mmSight: Towards Robust Millimeter-Wave Imaging on Handheld Devices

Abstract

We propose mmSight, a system that enables Synthetic Aperture Radar (SAR) imaging on handheld millimeter-wave (mmWave) devices. SAR imaging requires precise device self-localization or bulky motion controllers to reconstruct an image, but standard handheld devices suffer from various pose errors that cannot be addressed with traditional motion compensation methods. mmSight uses the time delay of the reflected mmWave signals across separated antennas to limit the pose error and perform improved mobile mmWave imaging. Since the mmWave signals are fundamentally limited by specularity and weak reflectivity, even a perfect pose correction may not yield a perceptible image. To this end, mmSight employs a generative learning model to learn the relationship between the imperceptible 3D image and a discernable 2D image and automatically classifies objects into several categories. We show that mmSight improves the structural quality of the mmWave images from 0.01 to 0.92, and it can be leveraged to identify several common hidden objects.