Vehicle Map Matching Using SD Maps and Trajectory Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing map matching and vehicle localization techniques in autonomous vehicles face inaccuracies due to limited satellite visibility and the computational expense of sensor fusion, especially in complex environments, and are compromised by outdated or incomplete high-definition map data.
Innovation Solution
A deep-learning based approach that encodes GNSS and inertial navigation system sensors for accurate map matching and vehicle localization using low-resolution Standard Definition maps, employing encoding and map matching networks to determine vehicle position on candidate road segments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sensor fusion techniques combining GNSS with cameras, LiDAR, and radar are used, then localization accuracy is improved, but computational cost increases and requires high-definition map data
Solution Approach 1:
The patent extracts and uses only the necessary sensor data (GNSS coordinates and inertial navigation data) rather than fusing all available sensor types. This selective extraction approach maintains localization accuracy while significantly reducing computational complexity and eliminating the requirement for high-definition map data.
Solution Approach 2:
The patent employs standard definition map data instead of expensive high-definition map data. The SD map serves as a sufficient substitute that provides the necessary road geometry information for accurate localization without the computational burden and data requirements of HD maps.
2Measurement precision
If high-definition map data is used for map matching, then map matching accuracy is improved, but practical applications are limited due to computational expense and data availability
Solution Approach 1:
The patent successfully uses standard definition map data, which is widely available, computationally efficient to process, and sufficiently accurate for localization purposes. This approach dramatically improves practical applicability while maintaining acceptable map matching accuracy through the deep learning model's ability to extract relevant features from lower-resolution data.
3Device complexity
If traditional map matching methods relying on GNSS are used, then system simplicity is maintained, but localization accuracy deteriorates in environments with limited satellite visibility
Solution Approach 1:
The patent introduces an intermediary deep learning-based map matching network that processes GNSS coordinates and inertial navigation data to determine vehicle position on road segments. This intermediary layer transforms simple sensor inputs into accurate localization outputs, bridging the gap between system simplicity and measurement precision in challenging environments.
4Adaptability or versatility
If deep learning models are used for map matching, then adaptability to handle multiple sensor inputs is improved, but computational requirements increase
Solution Approach 1:
The patent extracts and processes only the essential sensor inputs (GNSS coordinates and inertial navigation data) through the deep learning model, rather than integrating all possible sensor types. This selective approach maintains the model's adaptability and sensor fusion capabilities while significantly reducing computational energy consumption compared to comprehensive sensor fusion approaches.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The herein disclosed technology relates a method and device for determining a current position of a vehicle on a navigation map. The method (100) comprises: obtaining (5102) a driven ego-trajectory of the vehicle comprising a sequence of consecutive positions of the vehicle, wherein the sequence of consecutive positions comprises at least a last position of the vehicle; obtaining (S104) a set of candidate road segments of the navigation map, wherein the set of candidate road segments are selected based on the last position of the vehicle; processing (S108) the driven trajectory, and each candidate road segment of the set of candidate road segments through an encoding network, to generate a feature representation of the driven trajectory and each of the candidate road segments respectively; applying (S110) a map matching network to the feature representation of the driven trajectory and the feature representation of each candidate road segment, wherein the map matching network is trained to determine a similarity between a driven trajectory and a road segment; determining (5112) a current road segment, of the set of candidate road segments, that the vehicle is currently on, based on the application of the map matching network; and determining (5118) a current position of the vehicle along the current road segment by processing information indicative of the application of the map matching network and the feature representation of the driven ego-trajectory through a position prediction network.