Structured prediction crosswalk generation

1. A computing system comprising:

one or more processors programmed to perform operations comprising:

obtaining image data, the image data comprising: a plurality of image elements showing a roadway that includes a crosswalk; centerline data indicating a position of a centerline of the roadway; and an intersection polygon indicating a position of an intersection on the roadway;

generating a segmentation feature map using the image data, the segmentation feature map comprising a first portion of the plurality of image elements within a first boundary and a second portion of the plurality of image elements within a second boundary;

generating a classification feature map using the image data, the classification feature map describing a first distance between a first element of the plurality of image elements and the first boundary and a second distance between the first element of the plurality of image elements and the second boundary;

determining a position for the first boundary and a position for the second boundary using the segmentation feature map, the classification feature map, the centerline on the roadway, and the intersection polygon; and

providing map data associated with a map of the roadway, wherein the map data includes a position of the crosswalk on the roadway in the map, the position of the crosswalk indicated at least in part by the position for the first boundary and the position for the second boundary.

2. The computing system of claim 1 , wherein determining of the position for the first boundary and the position for the second boundary comprises applying a weight to at least one of the segmentation feature map or the classification feature map.

3. The computing system of claim 1 , the operations further comprising determining a first angle relating the first element of plurality of image elements to the centerline of the roadway, the determining of the position for the first boundary and the position of the second boundary based at least in part on the first angle.

4. The computing system of claim 3 , the operations further comprising:

determining, using the first angle and the image data, a first likelihood that the first element of the plurality of image elements is either on the first boundary or on the second boundary;

determining, using the first angle and the image data, a second likelihood that the first element of the plurality of image elements is between the first boundary and the second boundary; and

determining, using the first likelihood and the second likelihood, a first element value indicating whether the first element is part of the crosswalk, the determining of the position for the first boundary and the position for the second boundary based at least in part on the first element value.

5. The computing system of claim 3 , the determining of the first angle further comprising determining a dilated representation of a boundary angle including an x-value and a y-value with respect to the centerline of the roadway.

6. The computing system of claim 1 , the operations further comprising:

training, by the computing system, a computerized model to optimize a loss function of the computerized model, the loss function comprising a sum of a segmentation loss, an alignment loss, and a boundary loss; and

generating, by the computing system and using the image data, a plurality of images, each of the plurality of images being associated with at least one of a plurality of characteristics of the image data, the determining of the position for the first boundary and the position for the second boundary being based at least in part on the plurality of images.

7. The computing system of claim 6 , the segmentation loss being based at least in part on a comparison between a first probability that the first element depicts a crosswalk and a ground truth probability that any given element of the plurality of image elements depicts a crosswalk.

8. The computing system of claim 6 , the operations further comprising determining a first angle relating the first element of plurality of image elements to the centerline of the roadway, the alignment loss being based on comparing the first angle to a ground truth angle.

9. The computing system of claim 6 , the boundary loss being based on an inverse distance transform of the classification feature map.

10. A computer-implemented method for generating a map of a geographic area indicating crosswalks in the geographic area, the method comprising:

obtaining, by a computing system, image data by a computing system comprising one or more processors, the image data comprising: a plurality of image elements showing a roadway that includes a crosswalk; centerline data indicating a position of a centerline of the roadway; and an intersection polygon indicating a position of an intersection on the roadway;

generating, by the computing system, a segmentation feature map using the image data, the segmentation feature map comprising a first portion of the plurality of image elements within a first boundary and a second portion of the plurality of image elements within a second boundary;

generating, by the computing system, a classification feature map using the image data, the classification feature map having a first distance between a first element of the plurality of image elements and the first boundary and a second distance between the first element of the plurality of image elements and the second boundary;

determining, by the computing system, an position for the first boundary and an position for the second boundary using the segmentation feature map, the classification feature map, the centerline on the roadway, and the intersection polygon; and

providing, by the computing system, map data associated with a map of the roadway, wherein the map data includes the position of the crosswalk on the roadway in the map, the position of the crosswalk indicated at least in part by the position for the first boundary and the position for the second boundary.

11. The method of claim 10 , the determining of the position for the first boundary and the position for the second boundary comprising balancing the segmentation feature map and the classification feature map in relation to the centerline on the roadway and the intersection polygon.

12. The computer-implemented method of claim 10 , further comprising determining a first angle relating the first element of plurality of image elements to the centerline of the roadway, the determining of the position for the first boundary and the position of the second boundary based at least in part on the first angle.

13. The computer-implemented method of claim 12 , further comprising:

determining, using the first angle and the image data, a first likelihood that the first element of the plurality of image elements is either on the first boundary or on the second boundary;

determining, using the first angle and the image data, a second likelihood that the first element of the plurality of image elements is between the first boundary and the second boundary; and

determining, using the first likelihood and the second likelihood, a first element value indicating whether the first element is part of the crosswalk, the determining of the position for the first boundary and the position for the second boundary based at least in part on the first element value.

14. The computer-implemented method of claim 12 , the determining of the first angle further comprising determining a dilated representation of a boundary angle including an x-value and a y-value with respect to the centerline of the roadway.

15. The computer-implemented method of claim 10 , further comprising:

training, by the computing system, a computerized model to optimize a loss function of the computerized model, the loss function comprising a sum of a segmentation loss, an alignment loss, and a boundary loss; and

generating, by the computing system and using the image data, a plurality of images, each of the plurality of images being associated with at least one of a plurality of characteristics of the image data, the determining of the position for the first boundary and the position for the second boundary being based at least in part on the plurality of images.

16. The computer-implemented method of claim 15 , the segmentation loss being based at least in part on a comparison between a first probability that the first element depicts a crosswalk and a ground truth probability that any given element of the plurality of image elements depicts a crosswalk.

17. The computer-implemented method of claim 15 , further comprising determining a first angle relating the first element of plurality of image elements to the centerline of the roadway, the alignment loss being based on comparing the first angle to a ground truth angle.

18. The computer-implemented method of claim 15 , the boundary loss being based on an inverse distance transform of the classification feature map.

19. At least one non-transitory computer-readable medium including program instructions that, when executed by at least one processor, cause the at least one processor to execute operations comprising:

obtaining image data by a computing system comprising one or more processors, the image data comprising: a plurality of image elements showing a roadway that includes a crosswalk; centerline data indicating a position of a centerline of the roadway; and an intersection polygon indicating a position of an intersection on the roadway;

generating a segmentation feature map using the image data, the segmentation feature map comprising a first portion of the plurality of image elements within a first boundary and a second portion of the plurality of image elements within a second boundary;

generating a classification feature map using the image data, the classification feature map having a first distance between a first element of the plurality of image elements and the first boundary and a second distance between the first element of the plurality of image elements and the second boundary;

determining an position for the first boundary and an position for the second boundary using the segmentation feature map, the classification feature map, the centerline on the roadway, and the intersection polygon; and

providing map data associated with a map of the roadway, wherein the map data includes the position of the crosswalk on the roadway in the map, the position of the crosswalk indicated at least in part by the position for the first boundary and the position for the second boundary.

20. The at least one non-transitory computer-readable medium of claim 19 , the determining of the position for the first boundary and the position for the second boundary comprising balancing the segmentation feature map and the classification feature map in relation to the centerline on the roadway and the intersection polygon.