Method, system and apparatus for detecting item facings

Abstract

A method by an imaging controller of detecting item facings from image sensor data includes: obtaining, at the imaging controller, the image sensor data corresponding to a support structure containing at least one item; identifying, by a feature detector of the imaging controller, a set of matched keypoint pairs from keypoints of the image sensor data; determining, by a peak detector of the imaging controller, a separation distance between the keypoints of each matched keypoint pair; detecting, by the peak detector, a count of item instances represented in the image sensor data based on the separation distances; and presenting item facing detection output including the count of item instances.

Description

BACKGROUND

Environments in which objects are managed, such as retail facilities, warehousing and distribution facilities, and the like, may store such objects in regions such as aisles of shelf modules or the like. For example, a retail facility may include objects such as products for purchase, and a distribution facility may include objects such as parcels or pallets. A mobile automation apparatus may be deployed within such facilities to perform tasks at various locations. For example, a mobile automation apparatus may be deployed to capture data representing an aisle and corresponding products in a retail facility. However, the variability of the products in the facility, as well as variations in data capture conditions (e.g. lighting and the like) can prevent the accurate detection of individual products and their status from such data.

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

The accompanying figures, where like reference numerals refer to identical or functionally similar elements throughout the separate views, together with the detailed description below, are incorporated in and form part of the specification, and serve to further illustrate embodiments of concepts that include the claimed invention, and explain various principles and advantages of those embodiments.

FIG. 1 is a schematic of a mobile automation system.

FIG. 2 depicts a mobile automation apparatus in the system of FIG. 1.

FIG. 3 is a block diagram of certain internal components of the mobile automation apparatus in the system of FIG. 1.

FIG. 4 is a flowchart of a method of detecting product facings in the system of FIG. 1.

FIG. 5 is a diagram illustrating input data to the method of FIG. 4.

FIG. 6 is a diagram illustrating detection of image keypoints at block 410 of the method of FIG. 4.

FIG. 7 is a diagram illustrating an example performance of block 415 of the method of FIG. 4.

FIG. 8 is a diagram illustrating an example performance of block 420 of the method of FIG. 4.

FIG. 9 is a flowchart of a method for allocating keypoints to clusters at block 425 of the method of FIG. 4.

FIG. 10 is a diagram illustrating a performance of blocks 425 and 430 of the method of FIG. 4.

FIG. 11 is a diagram illustrating sequential performances of block 420 of the method of FIG. 4 in different dimensions.

Skilled artisans will appreciate that elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale. For example, the dimensions of some of the elements in the figures may be exaggerated relative to other elements to help to improve understanding of embodiments of the present invention.

The apparatus and method components have been represented where appropriate by conventional symbols in the drawings, showing only those specific details that are pertinent to understanding the embodiments of the present invention so as not to obscure the disclosure with details that will be readily apparent to those of ordinary skill in the art having the benefit of the description herein.

DETAILED DESCRIPTION

Examples disclosed herein are directed to a method by an imaging controller of detecting item facings from image sensor data includes: obtaining, at the imaging controller, the image sensor data corresponding to a support structure containing at least one item; identifying, by a feature detector of the imaging controller, a set of matched keypoint pairs from keypoints of the image sensor data; determining, by a peak detector of the imaging controller, a separation distance between the keypoints of each matched keypoint pair; detecting, by the peak detector, a count of item instances represented in the image sensor data based on the separation distances; and presenting item facing detection output including the count of item instances.

Additional examples disclosed herein are directed to a computing device, comprising: a feature detector configured to: obtain image sensor data corresponding to a support structure containing at least one item; and identify a set of matched keypoint pairs from keypoints of the image sensor data; a peak detector configured to: determine a separation distance between the keypoints of each matched keypoint pair; detect a count of item instances represented in the image sensor data based on the separation distances; and a boundary generator configured to present item facing detection output including the count of item instances.

Further examples disclosed herein are directed to a non-transitory computer-readable medium storing instructions executable by an imaging controller to configure the imaging controller to: obtain image sensor data corresponding to a support structure containing at least one item; identify a set of matched keypoint pairs from keypoints of the image sensor data; determine a separation distance between the keypoints of each matched keypoint pair; detect a count of item instances represented in the image sensor data based on the separation distances; and present item facing detection output including the count of item instances.

FIG. 1 depicts a mobile automation system 100 in accordance with the teachings of this disclosure. The system 100 includes a server 101 in communication with at least one mobile automation apparatus 103 (also referred to herein simply as the apparatus 103) and at least one client computing device 104 via communication links 105, illustrated in the present example as including wireless links. In the present example, the links 105 are provided by a wireless local area network (WLAN) deployed via one or more access points (not shown). In other examples, the server 101, the client device 104, or both, are located remotely (i.e. outside the environment in which the apparatus 103 is deployed), and the links 105 therefore include wide-area networks such as the Internet, mobile networks, and the like. The system 100 also includes a dock 106 for the apparatus 103 in the present example. The dock 106 is in communication with the server 101 via a link 107 that in the present example is a wired link. In other examples, however, the link 107 is a wireless link.

The client computing device 104 is illustrated in FIG. 1 as a mobile computing device, such as a tablet, smart phone or the like. In other examples, the client device 104 is implemented as another type of computing device, such as a desktop computer, a laptop computer, another server, a kiosk, a monitor, and the like. The system 100 can include a plurality of client devices 104 in communication with the server 101 via respective links 105.

