Child seat detection for a seat occupancy classification system

1 . A method comprising:

receiving a current image showing a region of a vehicle seat currently captured inside of a vehicle;

retrieving one or more reference images, wherein at least one reference image is a previously stored first reference image of the vehicle seat being unoccupied;

determining a seat state of the vehicle seat by:

processing the current image and one or more reference images with a machine learning classification network that has been pre-trained, and

determining, based on an output from the machine learning classification network, the seat state indicating at least whether the vehicle seat in the current image comprises a mounted child seat;

forwarding the seat state to a seat occupancy classification system; and

in response to determining that no decision on the seat state is possible, storing the current image as an updated image for a reference image update process;

wherein the reference image update process includes:

receiving a new current image, wherein the new current image depicts a vehicle seat without a mounted child seat being unoccupied or a vehicle seat with a mounted child seat being unoccupied;

applying a Siamese neural network to compare the new current image with the updated image; and

in response to a Euclidean distance between embeddings of the new current image and the updated image being smaller than a second threshold, storing the new current image as one of the one or more reference images or replacing one of the one or more reference images with the new current image.

2 . The method of claim 1 , wherein the one or more reference images further comprise:

a second reference image being at least one of a previously stored image of a last detected child seat mounted on the vehicle seat, another previously detected child seat mounted on the vehicle seat, or an example child seat in a similar vehicle.

3 . The method of claim 2 , wherein further reference images of the reference images are previously stored images of other child seats mounted on the vehicle seat.

4 . The method of claim 1 ,

wherein the machine learning classification network is a neural network with a first input channel configured to receive the current image and one or more reference input channels each configured to receive a reference image of the one or more reference images, and

wherein the output of the neural network is a confidence score indicative of whether a child seat is mounted on the vehicle seat.

5 . The method of claim 1 , wherein the machine learning classification network is based on a Siamese neural network configured to compare the current image to the one or more reference images.

6 . The method of claim 5 , wherein the Siamese neural network is based on a plurality of convolutional neural networks that determine embeddings of the current image and the one or more reference images.

7 . The method of claim 6 , wherein the Siamese neural network was trained with a plurality of triples, wherein a triple comprises an anchor image, a positive match image, and a negative match image, wherein the triples comprise at least one of:

the anchor image is an example vehicle seat being occupied, the positive match image is the example vehicle seat being unoccupied, and the negative match image is the example vehicle seat with a mounted child seat being unoccupied;

the anchor image is an example vehicle seat with a mounted child seat being occupied, the positive match image is the example vehicle seat with the mounted child seat being unoccupied, and the negative match image is the example vehicle seat being unoccupied;

the anchor image is an example vehicle seat being unoccupied, the positive match image is the example vehicle seat being occupied, and the negative match image is the example vehicle seat with a mounted child seat being occupied; and

the anchor image is an example vehicle seat with a mounted child seat being unoccupied, the positive match image is the example vehicle seat with the mounted child seat being occupied, and the negative match image is the example vehicle seat being occupied.

8 . The method of claim 7 , wherein an image of a seat being occupied is a real image of the occupation by at least one of a child, an adult, or an object, or wherein an image of a seat being occupied is generated by modifying an image of an unoccupied seat to simulate an occupation.

9 . The method of claim 6 , wherein determining the seat state comprises:

determining one or more Euclidean distances between the embeddings of current image and the one or more reference images;

selecting a minimum value of the one or more Euclidean distances; and

in response to the minimum value of the Euclidean distance being smaller than a first threshold, determining that the respective reference image is a match, wherein the seat state is determined based on the match of the reference image.

10 . The method of claim 9 , wherein the Siamese neural network was trained with a plurality of triples, wherein a triple comprises an anchor image, a positive match image, and a negative match image, wherein the triples comprise at least one of:

the anchor image is an example vehicle seat being occupied, the positive match image is the example vehicle seat being unoccupied, and the negative match image is the example vehicle seat with a mounted child seat being unoccupied;

the anchor image is an example vehicle seat with a mounted child seat being occupied, the positive match image is the example vehicle seat with the mounted child seat being unoccupied, and the negative match image is the example vehicle seat being unoccupied;

the anchor image is an example vehicle seat being unoccupied, the positive match image is the example vehicle seat being occupied, and the negative match image is the example vehicle seat with a mounted child seat being occupied; and

the anchor image is an example vehicle seat with a mounted child seat being unoccupied, the positive match image is the example vehicle seat with the mounted child seat being occupied, and the negative match image is the example vehicle seat being occupied.

