← Back ICRA 2026

Real-Time Glass Detection and Reprojection Using Sensor Fusion Onboard Aerial Robots

Malakhi Hopkins, Varun Murali, Vijay Kumar, Camillo Jose Taylor

PDF

AI summary

Key figure (auto-extracted from paper)

A lightweight, CPU-only sensor fusion pipeline enables real-time detection and depth reprojection of transparent glass obstacles on a sub-300g quadrotor.

Glass detection ToF sensor sonar fusion aerial robotics transparent obstacle mapping SWaP constraints

Problem

Transparent obstacles like glass lack discernible features and cause conventional depth sensors to fail, leading to inaccurate maps and collision risks for autonomous aerial robots, particularly on low Size, Weight, and Power (SWaP) platforms.

Approach

The system fuses Time-of-Flight camera data with ultrasonic sonar measurements to detect specular reflection 'speckles' on glass, then uses custom convolution kernels and depth segmentation to reproject the glass plane's depth into empty space regions in real-time.

Key results

Real-time CPU-only glass detection and reprojection on a sub-300g quadrotor
Superior precision and mIoU compared to state-of-the-art RGB segmentation methods
Ultrasonic gating boosts segmentation mIoU from ~40% to over 74% in cluttered environments
Analytical validation of safety margins and successful mapping of glass planes at 0–15 degree tilts

Why it matters

Enables safe, real-time autonomous navigation for lightweight aerial robots in indoor environments where traditional perception systems fail on transparent surfaces.

Abstract

No abstract on file.

Index terms

Sensor Fusion Mapping Aerial Systems: Perception and Autonomy