Risk-guided depth reconstruction for single-photon LiDAR in mixed-pixel high-risk regions

Single-photon LiDAR is important for long-range weak-signal imaging. However, beam divergence and the effective field of view (FoV) can induce high-risk regions with neighborhood return mixing. In high-risk regions, especially near depth boundaries, this mixing causes multiple-return interference, peak broadening, and main-depth readout distortion. These effects weaken effective main-peak responses and introduce main-depth readout bias. To address this problem, this work proposes a robust single-photon LiDAR depth reconstruction method for high-risk regions. The method introduces a mixed-pixel risk guide to estimate spatially varying risk cues associated with multi-peak and broadened responses, and uses risk-guided residual shrinkage to reduce over-smoothing while preserving weak peaks and structural details. At the output stage, histogram distribution and peak evidence are jointly predicted to correct the main-depth readout and reduce the bias of soft-argmax-style weighted-average readout. A staged joint optimization objective further stabilizes temporal-distribution recovery and strengthens high-risk region supervision. Experiments on synthetic data show that the proposed method achieves robust performance across varying photon conditions and effective FoV settings, with clear advantages under strong mixed-pixel degradation and competitive performance under the extreme 2:500 condition. Ablation results verify the complementarity between high-risk response recovery and main-peak-discriminative readout. Results on real measurements further demonstrate lower reconstruction errors and better preservation of multi-depth boundaries, fine structures, and local depth continuity.