Granted Patent B1

US 12,573,237 · App. 18/344,519 · Granted Mar 10, 2026

Inventors: Tian Lan (Seattle, WA); Hui Liang (Issaquah, WA); Samuel Nathan Hallman (Marina del Rey, CA); Vijaya Naga Jyoth Sumanth Chennupati (Bothell, WA); Chuhang Zou (Seattle, WA); Hao Pan (Seattle, WA); Suyu Sang (San Jose, CA)

Assignee: Amazon Technologies, Inc.

G06V40/28G06T7/10G06T7/20G06T7/70G06T2207/20132G06T2207/30196

Abstract

Claims (131)

1 . A system comprising:

a camera comprising at least one processor and at least one memory component, wherein the camera has a field of view including at least a portion of a fixture having a plurality of product spaces within the field of view; and

a computer system in communication with at least the camera, wherein the computer system has positions of at least one point corresponding to each of the plurality of product spaces stored thereon, and wherein the computer system is programmed with one or more sets of instructions that, when executed by the computer system, cause the computer system to execute a method comprising:

receiving a plurality of images from the camera, wherein each of the plurality of images was captured over a period of time;

determining positions of at least one hand of an actor over at least a portion of the period of time;

identifying a first point of an image plane of the camera corresponding to a portion of a first product space;

identifying a second point of the image plane corresponding to a portion of a second product space;

calculating, for each of the plurality of images, a first score representative of an overlap between the positions of the at least one hand and the first point;

calculating, for each of the plurality of images, a second score representative of an overlap between the positions of the at least one hand and the second point;

selecting a crop center for a first clip of images based at least in part on the first scores calculated for the images of the first clip, the second scores calculated for the images of the first clip, the first point and the second point, wherein each of the images of the first clip is one of the plurality of images;

cropping each of the first clip of images by a crop window about the crop center;

providing the cropped first clip of images as inputs to a model, wherein the model comprises:

a feature encoder having a convolutional neural network backbone, an action encoder, a region encoder and a hand encoder;

a feature queue; and

a sequence head comprising an action head, an item head, a hand head and a quantity head;

receiving outputs from the model in response to the inputs;

determining that the actor executed at least one of a taking event, a return event or an event that is neither the taking event nor the return event with an item associated with the product space based at least in part on the outputs received from the model in response to the inputs; and

storing an indication that the actor executed the at least one of the taking event, the return event or the event that is neither the taking event nor the return event in association with the actor to an external system in communication with the camera over one or more networks.

2 . The system of claim 1 , wherein each of the first clip of images comprises a plurality of pixels, and

wherein the method further comprises:

prior to providing the cropped first clip of images as inputs to the model,

stacking each of the cropped first clip of images with channels representing hands and items,

wherein each of the plurality of pixels is represented by:

a plurality of color channels representing a color of one of the plurality of pixels;

a channel indicating whether the one of the plurality of pixels depicts a portion of a hand; and

a channel indicating whether the one of the plurality of pixels depicts an item within the portion of the hand.

3 . The system of claim 1 , wherein an area of each of the plurality of images is approximately twelve times greater than an area of the crop window.

4 . A method comprising:

capturing at least a first plurality of images by a first camera having a first field of view, wherein at least a portion of a fixture comprising a first product space and a second product space is within the first field of view;

determining positions of at least one hand of a first actor over at least a first period of time;

selecting a first center point for cropping at least some of the first plurality of images captured over the first period of time, wherein the first center point is selected based at least in part on a first position corresponding to the first product space, a second position corresponding to the second product space and the positions of the at least one hand of the first actor over the first period of time;

generating a first clip of images, wherein each of the images of the first clip is one of the first plurality of images captured over the first period of time cropped by a window about the first center point;

generating a first hypothesis based at least in part on the first clip of images, wherein the first hypothesis identifies:

a first event, wherein the first event is one of a taking event, a return event or neither a taking event nor a return event;

the first actor; and

a first item associated with one of the first product space or the second product space; and

associating at least a first quantity of the first item with the first actor based at least in part on the first hypothesis.

5 . The method of claim 4 , wherein generating the first hypothesis comprises:

generating a first set of features from the first clip of images, wherein each of the first set of features is one of an action feature, a region feature or a hand feature;

determining a first classification of an event type depicted in the first clip of images; and

generating the first hypothesis based at least in part on the first set of features and the first classification.

