Real-Time Hand Gesture Classification Using Infrared Sensor Arrays-Based Wearable Bracelet and Efficient 1-D Convolutional Neural Network

Hand gesture recognition is pivotal for intuitive human–machine interaction, particularly in healthcare and assistive technologies, where traditional interfaces (e.g., keyboards) are impractical. Existing modalities, such as electromyography and inertial sensors (inertial measurement units), struggle with noise sensitivity, motion dependence, or limited resolution for fine gestures. This work proposes infrared sensing as a robust alternative, leveraging reflected light patterns to capture both macrogestures and microgestures without relying on muscle activity or pronounced arm movements. We conducted experiments and compared different architectures to ensure the correct classification of hand gestures with the lowest possible latency, when targeting real-time processing. Experimental results demonstrate that shallow two-layer 1-D convolutional neural networks (CNNs) achieve rapid inference (3 ms) and minimal memory (56 kB) but suffer from low accuracy (81.45%), while deeper 12-layer CNNs attain 98.29% accuracy at prohibitive cost (176 ms latency, 17.4 MB memory). A six-layer 1-D CNN strikes an optimal balance, delivering 95.97% accuracy with moderate resources (56 ms latency, 640 kB memory), outperforming similarly accurate long short-term memory (94.88%, 136 ms), and recurrent neural network (96.59%, 102 ms) models. Confusion matrix analysis confirms consistent performance across seven gestures, including nuanced distinctions, such as thumb–index versus thumb–pinky pinches. By optimizing architectural depth and sensor integration, this work enables real-time operation on microcontrollers, such as the STM32F7, advancing applications in touchless medical interfaces and assistive devices for users with motor impairments.