulating the raw data for that gesture.

## Radar Configuration
The radar system utilized a sawtooth chirp pattern spanning from 58.5 GHz to 63.5 GHz, with a bandwidth of 5 GHz and a chirp slope of 450 MHz µs−1. Each chirp was spaced by 700 µs within a sequence of 96 chirps, resulting in a frame rate of 13.5 FPS. With only 16 samples acquired per chirp at a 2 MHz sampling rate, and considering hardware overhead, the total chirp time amounted to 13 µs.

## Data Format
The .npy binary files contain the raw output from the sensor. The format of the ndarrya is the following (N, F, C, S, A) with
- I: Instance. Each subject recorded 50 instances per gesture, with the exception of the first subject (*person_00*) which recorded 700 instances per gesture.
- F: Frames. Each sample is a sequence of 16 frames with timing figure as described in the configuration section above.
- C: Chirps: Each frame contains a sequence of 96 chirps, with no extra time delay the end of the sequence/frame.
- S: Samples: The ADC was configured to sample the IF signal with 16 samples per chirp.
- A: Antennas: All three available receiving antennas were used during the acquisition.

This results in each binary file being (50, 16, 96, 16, 3) for all subjects except the first one where instead we have (700, 16, 96, 16, 3)