Process of and apparatus for identifying livestock animals

Abstract

A process of identifying livestock animals in a livestock habitat (12). The process comprises a training phase and an operative phase which follows the training phase. The training phase comprises acquiring plural radio frequency identification (RFID) code and image pairs for each of plural livestock animals, each RFID code acquired by reading with an RFID reader (18) a passive RFID tag (20) attached to the respective livestock animal, and each image of the respective livestock animal acquired by a camera (16). The training phase also comprises forming an image classification model for each of the plural livestock animals in dependence on the plural RFID code and image pairs for the respective livestock animal. The training phase further comprises linking the RFID code of the plural RFID code and image pairs for each of the plural livestock animals to the image classification model formed for the respective livestock animal. The operative phase comprises acquiring an image of a livestock animal of the plural livestock animals, the image acquired by a camera (16) installed in the livestock habitat. The operative phase also comprises acquiring an RFID code from the livestock animal, the RFID code acquired by reading with an RFID reader 18 the passive RFID tag (20) attached to the livestock animal. The operative phase further comprises making a first identification by applying the image of the livestock animal to at least one of the image classification models, making a second identification in dependence on the acquired RFID code, and selecting one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other.

Description

[0001] PROCESS OF AND APPARATUS FOR IDENTIFYING LIVESTOCK ANIMALS

[0002] Field of the Invention

[0003] The present invention relates to processes of identifying livestock animals, in particular but not exclusively by reading of an RFID code from an RFID tag attached to each livestock animal. The present invention also relates to apparatus for identifying livestock animals comprising an RFID tag attached to each livestock animal, wherein an RFID code is read from each RFID tag.

[0004] Background Art

[0005] It is known to identify livestock, such as cattle, by way of a passive Radio-frequency identification (RFID) tag attached to each livestock animal. Each livestock animal is identified when its passive RFID tag is read by an RFID reader, such as when the livestock animal passes through a gate, is at a feeding station, or enters a milking parlour. Reading of each RFID tag yields a unique code which is used to address a record corresponding to the identified livestock animal. Records for livestock animals may contain the like of milk yield data, feeding data, and physiological data.

[0006] The known RFID approach to livestock identification is prone to error. For example, the RFID tag of a livestock animal adjacent an intended livestock animal may be read in error instead of the RFID tag of the intended livestock animal whereby the adjacent livestock animal is identified twice, i.e. once in error and once properly, and the intended livestock animal is missed. By way of further example, the RFID tag of a livestock animal may be misread. The present inventors have become appreciative that a minimum percentage of successful livestock animal identifications is needed in certain circumstances with the known RFID approach falling short of achieving the minimum percentage.

[0007] It is therefore an object for the present invention to provide an improved process of identifying livestock animals in dependence on acquisition of an RFID code from each livestock animal, the RFID code acquired by reading with an RFID reader an RFID tag attached to the livestock animal. It is a further object for the present invention to provide an improved apparatus for identifying livestock animals comprising an RFID tag attached to each livestock animal and an RFID reader, the RFID reader reading an RFID code from each RFID tag.

[0008] Statement of Invention

[0009] According to a first aspect of the present invention there is provided a process of identifying livestock animals in a livestock habitat, the process comprising a training phase and an operative phase, the training phase comprising operating a processor to: control acquisition of plural radio frequency identification (RFID) code and image pairs for each of plural livestock animals, each RFID code acquired by reading with an RFID reader a passive RFID tag attached to the respective livestock animal, and each image of the respective livestock animal acquired by a camera; form an image classification model for each of the plural livestock animals in dependence on the plural RFID code and image pairs for the respective livestock animal; and link the RFID code of the plural RFID code and image pairs for each of the plural livestock animals to the image classification model formed for the respective livestock animal, the operative phase following the training phase and comprising operating the processor to: control acquisition of an image of a livestock animal of the plural livestock animals, the image acquired by a camera installed in the livestock habitat; control acquisition of an RFID code from the livestock animal, the RFID code acquired by reading with an RFID reader the passive RFID tag attached to the livestock animal; make a first identification by applying the image of the livestock animal to at least one of the image classification models; make a second identification in dependence on the acquired RFID code; and select one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other.

[0010] The process of identifying livestock animals, such as cattle, in a livestock habitat, such as a barn, comprises a training phase and an operative phase. The training phase comprises operating a processor to control acquisition of plural radio frequency identification (RFID) code and image pairs for each of plural livestock animals. Each RFID code is acquired by reading with an RFID reader a passive RFID tag attached to the respective livestock animal, such as a passive RFID tag attached to an ear of the respective cow. Each of the passive RFID tags attached to the plural livestock animals yields a unique RFID code whereby the plural livestock animals can be distinguished from one another. More specifically, each RFID code may be unique amongst the RFID codes yielded by the RFID tags attached to the plural livestock animals.

