Hierarchical spatio-temporal denoising for photon-efficient 3D LiDAR imaging under low SBR

Single-photon light detection and ranging (LiDAR) enables long-range, high-precision 3D imaging with single-photon sensitivity and picosecond timing resolution, yet imaging under extremely low signal-to-background ratio (SBR) conditions remains challenging due to severe noise contamination and signal loss. In this work, a coaxial 1550-nm single-photon LiDAR system is constructed, and a hierarchical spatio-temporal 3D collaborative denoising algorithm is developed to address this issue. The method exploits voxel-level spatial–temporal correlations through a three-stage filtering strategy: extreme-value rejection for outlier suppression, anisotropic Gaussian-kernel neighborhood reinforcement for edge-preserving enhancement, and sliding-window adaptive integration for residual noise smoothing. Simulations using a custom IOE-letter scene at SBR = 0.01 demonstrate that the proposed method achieves an RMSE of 0.140 and SSIM of 0.959, outperforming 3D deconvolution (RMSE = 0.311, SSIM = 0.789) and maximum likelihood estimation (RMSE = 1.833, SSIM = 0.004). In 4.0-km building experiments, with the signal attenuated to 10% and corrupted by Poisson noise, the method maintains high structural fidelity (SSIM = 0.492) compared to MLE (SSIM = 0.058). These results confirm that integrating spatial–temporal neighborhood information significantly enhances robustness and depth recovery accuracy for photon-efficient, long-range 3D imaging in photon-starved conditions.