The system 100 is deployed, in the illustrated example, in a retail facility including a plurality of support structures such as shelf modules 110-1, 110-2, 110-3 and so on (collectively referred to as shelf modules 110 or shelves 110, and generically referred to as a shelf module 110 or shelf 110—this nomenclature is also employed for other elements discussed herein). The support structures can have various other forms in some examples, including tables, peg boards and the like. Each shelf module 110 supports a plurality of items such as products 112. Each shelf module 110 includes a shelf back 116-1, 116-2, 116-3 and a support surface (e.g. support surface 117-3 as illustrated in FIG. 1) extending from the shelf back 116 to a shelf edge 118-1, 118-2, 118-3.

The shelf modules 110 (also referred to as sub-regions of the facility) are typically arranged in a plurality of aisles (also referred to as regions of the facility), each of which includes a plurality of modules 110 aligned end-to-end. In such arrangements, the shelf edges 118 face into the aisles, through which customers in the retail facility, as well as the apparatus 103, may travel. As will be apparent from FIG. 1, the term “shelf edge” 118 as employed herein, which may also be referred to as the edge of a support surface (e.g., the support surfaces 117) refers to a surface bounded by adjacent surfaces having different angles of inclination. In the example illustrated in FIG. 1, the shelf edge 118-3 is at an angle of about ninety degrees relative to the support surface 117-3 and to the underside (not shown) of the support surface 117-3. In other examples, the angles between the shelf edge 118-3 and the adjacent surfaces, such as the support surface 117-3, is more or less than ninety degrees.

The apparatus 103 is equipped with a plurality of navigation and data capture sensors 108, such as image sensors (e.g. one or more digital cameras) and depth sensors (e.g. one or more Light Detection and Ranging (LIDAR) sensors, one or more depth cameras employing structured light patterns, such as infrared light, or the like). The apparatus 103 is deployed within the retail facility and, via communication with the server 101 and use of the sensors 108, navigates autonomously or partially autonomously along a length 119 of at least a portion of the shelves 110.

While navigating among the shelves 110, the apparatus 103 can capture images (also referred to as image sensor data), depth measurements and the like, representing the shelves 110 (generally referred to as shelf data or captured data). Navigation may be performed according to a frame of reference 102 established within the retail facility. The apparatus 103 therefore tracks its pose (i.e. location and orientation) in the frame of reference 102.

The server 101 includes a special purpose controller, such as a processor 120, specifically designed to control and/or assist the mobile automation apparatus 103 to navigate the environment and to capture data. The processor 120 is also specifically designed, as will be discussed in detail herein, to process image data and/or depth measurements captured by the apparatus 103 representing the shelf modules 110, in order to detect product facings on the shelf modules 110. As will be apparent to those skilled in the art, a product facing is a single instance of a product facing into the aisle. Thus, if a support surface 117 carries three identical products adjacent to one another, the products represent three distinct facings. The resulting detected product facings can be provided to product status detection mechanisms (which may also be implemented by the processor 120 itself).

The processor 120 is interconnected with a non-transitory computer readable storage medium, such as a memory 122. The memory 122 includes a combination of volatile memory (e.g. Random Access Memory or RAM) and non-volatile memory (e.g. read only memory or ROM, Electrically Erasable Programmable Read Only Memory or EEPROM, flash memory). The processor 120 and the memory 122 each comprise one or more integrated circuits. In some embodiments, the processor 120 is implemented as one or more central processing units (CPUs) and/or graphics processing units (GPUs).

The memory 122 stores computer readable instructions for performing various actions, including control of the apparatus 103 to navigate the shelf modules 110 and capture shelf data, as well as post-processing of the shelf data. The execution of the above-mentioned instructions by the processor 120 configures the server 101 to perform various actions discussed herein. The applications stored in the memory 122 include an item facing detection application 124 (also simply referred to as the application 124). In general, via execution of the application 124 or subcomponents thereof and in conjunction with other components of the server 101, the processor 120 performs various actions to detect, in image data and depth measurements representing the shelves 110 (e.g. data captured by the apparatus 103), individual product facings, for use in downstream processing to detect product status information (e.g. whether products are out of stock, misplaced or the like).

Certain example components of the application 124 are shown in FIG. 1, including a feature detector 126, a peak detector 128, and a boundary generator 130. The feature detector 126 selects keypoints from image data depicting the products (e.g. distinct, repeatably identifiable regions in the image) and identifies pairs of such keypoints with matching properties. That is, the feature detector 126 detects portions of the image that appear to depict similar features of distinct instances of a given product 112. The peak detector 128 detects a count of the instances of the above-mentioned product 112 appearing in the image data, based on the output of the feature detector 126. The boundary generator 130, in turn, employs the output of either or both of the feature detector 126 and the peak detector 128 to generate boundaries for each instance of the product 112 in the image data.

In other embodiments, the application 124 may be implemented as a suite of logically distinct application, each implementing a suitable portion of the functionality discussed below. For example, the detectors 126 and 128, as well as the boundary generator 130, may be implemented as separate applications.

The memory 122 can also store data for use in the above-mentioned control of the apparatus 103, such as a repository 132 containing a map of the retail environment and any other suitable data (e.g. operational constraints for use in controlling the apparatus 103, data captured by the apparatus 103, and the like).

