Ego-Vehicle Localization Using Semantic Map Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing localization methods for autonomous vehicles face challenges in maintaining precise and robust localization due to variations in lighting conditions, weather, and seasonal changes, which can lead to performance losses, especially when relying heavily on visual information.
Innovation Solution
A method utilizing a global semantic map with semantic sensor data from cameras, radars, or lidars, processed through a Monte Carlo localization framework, where semantic classes are matched and fused using Bayesian recursive state estimation, allowing for robust localization without post-processing, and using boundaries and corners of objects as landmarks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If visual localization methods using feature points or 3D reconstruction are used, then localization capability is achieved, but performance losses occur due to lighting conditions, weather, and seasonal changes
Solution Approach 1:
The patent transforms the localization approach from relying on visual appearance parameters (color, texture, lighting) to relying on geometric and semantic parameters (object boundaries, corners, semantic classes). This parameter transformation makes localization invariant to lighting, weather, and seasonal changes while maintaining accuracy.
Solution Approach 2:
The patent replaces the mechanical/physical visual feature matching system with a semantic understanding system. Instead of matching visual patterns that change with environment, the system uses semantic class identification and geometric structure matching, which are invariant to environmental conditions.
2Measurement precision
If sensor-heavy vehicles with multiple sensors are used for map recording, then localization precision is improved, but device complexity and cost increase
Solution Approach 1:
The patent enables standard vehicles with basic cameras to create and use semantic maps independently. The system processes visual data through semantic segmentation to extract meaningful geometric and semantic information, allowing ordinary vehicles to achieve high-precision localization without specialized heavy sensor equipment.
Solution Approach 2:
The patent extracts geometric parameters (boundaries, corners) and semantic parameters (object classes) from standard camera data, transforming ordinary visual information into precise localization data. This parameter extraction approach achieves high precision using simple, widely-available camera systems rather than complex sensor arrays.
Data Source
Figure 1~2
Figure 3
AI summary
The invention relates to a method for localizing an ego-vehicle (Ev) in a map comprising the following steps: - Receiving (S1) a global semantic map for an environment of the ego-vehicle (Ev) - Capturing (S2) the surroundings of the ego-vehicle (Ev) by means of at least one environmental sensor generating semantic sensor data; - Matching (S3) the generated semantic sensor data with the received global semantic map; - Estimating (S4) the position of the ego-vehicle (Ev) based on the result of the matching.