11 . The method of claim 10 , wherein an image of a seat being occupied is a real image of the occupation by at least one of a child, an adult, or an object, or wherein an image of a seat being occupied is generated by modifying an image of an unoccupied seat to simulate an occupation.

12 . The method of claim 1 , wherein the Siamese neural network was trained with a plurality of triples, wherein a triple comprises an anchor image, a positive match image, and a negative match image, wherein the triples comprise at least one of:

the anchor image is an example vehicle seat being occupied, the positive match image is the example vehicle seat being unoccupied, and the negative match image is the example vehicle seat with a mounted child seat being unoccupied;

the anchor image is an example vehicle seat with a mounted child seat being occupied, the positive match image is the example vehicle seat with the mounted child seat being unoccupied, and the negative match image is the example vehicle seat being unoccupied;

the anchor image is an example vehicle seat being unoccupied, the positive match image is the example vehicle seat being occupied, and the negative match image is the example vehicle seat with a mounted child seat being occupied; and

the anchor image is an example vehicle seat with a mounted child seat being unoccupied, the positive match image is the example vehicle seat with the mounted child seat being occupied, and the negative match image is the example vehicle seat being occupied.

13 . The method of claim 12 , wherein an image of a seat being occupied is a real image of the occupation by at least one of a child, an adult, or an object, or wherein an image of a seat being occupied is generated by modifying an image of an unoccupied seat to simulate an occupation.

14 . The method of claim 1 , wherein the method is initiated by the seat occupancy classification system in response to the seat occupancy classification system determining an uncertain state with regard to whether a person is seated on a mounted child seat.

15 . The method of claim 1 , wherein the seat occupancy classification system serves to control one or more safety systems in the vehicle, wherein the one or more safety systems comprises at least one of airbag regulation or seatbelt tensioning control.

16 . A computing system comprising computer hardware components configured to:

receive a current image showing a region of a vehicle seat currently captured inside of a vehicle;

retrieve one or more reference images, wherein at least one reference image is a previously stored first reference image of the vehicle seat being unoccupied;

determine a seat state of the vehicle seat by:

processing the current image and one or more reference images with a pre-trained machine learning classification network, and

determining, based on an output from the machine learning classification network, the seat state indicating at least whether the vehicle seat in the current image comprises a mounted child seat;

forward the seat state to a seat occupancy classification system; and

in response to determining that no decision on the seat state is possible, store the current image as an updated image for a reference image update process;

wherein the reference image update process includes:

receiving a new current image, wherein the new current image depicts a vehicle seat without a mounted child seat being unoccupied or a vehicle seat with a mounted child seat being unoccupied;

applying a Siamese neural network to compare the new current image with the updated image; and

in response to a Euclidean distance between embeddings of the new current image and the updated image being smaller than a second threshold, storing the new current image as one of the one or more reference images or replacing one of the one or more reference images with the new current image.

17 . The computing system of claim 16 , further comprising:

a vehicle including:

a camera for taking the current image; and

the seat occupancy classification system.

18 . A computer program product comprising instructions, which, when executed on a computer, cause the computer to:

receive a current image showing a region of a vehicle seat currently captured inside of a vehicle;

retrieve one or more reference images, wherein at least one reference image is a previously stored first reference image of the vehicle seat being unoccupied;

determine a seat state of the vehicle seat by:

processing the current image and one or more reference images with a pre-trained machine learning classification network, and

determining, based on an output from the machine learning classification network, the seat state indicating at least whether the vehicle seat in the current image comprises a mounted child seat;

forward the seat state to a seat occupancy classification system; and

in response to determining that no decision on the seat state is possible, store the current image as an updated image for a reference image update process;

wherein the reference image update process includes:

receiving a new current image, wherein the new current image depicts a vehicle seat without a mounted child seat being unoccupied or a vehicle seat with a mounted child seat being unoccupied;

applying a Siamese neural network to compare the new current image with the updated image; and

in response to a Euclidean distance between embeddings of the new current image and the updated image being smaller than a second threshold, storing the new current image as one of the one or more reference images or replacing one of the one or more reference images with the new current image.