HIO-SDF: Hierarchical Incremental Online Signed Distance Fields

TL;DR

HIO-SDF is introduced, a new method that represents the environment as a Signed Distance Field (SDF) using a hierarchical approach which employs a coarse voxel grid that captures the observed parts of the environment together with high-resolution local information to train a neural network.

Abstract

A good representation of a large, complex mobile robot workspace must be space-efficient yet capable of encoding relevant geometric details. When exploring unknown environments, it needs to be updatable incrementally in an online fashion. We introduce HIO-SDF, a new method that represents the environment as a Signed Distance Field (SDF). State of the art representations of SDFs are based on either neural networks or voxel grids. Neural networks are capable of representing the SDF continuously. However, they are hard to update incrementally as neural networks tend to forget previously observed parts of the environment unless an extensive sensor history is stored for training. Voxel-based representations do not have this problem but they are not space-efficient especially in large environments with fine details. HIO-SDF combines the advantages of these representations using a hierarchical approach which employs a coarse voxel grid that captures the observed parts of the environment together with high-resolution local information to train a neural network. HIO-SDF achieves a 46% lower mean global SDF error across all test scenes than a state of the art continuous representation, and a 30% lower error than a discrete representation at the same resolution as our coarse global SDF grid. Videos and code are available at: https://samsunglabs.github.io/HIO-SDF-project-page/

Figures (7)

• Figure 1: Overview of approach: Online discrete Signed Distance Field (SDF) representations can be very accurate but do not include any information in unobserved areas. Existing continuous SDF representations solve that problem but cannot easily capture fine details in the environment. Our proposed method (HIO-SDF) uses a hierarchy of discrete and continuous algorithms for accurate online 3D SDF reconstruction by a global neural network using periodic activation functions.
• Figure 2: Overview of our hierarchical incremental online global SDF architecture (HIO-SDF): The input to the algorithm is a stream of posed point cloud measurements (dense or sparse) and the output is a global implicit 3D SDF map of the environment that can be used directly for downstream tasks. At each time step, the input is consumed by two different modules that produce SDF data (i.e., tuples of 3D coordinates and associated SDF values) for the supervision of a SIREN Sitzmann_SIREN_2020 representing our global SDF: 1) a discrete global SDF module fusing measurements in the background at a coarse resolution, and 2) a continuous local SDF module that does not need training and performs brute-force distance computations on the GPU between stratified raycasted samples and the currently observed point cloud. Training of the global SIREN is incremental and online, as new sensor measurements are received.
• Figure 3: Local SDF data generation method: The local SDF module uses raycasting from the sensor location to find points with positive (blue) and negative (red) SDF values before and after the currently observed point cloud (black), respectively. The absolute local SDF value of each point is the minimum distance (magenta) to the entire local point cloud.
• Figure 4: Global SDF slice comparison: The sensor trajectories for each scene are shown in red at the top row with ground truth meshes. SDF slices are computed using the final SDF at the end of each sequence. HIO-SDF is able to represent more detailed SDF information around furniture, compared to both iSDF and Voxfield. For Voxfield, invisible regions by the sensor are shown in white. Details captured by our method even though not captured by the underlying coarse-SDF Voxfield backbone are highlighted in yellow.
• Figure 5: Mesh reconstruction comparison: By virtue of using SIREN and utilizing both continuous local and coarse global information, HIO-SDF generates complete and smooth surfaces that also include finer geometric details over the entire environment compared to iSDF and Voxfield.