SignID: Acoustic-based Identification with Single Sign Gesture

Hui Xu,Tianran Wang,Xiaoyu Wang,Yubo Yan,Panlong Yang
DOI: https://doi.org/10.1109/bigcom53800.2021.00009
2021-01-01
Abstract:With the proliferation of Smart Home and Artificial Intelligence of Things (AIoT), humanized while effective HumanComputer Interface (HCI) has been the central focus of research to bridge user’s daily life and ever-growing hardware capability. In this context, ultrasound-based acoustic sensing is being widely adopted due to its nature of contact-less and platform-agnostic. In order to fully exploit the potential of ultrasonic bandwidth and kill two birds with one stone, we propose SignID, a multi-task learning boosted system capable of jointly recognizing user gestures and identification. The system only needs a pair of speaker and microphone, which is available on various smart devices, to emit sound wave and receive the Doppler-shifted echo. To handle ambient noise and lower the computation overheads, we adopt a bandpass filter to remove the frequency band out of interest, then apply undersampling to compress the measurement for the subsequent FFT processing. After segmentation and spectrogram extraction, a multi-task deep neural network based on cross-stitch units is responsible for jointly learning representation for both identification and gesture recognition. The proposed system is evaluated based on our implementation in a real-life scenario, and the overall accuracy is over 95%. We also test the impact of different environmental factors, including user distance, ambient noise, body gesture, and training data volume. The results prove our model’s robustness to various disturbances.
What problem does this paper attempt to address?