An Efficient Way for Active None-Line-of-Sight: End-to-End Learned Compressed NLOS Imaging

Non-line-of-sight imaging (NLOS) is an emerging detection technique that uses multiple reflections of a transmitted beam, capturing scenes beyond the user's field of view. Due to its high reconstruction quality, active transient NLOS imaging has been widely investigated. However, much of the existing work has focused on optimizing reconstruction algorithms but neglected the time cost during data acquisition. Conventional imaging systems use mechanical point-by-point scanning, which requires high time cost and not utilizing the sparsity of the NLOS objects. In this paper, we propose to realize NLOS in an efficient way, based upon the theory of compressive sensing (CS). To reduce data volume and acquisition time, we introduce the end-to-end CS imaging to learn an optimal CS measurement matrix for efficient NLOS imaging. Through quantitative and qualitative experimental comparison with SOTA methods, we demonstrate an improvement of at least 1.4 dB higher PSNR for the reconstructed depth map compared to using partial Hadamard sensing matrices. This work will effectively advance the real-time and practicality of NLOS.