Vehicle Localization via 3D Geometric Map Matching

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing vehicle localization methods using camera images face challenges in creating a map invariant to visual changes over time, such as lighting and weather variations, leading to difficulties in extracting constant features and maintaining accurate localization.

Innovation Solution

The method involves classifying objects in a pre-existing 3D-map and pixels in 2D-images using a predetermined set of semantic classes, and then matching these classified objects and pixels to determine the vehicle's location, thereby reducing the need for uniquely identifiable objects and improving robustness against appearance changes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If image features are extracted and a map is built based on these features for vehicle localization, then localization precision can be achieved, but the map becomes sensitive to visual changes over time such as lighting and weather variations

Engineering Contradiction:
Improvelocalization precisionVSAvoidrobustness against temporal changes
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent changes the parameter used for map representation from appearance-based image features to geometry-based 3D structures. By transforming the map from a 2D image feature map to a 3D geometric map with classified objects, the system achieves invariance to lighting and weather conditions while maintaining localization precision through geometric correspondence matching.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If continuously updated image features are added to keep the map current, then the map remains updated with changing appearances, but the map size grows indefinitely and requires frequent revisits

Engineering Contradiction:
Improvemap currencyVSAvoidmap size and update frequency
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent discards the approach of continuously accumulating image features and instead recovers localization capability through geometric object classification. The 3D map stores stable geometric structures that do not require continuous updates, as the geometric properties of objects remain constant over time unlike their appearance.

Inventive Principle:
Principle #34Discarding and recovering

3Reliability

If place recognition is used to get a rough idea of location, then robustness against temporal changes is improved, but localization precision becomes too low for self-driving cars

Engineering Contradiction:
Improverobustness against temporal changesVSAvoidlocalization precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent transitions from 2D image-based place recognition to 3D geometric map matching. By adding the third dimension and using classified geometric objects in 3D space, the system achieves both the robustness of place recognition and the precision required for self-driving localization through volumetric correspondence matching.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP3343431B1Method and system for vehicle localization from camera image
Publication Date: 2025.04.30 VOLVO CAR CORP
  • EP3343431B1 patent drawingFigure 1
  • EP3343431B1 patent drawingFigure 2~3A
  • EP3343431B1 patent drawingFigure 3B

AI summary

There is provided a method for determining a location of a vehicle, the method comprising acquiring a 2D-image using a vehicle-based camera, the image depicting the vehicle surroundings; classifying pixels in the image such that each classified pixel belong to one class of set of classes, thereby forming a classified 2D-image. The classification is performed by a learning algorithm. The method further comprises determining an initial estimated vehicle location and defining possible vehicle locations based on the initial location. Next, the method comprises matching the classified 2D-image with a 3D-map comprising a plurality of geometric objects, each geometric object belonging to one class of the set of classes, by comparing a classification of a geometric object in the 3D-map, for a possible vehicle location, with a classification of a corresponding at least one pixel in the classified 2D-image; determining a matching score for at least one possible vehicle location based on matched pixels of the 2D-image; and determining a vehicle location based on the score of the possible vehicle location.