IDSOR: Intensity- and Distance-Aware Statistical Outlier Removal for Weather-Robust LiDAR Point Clouds

TL;DR

IDSOR mitigates weather-induced LiDAR outliers by modeling a range-dependent distribution of weather returns with a Gamma PDF $f_r(r;k,\theta)$ and building an intensity- and distance-aware threshold $T_{\mathrm{IDSOR}} = s\,T_g \bigl(1-\alpha_i h_i\bigr)$, where $\alpha_i = \frac{\rho\, f_r(r_i)}{\rho\, f_r(r_i) + 1}$ and $h_i = 1 - i_{\mathrm{norm},i}$. A DROR-prior variant uses a data-driven estimate $\hat{f}_r(\cdot)$ from a weather-dominated subset to refine the threshold. Evaluation against DSOR, DROR, and DDIOR on simulation-augmented railway data and the Winter Adverse Driving Dataset shows IDSOR achieves a favorable precision–recall balance, with the DROR-prior variant offering a small precision gain and robust performance even under challenging weather. These results indicate IDSOR provides practical, robust LiDAR denoising for safety-critical applications without extensive manual tuning, achieving high recall while maintaining high precision on real and simulated adverse-weather data.

Abstract

LiDAR point clouds captured in rain or snow are often corrupted by weather-induced returns, which can degrade perception and safety-critical scene understanding. This paper proposes Intensity- and Distance-Aware Statistical Outlier Removal (IDSOR), a range-adaptive filtering method that jointly exploits intensity cues and neighborhood sparsity. By incorporating an empirical, range-dependent distribution of weather returns into the threshold design, IDSOR suppresses weather-induced points while preserving fine structural details without cumbersome manual parameter tuning. We also propose a variant that uses a previously proposed method to estimate the weather return distribution from data, and integrates it into IDSOR. Experiments on simulation-augmented level-crossing measurements and on the Winter Adverse Driving dataset (WADS) demonstrate that IDSOR achieves a favorable precision-recall trade-off, maintaining both precision and recall above 90% on WADS.

Figures (5)

• Figure 1: LiDAR point clouds at an experimental railway level crossing in Sweden. (a) Point cloud augmented with simulated rain at $50~\mathrm{mm/h}$ using teufel2022simulating. (b) DSOR result. (c) DROR result. Both methods suppress rain-induced reflections and preserve salient objects (vehicle and pedestrians), but degrade fine track details (red boxes; enlarged views in the upper-right corners), particularly DSOR; signal towers (yellow boxes) are removed by DSOR and partially degraded by DROR.
• Figure 2: Averaged histograms of weather-particle-induced LiDAR returns at different ranges under simulated rain and snow conditions, and the corresponding Gamma fit.
• Figure 3: Level-crossing dataset under simulated rain (50 mm/h): (a) DDIOR, (b) IDSOR, and (c) DROR-prior IDSOR. Red bounding boxes (with enlarged views in the upper-right corners) indicate the railway tracks, and yellow bounding boxes indicate the signal towers.
• Figure 4: Precision and recall comparison of DROR, DSOR, DDIOR, IDSOR, and DROR-prior IDSOR for falling-snow removal on the WADS dataset.
• Figure 5: Qualitative comparison on WADS kurup2021wads. The yellow points highlight the falling-snow class in the WADS annotations. From left to right and top to bottom: (a) original, (b) DSOR, (c) DROR, (d) DDIOR, (e) IDSOR, and (f) DROR-prior IDSOR.