System and method for large-scale lane marking detection using multimodal sensor data

1. A system comprising:

a data processor; and

a multimodal lane detection module, executable by the data processor, the multimodal lane detection module being configured to perform a multimodal lane detection operation configured to:

fuse image data and point cloud data to produce a set of lane marking points in three-dimensional (3D) space that correlate to the image data and the point cloud data based on a threshold, the threshold being based on a perspective depth of the lane marking points, the threshold being larger for lane marking points close to a position of an image generating device that produced the image data and smaller for lane marking points more distant from the position of the image generating device; and

generate a lane marking map from the set of lane marking points.

2. The system of claim 1 , wherein the multimodal lane detection module is further configured to receive vehicle metrics via a global positioning system or an inertial measurement unit to determine at least one of a location, an orientation, or a speed of a vehicle.

3. The system of claim 1 , wherein a neural network is used for identifying and labeling objects in the image data with object category labels on a per-pixel basis.

4. The system of claim 1 , wherein the image generating device includes an image camera or a motion video camera.

5. The system of claim 1 , wherein the point cloud data is produced by a laser range finder.

6. The system of claim 1 being configured to receive vehicle metrics related to an environment or a condition of a vehicle from a vehicle subsystem.

7. The system of claim 1 , wherein the fusion of the image data and point cloud data includes aligning and orienting the image data with a terrain map corresponding to a location and using a terrain map elevation data to transform the image data to the 3D space, wherein the location is a geographical location where a vehicle is located.

8. The system of claim 1 , wherein lane markings formed from the set of lane marking points are produced from each frame of the image data and the corresponding point cloud data.

9. A method comprising:

fusing image data and point cloud data to produce a set of lane marking points in three-dimensional (3D) space that correlate to the image data and the point cloud data based on a threshold, the threshold being based on a perspective depth of the lane marking points, the threshold being larger for lane marking points close to a position of an image generating device that produced the image data and smaller for lane marking points more distant from the position of the image generating device; and

generating a lane marking map from the set of lane marking points.

10. The method of claim 9 including receiving vehicle metrics via a global positioning system or an inertial measurement unit to determine a location or a speed of a vehicle.

11. The method of claim 9 including tracking lane markings formed from the set of lane marking points across a plurality of frames of the image data.

12. The method of claim 11 wherein a smoothing technique is used to fit smooth new curves for each lane marking across the plurality of frames.

13. The method of claim 9 wherein the fusing of the image data and point cloud data includes projecting 3D point cloud data on to two-dimensional (2D) image data, and adding a 3D point cloud point to the set of lane marking points if a distance between a position of the projected 3D point cloud point in 2D space and a position of at least one of the set of lane marking points is within a pre-determined threshold.

14. The method of claim 9 including receiving vehicle metrics via a Global Positioning System (GPS), an inertial measurement unit (IMU), or a radar, to determine at least one of a location, an orientation, or a speed of a vehicle.

15. The method of claim 9 including registering the point cloud data within a time range to a common coordinate space.

16. The method of claim 15 including generating an accumulated point cloud representing a collection of the point cloud data over time, wherein the point cloud data are aligned.

17. A non-transitory machine-useable storage medium embodying instructions which, when executed by a machine, cause the machine to:

fuse image data and point cloud data to produce a set of lane marking points in three-dimensional (3D) space that correlate to the image data and the point cloud data based on a threshold, the threshold being based on a perspective depth of the lane marking points, the threshold being larger for lane marking points close to a position of an image generating device that produced the image data and smaller for lane marking points more distant from the position of the image generating device; and

generate a lane marking map from the set of lane marking points.

18. The non-transitory machine-useable storage medium of claim 17 wherein a neural network is used for identifying and labeling objects in the image data with object category labels.

19. The non-transitory machine-useable storage medium of claim 18 wherein the neural network is trained by using training images, wherein the training images include contexts of at least one of environments, locations, weather conditions, and lighting conditions.

20. The non-transitory machine-useable storage medium of claim 19 wherein the training images include an object labeling created manually or by automated processes.