Annotated Pose Graph Optimisation for Place-Aware Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating pose graphs in mapping physical areas are computationally expensive and require additional steps to determine the type of location or environment, often leading to improper cluster segregation due to lack of consideration for both spatial constraints and feature information, especially in diverse environments.
Innovation Solution
A method that optimizes pose graphs by predicting place categories for each node based on sensor poses and feature vectors, using machine learning models like CNNs and MLPs, and annotating nodes with place category labels to generate an annotated optimised pose graph.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pose graph optimisation is performed using non-linear least-squares formulation, then the discrepancy between measurements and constructed model is minimized, but the computational cost becomes expensive
Solution Approach 1:
The patent divides the pose graph into multiple hierarchical levels (fine-level and coarse-level graphs). The fine-level graph captures detailed local geometry while the coarse-level graph represents broader spatial relationships. This segmentation allows optimization to be performed at different scales, reducing the overall computational burden while maintaining accuracy.
Solution Approach 2:
The patent introduces a hierarchical dimension to the traditional single-level pose graph by creating multiple levels of abstraction. This dimensional transformation enables the system to process information at different granularities, optimizing computational efficiency without sacrificing measurement precision.
2Measurement precision
If loop closure step is added to remove redundancies, then the accuracy of mapping is improved, but the difficulty of searching and validating potential loop closures increases
Solution Approach 1:
The patent segments the loop closure detection process into hierarchical levels. At the coarse level, potential loop closures are identified using simplified representations, and at the fine level, these candidates are validated with detailed measurements. This segmentation reduces the search space and validates loop closures more efficiently.
Solution Approach 2:
The patent adds a hierarchical dimension to loop closure detection by creating multiple levels of validation. This dimensional approach transforms the complex single-level search problem into a multi-level process, reducing computational complexity while improving detection accuracy.
3Ease of manufacture
If nodes are clustered based on node feature similarity without considering location identity, then the clustering process is simplified, but the clustering accuracy deteriorates due to ambiguous cluster assignments
Solution Approach 1:
The patent segments the clustering process into two distinct stages: geometric clustering based on spatial relationships and semantic clustering based on place categories. This segmentation allows each clustering type to specialize, improving overall accuracy while maintaining process simplicity through modular design.
Solution Approach 2:
The patent applies different clustering criteria to different aspects of node characteristics. Geometric clustering uses spatial features while semantic clustering uses place category information. This local quality approach ensures that each clustering dimension uses the most appropriate features for that specific purpose.
4Measurement precision
If existing methods use specific landmarks such as doors to generate accurate subgraphs, then the accuracy of subgraph generation is improved, but the adaptability to different environments deteriorates
Solution Approach 1:
The patent creates a universal place category system that can identify and categorize multiple types of locations (rooms, corridors, intersections, etc.) rather than relying on specific landmarks like doors. This universal categorization system can be applied across diverse environments, from indoor to outdoor settings, maintaining accuracy while improving adaptability.
Solution Approach 2:
The patent changes the fundamental parameters used for subgraph generation from specific landmark-based features to general place category parameters. This parameter transformation allows the system to generate accurate subgraphs in diverse environments by adapting to different location types rather than requiring specific landmark structures.
5Use of energy by moving object
If hierarchical pose graphs with multiple layers are generated, then the computational power is reduced, but the device complexity increases
Solution Approach 1:
The patent segments the pose graph into hierarchical levels with clear separation between fine-level and coarse-level representations. Each level has distinct optimization processes and data structures, which organizes the complexity into manageable segments rather than a monolithic complex system.
Solution Approach 2:
The patent introduces a hierarchical dimension to the pose graph structure, transforming a single-level complex problem into a multi-level structured system. This dimensional transformation organizes complexity vertically across levels rather than horizontally within a single layer, improving computational efficiency through structured organization.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a computer-implemented methods of generating an annotated optimised pose graph representative of a physical area, the method comprising: receiving a pose graph representative of a physical area, predicting a place category for each node based at least on the sensor pose and at least one feature vector associated with such node; optimising the pose graph based at least on the predicted place category; and annotating each node with a place category label corresponding to the predicted place category to generate an annotated optimised pose graph. The invention further relates to a machine-learning model to carry out the method, methods of generating a training dataset and training the machine-learning model, a training dataset used to train the machine-learning model, use of the generated annotated optimised pose graph, a data processing system, and a computer program, machine-readable storage medium, or a data carrier signal.