The processor 120, as configured via the execution of the application 124, is also referred to herein as an imaging controller 120, or simply as a controller 120. As will now be apparent, some or all of the functionality implemented by the imaging controller 120 described below may also be performed by preconfigured special purpose hardware controllers (e.g. one or more logic circuit arrangements specifically configured to optimize the speed of image processing, for example via FPGAs and/or Application-Specific Integrated Circuits (ASICs) configured for this purpose) rather than by execution of the application 124 by the processor 120.

The server 101 also includes a communications interface 134 interconnected with the processor 120. The communications interface 134 includes suitable hardware (e.g. transmitters, receivers, network interface controllers and the like) allowing the server 101 to communicate with other computing devices—particularly the apparatus 103, the client device 104 and the dock 106—via the links 105 and 107. The links 105 and 107 may be direct links, or links that traverse one or more networks, including both local and wide-area networks. The specific components of the communications interface 134 are selected based on the type of network or other links that the server 101 is required to communicate over. In the present example, as noted earlier, a wireless local-area network is implemented within the retail facility via the deployment of one or more wireless access points. The links 105 therefore include either or both wireless links between the apparatus 103 and the mobile device 104 and the above-mentioned access points, and a wired link (e.g. an Ethernet-based link) between the server 101 and the access point.

The processor 120 can therefore obtain data captured by the apparatus 103 via the communications interface 134 for storage (e.g. in the repository 132) and subsequent processing, e.g. via execution of the application 124 to detect product facings, as noted above). The server 101 may also transmit status notifications (e.g. notifications indicating that products are out-of-stock, in low stock or misplaced) to the client device 104 responsive to the determination of product status data. The client device 104 includes one or more controllers (e.g. central processing units (CPUs) and/or field-programmable gate arrays (FPGAs) and the like) configured to process (e.g. to display) notifications received from the server 101.

Turning now to FIG. 2, the mobile automation apparatus 103 is shown in greater detail. The apparatus 103 includes a chassis 201 containing a locomotive assembly 203 (e.g. one or more electrical motors driving wheels, tracks or the like). The apparatus 103 further includes a sensor mast 205 supported on the chassis 201 and, in the present example, extending upwards (e.g., substantially vertically) from the chassis 201. The mast 205 supports the sensors 108 mentioned earlier. In particular, the sensors 108 include at least one imaging sensor 207, such as a digital camera. In the present example, the mast 205 supports seven digital cameras 207-1 through 207-7 oriented to face the shelves 110.

The mast 205 also supports at least one depth sensor 209, such as a 3D digital camera capable of capturing both depth data and image data. The apparatus 103 also includes additional depth sensors, such as LIDAR sensors 211. In the present example, the mast 205 supports two LIDAR sensors 211-1 and 211-2. In other examples, the mast 205 can support additional LIDAR sensors 211 (e.g. four LIDARs 211). As shown in FIG. 2, the cameras 207 and the LIDAR sensors 211 are arranged on one side of the mast 205, while the depth sensor 209 is arranged on a front of the mast 205. That is, the depth sensor 209 is forward-facing (i.e. captures data in the direction of travel of the apparatus 103), while the cameras 207 and LIDAR sensors 211 are side-facing (i.e. capture data alongside the apparatus 103, in a direction perpendicular to the direction of travel). In other examples, the apparatus 103 includes additional sensors, such as one or more RFID readers, temperature sensors, and the like.

The mast 205 also supports a plurality of illumination assemblies 213, configured to illuminate the fields of view of the respective cameras 207. That is, the illumination assembly 213-1 illuminates the field of view of the camera 207-1, and so on. The cameras 207 and LIDARs 211 are oriented on the mast 205 such that the fields of view of the sensors each face a shelf 110 along the length 119 of which the apparatus 103 is traveling. As noted earlier, the apparatus 103 is configured to track a pose of the apparatus 103 (e.g. a location and orientation of the center of the chassis 201) in the frame of reference 102, permitting data captured by the apparatus 103 to be registered to the frame of reference 102 for subsequent processing.

Referring to FIG. 3, certain components of the mobile automation apparatus 103 are shown, in addition to the cameras 207, depth sensor 209, LIDARs 211, and illumination assemblies 213 mentioned above. The apparatus 103 includes a special-purpose controller, such as a processor 300, interconnected with a non-transitory computer readable storage medium, such as a memory 304. The memory 304 includes a suitable combination of volatile memory (e.g. Random Access Memory or RAM) and non-volatile memory (e.g. read only memory or ROM, Electrically Erasable Programmable Read Only Memory or EEPROM, flash memory). The processor 300 and the memory 304 each comprise one or more integrated circuits.

The memory 304 stores computer readable instructions for execution by the processor 300. In particular, the memory 304 stores an apparatus control application 308 which, when executed by the processor 300, configures the processor 300 to perform various functions related to navigating the facility and controlling the sensors 108 to capture data, e.g. responsive to instructions from the server 101. Those skilled in the art will appreciate that the functionality implemented by the processor 300 via the execution of the application 308 may also be implemented by one or more specially designed hardware and firmware components, such as FPGAs, ASICs and the like in other embodiments.

The memory 304 may also store a repository 312 containing, for example, a map of the environment in which the apparatus 103 operates, for use during the execution of the application 308. The apparatus 103 also includes a communications interface 316 enabling the apparatus 103 to communicate with the server 101 (e.g. via the link 105 or via the dock 106 and the link 107), for example to receive instructions to navigate to specified locations and initiate data capture operations.

