Abstract:
Current photon time-of-flight-based non-line-of-sight (NLOS) imaging systems face inherent resolution limitations imposed by hardware temporal jitter, severely degrading performance in closely spaced multi-surface scenarios. Due to this, sub-resolution imaging, which resolves target features below system-constrained resolution thresholds, has remained unexplored in NLOS research. To transcend this limitation, we propose a transient waveform decomposition framework to achieve sub-resolution NLOS imaging. Building on the confocal NLOS architecture and the returned transient signals' statistical characteristics, the reflections of multiple hidden surfaces are modeled as Gaussian mixture distributions. An L1-L2 hybrid-regularized deconvolution algorithm is then developed to suppress inaccuracy in pulse initialization caused by time jitter and Poisson noise. A modified Levenberg-Marquardt algorithm with non-negative and anti-merge constraints subsequently decomposes surface-specific photon histograms, enabling individual surface reconstructions via light-cone transform. Experimental results demonstrate incredible sub-resolution discrimination, reducing axial/lateral discrimination limits from 5.25 cm/9.74 cm to 1 cm for the constructed resolution bar at 3 m invisible distance. Moreover, this method also maintains adaptability under mutual occlusion and scattering occlusion. The proposed method establishes a new avenue for high-resolution NLOS imaging, providing significant insight into the practical applications of the NLOS technique.
