Non-line-of-sight imaging and location determination using deep learning

Abstract

Non-line-of-sight (NLOS) imaging methods have been proposed and rapidly developed to address the challenge of detecting objects hidden from the direct line of sight. One critical issue in NLOS imaging is acquiring the spatial position of hidden objects. Active NLOS imaging methods solve this problem using complex designs and sophisticated equipment while the conventional passive NLOS imaging scheme, which does not have a controllable light source and time-gating detector, rarely captures the spatial information of objects. Using speckle patterns and the assistance of a deep learning method, we experimentally demonstrated that the passive NLOS imaging system could simultaneously visualize and locate objects hidden in corners like an active NLOS imaging system could. High-fidelity reconstructed images and high-accuracy location recognition were realized using the proposed network. The network could provide 99% recognition accuracy for the object's position with an axis spacing of 36 μm. These results were significant for NLOS imaging with high-accuracy positioning requirements.