The actions performed by the server 101, and specifically by the processor 120 as configured via execution of the application 124, to detect product facings from captured data representing the shelves 110 (e.g. images and depth measurements captured by the apparatus 103) will now be discussed in greater detail with reference to FIG. 4.

FIG. 4 illustrates a method 400 of detecting product facings. The method 400 will be described in conjunction with its performance in the system 100, and in particular by the server 101, with reference to the components illustrated in FIG. 1. As will be apparent in the discussion below, in other examples, some or all of the processing described below as being performed by the server 101 may alternatively be performed by the apparatus 103.

At block 405, the server 101, and particularly the feature detector 126, obtains at least an image corresponding to an area of a shelf 110 that contains at least one product 112. In particular, the area in the examples discussed below can contain as few as one product 112 of a given type, or multiple contiguous products 112 of that type. As will be apparent to those skilled in the art, products 112 that are arranged contiguously (i.e. immediately adjacent to one another on the shelf module 110, including being in physical contact with one another) may be difficult to segment by processing image and/or depth data according to other techniques such as edge detection.

The area mentioned above may also be referred to as a facing block, in that the image represents a block of at least one instance of a product 112, up to an unknown number of instances of the product 112. The server 101 may also obtain, at block 405, a set of depth measurements corresponding to the facing block. The depth measurements can be, for example, a plurality of sets of coordinates in the frame of reference 102. The image and depth data noted above may have been previously acquired by the apparatus 103 and provided to the server 101 for further processing.

The image data and depth data obtained at block 405 can be obtained in various ways. For example, turning to FIG. 5, the server 101 can extract the image and depth data from a larger data set based on a region of interest indicator defining the facing block area. In particular, FIG. 5 illustrates a point cloud 500 containing a plurality of depth measurements representing a shelf module 110 supporting various products 112. In particular, the point cloud 500 depicts three facings of a first product 112a, two facings of a second product 112b, and three facings of a third product 112c.

FIG. 5 also illustrates an image 504 of the shelf module 110 depicted by the point cloud 500. The image may be, for example, an RGB image captured by the apparatus 103 substantially simultaneously with the capture of the point cloud 500. The image 504 also depicts (albeit in two dimensions, rather than three dimensions) the shelf edges 118, shelf back 116, and products 112a, 112b and 112c as the point cloud 500. Finally, FIG. 5 illustrates a set of ROI indicators, or facing block areas, 508a, 508b and 508c indicating positions of product blobs detected, e.g. by the server 101 or another computing device. Detection of the ROI indicators can be performed based on any suitable combination of depth measurements and images, including the point cloud 500 and the image 504. The ROI indicators are illustrated in two dimensions, but can also be defined in three dimensions, according to the frame of reference 102. As is evident from FIG. 5, the ROI indicators 508 correspond to the respective positions of the products 112a, 112b and 112c, but do not distinguish between the individual product facings shown in the point cloud 500 and the image 504. The spaces between products 112 on the shelves 110 may be small enough that detecting individual product facings, in the absence of the processing techniques described below, is inaccurate or overly computationally expensive.

From the input data set out above, at block 405 the server can extract an image of a particular facing block for further processing. For example, an image 512 corresponding to the facing block 508a (i.e. the ROI indicator 508a) is shown separately from the image 504. The process described below can later be repeated for the other facing block areas 508. In other examples, the above processing to extract the relevant image and depth data can be performed previously, or by another computing device, such that at block 405 the server 101 simply retrieves the image 512, e.g. from the repository 132.

Returning to FIG. 4, at block 410 the server 101, and in particular the feature detector 126, is configured to generate feature descriptors for a set of keypoints, or interest points, of the image 512. The server 101 can implement any of a wide variety of keypoint selection and feature descriptor generation mechanisms. Examples of such mechanisms include an Oriented FAST and Rotated BRIEF (ORB) feature generator, a Scale Invariant Feature Transform (SIFT) feature generator, and the like. In general, such mechanisms select certain points in the image 512 that have sufficiently distinctive surrounding regions (a surrounding region can be a patch of 10×10 pixels centered on the keypoint, for example; various other surrounding region shapes and sizes may also be employed) that the keypoint can be readily identified in another image or another portion of the same image. The nature of the feature descriptors generated at block 410 varies with the mechanism implemented by the server 101. For example, if the server 101 implements an ORB feature generator at block 410, the feature descriptors are, for each keypoint, a one-dimensional binary vector of configurable length (e.g. 128 bits).

Turning to FIG. 6, the image 512 is shown, with several example keypoints 604 (specifically, eight sample keypoints 604-1 through 604-8). For each keypoint 604, the server 101 generates a feature descriptor 608, such as the above-mentioned vector 608. Thus, eight feature descriptors 608-1 through 608-8 are shown in FIG. 6. As will be apparent to those skilled in the art, the process of identifying keypoints and generating feature descriptors may yield a greater number of keypoints and corresponding feature descriptors than the set of eight shown in FIG. 6.

Referring again to FIG. 4, having generated feature descriptors at block 410, at block 415 the server 101 identifies a set of matched keypoint pairs based on the feature descriptors. Returning to FIG. 6, it will be apparent that several of the keypoints 604 depict repeating image features. For example, the keypoints 604-1, 604-2 and 604-3 each correspond to the apex of the letter “A” on separate instances of a product 112a. Similarly, the keypoints 604-4, 604-5 and 604-6 correspond to distinct instances of the same image feature, as do the keypoints 604-7 and 604-8. The feature descriptors 608 of such keypoints are therefore expected to be similar (although not necessarily identical).

