High-definition maps and localization for road vehicles

1. A method comprising:

providing aerially captured images of a geographical area to a machine learning model trained to recognize features of objects depicted in the images;

obtaining, using the machine learning model, feature matrices that respectively indicate the features;

determining three-dimensional geometric representations of the features based at least on the feature matrices as obtained using the machine learning model, the three-dimensional geometric representations including one or more of: one or more points, one or more polylines, or one or more polygons that represent the features;

annotating the images with the three-dimensional geometric representations;

generating map data corresponding to the geographical area based at least on the images as annotated; and

providing the map data to a vehicle such that the map data is used by the vehicle in performance of one or more driving operations.

2. The method of claim 1 , wherein the three-dimensional geometric representations include global positioning coordinates corresponding to the features.

3. The method of claim 1 , further comprising updating one or more parameters of the machine learning model using a previously annotated image, wherein training the machine learning model includes:

projecting an annotation of the previously annotated image onto a camera plane corresponding to a camera used to capture the previously annotated image;

obtaining, using the annotation as projected, a two-dimensional projection of a particular three-dimensional geometric representation of a feature that is included in the annotation;

transforming the two-dimensional projection into a particular feature matrix; and

training the machine learning model using the particular feature matrix.

4. The method of claim 3 , wherein the projecting the annotation onto the camera plane is based at least on a pose of the camera at a time that the camera was used to capture the previously annotated image.

5. The method of claim 1 , wherein the map data includes at least one of: lane graph map data or localization prior map data.

6. The method of claim 1 , wherein the generating the map data includes:

defining geometries of one or more lane segments from the three-dimensional geometric representations of the features; and

connecting the one or more lane segments according to their geometries to define one or more traffic lanes.

7. The method of claim 6 , further comprising:

defining geometries of one or more traffic lights from the three-dimensional geometric representations of the features; and

assigning the one or more traffic lights to the one or more traffic lanes.

8. The method of claim 1 , further comprising stitching two or more images of the images together based at least on one or more poses of one or more respective cameras used to capture the images, wherein the generating the map data is performed based at least on the stitched together images.

9. The method of claim 1 , wherein the generating the map data includes:

discretizing the three-dimensional geometric representations of the features into aligned point clouds;

assigning a normalized vector to each point of the aligned point clouds, the normalized vector indicating a direction of a polyline from which the point was discretized; and

indexing the aligned point clouds and the normalized vectors into an encoded form that allows for use in navigation within the geographical area.

10. A system comprising:

one or more processors to perform operations, the operations comprising:

obtaining, using a machine learning model, feature classifications that correspond to features of objects depicted in images of a geographical area in which the images are provided to the machine learning model, one or more of the feature classifications being based at least on one or more epipolar geometrical relations corresponding to two or more of the images;

annotating the images with three-dimensional geometric representations that are based at least on the feature classifications as obtained; and

generating map data corresponding to the geographical area based at least on the images as annotated.

11. The system of claim 10 , wherein the three-dimensional geometric representations include one or more of: one or more points, one or more polylines, or one or more polygons that represent the features.

12. The system of claim 10 , wherein the feature classifications include feature matrices that indicate the features.

13. The system of claim 10 , the operations further comprising updating one or more parameters of the machine learning model using a previously annotated image, wherein training the machine learning model includes:

projecting an annotation of the previously annotated image onto a camera plane corresponding to a camera used to capture the previously annotated image;

obtaining, using the annotation as projected, a two-dimensional projection of a particular three-dimensional geometric representation of a feature that is included in the annotation; and

training the machine learning model based at least on the two-dimensional projection.

14. The system of claim 10 , wherein the map data includes at least one of: lane graph map data or localization prior map data.

15. The system of claim 10 , wherein the generating the map data includes:

defining geometries of one or more lane segments from the three-dimensional geometric representations of the features;

connecting the one or more lane segments according to their geometries to define one or more traffic lanes;

defining geometries of one or more traffic lights from the three-dimensional geometric representations of the features; and

assigning the one or more traffic lights to the one or more traffic lanes.

16. The system of claim 10 , the operations further comprising stitching two or more images of the images together based at least on one or more poses of one or more respective cameras used to capture the images, wherein the generating the map data is based at least on the stitched together images.

17. The system of claim 10 , wherein the generating the map data includes:

discretizing the three-dimensional geometric representations of the features into aligned point clouds;

assigning a normalized vector to each point of the aligned point clouds, the normalized vector indicating a direction of a polyline from which the point was discretized; and

indexing the aligned point clouds and the normalized vectors into an encoded form that allows for use in navigation within the geographical area.

18. The system of claim 10 , wherein the images include aerially captured images.

19. A system comprising:

one or more processors to perform operations comprising:

obtaining, using a machine learning model, feature classifications that correspond to features of objects depicted in images of a geographical area in which the images are provided to the machine learning model;

annotating the images with three-dimensional geometric representations that are based at least on the feature classifications as obtained, the three-dimensional geometric representations including one or more of: one or more points, one or more polylines, or one or more polygons that represent the features;

generating nna p data corresponding to the geographical area based at least on the images as annotated; and

providing the map data to a vehicle such that the map data is used by the vehicle in performance of one or more driving operations.

20. The system of claim 19 , wherein one or more of the feature classifications is based at least on one or more epipolar geometrical relations corresponding to two or more of the images.