Accurate depth perception is essential for understanding crowd dynamics in evacuation and disaster-response scenarios. However, non-repetitive scanning LiDAR, despite its affordability and wide field of view, produces extremely sparse and irregular depth measurements that are difficult to interpret and limit its applicability for real-time monitoring. This study addresses this challenge by proposing an RGB-guided sparse depth completion method tailored to the sampling characteristics of non-repetitive LiDAR. The proposed framework integrates color-guided bilateral filtering with confidence-aware weighting and a masked formulation that explicitly handles invalid LiDAR returns. In addition, a scene-wise parameter optimization strategy based on leave-one-out consistency is introduced, enabling self-supervised tuning of the spatial and color scales without requiring dense ground-truth depth. This combination allows the method to reconstruct continuous depth structures while suppressing over-smoothing and overfilling, even under severe sparsity. Experiments on a simulated dataset, generated from dense human depth maps with a non-repetitive LiDAR sampling model, show that the proposed method clearly outperforms conventional spatial filtering baselines in depth reconstruction accuracy and structural preservation, while reducing spurious filling in empty regions. The proposed method provides a lightweight and interpretable depth-completion module suitable for edge deployment. It enhances the reliability of depth information for crowd monitoring and contributes practical value for situational awareness and evacuation analysis in disaster-response environments. • RGB-guided sparse depth completion for irregular non-repetitive LiDAR sampling. • Confidence-aware bilateral filtering with self-consistent parameter optimization. • Simulation pipeline reproducing non-repetitive LiDAR sampling for scalable testing. • Improved depth accuracy and structural consistency on 30,000 simulated samples. • Lightweight depth reconstruction supporting real-time evacuation guidance.
Zhang et al. (Thu,) studied this question.