aiqbench analysis
This research introduces a per-pixel depth reliability mechanism that significantly reduces false obstacle detection in RGB-D sensors under severe glare conditions, directly enhancing the safety and robustness of autonomous navigation systems. By explicitly modeling sensor trustworthiness before occupancy integration, the method maintains high task success rates while minimizing phantom obstacles caused by reflective surfaces. The approach offers a critical advancement for deploying vision-based robots in complex indoor environments where traditional stereo confidence filtering often fails.
Reliability-Guided RGB-D Sensor Fusion for Glare-Resilient Navigation Costmaps
arXiv:2604.12753v2 Announce Type: replace Abstract: Specular glare on reflective floors, glass boundaries, and glossy indoor surfaces can corrupt active-stereo RGB-D measurements, producing holes and spikes that persist as phantom obstacles in navigation costmaps. This article presents a glare-resilient RGB-D sensor-fusion method based on explicit per-pixel depth reliability. A lightweight Depth Reliability Map network (DRM-Net) predicts sensor trustworthiness, and reliability-guided fusion (RGF) combines continuous weighting with a minimum gate before occupancy integration. Rejected measurements generate neither obstacle insertion nor free-space clearing; affected cells remain unknown or retain prior evidence. Training targets are built from a five-frame, pose-aligned multiview buffer using independent LiDAR/AMCL poses, occlusion-aware aggregation, and a calibrated range-dependent depth-uncertainty model. The evaluation includes tuned nvblox TSDF, Intel RealSense SDK postprocessing, high-threshold stereoconfidence filtering, TDCNet, and HDCNet baselines, together with statistical, safety, generalization, and embedded-runtime analyses. Under severe glare, Depth Reliability Map (DRM)+RGF achieves false obstacle rate (FOR) 0.056 +/- 0.012, free-space recall (FSR) 0.897 +/- 0.045, FNOR 0.018, 1.00 +/- 0.00 degraded-mode safety interventions per 10 m, and 91.4% task success while operating at 16.5 ms per frame. Across the retained reflective-scene trials (Baseline N = 98; DRM+RGF N = 105), collisions decrease from 14 to 1. These results support RGF as a favorable safety-utility tradeoff relative to aggressive filtering and dense completion for glare-affected indoor navigation.
Abstract:Specular glare on reflective floors, glass boundaries, and glossy indoor surfaces can corrupt active-stereo RGB-D measurements, producing holes and spikes that persist as phantom obstacles in navigation costmaps. This article presents a glare-resilient RGB-D sensor-fusion method based on explicit per-pixel depth reliability. A lightweight Depth Reliability Map network (DRM-Net) predicts sensor trustworthiness, and reliability-guided fusion (RGF) combines continuous weighting with a minimum gate before occupancy integration. Rejected measurements generate neither obstacle insertion nor free-space clearing; affected cells remain unknown or retain prior evidence. Training targets are built from a five-frame, pose-aligned multiview buffer using independent LiDAR/AMCL poses, occlusion-aware aggregation, and a calibrated range-dependent depth-uncertainty model. The evaluation includes tuned nvblox TSDF, Intel RealSense SDK postprocessing, high-threshold stereoconfidence filtering, TDCNet, and HDCNet baselines, together with statistical, safety, generalization, and embedded-runtime analyses. Under severe glare, Depth Reliability Map (DRM)+RGF achieves false obstacle rate (FOR) 0.056 +/- 0.012, free-space recall (FSR) 0.897 +/- 0.045, FNOR 0.018, 1.00 +/- 0.00 degraded-mode safety interventions per 10 m, and 91.4% task success while operating at 16.5 ms per frame. Across the retained reflective-scene trials (Baseline N = 98; DRM+RGF N = 105), collisions decrease from 14 to 1. These results support RGF as a favorable safety-utility tradeoff relative to aggressive filtering and dense completion for glare-affected indoor navigation.
Substantially revised version updated to match the final peer-reviewed article published in IEEE Sensors Journal. The experiments and analyses were substantially rerun and revised, and the author list has been updated to reflect contributions to the current version
Robotics (cs.RO)
arXiv:2604.12753 [cs.RO]
(or arXiv:2604.12753v2 [cs.RO] for this version)
https://doi.org/10.1109/JSEN.2026.3727559 DOI(s) linking to related resources
Full text under the publisher's open licence · 266 words
Related coverage
Reliability-Guided Depth Fusion for Glare-Resilient Navigation Costmaps
arXiv · Jun 3, 2026
GlassFormer: Learning Real-time Glass Segmentation using Radar-Depth Fusion
arXiv · Sep 30, 2026
Digital States 2026: AI Moves From Experiment to Infrastructure
govtech.com · Sep 24, 2026
Dirac Labs Raises $1.8M for Quantum Navigation Sensors
The Quantum Insider · Aug 20, 2026