End-to-end learned lane boundary detection based on a transformer

1. A method for an end-to-end boundary lane detection system, comprising:

gridding a red-green-blue (RGB) image of a roadway captured by a camera sensor mounted on an ego vehicle into a plurality of image patches of a grid image;

generating, using a transformer self-attention network, different image patch embeddings to learn correlations and exchange a context between the different image patch embeddings;

decoding the different image patch embeddings into predetermined categories of lane boundaries, grid offsets from a center of the grid image, and instance identifications of the different image patch embeddings; and

generating lane boundary keypoints corresponding to at least one lane of the RGB image of the roadway based on the decoding into predetermined categories of lane boundaries or background, grid offsets from a center of the grid image, and instance identifications of the different image patch embeddings.

2. The method of claim 1 , further comprising:

planning a trajectory of the ego vehicle based on the lane boundary keypoints corresponding to the at least one lane of the RGB image of the roadway; and

operating the ego vehicle along the trajectory through the at least one lane of the roadway according to the lane boundary keypoints corresponding to the at least one lane.

3. The method of claim 1 , in which the generating of the different image patch embeddings comprises providing the plurality of image patches to a transformer encoder to generate the different image patch embeddings to learn the correlations and exchange the context between the different image patch embeddings using the transformer self-attention network.

4. The method of claim 3 , in which the ego vehicle comprises an autonomous vehicle.

5. The method of claim 1 , in which decoding the different image patch embeddings comprises:

providing the different image patch embeddings to a multi-layer perceptron (MLP) block; and

generating, by the MLP block, the predetermined categories, the grid offsets, and the instance identifications.

6. The method of claim 1 , in which decoding the different image patch embeddings comprises:

providing the different image patch embeddings to a recognition head, a localization head, and an instance head;

generating, by the recognition head, the predetermined categories;

generating, by the localization head, the grid offsets; and

generating, by the instance head, the instance identifications.

7. The method of claim 1 , in which generating the lane boundary keypoints comprises:

generating left boundary lane keypoints for a lane of a roadway;

generating right boundary lane keypoints for the lane of the roadway; and

generating a center boundary lane keypoint for the lane of the roadway.

8. The method of claim 1 , further comprising:

generating left boundary lane keypoints for an exit lane of a roadway; and

generating right boundary lane keypoints for the exit lane of the roadway.

9. A non-transitory computer-readable medium having program code recorded thereon for end-to-end boundary lane detection, the program code being executed by a processor and comprising:

program code to grid a red-green-blue (RGB) image of a roadway captured by a camera sensor mounted on an ego vehicle into a plurality of image patches of a grid image;

program code to generate different image patch embeddings to learn correlations and exchange a context between the different image patch embeddings using a transformer self-attention network;

program code to decode the different image patch embeddings into predetermined categories of lane boundaries, grid offsets from a center of the grid image, and instance identifications of the different image patch embeddings; and

program code to generate lane boundary keypoints corresponding to at least one lane of the RGB image of the roadway based on the decoding into predetermined categories of lane boundaries, grid offsets from the center of the grid image, and instance identifications of the different image patch embeddings.

10. The non-transitory computer-readable medium of claim 9 , further comprising:

program code to plan a trajectory of the ego vehicle based on the lane boundary keypoints of the RGB image; and

program code to operate the ego vehicle along the trajectory through the at least one lane of the roadway according to the lane boundary keypoints corresponding to the at least one lane.

11. The non-transitory computer-readable medium of claim 9 , in which the program code to generate the different image patch embeddings comprises program code to provide the plurality of image patches to a transformer encoder to generate the different image patch embeddings to learn the correlations and exchange the context between the different image patch embeddings using the transformer self-attention network.

12. The non-transitory computer-readable medium of claim 11 , in which the ego vehicle comprises an autonomous vehicle.

13. The non-transitory computer-readable medium of claim 9 , in which the program code to decode the different image patch embeddings comprises:

program code to provide the different image patch embeddings to a multi-layer perceptron (MLP) block; and

program code to generate, by the MLP block, the predetermined categories, the grid offsets, and the instance identifications.

14. The non-transitory computer-readable medium of claim 9 , in which the program code to decode the different image patch embeddings comprises:

program code to provide the different image patch embeddings to a recognition head, a localization head, and an instance head;

program code to generate, by the recognition head, the predetermined categories;

program code to generate, by the localization head, the grid offsets; and

program code to generate, by the instance head, the instance identifications.

15. The non-transitory computer-readable medium of claim 9 , in which the program code to generate the lane boundary keypoints comprises:

program code to generate left boundary lane keypoints for a lane of a roadway;

program code to generate right boundary lane keypoints for the lane of the roadway; and

program code to generate a center boundary lane keypoint for the lane of the roadway.

16. The non-transitory computer-readable medium of claim 9 , further comprising:

program code to generate left boundary lane keypoints for an exit lane of a roadway; and

program code to generate right boundary lane keypoints for the exit lane of the roadway.

17. A system for end-to-end boundary lane detection, the system comprising:

an image patch generation module to grid a red-green-blue (RGB) image of a roadway captured by a camera sensor mounted on an ego vehicle into a plurality of image patches;

an image patch embedding module to generate different image patch embeddings to learn correlations and exchange a context between the different image patch embeddings using a transformer self-attention network;

an image patch decoding module to decode the different image patch embeddings into predetermined categories of lane boundaries, grid offsets from a center of the grid image, and instance identifications of the different image patch embeddings;

a keypoint generation module to generate lane boundary keypoints corresponding to at least one lane of the RGB image of the roadway based on the decoding into the predetermined categories of lane boundaries, the grid offsets from the center of the grid image, and the instance identifications of the different image patch embeddings; and

a controller circuitry to operate the ego vehicle along a trajectory through the at least one lane of the roadway according to the generated lane boundary keypoints corresponding to the lane.

18. The system of claim 17 , further comprising a planner to plan the trajectory of the ego vehicle based on the lane boundary keypoints corresponding to the at least one lane of the RGB image of the roadway.

19. The system of claim 17 , in which the image patch embedding module further comprises a transformer encoder to generate the different image patch embeddings to learn the correlations and exchange the context between the different image patch embeddings using the transformer self-attention network.

20. The system of claim 17 , in which the image patch decoding module comprises a multi-layer perceptron (MLP) block to generate the predetermined categories, the grid offsets, and the instance identifications from the different image patch embeddings.