Granted Patent B2

US 12,506,948 · App. 18/515,347 · Granted Dec 23, 2025

Inventors: Noelia Vállez Enano (Dublin, IE); Aubrey Dunne (Dublin, IE); José Luis Espinosa Aranda (Dublin, IE)

Assignee: UBOTICA TECHNOLOGIES LIMITED

H04N23/61G06T7/248G06T7/292G06T7/74G06T2207/10032G06T2207/20081G06T2207/20084

Abstract

Claims (91)

1 . A method of capturing images of objects that are moving relative to a camera, the method comprising:

capturing, using the camera, a frame containing a target object, wherein the camera is controlled by an AX algorithm that uses partial image statistics from a part of a frame contained within a bounding box of a previous frame;

generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box;

determining a candidate object feature vector for a part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box;

computing the distance between each candidate object feature vector and a target object feature vector;

selecting the candidate object bounding box associated with the candidate object feature vector that is closest to the target object feature vector;

generating partial image statistics on the part of the frame contained within the selected candidate object bounding box; and

providing the partial image statistics as input to the AX algorithm that controls the camera.

2 . The method of claim 1 , wherein generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box comprises:

using a kinetic model to predict the target object's location within the frame; and

generating at least one candidate object bounding box containing the target object's predicted location.

3 . The method of claim 1 , wherein generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box comprises generating at least one candidate object bounding box with a size that is different to the size of the previously generated bounding box.

4 . The method of claim 1 , wherein generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box comprises:

using a Siamese network to localize the central point of the target object; and

generating at least one candidate object bounding box that is centered on the central point of the target object.

5 . The method of claim 1 , wherein determining a candidate object feature vector for the part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box comprises using a fixed feature extraction neural network.

6 . The method of claim 1 , wherein computing the distance between each candidate object feature vector and a target object feature vector comprises computing the distance between each candidate object feature vector and a combination of the target object feature vector and at least one previous candidate object feature vector that was closest to the target object feature vector.

7 . The method of claim 1 , wherein generating partial image statistics on the part of the frame contained within the selected candidate object bounding box comprises generating partial image statistics on the part of the frame contained within the selected candidate object bounding box of a previous frame.

8 . The method of claim 1 , wherein generating partial image statistics on the part of the frame contained within the selected candidate object bounding box comprises:

using a neural network to create a mask that covers the background pixels in the part of the frame contained within the selected candidate object bounding box; and

generating partial image statistics on the unmasked pixels in the part of the frame contained within the selected candidate object bounding box.

9 . The method of claim 1 , further comprising:

generating and selecting a second candidate object bounding box for a second target object in the captured frame; and

generating partial image statistics on the parts of the frame contained within the selected candidate object bounding boxes.

10 . The method of claim 1 , further comprising:

receiving, in a second camera, a target object feature vector from the first camera;

capturing a frame containing the target object using the second camera;

generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box;

determining a candidate object feature vector for the part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box; and

computing the distance between each candidate object feature vector and the received target object feature vector.

11 . A computing device, comprising:

a processing system configured to:

capture, using a camera, a frame containing a target object, wherein the camera is controlled by an AX algorithm that uses partial image statistics from a part of a frame contained within a bounding box of a previous frame;

generate at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box;

determine a candidate object feature vector for a part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box;

compute the distance between each candidate object feature vector and a target object feature vector;

select the candidate object bounding box associated with the candidate object feature vector that is closest to the target object feature vector;

generate partial image statistics on the part of the frame contained within the selected candidate object bounding box; and

provide the partial image statistics as input to the AX algorithm that controls the camera.

12 . The computing device of claim 11 , wherein the processing system is configured to generate the at least two candidate object bounding boxes for the target object in the captured frame using the previously generated bounding box by:

using a kinetic model to predict the target object's location within the frame; and

generating at least one candidate object bounding box containing the target object's predicted location.

13 . The computing device of claim 11 , wherein the processing system is configured to generate at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box by generating at least one candidate object bounding box with a size that is different to the size of the previously generated bounding box.

14 . The computing device of claim 11 , wherein the processing system is configured to generate at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box by:

using a Siamese network to localize the central point of the target object; and

generating at least one candidate object bounding box that is centered on the central point of the target object.

