The authors present a system called Dynamic Open Vocabulary Enhanced Safe-landing with Intelligence (DOVESEI) that aims to enable safe landing of autonomous UAVs in urban environments. The key components of the system are:
Landing Heatmap Generation Service: This module uses an open vocabulary semantic segmentation model (CLIPSeg) to generate a heatmap of optimal landing locations in the current camera frame.
Main Processing Node: This module processes the raw segmentation heatmap and applies a "dynamic focus" masking mechanism to guide the UAV towards the best landing spot. The dynamic focus adjusts based on the current state of the system (searching, aiming, landing, etc.).
The authors conducted experiments using high-resolution satellite images of Paris, France, and found that the inclusion of the dynamic focus mechanism significantly improved the success rate of safe landings compared to using the raw segmentation heatmap alone. The system was able to successfully land the UAV at altitudes as low as 20 meters, enabling the use of lightweight stereo cameras and conventional 3D path planning for the final descent.
The key advantages of the proposed approach are its adaptability to different environments, the use of only a monocular camera and onboard computational resources, and the ability to bypass the need for extensive data collection or recalibration. The authors envision this system as a compact, lightweight, and onboard external controller that can be integrated with commercial UAVs to enable safe landings even in scenarios with internal navigational or sensory system issues.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Haechan Mark... at arxiv.org 05-07-2024
https://arxiv.org/pdf/2308.11471.pdfDeeper Inquiries