GAN Based Noise Generation to Aid Activity Recognition when Augmenting Measured WiFi Radar Data with Simulations
Shelly Vishwakarma, Chong Tang, Wenda Li, K. Woodbridge, Raviraj Adve, Kevin Chetty
Abstract
This work considers the use of a passive WiFi radar (PWR) to monitor human activities. Real-time uncooperative monitoring of people has numerous applications ranging from smart cities and transport to IoT and security. In e-healthcare, PWR technology could be used for ambient assisted living and early detection of chronic health conditions. Large training datasets could drive forward machine-learning-focused research in the above applications. However, generating and labeling large volumes of high-quality, diverse radar datasets is an onerous task. Therefore, we present an open-source motion capture data-driven simulation tool, SimHumalator, that can generate large volumes of human micro-Doppler radar data at multiple IEEE WiFi standards(IEEE 802.11g, n, and ad). We qualitatively compare the micro-Doppler signatures generated through SimHumalator with the measured signatures. To create a more realistic training dataset, we artificially add noise to our clean simulated spectrograms. A noise distribution is directly learned from real radar measurements using a Generative Adversarial Network (GAN). We observe improvements in the classification performances between 3 to 8%. Our results suggest that simulation data can be used to make adequate training data when the available measurement training support is low.