Lightweight Online Semisupervised Learning for Ultrasonic Radar-Based Dynamic Hand Gesture Recognition

Ultrasonic dynamic hand gesture recognition (U-DHGR) is a promising approach of human–computer interaction (HCI) for a broad range of emerging applications. Cross-user robustness is a key challenge for its classification algorithms. Due to the highly personalized gesture characteristics and the limited number of training samples, traditional offline supervised learning algorithms usually fail to achieve as high accuracy on new users as on the training participants. In this article, we propose a lightweight online semisupervised learning algorithm for U-DHGR that aims to manipulate the incrementally collected unlabeled user samples to further improve the personalized recognition performance and reduce the data preparation cost. For each of the defined gesture classes, a tree-ensemble-based binary classifier is deployed, whose partial structures are offline trained to predict pseudo-labels for the incremental training of the remaining structures. Experiments show that by combining the predictions of both the offline and incrementally trained parts of the classifier, improvements ranging from 3.8% to 7.9% are achieved in the recognition accuracies for different users. The average accuracy of the eight defined gestures reaches 96.7%, which is 3.1% higher than its nearest offline supervised competitor. The model size of the classifier is only 0.5 MiB after finishing the incremental training. The average inference time and incremental sample process time on a laptop are 11.7 and 37.9 ms, respectively, which shows the potential to be integrated into resource-constrained platforms.