Vehicle Positioning via Trajectory Map Matching on SD Road Segments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional map matching and vehicle localization methods in navigation systems face inaccuracies due to limited satellite visibility and require high-definition map data, which are computationally expensive and challenging to maintain, especially in complex urban environments.
Innovation Solution
A deep-learning based approach that encodes GNSS and inertial navigation system data for accurate map matching and vehicle localization using encoding and map matching networks, enabling precise localization on Standard Definition maps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional GNSS-based map matching methods are used, then the system is simple to implement, but localization accuracy deteriorates in environments with limited satellite visibility
Solution Approach 1:
The patent combines multiple data sources (GNSS positions, INS data, map data) and multiple processing approaches (traditional map matching, deep learning-based map matching, position smoothing) into an integrated localization system. The fusion network merges features from different inputs to produce a refined vehicle position that overcomes the limitations of individual methods, particularly in urban canyons and tunnel environments where GNSS signals are weak or unavailable.
Solution Approach 2:
The patent introduces intermediate processing components including an encoding network that transforms raw inputs into feature representations, a fusion network that acts as an intermediary to combine traditional and deep learning-based map matching results, and a position smoothing component that mediates between discrete position estimates to produce continuous, refined trajectories. These intermediaries enable the system to achieve high accuracy without requiring direct, complex processing of all raw inputs.
2Measurement precision
If sensor fusion techniques with high-definition map data are used, then localization accuracy is improved, but computational cost and data requirements increase
Solution Approach 1:
The patent applies partial action by using only the necessary portion of map data and sensor inputs required for accurate localization. The system processes GNSS positions and INS data through selective encoding and fusion operations, applying deep learning methods only where they provide marginal gains over traditional methods. This selective application reduces computational overhead while maintaining accuracy, avoiding the excessive processing required by full sensor fusion with high-definition maps.
Solution Approach 2:
The patent employs computationally efficient approximations and simplified data representations that can be processed quickly and discarded after use. The encoding network uses lightweight feature extraction that does not require maintaining complex, persistent data structures. The system processes and discards intermediate feature representations after they serve their purpose in the fusion network, reducing memory and computational requirements compared to systems that maintain extensive high-definition map data and complex sensor fusion state.
3Measurement precision
If deep learning based map matching is used, then localization accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies preliminary action through the encoding network that pre-processes and transforms raw GNSS positions, INS data, and map data into compressed feature representations before they enter the main map matching and fusion processes. This preliminary encoding reduces the dimensionality and complexity of the data, enabling faster subsequent processing by the deep learning components. The system performs this encoding in advance, so that when real-time map matching is required, the computationally intensive deep learning operations work with already-prepared, compact features rather than raw data.
Data Source
AI summary
A method and device for determining a current position of a vehicle on a navigation map is disclosed. First, a driven ego-trajectory of the vehicle, which is a sequence of the vehicle's recent positions, is obtained. Based on the last position of the vehicle, a set of candidate road segments on the navigation map is selected. The driven ego-trajectory and each candidate road segment are then processed by an encoding network to generate respective feature representations. A map matching network analyzes these feature representations to determine the similarity between the driven trajectory and each road segment. Using this analysis, the specific road segment on which the vehicle is currently traveling is identified. Finally, information from the map matching 10 network, together with the feature representation of the driven ego-trajectory, is processed through a position prediction network to determine the current position of the vehicle.