Lane line creation for high definition maps for autonomous vehicles
An HD map system represents landmarks on a high definition map for autonomous vehicle navigation, including describing spatial location of lanes of a road and semantic information about each lane, and along with traffic signs and landmarks. The system generates lane lines designating lanes of roads based on, for example, mapping of camera image pixels with high probability of being on lane lines into a three-dimensional space, and locating/connecting center lines of the lane lines. The system builds a large connected network of lane elements and their connections as a lane element graph. The system also represents traffic signs based on camera images and detection and ranging sensor depth maps. These landmarks are used in building a high definition map that allows autonomous vehicles to safely navigate through their environments.
1. A method for generating lane lines in a high definition map comprising: receiving image data captured by sensors of autonomous vehicles driving on a road, the image data representing a portion of the road;
classifying pixels of the image data with respect to whether the pixels correspond to one or more lane line segments of the portion of the road;
identifying a set of points that lie on at least one of the one or more lane line segments based on the classification of the pixels;
grouping the set of points into lane line clusters based on proximity of the points of the set of points with respect to each other, wherein each lane line cluster is determined to be associated with a respective lane line segment;
locating a corresponding center line for each lane line cluster;
determining a lane line by connecting a plurality of center lines of respective lane line clusters, the connection associating endpoints of at least two center lines and connecting respective lane line segments that correspond to the connected lane line clusters; and
generating a high definition map based on the determined lane lines, the high definition map for use in driving of autonomous vehicles.
2. The method of claim 1 , wherein classifying the pixels is based on a deep learning based classification model.
3. The method of claim 1 , wherein identifying the set of points comprises computing a likelihood that a respective point of the set of points is in the middle of a respective lane line segment, the probability determined based on the relative placement of the respective point to the respective lane line segment.
4. The method of claim 1 , wherein identifying the set of points comprises mapping between a three-dimensional representation of the portion of the road and a two-dimensional representation of the portion of the road.
5. The method of claim 4 , wherein each identified two-dimensional pixel is mapped to a corresponding 3D voxel, wherein the corresponding 3D voxel is a representation of a corresponding point in a three-dimensional space.
6. The method of claim 5 , wherein mapping between the 2D pixels and 3D voxels comprises iterating through a plurality of 3D voxels to identify a corresponding 2D pixel.
7. The method of claim 4 , wherein mapping between a three-dimensional representation and a two dimensional representation further comprises:
identifying a three-dimensional point;
converting the three-dimensional point to camera coordinates;
projecting the converted three-dimensional point to an image captured by a camera mounted on a vehicle, wherein the converted point is mapped to a pixel; and
associating the pixel with the three-dimensional point based on a probability that the pixel lies on a lane line.
8. The method of claim 4 , wherein mapping between a three-dimensional representation and a two dimensional representation further comprises:
computing a probability that a three-dimensional voxel is on the lane line segment, based on the probability that the corresponding identified pixel is on the lane line segment.
9. The method of claim 1 , wherein grouping the set of points into the lane line clusters comprises forming a group of points that are located within a threshold distance of each other, wherein the threshold distance is greater than a lane line width and significantly less than a distance between two lane lines.
10. The method of claim 1 , wherein grouping the set of points into the lane line clusters comprises identifying skeleton points within each group of points, skeleton points referring to points at the center of a corresponding lane line segment.
11. The method of claim 10 , wherein identifying skeleton points further comprises:
identifying one or more subclusters that cover in entirety the group of points; and
locating a geometric center of each of the one or more subclusters.
12. The method of claim 1 , wherein grouping the set of points into the lane line clusters further comprises:
distinguishing between multiple different lane lines within a specified proximity of each other;
computing a local slope of each skeleton point using neighboring skeleton points within the same lane line segment; and
grouping skeleton points into the lane line clusters, the grouping of skeleton points performed with consideration as to distance between the skeleton points and degree of difference between the local slopes.
13. The method of claim 1 , wherein locating a corresponding center line for a particular lane line cluster comprises:
generating a polyline representation of the particular lane line cluster, the polyline representing a connected sequence of skeleton points forming a single lane line cluster;
identifying outlier points, if any, from the polyline representation, outlier points representing points that falsely suggest a change in the direction of the particular lane line cluster;
removing the identified outlier points; and
removing skeleton points between the endpoints of the particular lane line cluster, the removed points having no effect on changes in the direction of the particular lane cluster.
