Abstract
The generation of 3D city models is a very active field of research. Modeling environments as point clouds may be fast, but has disadvantages. These are easily solvable by using volumetric representations, especially when considering selective data acquisition, change detection and fast changing environments. Therefore, this paper proposes a framework for the volumetric modeling and visualization of large scale urban environments. Beside an architecture and the right mix of algorithms for the task, two compression strategies for volumetric models as well as a data quality based approach for the import of range measurements are proposed. The capabilities of the framework are shown on a mobile laser scanning dataset of the Technical University of Munich. Furthermore the loss of the compression techniques is evaluated and their memory consumption is compared to that of raw point clouds. The presented results show that generation, storage and real-time rendering of even large urban models are feasible, even with off-the-shelf hardware.
Introduction
The demand for high-fidelity 3D city models has surged with the rise of autonomous driving, urban planning, and digital twin technologies. Traditionally, these models are represented as point clouds—collections of 3D coordinates captured via LiDAR or photogrammetry. While point clouds are straightforward to generate, they possess inherent limitations. They do not explicitly distinguish between "free" and "unexplored" space, making navigation and change detection computationally expensive. Furthermore, point clouds lack a fixed memory upper bound, as memory consumption scales linearly with the number of points captured, regardless of the scene's complexity.
Volumetric representations, such as voxels, address these shortcomings by partitioning space into a regular grid. This allows for explicit occupancy modeling and efficient spatial queries. However, scaling voxel grids to city-wide or global levels introduces significant challenges in memory management, data integration, and real-time visualization. This paper presents a scalable framework designed to overcome these hurdles, enabling global-scale volumetric modeling on standard hardware.
System Architecture and Methodology
The proposed framework is built upon a modular architecture that decouples data ingestion, storage, and visualization. At its core is a Global Index that manages the transition from planetary scales down to individual voxels.
Global Indexing and Spatial Partitioning
To achieve global scalability, the framework utilizes the Earth-Centered, Earth-Fixed (ECEF) coordinate system. The Earth's surface is tessellated into a grid of "tiles," which serve as the primary unit for data persistence and memory management. Each tile is uniquely addressed using Morton coding (Z-order curve), which preserves spatial locality and allows for efficient range queries. This hierarchical indexing enables the system to dynamically load and unload specific urban areas based on the observer's position, preventing RAM exhaustion.
Probabilistic Octrees and Uncertainty-Aware Import
Within each tile, the environment is represented using a probabilistic octree, specifically an implementation based on OctoMap. Unlike binary occupancy grids, probabilistic octrees maintain a likelihood value for each cell, allowing the system to integrate multiple measurements over time and handle sensor noise gracefully.
A key innovation in the framework is the "Uncertainty-Aware Import" strategy. Rather than treating every laser scan point as an absolute coordinate, the system considers the sensor's standard deviation ($\sigma$). Data is integrated at a depth in the octree that corresponds to the sensor's precision. High-precision data creates fine-grained voxels, while noisy data is integrated at coarser levels, naturally filtering out "ghost" objects and discretization artifacts.
Compression Strategies
To manage the massive data volumes associated with urban modeling, the authors propose two compression techniques:
- Lossless Leaf Optimization: By recognizing that approximately 87.5% of nodes in a fully populated octree are leaf nodes, the framework optimizes their storage. By removing redundant child pointers and metadata from these terminal nodes, a 15.6% reduction in memory is achieved without any loss of fidelity.
- Lossy 9-bit Quantization: For scenarios where memory is at a premium, the framework offers a lossy mode. Occupancy and color information are quantized into a 9-bit encoding. This results in a staggering 80.5% compression rate, with minimal impact on visual quality or model accuracy.
Experimental Evaluation
The framework was rigorously tested using a mobile laser scanning (MLS) dataset from the Technical University of Munich (TUM). The dataset covers approximately 0.2 $km^2$ of the inner-city campus and includes over 715 million points captured by Velodyne HDL-64E scanners and Ladybug5 cameras.
The results demonstrate significant efficiency gains. While the raw point cloud data occupied 17.1 GiB, the compressed volumetric model was substantially smaller. In terms of performance, the software-based "Splatting" renderer achieved real-time frame rates (~47ms per frame) at a resolution of 1.6 meters. For high-resolution tasks (10cm), the system employs an iterative refinement approach, allowing users to navigate a low-resolution proxy that progressively gains detail when the camera remains stationary.
Conclusion
This research demonstrates that global-scale volumetric modeling is not only theoretically possible but practically feasible on consumer-grade hardware. By combining hierarchical global indexing with uncertainty-aware data integration and aggressive compression, the framework provides a robust foundation for next-generation 3D city models. Future work will focus on semantic extraction to distinguish between static infrastructure and dynamic objects, further enhancing the utility of these models for real-time urban monitoring.
Summary updated