Hidden source localization via learning-based reconstruction of thermal light-fields

We develop a framework for thermal passive non-line-of-sight (NLOS) imaging based on the reconstruction of noisy scattered light fields using neural networks. Thermal NLOS imaging is undermined by the extremely low reflectivity and relatively diffuse nature of many surfaces in the long-wave-infrared (LWIR) domain. Previous approaches in thermal NLOS imaging have relied on linear methods to denoise and deblur measurements. We show that a simple convolutional-neural-network trained using synthetic data is capable of recovering extremely weak signals from real-life noisy thermal light fields. Our proposed framework does not require knowledge of the scattering properties of the surface or expected signal-to-noise ratio of measurements. Experimental data was captured in a miniaturized thermal NLOS imaging studio. When localizing multiple human-temperature objects using a diffuse drywall scatterer, our method consistently improves the standard deviation in source depth estimate by more than 50% from comparable methods. It also outperforms methods that use additional prior knowledge of the scattering surface, improving the depth standard deviation by around 25%.