14. The method of claim 1 , wherein connecting two or more center lines for lane line clusters into a complete lane line comprises:
identifying neighboring endpoints of additional lane line segments within a specified distance of either endpoint of the identified lane line segment;
for each identified neighboring point, computing a connectivity score, the connectivity score relating to the closest endpoint of the identified lane line segment;
ranking all connectivity scores;
for each connection with a connectivity score, confirming that both endpoints are available to connect;
confirming that the connection is free of intersections with any existing lane line connections or existing lane lines;
selecting the connection with both confirmed conditions and the highest connectivity score, as indicated by the ranking of the connectivity scores; and
recording the connection between the at least two lane line segments.
15. The method of claim 14 , wherein the connectivity score is based on one or more of:
the distance between the identified lane line segment and the lane line segment containing the neighboring point, or
the degree of change in direction between the identified lane line segment and the lane line segment containing the neighboring point.
16. The method of claim 14 , wherein the connection is removed from the ranked list of connectivity scores responsive to one or more of:
a lack of confirmation that both endpoints are available to connect, or
a lack of confirmation that the connection is free of intersections with any existing lane line connections.
17. The method of claim 1 , wherein classifying the pixels is based on a machine learning model applied to an image generated from intensity values obtained from a set of 3D voxels, wherein each of the set of 3D voxels is obtained from a point cloud representation of a region and represents the lowest 3D voxel in a column of 3D voxels.
18. The method of claim 1 , wherein classifying the pixels comprises:
receiving an occupancy map for a portion of a geographical region, the occupancy map comprising a plurality of 3D voxels, wherein each voxel is associated with an intensity value and a color value, the plurality of 3D voxels organized as one or more columns;
for each column from the one or more columns of 3D voxels, selecting the lowest 3D voxel in the column;
creating images based on the lowest 3D voxels of the one or more columns of the 3D voxels, the images comprising: a first image storing the intensity values of the lowest 3D voxels, and a second image storing color values of the lowest 3D voxels; and
determining whether each pixel of the created images belongs to a lane line or does not belong to a lane line based on a machine learning model.
19. One or more non-transitory computer readable storage media having instructions encoded thereon that, in response to being executed by one or more processors, cause a system to perform operations, the operations comprising:
receiving image data captured by sensors of autonomous vehicles driving on a road, the image data representing a portion of the road;
classifying pixels of the image data with respect to whether the pixels correspond to one or more lane line segments of the portion of the road;
identifying a set of points that lie on at least one of the one or more lane line segments based on the classification of the pixels;
grouping the set of points into lane line clusters based on proximity of the points of the set of points with respect to each other, wherein each lane line cluster is determined to be associated with a respective lane line segment;
locating a corresponding center line for each lane line cluster;
determining a lane line by connecting a plurality of center lines of respective lane line clusters, the connection associating endpoints of at least two center lines and connecting respective lane line segments that correspond to the connected lane line clusters; and
generating a high definition map based on the determined lane lines, the high definition map for use in driving of autonomous vehicles.
20. A computer system comprising:
one or more processors; and
one or more non-transitory computer readable storage media having instructions encoded thereon that, in response to being executed by the one or more processors, cause the system to perform operations, the operations comprising:
receiving image data captured by sensors of autonomous vehicles driving on a road, the image data representing a portion of the road;
classifying pixels of the image data with respect to whether the pixels correspond to one or more lane line segments of the portion of the road;
identifying a set of points that lie on at least one of the one or more lane line segments based on the classification of the pixels;
grouping the set of points into lane line clusters based on proximity of the points of the set of points with respect to each other, wherein each lane line cluster is determined to be associated with a respective lane line segment;
locating a corresponding center line for each lane line cluster;
determining a lane line by connecting a plurality of center lines of respective lane line clusters, the connection associating endpoints of at least two center lines and connecting respective lane line segments that correspond to the connected lane line clusters; and
generating a high definition map based on the determined lane lines, the high definition map for use in driving of autonomous vehicles.