An End to End Wearable Device and System for Indefinite, Continuous, Real Time Gesture Recognition of Directional and Shape-Based Arm Gestures

Human-Computer Interfaces designate how individuals interact with digital devices and systems. Often these interfaces heavily involve fine motor control via keyboards, controllers, and touch screens. These modalities often exclude users with disabilities, such as amputees or those affected by carpal tunnel syndrome. This paper proposes the creation of a wearable gesture recognition system that enables hands-free, device interaction through the detection of user movements and gestures. The system comprises of a wearable device with an onboard inertial measurement unit (IMU), a complementary filter for orientation estimation, and a gesture classification pipeline that recognizes a set of directional, rotational, and shape based gestures. We evaluate two models: an XGBoost classifier using IMU data preprocessed with a Mahony filter, and Wavelet Transform feature extraction, and a Convolutional Neural Network (CNN) directly using raw IMU data and Mahony-based orientation estimation. On the benchmark 6DMG dataset, our models achieve an accuracy of 99.55% and 99.18% respectively, while our custom-collected dataset yields 94.81% and 96.28%. In real-time, continuous gesture tracking, the system achieves an average recognition accuracy of 92.33% with an average sensor to prediction latency of 20 milliseconds. Our system is actualized through a modular API and graphical user interface, enabling real-time software interaction through movements and recognized gestures. Compared to existing methods, our approach achieves high accuracy, continuous, real-time performance with a low-latency, scalable integration with external software.