6 . The method of claim 5 , further comprising:

determining positions of the at least one hand of the first actor over at least a second period of time;

generating a second clip of images, wherein each of the images of the second clip is one of the first plurality of images captured over the second period of time cropped by the window about the first center point;

generating a second set of features from the second clip of images, wherein each of the second set of features is one of an action feature, a region feature or a hand feature;

determining a second classification of an event type depicted in the second clip of images;

determining that the first classification is consistent with the second classification; and

in response to determining that the first classification is consistent with the second classification,

generating a sequence of features comprising the first set of features and the second set of features,

wherein the first hypothesis is generated based at least in part on the sequence of features, the first classification and the second classification.

7 . The method of claim 6 , wherein determining the first classification of at least the first clip of images comprises:

determining a first plurality of scores based at least in part on the first clip of images, wherein each of the first plurality of scores represents one of a probability that the first clip of images depicts a taking event, a probability that the first clip of images depicts a return event, or a probability that the first clip of images does not depict a taking event or a return event; and

determining the first classification based at least in part on a greatest one of the first plurality of scores,

wherein determining the second classification of at least the second clip of images comprises:

determining a second plurality of scores based at least in part on the second clip of images, wherein each of the second plurality of scores represents one of a probability that the second clip of images depicts a taking event, a probability that the second clip of images depicts a return event, or a probability that the second clip of images does not depict a taking event or a return event; and

determining the second classification based at least in part on a greatest one of the second plurality of scores, and

wherein generating the sequence of features comprises:

concatenating at least the first set of features and the second set of features,

wherein the first hypothesis is generated based at least in part on the concatenated first set of features and second set of features.

8 . The method of claim 5 , wherein generating the first set of features from the first clip of images comprises:

providing each of the images of the first clip as inputs to an encoder comprising a convolutional neural network backbone; and

receiving outputs from the encoder in response to the inputs, wherein each of the outputs is a feature map corresponding to one of the images of the first clip, and

wherein each of the first set of features is generated based at least in part on one of the feature maps.

9 . The method of claim 8 , wherein the encoder further comprises a hand attention module, and

wherein one of the first set of features is an action feature generated based at least in part on the feature maps corresponding to the images of the first clip and the positions of the at least one hand of the first actor over the first period of time.

10 . The method of claim 8 , further comprising:

identifying portions of the first plurality of images depicting at least the first product space and the second product space,

wherein one of the first set of features is a region feature generated based at least in part on the feature maps corresponding to the images of the first clip and the portions of the first plurality of images depicting at least the first product space and the second product space, and

wherein the region feature represents an action performed at one of the first product space or the second product space.

11 . The method of claim 8 , further comprising:

generating a trajectory of the at least one hand based at least in part on the positions of the at least one hand over the first period of time,

wherein one of the first set of features is a hand feature generated based at least in part on the feature maps corresponding to the images of the first clip and the trajectory of the at least one hand.

12 . The method of claim 4 , wherein selecting the first center point comprises:

calculating first scores for each of the first plurality of images captured over the first period of time based at least in part on the positions of the at least one hand of the first actor over the first period of time, wherein each of the first scores is representative of an overlap of the at least one hand and the first product space;

calculating second scores for each of the first plurality of images captured over the first period of time based at least in part on the positions of the at least one hand of the first actor over the first period of time, wherein each of the second scores is representative of an overlap of the at least one hand and the second product space; and

determining the first center point based at least in part on the first scores, the second scores, the first position and the second position.

13 . The method of claim 4 , wherein each of the first plurality of images is six hundred forty pixels by four hundred eighty pixels, and

wherein the window is one hundred sixty pixels by one hundred sixty pixels.

14 . The method of claim 4 , wherein each of the first clip of images comprises a plurality of pixels, and

wherein the method further comprises:

stacking each of the first clip of images with channels representing hands and items,

wherein each of the plurality of pixels is represented by:

a plurality of color channels representing a color of one of the plurality of pixels;

a channel indicating whether the one of the plurality of pixels depicts a portion of a hand; and

a channel indicating whether the one of the plurality of pixels depicts an item within the portion of the hand.

15 . The method of claim 5 , wherein generating the first set of features comprises:

providing each of the first clip of images as inputs to a model comprising:

a feature encoder having a convolutional neural network backbone, an action encoder, a region encoder and a hand encoder;

a feature queue configured to define a sequence of clips based at least in part on features received from the feature encoder; and

a sequence head comprising an action head, an item head, an actor head and a quantity head, wherein the sequence head is configured to generate a hypothesis based on a sequence of clips defined by the feature queue.