[0011] Each image of the respective livestock animal is acquired by a camera. The RFID reader and camera used during the training phase may also be used during the operative phase. Alternatively, the RFID reader and camera used during the training phase may differ from the RFID reader and camera used during the operative phase. Furthermore, the RFID reader and camera may be used during the operative phase in an environment different to the livestock habitat of the operative phase. The present inventors have found that at least five and more preferably at least ten RFID code and image pairs may constitute a training data set of sufficient number to provide for accurate livestock animal identification during the operative phase and before whatever further training there may be during the operative phase.

[0012] The training phase further comprises operating a processor to form an image classification model, such as a convolutional neural network model as described below, for each of the plural livestock animals and in dependence on the plural RFID code and image pairs for the respective livestock animal. The image classification model may be formed from the images in the plural RFID code and image pairs. The RFID code in the plural RFID code and image pairs may provide for identification of the respective one of the image classification models as the training phase progresses. The plural image classification models may be constituted separately with each of the plural image classification models having a single output layer. Alternatively, at least two of the plural image classification models may be constituted as a composite image classification model having plural output layers with each of the plural output layers corresponding to a respective livestock animal.

[0013] The training phase further comprises operating a processor to link the RFID code of the plural RFID code and image pairs for each of the plural livestock animals to the image classification model formed for the respective livestock animal. The present step may be performed at the outset of the step of forming the image classification model. The RFID code for each of the plural livestock animals may, for example, be linked to the respective image classification model by a pointer between the RFID code and the image classification model or part thereof. Where there is a composite image classification model, the RFID code for each of the plural livestock animals may, for example, be linked to the composite image classification model by a pointer between the RFID code and the respective output layer of the composite image classification model.

[0014] When the training phase is complete, the process of identifying livestock animals progresses to the operative phase. The operative phase comprises operating the processor to control acquisition of an image of a livestock animal of the plural livestock animals. The image is acquired by a camera installed in the livestock habitat, such as a camera installed in an overhead location adjacent a milking parlour where the livestock animals are cattle.

[0015] The operative phase also comprises operating the processor to control acquisition of an RFID code from the livestock animal. The RFID code is acquired by reading with an RFID reader the passive RFID tag attached to the livestock animal.

[0016] The operative phase further comprises operating the processor to make a first identification by applying the image of the livestock animal to at least one of the image classification models. The step of making the first identification may comprise applying the image to a first one of the image classification models. If the first one of the image classification models returns a positive outcome, the first identification may correspond to the livestock animal modelled by the first one of the image classification models. If the first one of the image classification models returns a negative outcome, the image may be applied to at least one further image classification model until an image classification model returns a positive outcome. The first identification may correspond to the livestock animal modelled by the image classification model returning a positive outcome. Alternatively, the image may be applied to all of the image classification models whereby the first identification may correspond to the livestock animal modelled by the image classification model which returns a positive outcome. Where there is a composite image classification model, one of the output layers of the composite image classification model may return a positive outcome. For example, if the composite image classification model models all of the plural livestock animals, one of the output layers should return a positive outcome. By way of further example, if the composite image classification model models fewer than all of the plural livestock animals and if none of the output layers returns a positive outcome, the image may be applied to a further composite image classification model or image classification model until a positive outcome is returned.

[0017] The first identification may correspond to the livestock animal modelled by the image classification model which has returned a positive outcome. The first identification may comprise the RFID code linked to the image classification model or the output layer of a composite image classification model which has returned a positive outcome.

[0018] The operative phase further comprises operating the processor to make a second identification in dependence on the acquired RFID code. The second identification may correspond to the livestock animal linked to the acquired RFID code. The second identification may comprise the acquired RFID code.

[0019] The operative phase yet further comprises operating the processor to select one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other. Making of the second identification corresponds to the known process described in the background to the invention. The present process comprises making a further identification in the form of the first identification which depends on images of livestock animals and as described above. Selection of one of the first and second identifications as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other addresses the aforementioned problems with the known process.

[0020] The step of selecting one of the first and second identifications as the identity of the livestock animal may comprise selecting the first identification as the identity of the livestock animal if the first and second identifications correspond with each other. Alternatively, the second identification may be selected as the identity of the livestock animal because selection of the second identification has the same outcome as selection of the first identification in view of the first and second identifications corresponding with each other. Correspondence between the first and second identifications may indicate a high degree of confidence of the livestock animal being identified correctly. Where each of the first and second identifications comprises an RFID code, one or other of the first and second identifications may be selected as the identity when the RFID codes are the same.

