Map partition system for autonomous vehicles

1. A computer-implemented method to partition a point cloud map for an autonomous driving vehicle (ADV), the method comprising:

identifying a road to be navigated by the ADV, the road represented by one or more point clouds captured by one or more LIDAR sensors;

extracting road marking information of the identified road from the one or more point clouds, the road marking information describing one or more road markings of the identified road;

partitioning the road into one or more road partitions based on the one or more road markings; and

generating the point cloud map based on the road partitions, wherein the point cloud map describes a driving environment surrounding the ADV using at least a portion of the one or more point clouds synchronized with one or more images captured by the ADV and comprises a static point cloud map generated by removing one or more dynamic objects in the point cloud map, wherein the one or more dynamic objects correspond to a number of voxels appear or disappear between two consecutive frames of point clouds captured by the one or more LIDAR sensors.

2. The method of claim 1 , wherein the one or more dynamic objects are removed by gathering two or more point cloud observations at different points in time for a same location and detecting a difference between the two or more point cloud observations.

3. The method of claim 1 , wherein identifying the road comprises:

applying an elevation threshold to the point clouds to extract a plurality of likely ground points;

applying a fitting algorithm to fit one or more planar objects to the likely ground points as a road object; and

selecting a planar object with the largest surface area having a normal direction approximating a direction of a vertical axis of the LIDAR sensors as the road object.

4. The method of claim 3 , further comprising identifying point clouds corresponding to the road object as road surface points.

5. The method of claim 3 , wherein extracting the road markings comprises:

applying a filter to remove noise for road surface points;

determining a directional gradient for road surface point clouds; and

applying a k-means clustering algorithm to the road surface point clouds based on a threshold for the directional gradient to divide the road surface point clouds into road markings points and other road surface points.

6. The method of claim 3 , wherein the road markings include markings for at least a pedestrian crossing and wherein the road is partitioned based on the markings for the pedestrian crossing.

7. The method of claim 1 , further comprising:

extracting curbs features from the point clouds representing road curbs;

identifying curb boundaries for the road based on the curbs features;

labeling points of the point clouds corresponding to the curb boundaries; and

partitioning the road into the one or more road partitions based on the road markings and the labeled road curbs points.

8. The method of claim 7 , wherein extracting curbs features for the point clouds comprising, for each point in the point clouds, extracting one or more multi-scale directional difference features to identify the road curbs.

9. The method of claim 8 , wherein identifying curbs boundaries for the road based on the curbs features comprises:

for each point in the point clouds, applying a machine learning algorithm to the extracted features to identify the point as an initial curb point; and

for each initial curb point,

determining one or more directional characteristics of the initial curb point,

generating a point cluster based on the initial curb point,

determining one or more characteristics for the point cluster, and

identifying the point cluster as a portion of curb boundaries if the one or more characteristics satisfy a predetermined condition.

10. The method of claim 9 , further comprising generating a curve to represent each of the point cluster, wherein the curve comprises a Cubic Bezier curve, and wherein the road is partitioned into the one or more road partitions based on the road markings and the generated Cubic Bezier curve.

11. A non-transitory machine-readable medium having instructions stored therein, which when executed by a processor, cause the processor to perform operations, the operations comprising:

identifying a road to be navigated by an autonomous driving vehicle (ADV), the road represented by one or more point clouds captured by one or more LIDAR sensors;

extracting road marking information of the identified road from the one or more point clouds, the road marking information describing one or more road markings of the identified road;

partitioning the road into one or more road partitions based on the one or more road markings; and

generating the point cloud map based on the road partitions, wherein the point cloud map describes a driving environment surrounding the ADV using at least a portion of the one or more point clouds synchronized with one or more images captured by the ADV and comprises a static point cloud map generated by removing one or more dynamic objects in the point cloud map, wherein the one or more dynamic objects correspond to a number of voxels appear or disappear between two consecutive frames of point clouds captured by the one or more LIDAR sensors.

12. The non-transitory machine-readable medium of claim 11 , wherein identifying the road comprises:

applying an elevation threshold to the point clouds to extract a plurality of likely ground points;

applying a fitting algorithm to fit one or more planar objects to the likely ground points as a road object; and

selecting a planar object with the largest surface area having a normal direction approximating a direction of a vertical axis of the LIDAR sensors as the road object.

13. The non-transitory machine-readable medium of claim 12 , wherein extracting the road markings comprises:

applying a filter to remove noise for road surface points;

determining a directional gradient for road surface point clouds; and

applying a k-means clustering algorithm to the road surface point clouds based on a threshold for the directional gradient to divide the road surface point clouds into road markings points and other road surface points.

14. The non-transitory machine-readable medium of claim 11 , further comprising:

extracting curbs features from the point clouds representing road curbs;

identifying curb boundaries for the road based on the curbs features;

labeling points of the point clouds corresponding to the curb boundaries; and

partitioning the road into the one or more road partitions based on the road markings and the labeled road curbs points.

15. The non-transitory machine-readable medium of claim 14 , wherein extracting curbs features for the point clouds comprising, for each point in the point clouds, extracting one or more multi-scale directional difference features to identify the road curbs.

16. A data processing system, comprising:

a processor; and

a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations, the operations including:

identifying a road to be navigated by an autonomous driving vehicle (ADV), the road represented by one or more point clouds captured by one or more LIDAR sensors;

extracting road marking information of the identified road from the one or more point clouds, the road marking information describing one or more road markings of the identified road;

partitioning the road into one or more road partitions based on the one or more road markings; and

generating a point cloud map based on the road partitions, wherein the point cloud map describes a driving environment surrounding the ADV using at least a portion of the one or more point clouds synchronized with one or more images captured by the ADV and comprises a static point cloud map generated by removing one or more dynamic objects in the point cloud map, wherein the one or more dynamic objects correspond to a number of voxels appear or disappear between two consecutive frames of point clouds captured by the one or more LIDAR sensors.

17. The system of claim 16 , wherein identifying the road comprises:

applying an elevation threshold to the point clouds to extract a plurality of likely ground points;

applying a fitting algorithm to fit one or more planar objects to the likely ground points as a road object; and

selecting a planar object with the largest surface area having a normal direction approximating a direction of a vertical axis of the LIDAR sensors as the road object.

18. The system of claim 17 , wherein extracting the road markings comprises:

applying a filter to remove noise for road surface points;

determining a directional gradient for road surface point clouds; and

applying a k-means clustering algorithm to the road surface point clouds based on a threshold for the directional gradient to divide the road surface point clouds into road markings points and other road surface points.

19. The system of claim 16 , wherein the operations further comprise:

extracting curbs features from the point clouds representing road curbs;

identifying curb boundaries for the road based on the curbs features;

labeling points of the point clouds corresponding to the curb boundaries; and

partitioning the road into the one or more road partitions based on the road markings and the labeled road curbs points.

20. The system of claim 19 , wherein extracting curbs features for the point clouds comprising, for each point in the point clouds, extracting one or more multi-scale directional difference features to identify the road curbs.