Data collection

The most important part of any machine learning model is the correctly labeled data. The amount of data required for an effective model depends on the use-case maturity requirements. For a Proof of Concept (PoC), you can collect 3-5 minutes of data per each gesture with 1-5 unique individuals. A production grade model will require significantly more data in terms of variability. The more unique users contribute to the training dataset, the better the model generalizes to new, unseen users.

For capturing data and monitoring the operations, use the Serial Monitor tool (Serial Terminal app) of the nRF Connect for Desktop app suite.

The serial terminal applies its own default buffering settings, which may need adjustment depending on your sampling rate and the duration of your recording session. You can modify these settings in the Settings tab.

For reference, the demo project uses a sampling frequency of 100 Hz. With a buffer limit of 100,000 lines, the terminal can store approximately 1000 seconds of data before it starts overwriting older samples.

1. Set Baud rate to 115200 in a Serial settings tab on the left side of the window.

   

2. With the port disconnected, Clear the console to ensure no residual data remains.

3. Connect to the port and begin performing the gestures.

   

   Perform the gesture repeatedly for 5 minutes, and follow these guidelines to ensure high-quality data collection:

   • Hold still for 3–5 seconds after starting.
   • Perform the gesture for 1–2 seconds.
   • Pause for approximately 1 second between gestures.
   • Vary speed, orientation, and intensity.

4. Once finished, click Disconnect from the port, and Write to File to store the captured data. Save the recorded data to a CSV file.

5. Repeat these steps for each gesture.