Lightweight Online Semi-Supervised Learning Algorithm for Ultrasonic Gesture Recognition

This paper presents a lightweight online semi-supervised learning algorithm that utilizes the sequentially arrived unlabeled user gesture samples to improve recognition performance for ultrasonic gesture recognition systems. For each class of gesture, a binary extreme random forest is pre-trained in the offline manner, by which the newly arrived unlabeled samples are given pseudo-labels before they are filtered according to probability and then fed to an initially unexpanded forest to grow its trees incrementally. To limit complexity, feature cache and index pool are introduced. Experiments show that the incrementally grown forests identify 8 gestures and 4 micro gestures with average accuracies of 95.8% and 93.6%, respectively exceeding its nearest offline supervised competitor by 2.1% and 4.1%.