Reliability-Guided Depth Fusion for Glare-Resilient Navigation Costmaps

Abstract: Specular glare on reflective floors and glass surfaces frequently corrupts RGB-D depth measurements, producing holes and spikes that accumulate as persistent phantom obstacles in occupancy-grid costmaps. This paper proposes a glare-resilient costmap construction method based on explicit depth-reliability modeling. A lightweight Depth Reliability Map (DRM) estimator predicts per-pixel measurement trustworthiness under specular interference, and a Reliability-Guided Fusion (RGF) mechanism uses this signal to modulate occupancy updates before corrupted measurements are accumulated into the map. Experiments on a real mobile robotic platform equipped with an Intel RealSense D435 and a Jetson Orin Nano show that the proposed method substantially reduces false obstacle insertion and improves free-space preservation under real reflective-floor and glass-surface conditions, while introducing only modest computational overhead. These results indicate that treating glare as a measurement-reliability problem provides a practical and lightweight solution for improving costmap correctness and navigation robustness in safety-critical indoor environments.

• The paper proposes a lightweight Depth Reliability Map (DRM) and Reliability-Guided Fusion (RGF) approach to mitigate glare-related sensor errors.
• Empirical results demonstrate a nearly 90% reduction in false obstacle rates and substantial improvements in navigation success under severe glare.
• The method integrates with existing costmap planners with minimal computational overhead, supporting real-time operation on embedded robotic platforms.

Reliability-Guided Depth Fusion for Glare-Resilient Navigation Costmaps: Technical Analysis

Introduction

The paper "Reliability-Guided Depth Fusion for Glare-Resilient Navigation Costmaps" (2604.12753) addresses the persistent challenge of specular glare corrupting RGB-D depth measurements in mobile robot navigation. Active-stereo depth sensors, such as the Intel RealSense D435, suffer significantly from specular and reflective indoor surfaces, leading to depth holes and spikes that, when accumulated in occupancy-grid costmaps, yield false obstacle insertions and loss of navigable free space. This work systematically reframes the problem as a measurement reliability issue and introduces a real-time, lightweight Depth Reliability Map (DRM) and Reliability-Guided Fusion (RGF) approach.

Problem Formulation and Prior Art

Active-stereo depth sensing in the presence of specular reflections produces structured measurement corruptions, notably persistent holes and spurious spikes, which standard occupancy-grid fusion techniques interpret as real obstacles. Prior work has explored three mitigation directions: (i) map-level structural inference [2], (ii) reflection-aware SLAM with outlier removal [3], and (iii) explicit depth-error modeling with probabilistic filtering [4]. However, these approaches insufficiently address the costmap accumulation mode, which is critical for robust downstream navigation. Naïve methods and heuristic outlier filters (spatial, temporal, or gating-based) fail to distinguish between genuinely unsafe observations and reflective artifacts, especially in safety-critical deployments.

Methodology

Depth Reliability Map (DRM) Network

The core contribution is a lightweight U-Net-inspired DRM estimator, comprising 61,936 parameters and requiring approximately 0.69 GFLOPs per inference at 320×240 resolution. Inputs consist of 5 channels (RGB, Depth, and an optional temporal depth difference), producing per-pixel reliability values $R_t(u) \in [0,1]$ upsampled for costmap fusion. Ground-truth reliability is constructed by comparing raw measurements with temporally filtered, artifact-suppressed reference depth, using both hard and soft reliability targets that reflect range-dependent uncertainty. Supervision employs an $L_1$ regression loss, ensuring robustness to outlier-prone reflective regions.

Reliability-Guided Fusion (RGF)

Occupancy integration leverages the per-pixel reliability estimates either as fusion weights or gating masks. For grid cell $g$ mapped from pixel $u$:

$$p_t(g) = \lambda p_{t-1}(g) + (1-\lambda) w_t(u) \cdot \text{obs}_t(g)$$

where $w_t(u) = R_t(u)$ and $\lambda$ is a temporal forgetting factor (set to 0.85). Low-reliability measurements contribute little to obstacle evidence, effectively suppressing phantom insertions induced by glare. An optional gating variant can further filter measurements below a reliability threshold.

Deployment and Evaluation Protocol

Experiments are performed on a real robotic platform (Intel RealSense D435 + Jetson Orin Nano), leveraging nvblox TSDF integration and Nav2 costmap planners under controlled glare scenarios across diverse reflective surfaces. Baseline methods include naïve fusion, validity/range-based gating, spatial median filtering, and temporal outlier rejection, all under identical costmap and mapping settings to isolate the impact of corruption mitigation.

Empirical Results

Sensor-Level Metrics

Under severe glare (L2), naïve TSDF fusion yields a hole rate of 45.2%, spike rate of 27.6%, and RMSE of 0.840 m. With DRM+RGF, these metrics improve dramatically to 5.4% hole rate, 3.1% spike rate, and 0.082 m RMSE—representing an order-of-magnitude reduction in gross sensor corruption.

Costmap Correctness and Navigation Outcomes

The practical impact is reflected in costmap-level and navigation metrics:

Glare Level	Method	FOR	FSR	Emerg. Stops/10m	Success Rate (%)
L2	Naïve Fusion	0.432	0.515	6.84	28.6
L2	DRM+RGF (proposed)	0.056	0.897	0.85	91.4

DRM+RGF reduces the False Obstacle Rate by nearly 90% under severe glare, elevates free-space recall by 74%, and elevates successful navigation trials from 28.6% to 91.4%. Emergency stops per unit traversal distance are similarly suppressed.

Computational Overhead

Despite performing learned reliability inference, DRM+RGF imposes only a modest overhead (6.5 ms/device frame), sustaining 60+ frames per second throughput and presenting a practical solution for real-time embeddable systems. This is a clear contrast to heavier depth-completion networks that are infeasible on edge platforms.

Comparative Performance

Heuristic filters (range gating, spatial/temporal methods) achieve relatively minor improvements under mild conditions but fail under L2 glare, where only the reliability-based paradigm preserves both safety (by preventing false obstacle insertions) and efficiency (by minimizing unnecessary detours and maintaining nearly optimal path ratios). The method's design does not require alterations to planners/controllers, facilitating drop-in deployment.

Implications and Outlook

From a systems perspective, the paper demonstrates the criticality of reliability modeling at the sensor fusion stage—rendering downstream mapping and planning components invariant to structured sensor failures endemic to reflective environments. Unlike dense reconstruction or visually plausible (but navigation-agnostic) depth completion, DRM+RGF specifically optimizes the tradeoff between free-space conservatism and false-positive obstacle suppression, which is paramount for safety-critical indoor autonomy.

The approach is particularly notable for its balance between robustness and computational frugality. This opens avenues for its application to broader sensor suites (stereo, LiDAR, mixed-modality systems) and dynamic deployment contexts, such as wet outdoor surfaces or cross-modality fusion in hybrid environments. Limitations remain in environments dominated by pervasive mirror reflections or consistent low reliability, where additional mirror segmentation or multi-view consistency may be needed.

Conclusion

The reliability-guided depth fusion paradigm articulated in this work offers a compelling, theoretically grounded, and practically viable solution to the persistent problem of glare-induced mapping corruption in mobile robotics operating in reflective environments. The explicit quantification and use of per-pixel depth reliability enable both substantial sensor-level and system-level robustness improvements, all within the real-time and computation constraints of embedded robotic platforms. Future work should explore multi-modal reliability integration, reflection segmentation, and enhanced temporal/multi-view coherence strategies to generalize this robustness to an even broader array of challenging sensing scenarios.