Pose Metric Network for Vision-Based Localization on HD Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current localization methods for autonomous driving, such as GPS and inertial measurement units, lack sufficient accuracy in complex urban environments, leading to errors that can exceed the width of vehicle lanes, making safe route planning challenging.
Innovation Solution
A pose metric network, trained using deep learning techniques, determines the similarity between captured images and map projections, enabling real-time accurate localization by analyzing features and generating operational parameters for improved vehicle navigation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If GPS and IMU sensors are used for localization, then real-time location data is obtained, but localization accuracy deteriorates to exceed lane width errors
Solution Approach 1:
The patent introduces map features as an intermediary between the vehicle and GPS/IMU sensors. The system projects map features onto the camera image plane and compares them with detected image features to compute a pose metric. This intermediary comparison mechanism enables accurate localization by bridging the gap between coarse sensor data and precise vehicle position determination.
Solution Approach 2:
The patent replaces the mechanical sensor-based localization system (GPS/IMU) with an optical-mechanical system using camera images and map projections. Instead of relying solely on physical sensors that drift, the system uses visual feature matching between captured images and projected map features, substituting mechanical measurement with optical field comparison.
2Loss of information
If feature-matching techniques are used with map data, then possible vehicle locations are determined, but the ability to quantify feature error in terms of pose error is lost
Solution Approach 1:
The patent transforms the feature matching problem from 2D image space to 3D pose space by projecting map features onto the image plane using candidate pose parameters. This dimensional transformation enables the system to evaluate multiple pose hypotheses simultaneously and quantify pose errors by comparing projected features with detected features across different spatial dimensions.
Solution Approach 2:
The patent changes the evaluation parameter from simple feature presence/absence to a continuous pose metric that quantifies similarity between projected map features and detected image features. By computing a pose metric based on feature matching quality, the system converts discrete feature matches into continuous pose error measurements, enabling precise localization feedback.
Data Source
AI summary
A captured image and corresponding observed pose data are captured or received. The captured image is captured by an image capturing device onboard a vehicle. The observed pose data is determined by a location sensor onboard the vehicle. The observed pose data indicates the pose of the vehicle when the image was captured. A map projection is generated based on map data and the observed pose data. The captured image and the map projection are analyzed by a pose metric network. The pose metric network determines a pose metric based on the analysis of the captured image and the map projection. The pose metric describes how similar the captured image is to the map projection.