Real-Time Recognition of Boxing Head Gestures with IMU-Earables: Machine Learning and Dynamic Time Warping

The rising prominence of earables, wearables meant to be worn around the ear, represents opportunities for novel applications. Previous research showcases the potential of earables in the context of sports; however, a gap is present for boxing, more specifically in recognition of defensive manoeuvres, even outside the realm of earable development. Thus, this paper explores the capability of real-time, IMU-based boxing head gesture recognition using the open-source OpenEarable framework through classical machine learning and dynamic time-warping approaches. A dataset was collected consisting of approximately 460 samples of left/right slips, left/right rolls, and pullbacks, by a hobby-level boxer. The results revealed that utilizing dynamic time warping in combination with templates based on barycenter averaging achieves effective results in gesture recognition. During the testing phase, the implemented algorithm achieved a high accuracy score of 99% on the collected dataset. This performance was further validated in a deployed real-world scenario, where the algorithm maintained an overall accuracy of 96% across 50 repetitions per gesture. Additionally, the system demonstrated robustness against variations in gesture execution speed and intensity.