Data augmentation for driver monitoring

1. A method for augmenting training images used for generating a model for autonomously monitoring vehicle drivers, comprising:

at a computer system including one or more processors and memory:

obtaining a first image of a first driver in an interior of a first vehicle;

separating, from the first image, a first driver image from a first background image of the interior of the first vehicle;

obtaining a second background image, wherein the second background image includes an image of an interior of a second vehicle;

generating a second image by overlaying the first driver image onto the second background image; and

adding the second image to a corpus of training images to be used by a machine learning system to generate a model for autonomously monitoring vehicle drivers.

2. The method of claim 1 , wherein the separating comprises:

applying a segmentation model to generate a segmentation mask that (1) associates a plurality of first pixels of the first image with the first driver image or (2) associates a plurality of second pixels of the first image with the first background image.

3. The method of claim 2 , wherein the segmentation model includes a U-Net that is based on a fully convolutional network.

4. The method of claim 1 , further comprising:

training the model for autonomously monitoring vehicle drivers to determine whether a vehicle driver is looking forward at a road ahead of a vehicle.

5. The method of claim 1 , further comprising:

training the model for autonomously monitoring vehicle drivers to determine whether a vehicle driver is looking forward at a road, looking to the left, looking to the right, looking down, closing eyes, or talking.

6. The method of claim 1 , further comprising, prior to overlaying the first driver image onto the second background image, implementing one or more of:

adjusting one or more image properties of at least one of the first driver image and the second background image to match lighting conditions of the first driver image and the second background image;

normalizing at least one of the first driver image and the second background image to match average brightness levels of the first driver image and the second background image; and

scaling at least one of the first driver image and the second background image.

7. The method of claim 1 , further comprising:

obtaining a third image of a second driver, where the first and second drivers are distinct from each other;

separating, from the third image, a second driver image from a respective background image;

generating a fourth image by overlaying the second driver image onto the second background image; and

adding the fourth image to the corpus of training images to be used by the machine learning system to generate the model for monitoring vehicle drivers.

8. The method of claim 1 , further comprising:

collecting a plurality of background images;

mapping each of the background images onto a respective point in a multidimensional space having a distance metric d;

clustering the plurality of background images using the distance metric d to form a plurality of image clusters;

for each of the image clusters, identifying one or more background images in the image cluster that are most distant according to the distance metric d;

forming a set of candidate background images comprising the identified one or more most distant background images in each of the image clusters; and

selecting the second background image from the set of candidate background images.

9. The method of claim 8 , wherein clustering the plurality of background images comprises:

selecting a positive integer number K;

selecting K cluster centers; and

for each of the plurality of background images:

determining a distance of the respective background image from each of the cluster centers; and

assigning the respective background image to a respective image cluster associated with a respective cluster center to which the respective background image has a shortest distance.

10. The method of claim 1 , further comprising:

collecting a plurality of background images;

mapping each of the background images onto a respective point in a multidimensional space having a distance metric d;

clustering the plurality of background images using the distance metric d to form a plurality of image clusters;

for each of the plurality of background images, determining, on a clustering plot, a respective distance between the respective background image and a corresponding cluster center of an image cluster to which the respective background image belongs, wherein the second background image is selected from the plurality of background images based on the respective distance between the second background image and the corresponding cluster center.

11. The method of claim 1 , wherein the first driver image is overlaid at a first location of the second background image, further comprising:

generating an alternative image by overlaying the first driver image at a second position of the second background image; and

adding the alternative image into the corpus of training images jointly with the second image.

12. The method of claim 1 , further comprising training the model for autonomously monitoring vehicle drivers to determine:

whether a vehicle driver is sitting on a driver seat or a passenger seat; and

in accordance with a determination whether the vehicle driver is sitting on the driver seat or the passenger seat, classify the vehicle driver as a distracted driver or a distracted passenger.

13. A computer system, comprising:

one or more processors; and

memory storing one or more programs configured for execution by the one or more processors, the one or more programs comprising instructions for:

obtaining a first image of a first driver in an interior of a first vehicle;

separating, from the first image, a first driver image from a first background image of the interior of the first vehicle;

obtaining a second background image, wherein the second background image includes an image of an interior of a second vehicle;

generating a second image by overlaying the first driver image onto the second background image; and

adding the second image to a corpus of training images to be used by a machine learning system to generate a model for autonomously monitoring vehicle drivers.

14. The computer system of claim 13 , wherein the instructions for separating the first driver image from the first background image further comprises instructions for:

applying a segmentation model to generate a segmentation mask that (1) associates a plurality of first pixels of the first image with the first driver image or (2) associates a plurality of second pixels of the first image with the first background image.

15. The computer system of claim 13 , the one or more programs further comprising instructions for:

training the model for autonomously monitoring vehicle drivers to determine whether a vehicle driver is looking forward at a road ahead of a vehicle.

16. The computer system of claim 13 , wherein the one or more programs further comprises instructions for, prior to overlaying the first driver image onto the second background image, implementing one or more of:

adjusting one or more image properties of at least one of the first driver image and the second background image to match lighting conditions of the first driver image and the second background image;

normalizing at least one of the first driver image and the second background image to match average brightness levels of the first driver image and the second background image; and

scaling at least one of the first driver image and the second background image.

17. A non-transitory computer-readable storage medium storing one or more programs configured for execution by one or more processors of a computer system, the one or more programs comprising instructions for: obtaining a first image of a first driver in an interior of a first vehicle;

separating, from the first image, a first driver image from a first background image of the interior of the first vehicle;

obtaining a second background image, wherein the second background image includes an image of an interior of a second vehicle;

generating a second image by overlaying the first driver image onto the second background image; and

adding the second image to a corpus of training images to be used by a machine learning system to generate a model for autonomously monitoring vehicle drivers.

18. The non-transitory computer-readable storage medium of claim 17 , further comprising instructions for:

training the model for autonomously monitoring vehicle drivers to determine whether a vehicle driver is looking forward at a road, looking to the left, looking to the right, looking down, closing eyes, or talking.

19. The non-transitory computer-readable storage medium of claim 17 , further comprising instructions for, prior to overlaying the first driver image onto the second background image, implementing one or more of:

adjusting one or more image properties of at least one of the first driver image and the second background image to match lighting conditions of the first driver image and the second background image;

normalizing at least one of the first driver image and the second background image to match average brightness levels of the first driver image and the second background image; and

scaling at least one of the first driver image and the second background image.

20. The non-transitory computer-readable storage medium of claim 17 , wherein the instructions for separating the first driver image from the first background image further comprises instructions for:

applying a segmentation model to generate a segmentation mask that (1) associates a plurality of first pixels of the first image with the first driver image or (2) associates a plurality of second pixels of the first image with the first background image.