Vehicle Positioning With Image Segmentation Graphs for GNSS-Limited Roads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle localization methods in autonomous driving face challenges in achieving high accuracy and robustness in environments with limited GNSS coverage, particularly due to accumulated errors in inertial navigation and the limitations of LiDAR and camera sensors, which are influenced by light and weather conditions.
Innovation Solution
A method utilizing a localization framework that integrates deep learning techniques, including Convolutional Neural Networks (CNN) for visual image feature extraction, Graph Convolutional Networks (GCN) for sensor data integration, and Graph Attention Networks (GAT) for data association, to construct and compare graph topological landmark maps, enabling precise vehicle localization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If inertial navigation system with IMU and wheel odometry is used for vehicle localization, then cost is reduced, but accumulated errors increase over time
Solution Approach 1:
The patent combines multiple localization methods including inertial navigation, GNSS, and visual/LiDAR-based map matching into a unified localization system. The system integrates data from IMU, wheel odometry, camera, and LiDAR sensors to compensate for the accumulated errors of inertial navigation while maintaining cost-effectiveness.
Solution Approach 2:
The system implements feedback mechanisms where localization results from map matching and sensor fusion are used to correct and reset the accumulated errors from inertial navigation. The filter continuously updates position estimates based on feedback from multiple sensor sources to maintain accuracy.
2Measurement precision
If advanced GNSS technology like DGPS and RTK is used to improve localization accuracy, then measurement precision is improved, but device complexity and cost increase
Solution Approach 1:
The patent employs standard GNSS receivers rather than expensive advanced GNSS systems like RTK or DGPS. The system achieves high accuracy through alternative means (visual and LiDAR map matching) instead of relying on costly advanced GNSS technology, effectively using cheaper components to achieve the same goal.
Solution Approach 2:
The system uses multi-functional sensors (camera and LiDAR) that serve both localization and environmental perception functions. This eliminates the need for dedicated expensive localization hardware, reducing overall system complexity while maintaining high accuracy.
3Measurement precision
If LiDAR and camera sensors are used for localization, then measurement precision is improved, but reliability decreases due to sensitivity to light and weather conditions
Solution Approach 1:
The patent fuses data from multiple sensor types (camera, LiDAR, IMU, wheel odometry) to create a composite localization solution. Each sensor type compensates for the weaknesses of others, with LiDAR providing depth information robust to lighting conditions and camera providing rich visual features, achieving both accuracy and reliability.
Solution Approach 2:
The system dynamically adjusts its reliance on different sensors based on environmental conditions. When light or weather conditions degrade camera or LiDAR performance, the system increases weighting on inertial navigation and other more reliable sensors, maintaining consistent localization reliability across varying conditions.
4Device complexity
If traditional feature extraction methods like SIFT, SURF, FAST, or ORB are used, then device complexity is reduced, but measurement precision and robustness to environmental changes decrease
Solution Approach 1:
The patent replaces traditional hand-crafted feature extraction methods with deep learning-based feature extraction using neural networks. This substitution enables the system to learn robust feature representations directly from raw sensor data, achieving higher accuracy and environmental robustness while the computational complexity is managed through efficient network architectures.
5Measurement precision
If complex image data from camera sensors or point clouds data from LiDAR scanners are directly utilized in filtering methods, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent segments the complex sensor data processing into distinct stages: feature extraction from raw images and point clouds, feature matching between current and reference data, and position estimation through filtering. This segmentation allows each stage to process data in an optimized manner, managing complexity while preserving measurement precision.
Data Source
AI summary
A method of determining a position of a vehicle in which a reference map is provided. The reference map comprises segmentations of a reference image with landmarks. A measurement image of a vehicle environment is captured and segmentations of the measurement image and neighborhood graphs are determined to obtain a measurement map, wherein a segmentation is represented by a vertex and where a neighborhood graph comprises the vertex and edges containing information to identify neighboring vertices of the vertex. Segmentations of the reference image are compared with the segmentations represented by the vertices of the measurement image and the neighborhood graphs, and segmentations contained in the reference image and in measurement image are determined. The vehicle's position is estimated with reference to the reference map during its movement along a road based on a result of the comparison.