[0021] Furthermore, the process may further comprise re-entering the calibration phase and augmenting the image classification model with the image acquired during the operative phase when the first and second identifications correspond with each other. The image classification model may thus be improved on an ongoing basis.

[0022] If the first and second identifications do not correspond with each other, the step of selecting one of the first and second identifications as the identity of the livestock animal may comprise one of the following approaches.

[0023] According to one approach, the step of selecting one of the first and second identifications as the identity of the livestock animal may comprise selecting the second identification as the identity of the livestock animal. For example, where the second identification comprises the RFID code acquired by reading the passive RFID tag, the livestock animal may be identified by the acquired RFID code. There may be no re-entering of the calibration phase, in contrast to doing so as described above when there is correspondence between the first and second identifications, in view of there being a lower degree of confidence of the livestock animal being identified correctly.

[0024] According to another approach, the step of selecting one of the first and second identifications as the identity of the livestock animal may comprise applying the acquired image to the image classification model linked to the acquired RFID code. The step may further comprise determining an extent of dissimilarity between the acquired image and the linked image classification model and selecting the first identification as the identity of the livestock animal if the extent of dissimilarity is greater than a dissimilarity threshold. There may be re-entering of the calibration phase, as described above when there is correspondence between the first and second identifications, in view of there being a reasonable degree of confidence of the livestock animal being identified correctly.

[0025] Otherwise, i.e. if the extent of dissimilarity is not greater than the dissimilarity threshold, the step may comprise selecting the second identification as the identity of the livestock animal. There may be no re-entering of the calibration phase, in contrast to doing so as described above when there is correspondence between first and second identifications, in view of there being a lower degree of confidence of the livestock animal being identified correctly.

[0026] According to another approach, selection of one of the first and second identifications as the identity of the livestock animal may be in dependence on at least one of an RFID code and an image acquired previously for a different livestock animal during the operational phase. More specifically, the step of selecting one of the first and second identifications may comprise comparing the currently acquired RFID code with an RFID code acquired previously during the operational phase. The step may further comprise selecting the first identification as the identity of the livestock animal if the acquired RFID code and the previously acquired RFID code are the same. More specifically, the first identification may be selected as the identity of the livestock animal if the acquired RFID code and the previously acquired RFID code are the same and if there was correspondence between first and second identifications for the livestock animal which was identified previously by way of the previously acquired RFID code. The previously identified livestock animal was identified to a higher degree of confidence in view of the correspondence between the first and second identifications whereby there may be a corresponding higher degree of confidence that the currently acquired RFID code belongs to a different livestock animal to the livestock animal of the currently acquired image. The first identification may therefore be selected in preference to the second identification as the identity of the livestock animal.

[0027] Furthermore, the previously acquired RFID code may, for example, have been acquired at a location nearby the location of acquisition of the currently acquired RFID code, such as from one of the passive RFID tags attached to cattle in adjacent stalls in a milking parlour. Selection of one of plural previously acquired RFID codes for comparison with the acquired RFID code may be in dependence on locations of acquisition of previously acquired RFID codes relative to the location of acquisition of the currently acquired RFID code.

[0028] Selection of one of the plural previously acquired RFID codes for comparison with the currently acquired RFID code may be in dependence on a spatial map of at least part of the livestock habitat. Locations of the plural previously acquired RFID codes relative to the currently acquired RFID code may be determined in dependence on the spatial map. For example, the spatial map may represent locations of plural stalls in a milking parlour and each of the plural previously acquired RFID codes and the currently acquired RFID code may be associated with a different one of the plural stalls. Locations of the plural previously acquired RFID codes relative to the currently acquired RFID code may be determined in dependence on the spatial map and further in dependence on the images linked to the plural previously acquired RFID codes and the currently acquired RFID code. More specifically, one or more features of the livestock habitat represented in each of the images may be correlated with the spatial map whereby the plural previously acquired RFID codes and the currently acquired RFID code may be located in the spatial map.

[0029] Reading of an RFID code from a passive RFID tag may fail. The process of identifying livestock animals may therefore further comprise waiting for a predetermined read time for the RFID reader to read an RFID code from a passive RFID tag. Furthermore, the process may further comprise identifying a livestock animal in dependence on an image of the livestock animal acquired by a camera if no RFID code is read before expiry of the predetermined read time. Identification of the livestock animal may thus be in dependence on the image alone and in view of failure to read an RFID code. There may be no re-entering of the calibration phase, in contrast to doing so as described above when there is correspondence between the first and second identifications, in view of there being a lower degree of confidence of the livestock animal being identified correctly.

