TL;DR
This paper presents a low-cost microdrone system for indoor mapping, particularly for emergency response in hard-to-reach areas. It utilizes a novel graph-based SLAM approach with 1D laser scanners to create detailed 3D maps, achieving comparable accuracy to larger, more expensive systems.
What Problem Does This Paper Address?
Mapping indoor environments, especially in disaster scenarios, is crucial for effective emergency response, facility management, and navigation. Traditional methods like terrestrial laser scanners (TLS) can be slow and costly, and existing building floor plans are often outdated or lack sufficient detail. While larger drones with 3D LiDARs can map interiors, their size and cost can be prohibitive for navigating confined or damaged spaces. This paper addresses the need for a compact, affordable, and capable indoor mapping solution that can operate in challenging environments.
Methods
The core of the system is a microdrone equipped with six 1D time-of-flight (TOF) laser rangefinders and an optical flow sensor. The mapping and localization are handled by a graph-based SLAM (Simultaneous Localization and Mapping) algorithm. This SLAM approach focuses on extracting planar features (walls, floors, ceilings) from the laser data. Key aspects of their method include:
- Feature-based SLAM: Instead of relying solely on point clouds, the system identifies and uses planar segments as primary features for mapping.
- Graph Optimization: The SLAM problem is formulated as a graph optimization problem, allowing for the refinement of the entire trajectory and map simultaneously.
- Loop Closure Detection and Correction: The system is designed to recognize previously visited areas and correct accumulated drift over time, which is crucial for maintaining map accuracy. Their loop closure specifically uses plane-to-plane matching.
- Spline-based Trajectory Representation: The drone's pose is modeled using B-splines, providing a continuous and more accurate representation of its trajectory compared to discrete poses.
- Sensor Fusion: Data from the laser rangefinders and optical flow sensor are fused to estimate the drone's pose and build the map.
Datasets
The authors evaluated their system on several indoor datasets collected at the University of Twente (ITC1, ITC2, ITC3) and from other institutions (Apartment dataset, DLR1, DLR2). These datasets represent various indoor environments, including corridors, irregularly shaped rooms, and connected spaces. The performance was compared against ground truth data from a RIEGL VZ-400 TLS, and point clouds from a macrodrone (MAX) and a vision-aided navigation helmet (IPS).
Key Insights and Ideas
- Compactness is Key for Inaccessibility: The microdrone's small size allows it to access spaces that larger drones or ground-based systems cannot, making it ideal for emergency response in damaged or confined buildings.
- 1D LiDARs Can Be Sufficient for Planar Environments: Contrary to the trend towards complex 3D LiDARs, this work demonstrates that a carefully designed system with multiple 1D scanners can effectively map indoor environments dominated by planar structures.
- Planar Features Offer Robustness: Using planar segments as the primary mapping features, rather than raw point clouds, simplifies data association and loop closure, especially in environments with repetitive structures.
- Graph SLAM with Loop Closure is Crucial: The ability to detect and correct loop closures significantly reduces accumulated drift, leading to more accurate and consistent maps, as evidenced by the comparison with and without SLAM.
Key Results
- The microdrone system successfully mapped all tested indoor environments.
- The accumulated drift without loop closure was on average 1.1% (0.35 m) over a 31-m trajectory.
- The generated point clouds showed comparable performance to a multilayer LiDAR-based macrodrone (MAX).
- Approximately 85% of cloud-to-cloud distances between the microdrone's map and the MAX drone's map were less than 10 cm.
- Distance differences compared to floor plans were generally less than 15 cm.
- The system demonstrated the ability to capture ceiling and floor details independently of flight height.
Limitations and Future Work
- Drift in Complex Environments: The paper notes that the drone's drift was significant in some scenarios, particularly when dealing with highly reflective surfaces (like kitchen floors) or when the optical flow sensor struggled with visual features. This led to a need for overly relaxed data association thresholds.
- Limited Range and Field of View: The 1D scanners have a limited range (up to 4m), which can affect coverage and the ability to capture details in larger spaces.
- Hardware Dependency: While the SLAM algorithm is hardware-independent, the planar segment extraction is tailored to the 1D scanner configuration.
- Future Work: The authors suggest integrating complementary sensors (e.g., thermal cameras for victim detection), adding propeller guards for safety, and incorporating additional navigational information (like a small IMU) to improve robustness. They also plan to run the SLAM online for more efficient drift correction and explore new multiranger decks with wider fields of view.
Summary updated