iMe: Incremental, Memory-Efficient Hand Gesture Recognition System With Wi-Fi

With the rapid development of Internet-of-Things (IoT) technology, gesture recognition based on wireless signals such as Wi-Fi has become an important method of human–computer interaction. However, existing gesture recognition methods rely on devices with rich storage resources, such as smartphones and laptops, to achieve high recognition accuracy. Due to cost reasons, most IoT devices are usually equipped with limited storage space, which makes it impossible to store sufficient gesture samples to train high-precision, multicategory gesture recognition models. In this article, we propose an incremental, memory-efficient gesture recognition system based on Wi-Fi, iMe, which only uses a small amount of data to achieve incremental recognition of multiple categories gestures and solves the challenge of implementing multicategory and high-precision gesture recognition systems on IoT devices. iMe uses a conditional variational autoencoder (CVAE) generative network with a dynamic updating structure to generate the old data required for gesture incremental recognition. An incremental learning framework based on knowledge distillation is adopted to address catastrophic forgetting during gesture incremental learning. The results demonstrate that iMe achieves optimal performance in both recognition accuracy and storage cost, outperforming state-of-the-art Fine-Tune, iCaRL, and EWC methods.