RF-Based Human Activity Recognition Using Signal Adapted Convolutional Neural Network

Human Activity Recognition (HAR) plays a critical role in wide range of real-world applications, and it is traditionally achieved via wearable sensing. Recently, to avoid the burden discomfort caused by devices, device-free approaches exploiting RF signals arise as promising alternative for HAR. Most latest require training large deep neural network model either time or frequency domain, entailing extensive storage contain intensive computations infer activities. Consequently, even with some major advances on HAR, current are still far from practical scenarios where computation resources possessed by, example, edge limited. Therefore, we introduce HAR-SAnet which novel RF-based HAR framework. It adopts an original signal adapted convolutional architecture: instead feeding handcraft features into classifier, fuses them adaptively both domains design end-to-end model. We apply point-wise grouped convolution depth-wise separable convolutions confine scale speed up inference execution time. The experiment results show that recognition accuracy outperforms state-of-the-art algorithms systems.