Abstract
In this paper, we present an approach for acquiring textured 3D models of room-sized indoor spaces using a quadrocopter. Such room models are for example useful for architects and interior designers as well as for factory planners and construction managers. The model is internally represented by a signed distance function (SDF) and the SDF is used to directly track the camera with respect to the model. Our solution enables accurate position control of the quadrocopter, so that it can automatically follow a pre-defined flight pattern. Our system provides live feedback of the acquired 3D model to the user. The final model consisting of a textured 3D triangle mesh can be saved in several standard CAD file formats.
1. Introduction
Creating accurate 3D models of indoor environments—such as factories, offices, or disaster zones—has traditionally been a manual and time-consuming task. While the advent of handheld RGB-D sensors like the Microsoft Kinect has simplified data acquisition, human operators are limited by physical reach and cannot safely access hazardous or hard-to-reach areas. Autonomous micro aerial vehicles (MAVs), specifically quadrocopters, offer a compelling solution for automated environmental scanning.
However, many existing drone-based mapping systems rely on sparse feature-based SLAM. While effective for navigation, these methods produce point clouds that lack the dense surface information required for high-fidelity architectural modeling or interior design. Conversely, dense reconstruction methods like KinectFusion provide high-quality models but are often too computationally intensive for the real-time, high-precision pose estimation required to control a flying robot. This paper addresses these challenges by proposing a system that utilizes a dense Signed Distance Function (SDF) for both robust camera tracking and high-quality textured mapping.
2. Methodology
The system architecture centers on an AscTec Pelican quadrocopter equipped with an Asus Xtion Pro Live RGB-D sensor. The core innovation lies in the dual utilization of the Signed Distance Function (SDF) for localization and environmental representation.
2.1 Direct SDF Tracking and Mapping
Unlike standard dense mapping approaches that use the Iterative Closest Point (ICP) algorithm to align synthetic point clouds, this system employs a direct tracking method. It minimizes the value of the SDF evaluated at the back-projected points of the depth image, which improves tracking accuracy and robustness by avoiding intermediate point cloud generation.
The environment is represented within a 3D voxel grid, where each voxel stores its color and its distance to the nearest surface, known as the Truncated Signed Distance Function (TSDF). To provide real-time feedback to the operator, the system executes a Marching Cubes algorithm every two seconds to generate a textured triangle mesh from the voxel data.
2.2 Truncation and Control
A critical parameter in this system is the truncation distance ($\delta$), which defines the region around a surface where distance values are maintained. This parameter balances the trade-off between tracking stability and geometric detail.
For flight control, visual pose estimates are transmitted to an external ground station equipped with an Nvidia GTX 560 GPU. The processed data is beamed back to the drone at 30 fps and fused with onboard inertial sensors via an Extended Kalman Filter (EKF). This allows for precise position control using an LQR controller, enabling the drone to follow pre-defined flight patterns automatically.
3. Experimental Evaluation
The authors validated the system's performance using the TUM RGB-D Benchmark and through real-world autonomous flights in three distinct indoor environments: a laboratory kitchen, a complex office space with multiple desks, and a "couch corner" for testing assisted flight modes.
3.1 Tracking and Flight Precision
Benchmark results demonstrated that the proposed direct SDF tracking significantly outperforms the Point Cloud Library (PCL) implementation of KinFu in terms of robustness, successfully tracking trajectories where KinFu failed. In real-world autonomous hovering tests, the system maintained the quadrocopter's position with an average standard deviation of just 2.1 cm.
3.2 Real-Time Performance
The entire tracking and mapping pipeline operates at approximately 27ms per frame. This efficiency allows for a full 30fps update rate, which is essential for the stability of the quadrocopter during autonomous flight. The resulting models are high-quality textured meshes compatible with standard CAD formats.
4. Discussion and Key Insights
The research highlights the SDF as a powerful mathematical tool for localization. By treating the map as a continuous function, the system can utilize standard Gauss-Newton optimization to achieve high-precision pose estimation. The authors found that a truncation distance of $\delta = 0.3m$ provided the optimal balance; smaller values led to tracking failures during sudden movements, while larger values obscured fine geometric details in the final model.
5. Conclusion and Future Work
This paper presents a robust approach for autonomous dense 3D mapping using a quadrocopter. By leveraging the SDF for both tracking and mapping, the system achieves the precision necessary for stable flight and high-quality output. Current limitations include the reliance on a ground-based GPU and the restriction to indoor environments due to infrared interference from sunlight. Future work will explore onboard CPU-based implementations, stereo-vision for outdoor use, and the integration of the generated 3D maps for active collision avoidance and autonomous exploration.
Summary updated