At block 415, the feature detector 126 is configured to identify pairs of feature descriptors that are equal or sufficiently similar, and store (e.g. in the memory 122 the pairs for further processing. Various search mechanisms can be employed to identify matched pairs of keypoints 604. For example, the feature detector 126 can select each feature descriptor 608 in turn and generate a difference metric between the selected feature descriptor 608 and every other feature descriptor 608. An example of a difference metric that may be employed with ORB feature descriptors is the Hamming distance.

In the example illustrated in FIG. 6, therefore, seven difference metrics can be computed for each feature descriptor 608. Matches can be selected for a given feature descriptor 608 by selecting the other feature descriptor 608 having the lowest difference metric. Because the method 400 seeks to identify multiple instances of a product 112, however, more than one pairing may exist for each feature descriptor 608. The server 101 can therefore select not only the single other feature descriptor 608 with the lowest difference metric, but a configurable number of other feature descriptors 608. For example, the server 101 can select, for each feature descriptor 608, the three closest feature descriptors 608 and therefore store three matched keypoint pairs.

Various optimizations can be applied to the performance of block 415. For example, rather than performing the above-mentioned brute force search, the server 101 may generate a k-d tree containing the feature descriptors 608 and perform searches for matching feature descriptors 608 using the k-d tree.

In addition, referring to FIG. 7, the server 101 can apply a threshold to the difference metric. FIG. 7 provides a graphical representation of the identification of matching feature descriptors 608 for the feature descriptor 608-1. Each other feature descriptor 608 is shown along with a graphical representation of a difference metric 700 between that feature descriptor 608 and the feature descriptor 608-1. Longer representations of difference metrics 700 indicate that the relevant feature descriptors 608 are less similar.

FIG. 7 also illustrates a selection threshold 704, presented as a circular region centered on the feature descriptor 608-1. The server 101 can be configured to select, as matches for the feature descriptor 608-1, up to three (or any other preconfigured number, e.g. selected to meet or exceed the maximum potential number of product instances in the image 512) nearest neighbors, based on the difference metrics, with the condition that any selected matching feature descriptor 608 has a difference metric 700 below the threshold 704. Therefore, in the present example, the feature descriptors 608-3 and 608-2 are selected as matches to the feature descriptor 608-1. In other words, two matched keypoint pairs are produced from the selection shown in FIG. 7: the pair consisting of the feature descriptors 608-1 and 608-2, and the pair consisting of the feature descriptors 608-1 and 608-3. The above process can be repeated to identify matches for each other feature descriptor 608.

The performance of block 415, therefore, results in a number of matched keypoint pairs. In the example shown in FIG. 6, the matched keypoint pairs resulting from block 415 may include: [604-1, 604-2], [604-1, 604-3], [604-4, 604-5], [604-4, 604-6], [604-7, 604-8]. As will be seen from FIG. 6, a third instance of the image feature corresponding to the keypoints 604-7 and 604-8 was not selected at block 410. Lighting conditions, image capture artifacts, or the like, may result in certain instances of repeating image features not being detected.

At block 420, the server 101 is configured to determine distances between the matched keypoint pairs from block 415, and to detect a number of product instances in the image from block 405 based on those distances. The distances determined at block 420 can be physical distances, for example based on the depth measurements obtained at block 405. That is, the peak detector 128 of the server 101 can be configured to register (e.g. via back-projection) each of the keypoints 604 to the depth measurements in the point cloud 500, to determine a three-dimensional location of each keypoint 604 according to the frame of reference 604. In the present example, the distances determined at block 420 are horizontal distances (e.g. distances along the X axis of the frame of reference 102 as shown in FIG. 1), reflecting the expectation that repeating instances of products are arranged substantially horizontally along the shelf modules 110. That is, to the extent that a given pair of matched keypoints 604 are not at the same elevation (i.e. position on the Z axis of the frame of reference 102), differences in elevation may be ignored at block 420 in the present example. Depth-wise spacing (i.e. along the Y axis shown in FIG. 1) may also be ignored at block 420.

Referring to FIG. 8, distances 800a and 800b are shown for the matched keypoint pairs [604-1, 604-2] and [604-1, 604-3]. To detect a number of product instances at block 420, the peak detector 128 can be configured to generate a histogram of distances. Specifically, the histogram can include a plurality of bins each corresponding to a selected range of distances. For example, the total width (in the X direction) of the area represented by the image 512 can be determined, and divided into a predefined number of bins. For each distance from the block 420, the corresponding bin is identified (i.e. the bin whose range encompasses the distance), and a score associated with that bin is incremented by one. For example, FIG. 8 illustrates an incomplete histogram 804 having ten bins corresponding to ten segments of the width of the image 512. The score associated with the third bin has been incremented by one for the distance 800a, and the score associated with the seventh bin has been incremented by one for the distance 800b.

The process shown in FIG. 8 for the distances 800a and 800b is repeated for each other matched keypoint pair identified at block 415. An example completed histogram 808 is also shown in FIG. 8, in which a greater number of keypoints than the sample set shown in FIGS. 6-8 have been matched and allocated to bins. To complete the performance of block 420, the peak detector 128 is configured to detect peaks in the histogram 808. Peak detection can be performed by implementing a suitable local maximum detection mechanism. Via such a local maximum detection mechanism, the peak detector 128 can detect peaks at the bins 812-1 and 812-2 in the illustrated example.