[0030] Before the step of controlling acquisition of an image of a livestock animal and the step of controlling acquisition of an RFID code from the livestock animal, the process may comprise detecting the livestock animal. More specifically, detecting the livestock animal may comprise acquiring at least one image of part of the livestock habitat and detecting a livestock animal in the image, such as by way of an object recognition algorithm. Furthermore, detecting the livestock animal may comprise acquiring plural images of part of the livestock habitat and more specifically acquiring video of part of the livestock habitat. In addition, the plural images or video may be analysed by way of an object recognition algorithm to detect a livestock animal. When a livestock animal has been detected, the operative phase of the process of identifying livestock animals may commence.

[0031] The step of forming an image classification model for each of the plural livestock animals may comprise forming an artificial neural network model and more specifically a convolutional neural network (CNN) model. The convolutional neural network (CNN) model may comprise an embeddings cluster. Alternatively or in addition, the step of forming an image classification model may comprise similarity learning and more specifically supervised similarity learning, such as by way of a K- nearest neighbour algorithm (KNN).

[0032] The camera may be a two-dimensional camera. A two-dimensional camera may provide more superior performance, such as in respect of resolution, than a three- dimensional camera. The camera may be an Internet Protocol (IP) camera.

[0033] The process of identifying livestock animals may involve use of plural cameras. As described above, at least one camera may be operative during the calibration phase and a different at least one camera may be operative during the operative phase. Plural cameras may be operative during the operative phase. Furthermore, the plural cameras may be installed at different respective locations in the livestock habitat, such as above a respective stall in a milking parlour.

[0034] In a form of the invention, plural cameras may be installed in the livestock habitat, such as above a stall in a milking parlour, whereby the plural cameras have different viewpoints of the same livestock animal. Having plural viewpoints of the same livestock animal may provide for improved identification of the livestock animal.

[0035] The process of identifying livestock animals may involve use of plural RFID readers. Each of the plural RFID readers may be installed at a different location whereby the plural RFID readers are capable of reading plural RFID tags at the same time. The processor may comprise at least one digital processor. According to a first approach, a digital processor may be disposed at each of plural cameras whereby at least part of the process is carried out locally to a camera. The first approach may be applied together with at least one of the second and third approaches described below. According to a second approach, a digital processor may be at a location in the livestock habitation spaced apart from plural cameras, such as on the ground inside a bam, the processor receiving image data from all of the plural cameras and RFID codes. According to a third approach, a digital processor may be located remotely from the livestock habitation, the digital processor receiving image data from all of plural cameras and RFID codes. Furthermore, image data from the plural cameras and the RFID codes may be conveyed to the remotely located digital processor by way of the Internet. The digital processor may be constituted by a cloud computing arrangement. The digital processor may therefore be of distributed form. The second and third approaches may be applied together with, for example, preliminary processing according to the second approach and completion of processing according to the third approach.

[0036] The processor is configured to perform one or more of the processes described herein. Apparatus may comprise structures and / or non-transitory memory having programmed instructions which are operated upon by the processor and may further comprise electronic circuitry to perform these processes.

[0037] The livestock habitat may be at least one of an indoor habitat and an outdoor habitat. For example, and where the livestock habitat is an indoor habitat, the livestock habitat may be a bam and more specifically a barn containing a milking parlour. Nevertheless, application of the present process may not be limited to cattle and may include livestock animals more generally where identification of livestock animals provides for monitoring of feeding, physiological health and the like.

[0038] The livestock animals may comprise poultry, such as hens and turkeys, and quadrupedal livestock animals, such as cattle, pigs or sheep. For certain applications and in view of greater need for record keeping for individual quadrupedal livestock animals, the livestock animals may be quadrupedal livestock animals and more specifically cattle.

[0039] The process of identifying livestock animals may further comprise modifying a record for the livestock animal identified by the step of selecting one of the first identification and the second identification as the identity of the livestock animal. The record for the livestock animal may be addressed by an RFID code for the identified livestock animal. The record may contain data pertaining to the livestock animal, such as the like of feeding and physiological health, and milk yield data where the livestock animal is a cow.

