Integrated design-sense-plan architecture for autonomous geometric-semantic mapping with UAVs.

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Autonomous Systems

Background:

• Autonomous mapping and inspection require efficient navigation and environment representation.
• Current methods often lack semantic understanding, limiting targeted exploration.
• Optimizing drone design for performance and efficiency is crucial.

Purpose of the Study:

• To present a complete solution for autonomous mapping and inspection using a lightweight multi-camera drone.
• To enhance autonomous navigation through efficient planning algorithms and improved environment representations.
• To integrate semantic information for biased exploration and optimize drone design.

Main Methods:

• Developed a novel sensor observation model and utility function for information gain.
• Proposed a reward function incorporating geometric and semantic information from Deep Convolutional Neural Networks (DCNNs).
• Implemented Next-Best-View (NBV) planning with DCNN-based semantic segmentation for real-time processing.
• Introduced a unified approach for UAV camera number selection optimizing performance and resource trade-offs.

Main Results:

• Experiments demonstrated improved reconstruction accuracy by biasing exploration towards task-relevant objects using semantic information.
• The system showed benefits over purely geometric methods in both virtual and real-world scenarios.
• The proposed drone design optimization approach is coupled with sense and plan algorithms.

Conclusions:

• The integrated system enhances autonomous mapping and inspection by leveraging semantic understanding.
• The approach offers a flexible solution applicable to diverse environments with available semantic data.
• Optimized drone design and planning algorithms contribute to improved exploration and mapping performance.