<![CDATA[Operating a smart board in class requires teachers to remain near the screen, limiting mobility while teaching. This study develops a wearable IMU gesture interface for hands-free presentation control using a Tiny Machine Learning (TinyML) approach. The system consists of an MPU6050 sensor (6-axis IMU) and an XIAO ESP32-S3 microcontroller mounted on a glove, recognizing five gesture classes: idle, flick_up, wave_left, wave_right, and flick_down. Three architectures—Random Forest (RF), Multi-Layer Perceptron (MLP), and 1D Convolutional Neural Network (CNN1D)—were evaluated on a dataset of 5,258 gesture windows from 7 users at a 100 Hz sampling frequency. Leave-One-Subject-Out (LOSO) cross-validation was used to measure cross-user generalization. RF achieved the highest accuracy of 93.91% (±8.98%), followed by MLP at 92.64% (±8.59%) and CNN1D at 88.95% (±14.91%). On the ESP32-S3, MLP recorded the best inference latency (0.14 ms), followed by RF (2.20 ms) and CNN1D (106.23 ms), with 100% stability over a 5-minute real-time test of 100 inferences. Gesture predictions were transmitted to the smart board application via Bluetooth Low Energy (BLE) Nordic UART Service. These results indicate the preliminary feasibility of wearable IMU-based TinyML for hands-free presentation control in smart classroom environments. ]]>
Copyrights © 2026