The number of product instances detected at block 420 is the number of peaks detected, increased by one. As will now be apparent, each peak indicates a recurring distance between matched keypoint pairs. That is, each peak indicates the presence of multiple matched keypoint pairs with similar spacing. For example, the matched keypoint pairs [604-1, 604-3] and [604-7, 604-8], as seen in FIG. 6, are separated by substantially the same horizontal distance. In other words, a peak in the histogram 808 indicates two distinct product instances. The total of two peaks in the histogram 808 indicates three distinct product instances.

The peak detector 128 can also apply a minimum threshold to peak detection, such that a local maximum is only identified as a peak if the score exceeds the threshold. In the event that the image data from block 405 depicts a single product instance, the feature detector 126 may detect few or no matched keypoint pairs, and the histogram generated at block 420 may therefore include scores that only represent a small proportion of the total number of identified keypoints. Under such conditions, local maxima in the histogram may still exist due to features matched within the same product instance, but such local maxima are likely to be smaller (i.e. have lower scores) than peaks resulting from repeated product instances. By applying a magnitude threshold to the peaks identified at block 420, the peak detector 128 may avoid detecting peaks that result from matches within a product instance rather than peaks that are indicative of distinct products instances.

Referring again to FIG. 4, at block 425 the boundary generator 130 of the server 101 is configured to allocate each keypoint 604 to one of a set of clusters equal in number to the count of product instances detected at block 420. The clusters may then, as will be discussed below, be employed to generate boundaries (e.g. bounding boxes in the frame of reference 102) for each product instance depicted in the image 512.

Allocation of keypoints 604 to clusters may be performed by the boundary generator 130 according to a suitable clustering mechanism, such as k-means clustering or the like. The count of product instances (three, in the present example performance of the method 400) detected at block 420 is provided to the clustering mechanism as an input. That is, the boundary generator 130 is configured to allocate the keypoints 604 to a specific number of clusters, given by the result of block 420.

In some examples, the allocation of keypoints 604 to clusters is based not only on the count of product instances detected at block 420, but also on the distances determined at block 420 and the results of block 415 (indicating which keypoints 604 are paired). In the present example, clustering of the keypoints 604 is performed at block 425 by implementing a spectral clustering algorithm, such as that described in Ng, A. Y. & Jordan, M. & Weiss, Y. (2001). On Spectral Clustering: Analysis and an Algorithm. Proceedings Adv. Neural Inf. Process Syst. 14. The spectral clustering mechanism accepts as inputs an expected number of clusters (i.e. the detected count of product instances from block 420), and an affinity matrix. The affinity matrix indicates a degree of similarity between each pair of keypoints 604 (not only the matched keypoint pairs mentioned above). That is, an affinity value may be assigned to each pair of keypoints, such as a value between zero and one.

In general, a pair of keypoints 604 with a greater affinity value (e.g. closer to one) are more likely to be placed in the same cluster, while a pair of keypoints with a smaller affinity value (e.g. closer to zero) are more likely to be placed in different clusters. When the boundary generator 130 implements spectral clustering at block 425, the affinity matrix can be initialized based on the results of blocks 415 and 420.

In particular, referring to FIG. 9, a method 900 is illustrated of initializing affinity values for use in spectral clustering at block 425. At block 905, the boundary generator 130 selects a pair of keypoints 604. The performance of the method 900 is repeated for every pair of keypoints 604, and not only for the matched keypoint pairs discussed earlier.

At block 910, the server 101 determines whether the selected pair of keypoints 604 is a matched pair, according to the results of block 415. For example, the determination at block 910 is affirmative for the keypoints 604-1 and 604-3. The server 101 therefore proceeds to block 915, at which the affinity value corresponding to the selected pair is set to zero. In other words, if the selected pair of keypoints 604 is a matched pair, the matched pair is assumed to depict two instances of the same product feature. Therefore, each member of the matched pair is associated with a different product instance, and is to be placed in a different cluster. Setting the affinity measurement to zero increases the likelihood that the keypoints 604 will be assigned to distinct clusters.

When the determination at block 910 is negative, as in the example of the keypoints 604-2 and 604-7, the server 101 proceeds to block 920 rather than block 915. At block 920, an affinity value is set for the selected pair of keypoints 604 based on the distance separating the selected keypoints 604. Specifically, the affinity value set at block 920 is inversely proportional to the distance separating the selected pair of keypoints 604, reflecting an assumption that keypoints 604 that are more distant from one another are more likely to correspond to distinct product instances.

FIG. 10 illustrates an example set of keypoints 604 following a performance of block 425. In particular, the keypoints 604 shown in short dashed lines have been assigned to a first cluster, the keypoints 604 shown in solid lines have been assigned to a second cluster, and the keypoints 604 shown in long dashed lines have been assigned to a third cluster. The server 101 can store, in association with each keypoint 604, a cluster identifier indicating which cluster the keypoint 604 was allocated to.

At block 430, the server 101 can be configured to generate item boundaries corresponding to detected product instances, based on the results of block 430. In particular, the boundary generator 130 can be configured to fit a boundary (e.g. a rectangular boundary, although other shapes of boundary may also be generated in other examples. Various boundary fitting mechanisms (e.g. RANSAC-based fitting, Minimum Volume Bounding Box, or the like) can be implemented by the boundary generator 130 at block 430 to generate the boundaries, for example as three-dimensional bounding boxes defined according to the frame of reference 102.