[0040] According to a second aspect of the present invention there is provided an apparatus for identifying livestock animals in a livestock habitat, the apparatus comprising: plural passive RFID tags each attached to a respective one of plural livestock animals; at least one RFID reader which reads an RFID code from each passive RFID tag; at least one camera which acquires an image of each of the plural livestock animals; and a processor which is configured to control a training phase of operation and an operative phase of operation which follows the training phase of operation, wherein the training phase comprises: acquiring plural radio frequency identification (RFID) code and image pairs for each of the plural livestock animals, each RFID code acquired by reading with the RFID reader the passive RFID tag attached to the respective livestock animal, and each image of the respective livestock animal acquired by a camera; forming an image classification model for each of the plural livestock animals in dependence on the plural RFID code and image pairs for the respective livestock animal; and linking the RFID code of the plural RFID code and image pairs for each of the plural livestock animals to the image classification model formed for the respective livestock animal, and wherein the operative phase comprises: acquiring an image of a livestock animal of the plural livestock animals, the image acquired by a camera installed in the livestock habitat; acquiring an RFID code from the livestock animal, the RFID code acquired by reading with the RFID reader the passive RFID tag attached to the livestock animal; making a first identification by applying the image of the livestock animal to at least one of the image classification models; making a second identification in dependence on the acquired RFID code; and selecting one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other.

[0041] Embodiments of the second aspect of the present invention may comprise one or more features of the first aspect of the present invention.

[0042] According to another aspect of the present invention there is provided a process of identifying livestock animals in a livestock habitat, the process comprising a training phase and an operative phase, the training phase comprising operating a processor to: control acquisition of plural radio frequency identification (RFID) code and image pairs for each of plural livestock animals, each RFID code acquired by reading with an RFID reader an RFID tag attached to the respective livestock animal, and each image of the respective livestock animal acquired by a camera; and form an image classification model for each of the plural livestock animals in dependence on the plural RFID code and image pairs for the respective livestock animal, the operative phase following the training phase and comprising operating the processor to: control acquisition of an image of a livestock animal of the plural livestock animals, the image acquired by a camera installed in the livestock habitat; control acquisition of an RFID code from the livestock animal, the RFID code acquired by reading with an RFID reader the RFID tag attached to the livestock animal; make a first identification by applying the image of the livestock animal to at least one of the image classification models; make a second identification in dependence on the acquired RFID code; and select one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other.

[0043] Embodiments of the present aspect of the present invention may comprise one or more features of the first aspect of the present invention.

[0044] According to another aspect of the present invention there is provided an apparatus for identifying livestock animals in a livestock habitat, the apparatus comprising: plural RFID tags each attached to a respective one of plural livestock animals; at least one RFID reader which reads an RFID code from each RFID tag; at least one camera which acquires an image of each of the plural livestock animals; and a processor which is configured to control a training phase of operation and an operative phase of operation which follows the training phase of operation, wherein the training phase comprises: acquiring plural radio frequency identification (RFID) code and image pairs for each of the plural livestock animals, each RFID code acquired by reading with the RFID reader the RFID tag attached to the respective livestock animal, and each image of the respective livestock animal acquired by a camera; and forming an image classification model for each of the plural livestock animals in dependence on the plural RFID code and image pairs for the respective livestock animal, and wherein the operative phase comprises: acquiring an image of a livestock animal of the plural livestock animals, the image acquired by a camera installed in the livestock habitat; acquiring an RFID code from the livestock animal, the RFID code acquired by reading with the RFID reader the RFID tag attached to the livestock animal; making a first identification by applying the image of the livestock animal to at least one of the image classification models; making a second identification in dependence on the acquired RFID code; and selecting one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other.

[0045] Embodiments of the present aspect of the present invention may comprise one or more features of the first aspect of the present invention. Brief Description of Drawings

[0046] Further features and advantages of the present invention will become apparent from the following specific description, which is given by way of example only and with reference to the accompanying drawings, in which:

[0047] Figure 1 is a block diagram representation of an apparatus according to an embodiment of the invention; and

[0048] Figure 2 is a flow chart representation of a process according to the invention;

[0049] Description of Embodiments

[0050] Apparatus 10 for identifying livestock animals and according to an embodiment of the present invention is shown in Figure 1 . The apparatus 10 is installed in a bam (which constitutes a livestock habitat) which contains a milking parlour 12. The milking parlour 12 comprises plural stalls 14. For simplicity, Figure 1 shows a small number of stalls 14. In a practical embodiment, the milking parlour 12 is of known herringbone or circular form and has a larger number of stalls. The apparatus comprises plural cameras 16 with each camera 16 installed above a respective stall 14 and oriented downwards whereby the camera acquires images of the stall. The cameras 16 are two-dimensional IP cameras of conventional form and function. The apparatus 10 also comprises plural RFID readers 18 with each RFID reader installed in or adjacent a respective stall 14. The apparatus 10 further comprises plural passive RFID tags 20 with each RFID tag attached to a respective cow (not shown) of a herd of cattle (which constitutes plural livestock animals). The RFID readers 18 and the RFID tags 20 are of conventional form and function. The apparatus 10 yet further comprises a central digital processor 22 of conventional form and function which is installed in a corner of the barn and which is in communication with the cameras 16 and the RFID readers 18. In a form of the embodiment, two cameras 16 are installed above each stall 14 in the milking parlour 12 and such that one of the two cameras has an overhead view of the livestock animal and the other of the two cameras has a perspective view of the livestock animal from behind and above the livestock animal. Having such two different viewpoints of the livestock animal may provide for improved identification of the livestock animal.

