Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Transformation & Translation Occupancy Grid Mapping: 2-Dimensional Deep Learning Refined SLAM

Leon Davies, Baihua Li, Mohamad Saada, Simon Sølvsten, Qinggang Meng

TL;DR

This work tackles the persistent gap between 3D SLAM advances and 2D occupancy grid map quality in large, dynamic environments. It proposes Transformation & Translation Occupancy Grid Mapping (TT-OGM), which converts 3D LiDAR odometry into accurate 2D OGMs via a $6$-DOF transformation and a 2D projection, then refines the map with a GAN-based observation completion module. A novel DRL-driven data generation pipeline supplies diverse, realistic SLAM errors for robust training, and a PatchGAN-based generator paired with a PatchNCE loss and a Query-Selected Attention module delivers real-time observation cleaning. Evaluations on the Haslegrave dataset and Radish maps show improved IoU and visual quality, with the approach generalizing to complex, large-scale indoor environments and offering practical benefits for floor-plan like mapping and autonomous navigation on embedded hardware.

Abstract

SLAM (Simultaneous Localisation and Mapping) is a crucial component for robotic systems, providing a map of an environment, the current location and previous trajectory of a robot. While 3D LiDAR SLAM has received notable improvements in recent years, 2D SLAM lags behind. Gradual drifts in odometry and pose estimation inaccuracies hinder modern 2D LiDAR-odometry algorithms in large complex environments. Dynamic robotic motion coupled with inherent estimation based SLAM processes introduce noise and errors, degrading map quality. Occupancy Grid Mapping (OGM) produces results that are often noisy and unclear. This is due to the fact that evidence based mapping represents maps according to uncertain observations. This is why OGMs are so popular in exploration or navigation tasks. However, this also limits OGMs' effectiveness for specific mapping based tasks such as floor plan creation in complex scenes. To address this, we propose our novel Transformation and Translation Occupancy Grid Mapping (TT-OGM). We adapt and enable accurate and robust pose estimation techniques from 3D SLAM to the world of 2D and mitigate errors to improve map quality using Generative Adversarial Networks (GANs). We introduce a novel data generation method via deep reinforcement learning (DRL) to build datasets large enough for training a GAN for SLAM error correction. We demonstrate our SLAM in real-time on data collected at Loughborough University. We also prove its generalisability on a variety of large complex environments on a collection of large scale well-known 2D occupancy maps. Our novel approach enables the creation of high quality OGMs in complex scenes, far surpassing the capabilities of current SLAM algorithms in terms of quality, accuracy and reliability.

Transformation & Translation Occupancy Grid Mapping: 2-Dimensional Deep Learning Refined SLAM

TL;DR

This work tackles the persistent gap between 3D SLAM advances and 2D occupancy grid map quality in large, dynamic environments. It proposes Transformation & Translation Occupancy Grid Mapping (TT-OGM), which converts 3D LiDAR odometry into accurate 2D OGMs via a -DOF transformation and a 2D projection, then refines the map with a GAN-based observation completion module. A novel DRL-driven data generation pipeline supplies diverse, realistic SLAM errors for robust training, and a PatchGAN-based generator paired with a PatchNCE loss and a Query-Selected Attention module delivers real-time observation cleaning. Evaluations on the Haslegrave dataset and Radish maps show improved IoU and visual quality, with the approach generalizing to complex, large-scale indoor environments and offering practical benefits for floor-plan like mapping and autonomous navigation on embedded hardware.

Abstract

SLAM (Simultaneous Localisation and Mapping) is a crucial component for robotic systems, providing a map of an environment, the current location and previous trajectory of a robot. While 3D LiDAR SLAM has received notable improvements in recent years, 2D SLAM lags behind. Gradual drifts in odometry and pose estimation inaccuracies hinder modern 2D LiDAR-odometry algorithms in large complex environments. Dynamic robotic motion coupled with inherent estimation based SLAM processes introduce noise and errors, degrading map quality. Occupancy Grid Mapping (OGM) produces results that are often noisy and unclear. This is due to the fact that evidence based mapping represents maps according to uncertain observations. This is why OGMs are so popular in exploration or navigation tasks. However, this also limits OGMs' effectiveness for specific mapping based tasks such as floor plan creation in complex scenes. To address this, we propose our novel Transformation and Translation Occupancy Grid Mapping (TT-OGM). We adapt and enable accurate and robust pose estimation techniques from 3D SLAM to the world of 2D and mitigate errors to improve map quality using Generative Adversarial Networks (GANs). We introduce a novel data generation method via deep reinforcement learning (DRL) to build datasets large enough for training a GAN for SLAM error correction. We demonstrate our SLAM in real-time on data collected at Loughborough University. We also prove its generalisability on a variety of large complex environments on a collection of large scale well-known 2D occupancy maps. Our novel approach enables the creation of high quality OGMs in complex scenes, far surpassing the capabilities of current SLAM algorithms in terms of quality, accuracy and reliability.
Paper Structure (18 sections, 11 equations, 9 figures, 4 tables, 1 algorithm)

This paper contains 18 sections, 11 equations, 9 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Simultaneous Localisation and Mapping
  4. Deep Learning & SLAM
  5. Methodology
  6. Overview of TT-OGM
  7. Transformation and Translation
  8. Transformation
  9. Translation
  10. Occupancy Grid Mapping
  11. Occupancy Grid Cleaing
  12. GAN Model
  13. Model Learning and Experiment Results
  14. Network Training
  15. Real-time SLAM: Haslegrave Dataset
  16. ...and 3 more sections

Figures (9)

  • Figure 1: Transformation & Translation Occupancy Grid Mapping - system overview. Accurate and noise free 2D SLAM from the singular input of consecutive 360° LiDAR scans. Transformation and Translation enables accurate 3D pose estimation algorithms to be relevant for 2D OGM. Error correction and artefact removal is applied through deep learning.
  • Figure 2: TT-OGM system diagram: Producing a high quality and accurate occupancy grid map in real-time. Main stages: 1) Input of consecutive LiDAR point clouds are translated to 2D LiDAR while the transformation between them in calculated through Scan-to-Scan (S2S) transforms and Generalised-iterative closest point (GICP). The position within the world is converted from LiDAR coordinate system using a Scan-To-Map (S2M) transformation. Ransac, box and voxel filters are used for outlier rejection and data decomplexification. 2) An estimated OGM is constructed with the pose produced by the transformation branch and the 2D LiDAR produced by the translation branch. 3) The OGM is filtered and processed to prepare it for the GAN model. 4) The OGM is cleaned by the Generator in the GAN model 5) The output is post processed into a clean OGM and provided to the robot or operator.
  • Figure 3: Results of box filter and voxel grid filtering incoming point clouds. (a) Raw point cloud data. (b) Filtered point cloud.
  • Figure 4: 3D LiDAR point cloud and translated 2D representation overlapped.
  • Figure 5: Diagram of filtering operation based on occupancy intensity.
  • ...and 4 more figures