Referring again to FIG. 10, three example boundaries 1000-1, 1000-2 and 1000-3 are shown as having been generated based on the keypoints 604 in each of the above-mentioned clusters. As will be apparent from FIG. 10, the boundaries 1000 indicate the presence, and the detected positions, of three distinct product instances, corresponding to the three products 112a shown in the image 512.

At block 435, the server 101 is configured to present item facing detection output, e.g. via a display connected to the server 101, and/or by transmitting output to the client device 104, and/or by providing the output to another component of the server 101 itself for further processing, or the like. The item facing detection output can include at least one of the count of item instances detected at block 420 and the boundaries generated at block 430. The item facing detection output can also be stored, e.g. in the memory 122.

Variations to the functionality described above are contemplated. For example, the server 101 can be configured to detect vertically stacked products 112 as well as horizontally arranged products 112 such as those shown in the image 512. In particular, as shown with a dashed line in FIG. 4 extending from block 430 to block 415, the performance of blocks 415-430 can be repeated for two physical dimensions associated with the support structures 110. More specifically, the server 101 can perform blocks 415-430 for a first physical dimension, such as the horizontal direction (e.g. the X axis of the frame of reference 102), and then repeat the performance of blocks 415-430 for a second physical dimension, such as the vertical direction (e.g. the Z axis of the frame of reference 102).

The performance of block 415 in such examples may be constrained to match keypoints separated by a minimum distance in the dimension currently being processed in the image sensor data. In addition, processing for the second physical dimension may include multiple performances of blocks 415-430, once for each cluster identified at block 430 for the first physical dimension. For example, a first performance of blocks 415-430 may segment the image sensor data into three horizontally-arranged clusters. Blocks 415-430 are then repeated three times, restricting each performance to the image sensor data within a respective horizontal cluster. In other examples, the above-mentioned order of processed dimensions can be reversed.

Turning to FIG. 11, an image 1100 is shown depicting six product instances, arranged in two contiguous stacks of three items each. Via a first performance of block 420, a histogram 1104 corresponding to the horizontal dimension is generated. As will be apparent from the image 1100, a peak 1108 in the histogram 1104 corresponds to matching keypoint pairs from horizontally-spaced items in the image 1100. Further, a peak 1112 corresponds to matching keypoint pairs from vertically-spaced items in the image 1100. Such matching keypoint pairs may have small or null horizontal distances separating them. The server 101 can therefore be configured to ignore the peak 1112 (or indeed any peak below a minimum distance threshold).

Additionally, a second iteration of block 420 yields a histogram 1116, in which a first peak 1120 results from matching keypoint pairs for horizontally-spaced items (which therefore have small vertical separation distances). The peak 1120, as with the peak 1112, may be ignored. In addition, the histogram 1116 presents two peaks 1124 and 1128. To generate a total count of item instances, the server 101 may multiply a detected horizontal count (two) by a detected vertical count (three).

In the foregoing specification, specific embodiments have been described. However, one of ordinary skill in the art appreciates that various modifications and changes can be made without departing from the scope of the invention as set forth in the claims below. Accordingly, the specification and figures are to be regarded in an illustrative rather than a restrictive sense, and all such modifications are intended to be included within the scope of present teachings.

The benefits, advantages, solutions to problems, and any element(s) that may cause any benefit, advantage, or solution to occur or become more pronounced are not to be construed as a critical, required, or essential features or elements of any or all the claims. The invention is defined solely by the appended claims including any amendments made during the pendency of this application and all equivalents of those claims as issued.

Moreover in this document, relational terms such as first and second, top and bottom, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. The terms “comprises,” “comprising,” “has”, “having,” “includes”, “including,” “contains”, “containing” or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises, has, includes, contains a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by “comprises . . . a”, “has . . . a”, “includes . . . a”, “contains . . . a” does not, without more constraints, preclude the existence of additional identical elements in the process, method, article, or apparatus that comprises, has, includes, contains the element. The terms “a” and “an” are defined as one or more unless explicitly stated otherwise herein. The terms “substantially”, “essentially”, “approximately”, “about” or any other version thereof, are defined as being close to as understood by one of ordinary skill in the art, and in one non-limiting embodiment the term is defined to be within 10%, in another embodiment within 5%, in another embodiment within 1% and in another embodiment within 0.5%. The term “coupled” as used herein is defined as connected, although not necessarily directly and not necessarily mechanically. A device or structure that is “configured” in a certain way is configured in at least that way, but may also be configured in ways that are not listed.

It will be appreciated that some embodiments may be comprised of one or more specialized processors (or “processing devices”) such as microprocessors, digital signal processors, customized processors and field programmable gate arrays (FPGAs) and unique stored program instructions (including both software and firmware) that control the one or more processors to implement, in conjunction with certain non-processor circuits, some, most, or all of the functions of the method and/or apparatus described herein. Alternatively, some or all functions could be implemented by a state machine that has no stored program instructions, or in one or more application specific integrated circuits (ASICs), in which each function or some combinations of certain of the functions are implemented as custom logic. Of course, a combination of the two approaches could be used.

Moreover, an embodiment can be implemented as a computer-readable storage medium having computer readable code stored thereon for programming a computer (e.g., comprising a processor) to perform a method as described and claimed herein. Examples of such computer-readable storage mediums include, but are not limited to, a hard disk, a CD-ROM, an optical storage device, a magnetic storage device, a ROM (Read Only Memory), a PROM (Programmable Read Only Memory), an EPROM (Erasable Programmable Read Only Memory), an EEPROM (Electrically Erasable Programmable Read Only Memory) and a Flash memory. Further, it is expected that one of ordinary skill, notwithstanding possibly significant effort and many design choices motivated by, for example, available time, current technology, and economic considerations, when guided by the concepts and principles disclosed herein will be readily capable of generating such software instructions and programs and ICs with minimal experimentation.