[0051] In a first form of the embodiment, image and data processing steps according to the invention are performed by a digital processor local to each camera 16. The local digital processor is a Jetson Xavier NX from NVIDIA, Trinity House, Cambridge Business Park, Cowley Road, Cambridge, CB4 OWZ, United Kingdom. Later steps according to the invention, such as updating of a record for the identified cow or analysis of fresh data for individual or plural cows, are performed in the central digital processor 22. In a second form of the embodiment, the image apparatus 10 comprises a cloud computing resource (not shown) of conventional form and function. The second form is as per the first form but with the later steps performed in the cloud computing resource instead of in the central digital processor 22. In a third form of the embodiment, the local digital processor is operative merely as a relay station to convey bare image and other data, which has been generated locally, to the cloud computing resource by way of the Internet or to the central digital processor 22 wherein steps according to the invention are performed. In a fourth form of the embodiment, substantially all image and data processing steps according to the invention are performed in a distributed fashion by way of a Jetson Xavier NX digital processor (not shown) local to each camera 16 whereby later steps, such as updating of a record for the identified cow or analysis of fresh data, are performed in the local digital processors.

[0052] A flow chart representation of a process 30 according to the invention is shown in Figure 2. The process makes use of the apparatus shown in Figure 1 . The process begins with a training phase 32 during which at least ten RFID code and image pairs are acquired for each cow in a herd of cattle 34. The RFID codes in the RFID code and image pairs are acquired with the camera 16 and the RFID reader 18 of one of the stalls 14 shown in Figure 1 or with a camera and a RFID reader at a different location which is convenient for conducting the training phase. Thereafter, the at least ten RFID code and image pairs for each cow are used to form a supervised similarity learning convolutional neural network (CNN) model having an embeddings cluster for the cow by way of a K-nearest neighbour algorithm (KNN) 36 in accordance with known practice. A CNN model is formed for each cow in the herd based on the respective RFID code and image pairs. The training phase further comprises linking the RFID code of the plural RFID code and image pairs for each cow to the embeddings cluster formed for the cow 38. The RFID code is linked to the respective embeddings cluster by a pointer between the RFID code and the CNN model or part thereof. In a form of the embodiment, a composite CNN model is formed in accordance with known practice. The composite CNN model has plural output layers with each of the plural output layers corresponding to a respective cow. When the training phase is complete, the process 30 progresses to the operative phase 40.

[0053] The operative phase of the process 30 will now be described in respect of identification of a cow in one of the stalls 14 shown in Figure 1 . The camera 16 above the stall in question 14 acquires video of the stall. The processor 22 applies frames from the acquired video to a CNN model, such as a CNN model formed during the training phase, to detect when a cow is in the stall 14 and more specifically to detect when the cow is properly positioned in the stall 42. When proper positioning of the cow has been detected, the camera 16 acquires an image of the cow 44. The acquired image is applied to each of the plural CNN models formed during the training phase until one of the plural CNN models returns a positive outcome reflecting recognition of the cow as one of the herd of cattle 46. Alternatively, and where a composite CNN model is used, the acquired image is applied to the composite CNN model and the output layer which returns a positive outcome is registered to thereby reflect recognition of the cow as one of the herd of cattle. The RFID code linked to the CNN model which has returned the positive outcome or the RFID code linked to the output layer which has returned a positive outcome is recorded as a first identification 48. In the meantime, the RFID reader 18 reads an RFID code from the RFID tag 20 attached to the cow in the stall 14. If reading of the RFID code has failed upon expiry of a predetermined read time 50, then the first identification is recorded as the identity of the cow 52. If the RFID code has been read before expiry of the predetermined read time 50, then the read RFID code is recorded as a second identification 54. The process then progresses to comparing the first and second identifications to determine if they are the same 56. If the first and second identifications are the same, then either the first identification or the second identification is recorded as the identity of the cow 58. Then the training phase is re-entered and the acquired image is used to augment the CNN model linked to the read RFID code 60. If the first and second identifications are different, then a decision algorithm is used to decide on the identity of the cow 62. Depending on circumstances, one of the following three decision algorithm is used.