15 . The computing device of claim 11 , wherein the processing system is configured to determine the candidate object feature vector for the part of the frame contained within each candidate object bounding box by using a fixed feature extraction neural network to perform inference on the part of the frame contained within the candidate object bounding box.

16 . The computing device of claim 11 , wherein the processing system is configured to compute the distance between each candidate object feature vector and the target object feature vector by computing the distance between each candidate object feature vector and the combination of the target object feature vector and at least one previous candidate object feature vector that was closest to the target object feature vector.

17 . The computing device of claim 11 , wherein the processing system is configured to generate partial image statistics on the part of the frame contained within the selected candidate object bounding box by generating partial image statistics on the part of the frame contained within the selected candidate object bounding box of a previous frame.

18 . The computing device of claim 11 , wherein the processing system is configured to generate partial image statistics on the part of the frame contained within the selected candidate object bounding box by:

using a neural network to create a mask that covers the background pixels in the part of the frame contained within the selected candidate object bounding box; and

generating partial image statistics on the unmasked pixels in the part of the frame contained within the selected candidate object bounding box.

19 . The computing device of claim 11 , wherein the processing system is further configured to:

generate and select a second candidate object bounding box for a second target object in the captured frame; and

generate partial image statistics on the parts of the frame contained within the selected candidate object bounding boxes.

20 . The computing device of claim 11 , wherein the processing system is further configured to:

receive, in a second camera, a target object feature vector from the first camera;

capture a frame containing the target object using the second camera;

generate at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box;

determine a candidate object feature vector for the part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box; and

compute the distance between each candidate object feature vector and the received target object feature vector.

21 . A non-transitory processor readable media having stored thereon processor-executable instructions configured to cause a processing system to perform operations for capturing images of objects that are moving relative to a camera, the operations comprising:

capturing, using the camera, a frame containing a target object, wherein the camera is controlled by an AX algorithm that uses partial image statistics from a part of a frame contained within a bounding box of a previous frame;

generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box;

determining a candidate object feature vector for a part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box;

computing the distance between each candidate object feature vector and a target object feature vector;

selecting the candidate object bounding box associated with the candidate object feature vector that is closest to the target object feature vector;

generating partial image statistics on the part of the frame contained within the selected candidate object bounding box; and

providing the partial image statistics as input to the AX algorithm that controls the camera.

22 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box comprises:

using a kinetic model to predict the target object's location within the frame; and

generating at least one candidate object bounding box containing the target object's predicted location.

23 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box comprises generating at least one candidate object bounding box with a size that is different to the size of the previously generated bounding box.

24 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box comprises:

using a Siamese network to localize the central point of the target object; and

generating at least one candidate object bounding box that is centered on the central point of the target object.

25 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that determining a candidate object feature vector for the part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box comprises using a fixed feature extraction neural network.

26 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that computing the distance between each candidate object feature vector and a target object feature vector comprises computing the distance between each candidate object feature vector and a combination of the target object feature vector and at least one previous candidate object feature vector that was closest to the target object feature vector.

27 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that generating partial image statistics on the part of the frame contained within the selected candidate object bounding box comprises generating partial image statistics on the part of the frame contained within the selected candidate object bounding box of a previous frame.

28 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations such that generating partial image statistics on the part of the frame contained within the selected candidate object bounding box comprises:

using a neural network to create a mask that covers the background pixels in the part of the frame contained within the selected candidate object bounding box; and

generating partial image statistics on the unmasked pixels in the part of the frame contained within the selected candidate object bounding box.

29 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations further comprising:

generating and selecting a second candidate object bounding box for a second target object in the captured frame; and

generating partial image statistics on the parts of the frame contained within the selected candidate object bounding boxes.

30 . The non-transitory processor readable media of claim 21 , wherein the stored processor-executable instructions are configured to cause the processing system to perform operations further comprising:

receiving, in a second camera, a target object feature vector from the first camera;

capturing a frame containing the target object using the second camera;

generating at least two candidate object bounding boxes for the target object in the captured frame using a previously generated bounding box;

determining a candidate object feature vector for the part of the frame contained within each candidate object bounding box by using a neural network to perform inference on the part of the frame contained within the candidate object bounding box; and

computing the distance between each candidate object feature vector and the received target object feature vector.

Assignments (1)

Continuity (2)

References Cited (13)