LGSDF: Continual Global Learning of Signed Distance Fields Aided by Local Updating

Yufeng Yue,Yinan Deng,Jiahui Wang,Yi Yang

2024-04-08

Abstract:Implicit reconstruction of ESDF (Euclidean Signed Distance Field) involves training a neural network to regress the signed distance from any point to the nearest obstacle, which has the advantages of lightweight storage and continuous querying. However, existing algorithms usually rely on conflicting raw observations as training data, resulting in poor map performance. In this paper, we propose LGSDF, an ESDF continual Global learning algorithm aided by Local updating. At the front end, axis-aligned grids are dynamically updated by pre-processed sensor observations, where incremental fusion alleviates estimation error caused by limited viewing directions. At the back end, a randomly initialized implicit ESDF neural network performs continual self-supervised learning guided by these grids to generate smooth and continuous maps. The results on multiple scenes show that LGSDF can construct more accurate ESDF maps and meshes compared with SOTA (State Of The Art) explicit and implicit mapping algorithms. The source code of LGSDF is publicly available at

Computer Vision and Pattern Recognition,Graphics,Robotics

What problem does this paper attempt to address?

### Problems Addressed by the Paper

This paper aims to address several key issues in the construction of Euclidean Signed Distance Fields (ESDF) and proposes a new algorithm—LGSDF (Local Grid Updating aided Continual Global Learning of Signed Distance Fields) to achieve more accurate and continuous map construction.

#### Main Issues
1. **Problems with Existing Methods**:
- Explicit mapping methods (such as Voxblox, FIESTA, VDB-EDT) typically rely on discrete grids with limited resolution, which cannot predict details in unobserved areas, leading to limited map performance.
- Implicit mapping methods (such as iSDF) can generate continuous maps, but severe conflicts in training data lead to distortion in the final map.

2. **Challenges in Continual Learning**:
- Catastrophic forgetting is prone to occur during continual learning, where learning new data may lead to the loss of old knowledge.

#### Proposed Method
- **Frontend Module**: Includes Active Sampling and Local Updating. The Active Sampling module samples at the pixel and point levels, while the Local Updating module dynamically updates the discrete axis-aligned grid by fusing sensor observations.
- **Backend Module**: Global Learning, which utilizes a randomly initialized neural network for self-supervised learning, guided by the updated grid to construct the implicit ESDF map.

#### Objectives
- Propose a new implicit ESDF mapping algorithm, LGSDF, which for the first time uses local grid updates to aid continual global learning.
- Design a novel pixel sampling strategy in the Active Sampling module to ensure that complex foreground instances receive more perceptual attention, improving overall detail quality.
- Design a novel grid selection strategy in the Global Learning module to ensure that the implicit map has both historical memory and dynamic sensitivity.

Through the above methods, LGSDF can achieve global continuity while ensuring local details, thereby constructing more accurate and smoother maps and grids. Experimental results validate the high accuracy and neatness of LGSDF in multiple public scenarios.