CNN Regression Localization via 2D Map Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CNN regression-based localization techniques using VSLAM are ineffective for large-scale structures due to limited visual information, high computational and memory requirements, drift errors, and incompatibility with human maps, especially in environments with significant glass surfaces.
Innovation Solution
Divide large structures into segments, gather visual data for camera pose estimation, generate 3D point clouds, correlate them with 2D maps, and train CNN regression models to predict coordinates within the 2D map framework, enabling efficient re-localization without additional instrument installation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If CNN regression-based localization using VSLAM is applied to large-scale structures, then localization capability is provided, but visual information becomes limited and computational requirements increase
Solution Approach 1:
The patent divides large-scale structures into multiple segments or regions, each with its own localized CNN model. This segmentation allows the system to process visual information in manageable chunks, preventing information loss while maintaining localization accuracy across the entire large-scale environment.
Solution Approach 2:
The patent transitions from traditional 2D image processing to 3D point cloud processing combined with 2D map coordinates. By incorporating depth information and spatial relationships in three dimensions, the system overcomes the limitations of visual information in large-scale 2D environments.
2Reliability
If CNN regression-based localization using VSLAM is applied to large-scale structures, then localization capability is provided, but computational and memory requirements increase
Solution Approach 1:
The patent segments the large-scale environment into multiple smaller regions, each processed by dedicated CNN models. This division reduces the computational and memory burden on individual processors compared to processing the entire large-scale environment in one unified model.
Solution Approach 2:
The patent processes only the relevant local region around the robot's current position in detail, while using coarser representations for distant areas. This partial processing approach reduces overall computational requirements while maintaining sufficient localization accuracy for the immediate environment.
3Reliability
If traditional localization methods are used, then localization is provided, but drift errors occur
Solution Approach 1:
The patent implements a feedback mechanism where the CNN regression model continuously compares predicted robot positions with actual sensor observations and map data. This closed-loop feedback corrects accumulated drift errors by referencing the pre-built 2D map and adjusting predictions accordingly.
Solution Approach 2:
The patent performs preliminary actions by pre-processing environmental data to create accurate 2D maps and pre-training CNN models with ground truth localization data. This preliminary preparation establishes a reference framework that prevents drift errors during actual operation.
4Measurement precision
If additional instruments are installed for localization, then localization accuracy is improved, but cost and system complexity increase
Solution Approach 1:
The patent makes the camera system multi-functional by using it for both visual exploration and precise localization tasks. By processing camera images through CNN regression models against pre-built 2D maps, the system achieves high localization accuracy without requiring separate dedicated localization sensors.
Solution Approach 2:
The patent enables the robot to perform self-localization using only its existing camera and processor resources. The system processes its own visual data through the CNN model and compares it with the 2D map to determine its position, eliminating the need for external localization instruments or infrastructure.
Data Source
AI summary
Embodiments described herein provide a processing apparatus comprising compute logic to train a convolutional neural network (CNN) to perform autonomous re-localization for a service robot or mobile device. In one embodiment the apparatus comprises an image processor to process visual data received via a sensor and a general purpose graphics processing engine perform camera pose estimation for image data and generate a transformation matrix to transform positions of camera pose estimations to positions within a human readable map of the location. The images and transformed positions are uses to train the CNN to perform re-localization.