Scan-Matching Visual SLAM Using Ground Plane Projection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current 2D SLAM methods for robotics struggle with feature detection and feature description in non-feature environments, leading to issues like getting lost, local maximum trapping, and difficulty in loop closure and pose recovery.
Innovation Solution
A scan-matching oriented visual SLAM method that combines 2D scan with color histogram/feature for mapping using RGBD cameras, involving ground plane detection, point cloud projection, color descriptor generation, and scan matching to improve localization and loop closure accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional 2D SLAM methods are used for feature detection and description, then the system can operate in general environments, but it fails in non-feature environments (e.g., white walls) leading to getting lost and local maximum trapping
Solution Approach 1:
The patent transitions from traditional 2D image-based feature detection to a hybrid 2D-3D approach by projecting depth information onto a 2D ground plane. This dimensionality transformation allows the system to leverage both 2D color descriptors and 3D spatial structure, enabling reliable localization in environments that lack sufficient 2D visual features.
Solution Approach 2:
The patent creates a composite representation by combining color descriptors (from RGB images) with depth information (from depth cameras) and projecting them onto a unified 2D ground plane map. This composite approach integrates multiple data types to overcome the limitations of using单一 feature type, improving both reliability and adaptability across different environments.
2Manufacturing precision
If dense 3D point cloud data is processed for accurate mapping, then mapping precision improves, but computational load increases significantly
Solution Approach 1:
The patent extracts only the essential information needed for SLAM by projecting 3D point cloud data onto a 2D ground plane, discarding redundant vertical dimension information. This extraction process maintains the critical spatial relationships for navigation while dramatically reducing the data volume and computational requirements for processing.
Solution Approach 2:
By transforming 3D point cloud data into a 2D projection on the ground plane, the system reduces computational complexity while preserving the essential spatial structure needed for accurate mapping and localization. This dimensionality reduction maintains mapping precision for ground-level navigation tasks.
3Measurement precision
If color descriptors are added to each grid cell for improved matching, then scan matching accuracy improves, but data storage requirements increase
Solution Approach 1:
The patent applies color descriptors selectively to grid cells that contain projected depth points, rather than to all grid cells in the map. This partial application approach provides sufficient color information for accurate matching in relevant areas while minimizing unnecessary data storage in empty or irrelevant regions of the map.
Data Source
AI summary
A system and method for constructing and updating a two dimensional (2D) grid cell map based on depth images and RGB images. The method includes: providing an RGB image and a depth image by a sensor of a computing device; determining scores of pixels in the RGB image representing possibility of the pixels being located in a ground plane; determining normal directions of pixels in the depth image; determining the ground plane based on the scores and the normal directions; projecting the pixels in the depth images to the ground plane; and generating and updating the 2D grid cell map by defining the ground plane into grid cells and attributing a color descriptor to each grid cell. The color descriptor of each grid cell is calculated based on the projected pixels within that grid cell.