Progressive compression of 3D point cloud for industrial digital twins

Overview & Synopsis

Point cloud sensing technologies have evolved rapidly, enabling their application across a broad range of domains such as engineering, cultural heritage, and environmental risk management. Once prohibitively expensive and requiring specialized expertise, these systems have become increasingly accessible, offering a cost-effective, fast, and easily updatable approach for realistic 3D modeling. Despite their advantages, point clouds present significant computational and storage challenges. Unlike mesh-based models, they lack explicit topological information, resulting in large and memory-intensive datasets. Such complexity impacts various algorithms, including rendering and transmission, where high processing power or limited memory capacity can severely constrain performance. These issues become particularly important for real-time interaction with the point cloud, and is exacerbated in cloud-based applications depending on networks quality and bandwidth constraints. In such contexts, efficient and progressive compression strategies are essential, not only to reduce file size, but also to enable incremental data transmission, where coarse representations can be used before transmission completion. This thesis proposes novel progressive compression techniques for both geometry and colors of point clouds, leveraging non-axis-aligned data structures. While conventional compression methods rely on axis-aligned spatial partitions for simplicity and predictability, these approaches often suffer from poor alignment with the intrinsic geometry of point clouds, limiting compression efficiency. Our method uses arbitrarily oriented quantized planar-based subdivision for tighter geometric fitting and improved rate-distortion performance. For color attributes, we introduce a palette quantization tree that balances color accuracy against bit-rate cost thus improving rate-distortion performance. Colors and opacity values are encoded within an octree structure guided by a nonaxis-aligned binary space partition (BSP) optimized for the spatial distribution of points. Finally, we design a volumetric rendering framework tailored for efficient visualization of this hybrid structure. The proposed visualization approach achieves high visual fidelity with minimal computational overhead.