[0054] When a first decision algorithm is used, the second identification, i.e. the read RFID code, is recorded as the identity of the cow. There is no re-entering of the calibration phase in view of there being a lower degree of confidence of the cow being identified correctly.

[0055] When a second decision algorithm is used, the acquired image is applied to the CNN model linked to the read RFID code. In addition, an extent of dissimilarity between the acquired image and the linked CNN model is determined. The first identification is selected as the identity of the cow if the extent of dissimilarity is greater than a dissimilarity threshold. The calibration phase is re-entered, as described above, in view of there being a reasonable degree of confidence of the cow being identified correctly. Otherwise, i.e. if the extent of dissimilarity is not greater than the dissimilarity threshold, the second identification i.e. the read RFID code, is selected as the identity of the cow. There is no re-entering of the calibration phase in view of there being a lower degree of confidence of the cow being identified correctly.

[0056] When the third decision algorithm is used, the currently read RFID code is compared with an RFID code read previously during the operational phase from an RFID tag 20 attached to a cow in at least one adjacent stall 14 in the milking parlour 12. The location of one or more adjacent RFID tags 20 are identified in dependence on a spatial map of the milking parlour 20. The spatial map is formed by combining a map of the milking parlour 20 with at least one feature in an acquired image of each stall 14 which has been recognised by an object recognition algorithm of conventional form and function. The at least one feature in each stall 14 is a feature of a cow in the stall 14 whereby the location of the RFID tag 20 attached to the cow is determined. The first identification is selected as the identity of the cow if the currently read RFID code and one adjacent previously read RFID code are the same and if there is identity of first and second identifications for the cow which was identified previously by way of the previously read RFID code. The previously identified cow was identified to a higher degree of confidence in view of the identity between the first and second identifications whereby there is a corresponding higher degree of confidence that the currently read RFID code belongs to a different cow to the cow of the currently acquired image. The first identification is therefore selected in preference to the second identification as the identity of the cow.

[0057] When the identity of the cow has been determined by one of the three decision algorithms described above, a record for the identified cow is addressed by way of the linked RFID code when the first identification is selected and the read RFID code when the second identification is selected. The record contains data pertaining to the cow, such as the like of feeding data, physiological health data and milk yield data. The addressed record is then modified to take account of freshly acquired data pertaining to the cow.

[0058] The operational phase described above is carried out for each cow in each stall 14 of the milking parlour 12.

Claims

Claims1 . A process of identifying livestock animals in a livestock habitat, the process comprising a training phase and an operative phase, the training phase comprising operating a processor to: control acquisition of plural radio frequency identification (RFID) code and image pairs for each of plural livestock animals, each RFID code acquired by reading with an RFID reader a passive RFID tag attached to the respective livestock animal, and each image of the respective livestock animal acquired by a camera; form an image classification model for each of the plural livestock animals in dependence on the plural RFID code and image pairs for the respective livestock animal; and link the RFID code of the plural RFID code and image pairs for each of the plural livestock animals to the image classification model formed for the respective livestock animal, the operative phase following the training phase and comprising operating the processor to: control acquisition of an image of a livestock animal of the plural livestock animals, the image acquired by a camera installed in the livestock habitat; control acquisition of an RFID code from the livestock animal, the RFID code acquired by reading with an RFID reader the passive RFID tag attached to the livestock animal; make a first identification by applying the image of the livestock animal to at least one of the image classification models; make a second identification in dependence on the acquired RFID code; and select one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other.

2. The process according to claim 1 , in which each image classification model is formed from the images in the plural RFID code and image pairs.

3. The process according to claim 1 or 2, in which the step of making the first identification comprises applying the image to a first one of the image classificationmodels, if the first one of the image classification models returns a positive outcome, the first identification corresponds to the livestock animal modelled by the first one of the image classification models, and if the first one of the image classification models returns a negative outcome, the image is applied to at least one further image classification model until an image classification model returns a positive outcome.

4. The process according to any preceding claim, in which the first identification corresponds to the livestock animal modelled by the image classification model which has returned a positive outcome, the first identification comprising the RFID code linked to the image classification model or an output layer of a composite image classification model which has returned the positive outcome.

5. The process according to any preceding claim, in which the second identification corresponds to the livestock animal linked to the acquired RFID code, the second identification comprising the acquired RFID code.