16 . The method of claim 4 , further comprising:

determining positions of at least one hand of a second actor over a second period of time;

selecting a second center point for cropping at least some of the first plurality of images captured over the second period of time, wherein the second center point is selected based at least in part on the first position, the second position and the positions of the at least one hand of the second actor over the second period of time;

generating a second clip of images, wherein each of the images of the second clip is one of the first plurality of images captured over the second period of time cropped by a window about the second center point;

generating a second hypothesis based at least in part on the second clip of images, wherein the second hypothesis identifies:

a second event, wherein the second event is one of a taking event, a return event or neither a taking event nor a return event;

the second actor; and

a second item associated with one of the first product space or the second product space; and

associating at least a second quantity of the second item with the second actor based at least in part on the second hypothesis.

17 . The method of claim 4 , further comprising:

capturing at least a second plurality of images by a second camera having a second field of view, wherein at least a portion of the fixture is within the second field of view;

determining positions of the at least one hand of the first actor over at least a second period of time;

selecting a second center point for cropping at least some of the second plurality of images captured over the second period of time, wherein the second center point is selected based at least in part on the first position, the second position and the positions of the at least one hand of the first actor over the second period of time;

generating a second clip of images, wherein each of the images of the second clip is one of the second plurality of images captured over the second period of time cropped by a window about the second center point; and

generating a second hypothesis based at least in part on the second clip of images, wherein the second hypothesis identifies:

the first event or a second event, wherein the second event is one of a taking event, a return event or neither a taking event nor a return event;

the first actor; and

the first item or a second item associated with one of the first product space or the second product space,

wherein the first quantity is associated with the first actor based at least in part on the first hypothesis and the second hypothesis.

18 . The method of claim 17 , further comprising:

determining, by a computer system, that the first actor executed one of the first event or the second based at least in part on the first hypothesis and the second hypothesis.

19 . The method of claim 4 , further comprising:

determining positions of the at least one hand of the first actor over a second period of time;

selecting a second center point for cropping at least some of the first plurality of images captured over the second period of time, wherein the second center point is selected based at least in part on the first position, the second position and the positions of the at least one hand of the first actor over the second period of time; and

generating a second clip of images, wherein each of the images of the second clip is one of the first plurality of images captured over the second period of time cropped by a window about the second center point;

wherein the first hypothesis is generated based at least in part on the first clip of images and the second clip of images.

20 . A computer system comprising at least one processor and at least one data store,

wherein the computer system is in communication with a plurality of cameras, and

wherein the computer system is programmed with one or more sets of instructions that, when executed by the at least one processor, cause the computer system to execute a method comprising:

generating a first sequence of features based at least in part on a first set of images,

wherein each of the images of the first set is cropped from a second set of images about a cropping window defined from a crop center,

wherein each of the images of the second set is captured by a first camera of the plurality of cameras,

wherein each of the images of the first set is a multi-channel image including, for each pixel of such images, a plurality of channels corresponding to color values, a channel corresponding to a mask for a hand, and a channel corresponding to a mask for a product, and

wherein each of the first set of images has been classified as depicting at least one event at one product space by a first model configured to generate a feature map based at least in part on a sequence of features derived from sets of images and a hypothesis of a type of an event and a location of the event based at least in part on a feature map and a first plurality of positional embeddings, and wherein each of the first plurality of positional embeddings corresponds to one of a plurality of product spaces;

generating a second sequence of features based at least in part on a third set of images,

wherein each of the images of the third set is cropped from a fourth set of images about a cropping window defined from a crop center,

wherein each of the images of the fourth set is captured by a second camera,

wherein each of the images of the third set is a multi-channel image including, for each pixel of such images, a plurality of channels corresponding to color values, a channel corresponding to a mask for a hand, and a channel corresponding to a mask for a product, and

wherein each of the third set of images has been classified as depicting at least one event at one product space by a second model configured to generate a feature map based at least in part on a sequence of features derived from sets of images and a hypothesis of a type of an event and a location of the event based at least in part on a feature map and a second plurality of positional embeddings, wherein each of the second plurality of positional embeddings corresponds to one of a plurality of product spaces;

determining that an actor executed at least one of a taking event, a return event or an event that is neither the taking event nor the return event with an item associated with a product space based at least in part on the first sequence of features and the second sequence of features; and

storing information regarding the at least one of the taking event, the return event or the event that is neither the taking event nor the return event in association with the actor in the at least one data store.

Assignments (1)

References Cited (205)