The Abstract of the Disclosure is provided to allow the reader to quickly ascertain the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. In addition, in the foregoing Detailed Description, it can be seen that various features are grouped together in various embodiments for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the claimed embodiments require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed embodiment. Thus the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as a separately claimed subject matter.

Claims

1. A method by an imaging controller of detecting item facings from image sensor data, the method comprising: obtaining, at the imaging controller, the image sensor data corresponding to a support structure containing at least one item;identifying, by a feature detector of the imaging controller, a set of matched keypoint pairs from keypoints of the image sensor data;determining, by a peak detector of the imaging controller, a separation distance between the keypoints of each matched keypoint pair;detecting, by the peak detector, a count of item instances represented in the image sensor data based on the separation distances; andpresenting item facing detection output including the count of item instances.

2. The method of claim 1, further comprising: obtaining point cloud data corresponding to the image sensor data;wherein determining the separation distance includes registering the matched keypoint pairs to the point cloud data.

3. The method of claim 1, wherein identifying the matched keypoint pairs comprises: generating, by the feature detector of the imaging controller, respective feature descriptors for a plurality of keypoints from the image sensor data; andcomparing the feature descriptors.

4. The method of claim 3, wherein identifying the matched keypoint pairs further comprises, for each of a plurality of keypoint pairs: determining a difference metric; andwhen the difference metric is below a threshold, identifying the keypoint pair as a matched keypoint pair.

5. The method of claim 1, further comprising: allocating, by a boundary generator of the imaging controller, each keypoint to a cluster; andgenerating an item boundary corresponding to the cluster;wherein the item facing detection output includes the item boundary.

6. The method of claim 5, wherein the allocating comprises: assigning an affinity value to each of a plurality of pairs of the keypoints based on (i) whether the pair is a matched keypoint pair, and (ii) the separation distance between the pair.

7. The method of claim 1, wherein detecting the count of item instances comprises: for each separation distance, incrementing a corresponding one of a set of histogram scores; anddetecting at least one peak in the set of histogram scores.

8. The method of claim 1, wherein the separation distance is a separation distance in a first physical dimension associated with the support structure.

9. The method of claim 8, further comprising: determining, by the peak detector, a further separation distance for each matched keypoint pair in a second physical dimension associated with the support structure; anddetecting, by the peak detector, a count of item instances represented in the image sensor data based on the separation distances in the first and second physical dimensions.

10. A computing device, comprising: a feature detector configured to: obtain image sensor data corresponding to a support structure containing at least one item; andidentify a set of matched keypoint pairs from keypoints of the image sensor data;a peak detector configured to: determine a separation distance between the keypoints of each matched keypoint pair;detect a count of item instances represented in the image sensor data based on the separation distances; anda boundary generator configured to present item facing detection output including the count of item instances.

11. The computing device of claim 10, wherein the feature detector is further configured to: obtain point cloud data corresponding to the image sensor data; andin order to determine the separation distance, register the matched keypoint pairs to the point cloud data.

12. The computing device of claim 10, wherein the feature detector is configured, in order to identify the matched keypoint pairs, to: generate respective feature descriptors for a plurality of keypoints from the image sensor data; andcompare the feature descriptors.

13. The computing device of claim 12, wherein the feature detector is further configured, in order to identify the matched keypoint pairs, to: for each of a plurality of keypoint pairs, determine a difference metric; andwhen the difference metric is below a threshold, identify the keypoint pair as a matched keypoint pair.

14. The computing device of claim 10, wherein the boundary generator is further configured to: allocate each keypoint to a cluster; andgenerate an item boundary corresponding to the cluster;wherein the item facing detection output includes the item boundary.

15. The computing device of claim 14, wherein the boundary generator is further configured, in order to allocate each keypoint to a cluster, to: assign an affinity value to each of a plurality of pairs of the keypoints based on (i) whether the pair is a matched keypoint pair, and (ii) the separation distance between the pair.

16. The computing device of claim 10, wherein the peak detector is further configured, in order to detect the count of item instances, to: for each separation distance, increment a corresponding one of a set of histogram scores; anddetect at least one peak in the set of histogram scores.

17. The computing device of claim 10, wherein the separation distance is a separation distance in a first physical dimension associated with the support structure.

18. The computing device of claim 17, wherein the peak detector is further configured to: determine a further separation distance for each matched keypoint pair in a second physical dimension associated with the support structure; anddetect a count of item instances represented in the image data based on the separation distances in the first and second physical dimensions.

19. A non-transitory computer-readable medium storing computer executable instructions that when executed by a processor cause the processor to perform the operations of: obtaining image sensor data corresponding to a support structure containing at least one item;identifying a set of matched keypoint pairs from keypoints of the image sensor data;determining a separation distance between the keypoints of each matched keypoint pair;detecting a count of item instances represented in the image sensor data based on the separation distances; andpresenting item facing detection output including the count of item instances.

20. The non-transitory computer-readable medium of claim 19, wherein the instructions further comprise: for each separation distance, incrementing a corresponding one of a set of histogram scores; anddetecting at least one peak in the set of histogram scores.