Learning lane geometry by aerial and ground perspectives

1 . A method for lane detection, the method comprises:

(a) determining, by a lane detection neural network (NN) and based on locations of one or more road lanes within an aerial image, locations of the one or more road lanes on a ground-vehicle-acquired (GVA) image; and

(b) responding to the determining of the locations of the one or more road lanes on the GVA image;

wherein the lane detection NN was trained to perform aerial-images-to-GVA-images conversion;

wherein the aerial image was generated, by a first NN, based in the GVA image and is not acquired by an aerial camera:

wherein the first NN was trained to perform GVA-images-to-aerial-images conversion;

wherein the GVA image was acquired by a two dimensional camera of a ground vehicle; and

wherein the locations of the one or more road lanes within the aerial image are provided by a second NN that was trained to perform lane detection in arial images;

wherein the first NN is a generative adversarial network (GAN) which uses (a) a generator NN to create the aerial image, and (b) a discriminator NN to provide a loss over a difference that is a function of (i) a discriminator loss the looks at the GVA image in its entirety and the aerial image in its entirety, and (ii) a lateral/horizontal distance between the original lane line within the GVA image and the constructed lane line within the aerial image; wherein the function is an average of (i) the discriminator loss the looks at the GVA image in its entirety and the aerial image in its entirety, and (ii) the lateral/horizontal distance between the original lane line within the GVA image and the constructed lane line within the aerial image.

2 . The method according to claim 1 , wherein the responding comprises providing information about the locations of the one or more road lanes on the GVA image for use in performing a driving related operation.

3 . The method according to claim 1 , comprising performing an autonomous driving related operation based on the one or more road lanes.

4 . The method according to claim 1 , comprising autonomously maintaining the vehicle within a current lane of the vehicle.

5 . The method according to claim 1 , wherein the first NN was trained by feeding the first NN with (i) GVA images acquired by two dimensional cameras one or more ground vehicles aerial, and (ii) aerial images acquired, in parallel to the GVA images, by one or more drones that followed the one or more ground vehicles.

6 . The method according to claim 1 , comprising training at least one NN of the first NN, the second NN or the lane detection NN.

7 . A non-transitory computer readable medium for lane detection, the non-transitory computer readable medium stores instructions for:

determining, by a lane detection neural network (NN) and based on locations of one or more road lanes within an aerial image, locations of the one or more road lanes on a ground-vehicle-acquired (GVA) image; and

responding to the determining of the locations of the one or more road lanes on the GVA image;

wherein the lane detection NN was trained to perform aerial-images-to-GVA-images conversion;

wherein the aerial image was generated, by a first NN, based in the GVA image and is not acquired by an aerial camera;

wherein the first NN was trained to perform GVA-images-to-aerial-images conversion;

wherein the GVA image was acquired by a two dimensional camera of a ground vehicle; and

wherein the locations of the one or more road lanes within the aerial image are provided by a second NN that was trained to perform lane detection in arial images; and

wherein the first NN is a generative adversarial network (GAN) which uses (a) a generator NN to create the aerial image, and (b) a discriminator NN to provide a loss over a difference that is a function of (i) a discriminator loss the looks at the GVA image in its entirety and the aerial image in its entirety, and (ii) a lateral/horizontal distance between the original lane line within the GVA image and the constructed lane line within the aerial image; wherein the function is an average of (i) the discriminator loss the looks at the GVA image in its entirety and the aerial image in its entirety, and (ii) the lateral/horizontal distance between the original lane line within the GVA image and the constructed lane line within the aerial image.

8 . The non-transitory computer readable medium according to claim 7 , comprising performing an autonomous driving related operation based on the one or more road lanes.

9 . The non-transitory computer readable medium according to claim 7 , comprising autonomously maintaining the vehicle within a current lane of the vehicle.

10 . The non-transitory computer readable medium according to claim 7 , wherein the first NN was trained by feeding the first NN with (i) GVA images acquired by two dimensional cameras one or more ground vehicles aerial, and (ii) aerial images acquired, in parallel to the GVA images, by one or more drones that followed the one or more ground vehicles.

11 . The non-transitory computer readable medium according to claim 7 , comprising training at least one NN of the first NN, the second NN or the lane detection NN.

12 . The method according to claim 1 , wherein the responding comprises displaying, on the GVA image, the one or more road lanes.

13 . The method according to claim 1 , wherein the responding comprises triggering a display, on the GVA image, the one or more road lanes.

14 . The method according to claim 1 , wherein the responding comprises providing information about the locations of the one or more road lanes on the GVA image.

15 . The non-transitory computer readable medium according to claim 7 wherein the responding comprises displaying, on the GVA image, the one or more road lanes.

16 . The non-transitory computer readable medium according to claim 7 , wherein the responding comprises providing information about the locations of the one or more road lanes on the GVA image for use in performing a driving related operation.