6. The process according to any preceding claim, in which the step of selecting one of the first and second identifications as the identity of the livestock animal comprises selecting the first identification as the identity of the livestock animal if the first and second identifications correspond with each other.

7. The process according to claim 6, in which each of the first and second identifications comprises an RFID code, and the first identification is selected as the identity when the RFID codes are the same.

8. The process according to claim 6 or 7 further comprising re-entering the calibration phase and augmenting the image classification model with the image acquired during the operative phase when the first and second identifications corresponded with each other.

9. The process according to any one of claims 1 to 5, in which the first and second identifications do not correspond with each other, and the step of selectingone of the first and second identifications as the identity of the livestock animal comprises selecting the second identification as the identity of the livestock animal.

10. The process according to any one of claims 1 to 5, in which the first and second identifications do not correspond with each other, and the step of selecting one of the first and second identifications as the identity of the livestock animal comprises: applying the acquired image to the image classification model linked to the acquired RFID code; and determining an extent of dissimilarity between the acquired image and the linked image classification model.11 . The process according to claim 10, in which the step of selecting one of the first and second identifications as the identity of the livestock animal comprises selecting the first identification as the identity of the livestock animal if the extent of dissimilarity is greater than a dissimilarity threshold.

12. The process according to claim 11 further comprising re-entering the calibration phase and augmenting the image classification model with the image acquired during the operative phase.

13. The process according to claim 10, in which the step of selecting one of the first and second identifications as the identity of the livestock animal comprises selecting the second identification as the identity of the livestock animal if the extent of dissimilarity is not greater than a dissimilarity threshold.

14. The process according to claim 10, in which the step of selecting one of the first and second identifications as the identity of the livestock animal comprises: comparing the currently acquired RFID code with at least one RFID code acquired previously during the operational phase; and selecting the first identification as the identity of the livestock animal if the acquired RFID code and the previously acquired RFID code are the same and if there was correspondence between first and second identifications for the livestock animal which was identified previously by way of the previously acquired RFID code.

15. The process according to claim 14, in which selection of one of plural previously acquired RFID codes for comparison with the currently acquired RFID code is in dependence on a spatial map of at least part of the livestock habitat, locations in the livestock habitat of the plural previously acquired RFID codes relative to a location in the livestock habitat of the currently acquired RFID code being determined in dependence on the spatial map.

16. The process according to any one of the preceding claims, and before the steps of controlling acquisition of an image of and an RFID code from the livestock animal, in which the process comprises detecting the livestock animal, the livestock animal detected by way of an object recognition algorithm.

17. The process according to any one of the preceding claims, in which the step of forming an image classification model for each of the plural livestock animals comprises forming a convolutional neural network (CNN) model.

18. The process according to claim 17, in which the convolutional neural network (CNN) model comprises an embeddings cluster.

19. The process according to any one of the preceding claims, in which the camera is a two-dimensional camera.

20. The process according to any one of the preceding claims, in which at least one camera is operative during the calibration phase and a different at least one camera is operative during the operative phase.21 . The process according to any one of the preceding claims further comprising modifying a record for the livestock animal identified by the step of selecting one of the first identification and the second identification as the identity of the livestock animal, the record for the livestock animal addressed by an RFID code for the identified livestock animal, the record containing data pertaining to the identified livestock animal.

22. An apparatus for identifying livestock animals in a livestock habitat, the apparatus comprising: plural passive RFID tags each attached to a respective one of plural livestock animals; at least one RFID reader which reads an RFID code from each passive RFID tag; at least one camera which acquires an image of each of the plural livestock animals; and a processor which is configured to control a training phase of operation and an operative phase of operation which follows the training phase of operation, wherein the training phase comprises: acquiring plural radio frequency identification (RFID) code and image pairs for each of the plural livestock animals, each RFID code acquired by reading with the RFID reader the passive RFID tag attached to the respective livestock animal, and each image of the respective livestock animal acquired by a camera; forming an image classification model for each of the plural livestock animals in dependence on the plural RFID code and image pairs for the respective livestock animal; and linking the RFID code of the plural RFID code and image pairs for each of the plural livestock animals to the image classification model formed for the respective livestock animal, and wherein the operative phase comprises: acquiring an image of a livestock animal of the plural livestock animals, the image acquired by a camera installed in the livestock habitat; acquiring an RFID code from the livestock animal, the RFID code acquired by reading with the RFID reader the passive RFID tag attached to the livestock animal; making a first identification by applying the image of the livestock animal to at least one of the image classification models; making a second identification in dependence on the acquired RFID code; and selecting one of the first identification and the second identification as the identity of the livestock animal in dependence on comparison of the first and second identifications with each other.

Images