Visual SLAM Pose Estimation Using 2D Key Frame Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Visual Simultaneous Localization and Mapping (VSLAM) technologies face challenges in accurately computing visual relative poses due to limitations in feature point matching and landmark creation, especially when image acquisition periods are long or image differences are large, leading to reduced success rates and accuracy in localization and mapping.
Innovation Solution
The method involves receiving images from a visual sensor, retrieving key frames from a database, matching images with key frames, computing visual relative poses based on two-dimensional coordinates of matching feature points, and updating the mobile device's pose and map using both visual and dead reckoning information, with processes including sorting key frames, selecting candidate frames, and performing graph optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If three-dimensional coordinates computation is used for visual relative pose calculation, then localization accuracy can be achieved, but the success rate decreases when image acquisition period is long or image differences are large
Solution Approach 1:
The patent extracts only the essential two-dimensional feature point coordinates from images, eliminating the need for complex three-dimensional coordinate computation. This extraction approach maintains sufficient information for pose calculation while avoiding the limitations that cause computation failures in challenging scenarios.
Solution Approach 2:
The patent replaces the traditional three-dimensional coordinate computation mechanism with a two-dimensional feature point matching mechanism. This substitution simplifies the computational model, removing the restrictive conditions required for 3D reconstruction and enabling successful pose calculation even when image acquisition periods are long or image differences are large.
2Loss of information
If three-dimensional coordinates computation is used, then complete spatial information is obtained, but computation failures occur more frequently in challenging scenarios
Solution Approach 1:
The patent extracts two-dimensional feature point coordinates directly from images without performing three-dimensional reconstruction. This extraction method retains the necessary spatial relationship information for pose calculation while avoiding the computational complexity and failure modes associated with 3D coordinate computation.
Solution Approach 2:
The patent uses simple two-dimensional feature point data instead of complex three-dimensional coordinate systems. This approach employs a simpler, more robust computational model that is less prone to failure, effectively replacing the fragile 3D reconstruction process with a more reliable 2D matching approach.
3Productivity
If feature point matching with pre-established landmarks is performed, then visual relative pose can be computed, but the number of usable landmarks decreases when image conditions are challenging
Solution Approach 1:
The patent makes the feature point matching system universal by removing the requirement for pre-established three-dimensional landmarks. The two-dimensional feature point matching approach can work with any image pair, making the system applicable in all scenarios without being limited by landmark availability or image acquisition conditions.
Solution Approach 2:
The patent extracts feature points directly from image data without requiring pre-established landmark databases. This extraction approach eliminates the bottleneck of landmark matching, allowing the system to compute poses using only the images themselves, thereby maintaining productivity even when traditional landmark-based methods fail.
Data Source
AI summary
A method includes acquiring an image through a visual sensor during a movement of a mobile device. The method includes matching the image with key frames stored in a key frame database. The key frames are created based on two-dimensional coordinates of feature points included in a plurality of images previously acquired through the visual sensor. The method also includes computing a visual relative pose based on two-dimensional coordinates of matching feature points included in both of the image and the one or more key frames that have been matched with the image. The method also includes computing relevant information of the visual relative pose based on the two-dimensional coordinates of the matching feature points. The method further includes updating an absolute pose of the mobile device and a map based on the relevant information of the visual relative pose and relevant information of a dead reckoning based relative pose.