Visual odometry and pairwise alignment for high definition map creation

Patent metadata

Field	Value
Publication number	US-10598489-B2
Application number	US-201916427037-A
Country	US
Kind code	B2
Filing date	May 30, 2019
Priority date	Dec 30, 2016
Publication date	Mar 24, 2020
Grant date	Mar 24, 2020

Abstract

As an autonomous vehicle moves through a local area, pairwise alignment may be performed to calculate changes in the pose of the vehicle between different points in time. The vehicle comprises an imaging system configured to capture image frames depicting a portion of the surrounding area. Features are identified from the captured image frames, and a 3-D location is determined for each identified feature. The features of different image frames corresponding to different points in time are analyzed to determine a transformation in the pose of the vehicle during the time period between the image frames. The determined poses of the vehicle are used to generate an HD map of the local area.

First claim

What is claimed is:

1. A method, comprising:
receiving a plurality of image frames depicting a local area surrounding a vehicle, each image frame corresponding to a point in time, and comprising at least a first image of the local area captured by a first camera and a second image of the local area captured by a second camera;
for each image frame of the plurality of image frames:
identifying a first set of feature points on the first image of the image frame, and a second set of feature points on the second image of the image frame;
determining features of the image frame associated with features within the local area, each feature corresponding to a first feature point of the first set of feature points matched with a second feature point of the second set of feature points; and
for each determined feature, determining a location of the feature based upon the first and second feature points of the feature;

determining a transformation between a first position of the vehicle at a first point in time corresponding to a first image frame of the plurality of image frames, and a second position of the vehicle at a second point in time corresponding to a second image frame of the plurality of image frames, by:
selecting a first subset of features of the first image frame;
identifying a second subset of features of the second image frame corresponding to the first subset of features, based upon a level of similarity between features of the first and second subsets;
determining the transformation based upon a transformation between the determined positions of features of the first and second subsets of features; and

generating a high definition map of the local area based on the transformation, the high definition map for use in driving by one or more autonomous vehicles.

2. The method of claim 1 , further comprising rectifying the first image and the second image of the image frame such that the first image and the second image are coplanar, and wherein the first and second feature points are not matched to correspond to a feature in response to a determination that a location of the first feature point on the first image and a location of the second feature point on the second image do not correspond to the same row of the first and second images.

3. The method of claim 1 , further comprising discarding a feature in response to a determination that the determined location of the feature corresponds to a distance from the vehicle exceeding a threshold distance.

4. The method of claim 1 , wherein determining the transformation between the first and second positions further comprises:
for each of a plurality of iterations:
selecting a different first subset of features of the first image frame;
identifying the second subset of features of the second image frame corresponding to the selected first subset of features;
determining a candidate transformation between the determined positions of corresponding features of the first and second subsets of features;
for each feature of first image frame, matching the feature with a feature of the second image frame, based upon applying the determined candidate transformation to the feature to generate a transformed feature; and
determining a metric for the determined candidate transformation, based upon the features of the first image frame and their determined matching features of the second image frame; and

selecting a candidate transformation of the plurality of iterations having a highest metric.

5. The method of claim 4 , wherein determining the transformation between the first and second positions further comprises:
recalculating the transformation between the first and second positions, based upon the features of the first image frame and their determined matching features associated with the selected candidate transformation.

6. The method of claim 4 , wherein matching the feature of the first image frame with a feature of the second image frame comprises determining a feature of the second image frame associated with a descriptor with at least a threshold similarity to a descriptor of the feature of the first image frame, that is nearest to a location the transformed feature.

7. The method of claim 4 , wherein the metric for the candidate transformation is based upon a number of features of the first image frame having a matching feature of the second image frame within a threshold distance.

8. The method of claim 4 , wherein the metric for the candidate transformation is based upon a volume of space covering locations of features of the first image frame that have a matching the second image frame within a threshold distance.

9. The method of claim 4 , wherein determining the transformation between the first and second positions further comprises:
identifying an initial guess of the transformation; and
for each candidate transformation of the plurality of iterations, discarding the determined candidate transformation if the candidate transformation deviates from the initial guess by over a threshold amount.

10. The method of claim 1 , wherein each feature of the first subset of features is selected to be at least a threshold distance from the remaining features of the first subset of features.

11. The method of claim 1 , wherein the transformation is determined using a PnP (Perspective-n-Point) algorithm.

12. A computer-readable storage medium storing instructions that when executed causes one or more processors to:
receive a plurality of image frames depicting a local area of a vehicle, each image frame corresponding to a point in time, and comprising at least a first image of the local area captured by a first camera and a second image of the local area captured by a second camera;
for each image frame of the plurality of image frames:
identify a first set of feature points on the first image of the image frame, and a second set of feature points on the second image of the image frame;
determine features of the image frame associated with features within the local area, each feature corresponding to a first feature point of the first set of feature points matched with a second feature point of the second set of feature points; and
for each determined feature, determine a location of the feature based upon the first and second feature points of the feature;

determine a transformation between a first position of the vehicle at a first point in time corresponding to a first image frame of the plurality of image frames, and a second position of the vehicle at a second point in time corresponding to a second image frame of the plurality of image frames, by:
selecting a first subset of features of the first image frame;
identifying a second subset of features of the second image frame corresponding to the first subset of features, based upon a level of similarity between features of the first and second subsets;
determining the transformation based upon a transformation between the determined positions of features of the first and second subsets of features; and

generate a high definition map of the local area based on the transformation, the high definition map for use in driving by one or more autonomous vehicles.

13. The computer-readable storage medium of claim 12 , wherein the instructions further cause the one or more processors to rectify the first image and the second image of the image frame such that the first image and the second image are coplanar, and wherein the first and second feature points are not matched to correspond to a feature in response to a determination that a location of the first feature point on the first image and a location of the second feature poi

Assignees

• Deepmap Inc

Inventors

• Zhang Ronghua
• Chen Chen
• Zeng Di
• Wheeler Mark Damon

Classifications

• G06T2207/30256

  Lane; Road marking · CPC title

• B60W40/06

  Road conditions · CPC title

• G06T7/55

  from multiple images · CPC title

• G01S19/46

  the supplementary measurement being of a radio-wave signal type · CPC title

• G06T7/593

  from stereo images · CPC title