Multi-modal segmentation network for enhanced semantic labeling in mapping

Patent metadata

Field	Value
Publication number	US-11527085-B1
Application number	US-202117553429-A
Country	US
Kind code	B1
Filing date	Dec 16, 2021
Priority date	Dec 16, 2021
Publication date	Dec 13, 2022
Grant date	Dec 13, 2022

Abstract

Provided are methods for enhanced semantic labeling in mapping with a semantic labeling system, which can include receiving, from a LiDAR sensor of a vehicle, LiDAR point cloud information including at least one raw point feature for a point, receiving, from a camera of the vehicle, image data associated with an image captured using the camera, generating at least one rich point feature for the point based on the image data, predicting, using a LiDAR segmentation neural network and based on the at least one raw point feature and the at least one rich point feature, a point-level semantic label for the point, and providing the point-level semantic label to a mapping engine to generate a map based on the point-level semantic label Systems and computer program products are also provided.

First claim

What is claimed is:

1. A vehicle comprising:
a camera configured to capture an image of an object proximate to the vehicle;
a LiDAR sensor configured to detect light reflected from the object proximate to the vehicle and generate LiDAR point cloud information based on the detected light, the LiDAR point cloud information comprising at least one raw point feature for a point;
at least one processor communicatively coupled to the camera and the LiDAR sensor; and
at least one memory storing instructions thereon that, when executed by the at least one processor, result in operations comprising:
receiving, from the LiDAR sensor, the at least one raw point feature for the point;
receiving, from the camera, image data associated with the image captured using the camera;
generating at least one rich point feature for the point based on the image data, the at least one rich point feature including a vector having vector values corresponding to a prediction score, the prediction score generated based on an application of a pixel-wise segmentation label to an enhanced pixel, the enhanced pixel generated by projecting the LiDAR point cloud information onto a pixel of the image data;
predicting, using a LiDAR segmentation neural network and based on the at least one raw point feature and the at least one rich point feature, a point-level semantic label for the point; and
providing the point-level semantic label to a mapping engine to generate a map based on the point-level semantic label.

2. The vehicle of claim 1 , wherein the at least one raw point feature comprises a vector having vector values corresponding to at least one of spatial information associated with the point, intensity information associated with the point, and depth information associated with the point.

3. The vehicle of claim 1 , wherein the pixel-wise segmentation label is predicted by providing the image data to an image segmentation neural network to cause the image segmentation neural network to generate the pixel-wise segmentation label.

4. The vehicle of claim 1 , wherein the prediction score represents a likelihood that the pixel-wise segmentation label corresponds to the point.

5. The vehicle of claim 4 , wherein the prediction score comprises a plurality of prediction scores; wherein the pixel-wise segmentation label comprises a plurality of pixel-wise segmentation labels; and wherein each prediction score of the plurality of prediction scores represents a likelihood that an associated pixel-wise segmentation label of the plurality of pixel-wise segmentation labels corresponds to the point.

6. The vehicle of claim 1 , wherein the at least one rich point feature is generated based on applying at least one post-processing technique to reduce re-projection error from the camera.

7. The vehicle of claim 1 , wherein the instructions that cause the at least one processor to generate the map cause the at least one processor to at least of: remove an object from a previous map, detect a landmark, compare semantic consistency between the map and the previous map, and annotate the map.

8. The vehicle of claim 1 , wherein the map comprises a LiDAR point cloud of the LiDAR point cloud information, and at least one point-level semantic label that is associated with at least one point of the LiDAR point cloud, the at least one point-level semantic label comprising the predicted point-level semantic label.

9. The vehicle of claim 1 , wherein the instructions cause the at least one rich point feature to be generated based on a first neural network; and wherein the LiDAR segmentation neural network is different from the first neural network.

10. The vehicle of claim 1 , wherein the operations further comprise:
receiving, by an image segmentation neural network and from the camera, the image data; and
predicting, based on the image data, a pixel-wise segmentation label.

11. The vehicle of claim 10 , wherein the operations further comprise:
projecting the LiDAR point cloud information onto a pixel of the image data to generate the enhanced pixel; and
applying the pixel-wise segmentation label from the image segmentation neural network to the enhanced pixel.

12. The vehicle of claim 11 , wherein the operations further comprise:
transmitting, to the LiDAR segmentation neural network, a vector having vector values corresponding to the enhanced pixel and the applied pixel-wise segmentation label.

13. The vehicle of claim 12 , wherein the vector values comprise a prediction score for the pixel-wise segmentation label applied to the enhanced pixel, the prediction score indicating a likelihood that the pixel-wise segmentation label corresponds to the point.

14. The vehicle of claim 11 , wherein the operations further comprise: applying, to the enhanced pixel, at least one post-processing technique configured to reduce a re-projection error from the camera.

15. A system, comprising:
at least one processor; and
at least one memory storing instructions thereon that, when executed by the at least one processor, result in operations comprising:
receiving, from a LiDAR sensor of a vehicle, LiDAR point cloud information comprising at least one coordinate of a point;
receiving, from a camera of the vehicle, image data associated with an image captured using the camera;
generating at least one rich point feature for the point based on the image data, the at least one rich point feature including a vector having vector values corresponding to a prediction score, the prediction score generated based on an application of a pixel-wise segmentation label to an enhanced pixel, the enhanced pixel generated by projecting the LiDAR point cloud information onto a pixel of the image data;
predicting, using a LiDAR segmentation neural network and based on the at least one coordinate and the at least one rich point feature, a point-level semantic label for the point; and
providing the point-level semantic label to a mapping engine to generate a map based on the point-level semantic label.

16. The system of claim 15 , wherein the at least one coordinate comprises a vector having vector values corresponding to at least one of spatial information associated with the point, intensity information associated with the point, and depth information associated with the point.

17. The system of claim 15 , wherein the pixel-wise segmentation label is predicted by providing the image data to an image segmentation neural network to cause the image segmentation neural network to generate the pixel-wise segmentation label.

18. The system of claim 15 , wherein the prediction score represents a likelihood that the pixel-wise segmentation label corresponds to the point.

19. The system of claim 18 , wherein the prediction score comprises a plurality of prediction scores; wherein the pixel-wise segmentation label comprises a plurality of pixel-wise segmentation labels; and wherein each prediction score of the plurality of prediction scores represents a likelihood that an associated pixel-wise segmentation label of the plurality of pixel-wise segmentation labels corresponds to the point.

20. The system of claim 15 , wherein the at least one rich point feature is generated based on applying at least one post-processing technique to reduce re-projection error from the camera.

21. A method, comprising:
receiving, with at least one processor and from a LiDAR sensor of a vehicle, LiDAR point cloud information comprising at least one raw point feature for a point;
receiving, with th

Assignees

• Motional Ad Llc

Inventors

• Widjaja Sergi Adipraja
• Sharma Dhananjai
• Liong Venice Erin B

Classifications

• G01S17/931 Primary

  of land vehicles · CPC title

• G01S17/89

  for mapping or imaging · CPC title

• G01S17/86

  Combinations of lidar systems with systems other than lidar, radar or sonar, e.g. with direction finders · CPC title

• G01S7/4802

  using analysis of echo signal for target characterisation; Target signature; Target cross-section · CPC title

• G01S7/4808

  Evaluating distance, position or velocity data · CPC title