System and method for determining at least one body parameter of a bovine subject based on range imaging

Abstract

The present disclosure relates to determination of at least one body parameter of a bovine subject in a shed environment based on vision technology, in particular range imaging. The body parameter can be any of the weight, the predicted slaughter (carcass) weight, the conformation class, the predicted carcass conformation class, the fat cover class, and the predicted carcass fat cover class of the bovine subject. A first embodiment relates to a method for determining at least one body parameter of a bovine subject, the method comprising the steps of acquiring at least one range image of at least one bovine subject, processing the at least one range image to generate at least one 3D image of a preselected bovine subject, segmenting the at least one 3D image to generate at least one point cloud representing the preselected bovine subject, and calculating the body parameter of said preselected bovine subject by modelling the at least one point cloud against a trained model employing artificial intelligence (AI), wherein the body parameter is selected from the group of: the weight, the predicted slaughter (carcass) weight, the predicted carcass conformation class, and the predicted carcass fat cover class.

Description

[0001] System and method for determining at least one body parameter of a bovine subject based on range imaging

[0002] The present disclosure relates to determination of at least one body parameter of a bovine subject in a shed environment based on vision technology, in particular range imaging 13D imaging. The body parameter can be any of the weight, the predicted slaughter (carcass) weight, the conformation class, the predicted carcass conformation class, the fat cover class, and the predicted carcass fat cover class of the bovine subject.

[0003] Background

[0004] When running a large agriculture with a large number of individual animals it is a challenge to constantly be aware of each and every animal’s well-being. The wellbeing of the animal is very important, both from the humane perspective of the farmer having the responsibility for the well-being of these animals, but also since healthy, well-fed animals ensure a greater productivity and are of greater value to the farmer. The wellbeing of the animal is highly correlated with the weight of the animal, especially the weight stability of a grown animal and the weight gain of an animal still growing. Hence, it is of great interest to frequently monitor the weight of an animal.

[0005] The process of weighing a farm animal is typically provided by physically placing an animal on a weight scale, and either manually or digitally noting the weight of that particular animal. This is often a laborious and long process as modern farms has a large number of animals, which means that each animal would have to be individually brought up on a weight scale, one by one, to log the individual weight of each animal. Furthermore, the animal will have to be identified in the process in order to ascribe the correct weight to the correct animal. This identification may for instance happen by the use of the identification tags on the animal, prolonging the process even further. Furthermore, if the animals are living in a shed or a field in which they are free to move around independently it is even more difficult to keep track of which animals have been weighed and which have not. Due to the complication of the process of weighing the animals, weighing is not in widespread use and the amount of data available when evaluating the weight development of each animal is limited or even absent. WO 2010 / 127023 describes a contactless system and method for estimating the volume, mass or weight of an animal. Preferably, the animal is imaged with a stereoscopic camera, and a 3D representation of the target animal is derived from the stereoscopic images. A software module is provided to reshape the virtual model, using a set of independently configurable shape variables, to approximately fit the spatial representation of the individual animal. Mass or weight of the animal is estimated as a function of shape variables characterizing the reshaped virtual model.

[0006] WO 2013 / 184321 discloses an image based method for estimating a weight of an animal. The image is compared to a plurality of models to select a best matching model in terms of size or shape of the animal, wherein each of the models has a known weight. By adjusting 1) the image relative to the selected model or 2) the selected model relative to the image a differential adjustment parameter s determined, and a weight of the animal is determined by adjusting the known weight of the selected model based upon the differential adjustment parameter.

[0007] WO 2014 / 026765 discloses an approach based on mobile 3D cameras for determining parameters about a livestock such as weight, portion weight or lean meat proportion. In the method a 3D image of the livestock is recorded to produce a 3D model which is analysed on the basis of previously determined reference values, in order to output and / or store the information obtained from the analysed livestock.

[0008] WO 2015 / 156833 describes a method for estimating bodyweight by images of the animal. The system uses markers to characterize the weight, health, and other parameters of the animal. The system is configured to log these parameters in a temporal database.

[0009] Nir et al. (Biosystem engineering, 173, p. 4-10 (2018)) describes a method for determining the weight of a cow from imaging data, and estimate the shape of the animal by an ellipse to calculate an approximatively weight of said animal.

[0010] The challenges in using 3D images in weight determination is often to have each animal singled out and immobilize it at the 3D imaging facility. Often this is done by placing the animal in a small stall in which the sidewalls or railing are confining the animal so that, basically, it is unable to move. In this case, the use of 3D imaging does not eliminate the need to have to handle each individual animal, as the animal will have to be individually shut in the small confined space and the process hence remains lengthy and laborious and somewhat manual. Furthermore, the setups are often complex and expensive making the weight determination a costly affair.

[0011] The weight of an animal, and especially the development over time of weight, is furthermore an important measure for the physical wellbeing of the animal in general. Hence, this is also an important parameter when evaluating the animal’s body condition score. The wellbeing of the animal is furthermore important for having productive animals in the sense that a well-fed animal can get pregnant from a young age and will produce a larger amount of milk and meat.

[0012] Summary

[0013] WO 2020 / 260631 from the current applicant relates to automatic determination of the weight of a bovine subject based on vision technology employing range imaging, by extracting data from a 3D image relating to the topology of the back of the animal, and calculating the weight of said animal by correlating the extracted data against a reference model comprising information of the topology of the back versus the weight for the breed of said animal. This approach was based on the novel finding that only a very limited number of topology predictors extracted from the 3D image is enough to accurately determine the weight of the animal.

[0014] The purpose of the present disclosure is to provide a system and method for automatic determination of various body parameters of an animal based on imaging of the animal, preferably exclusively based on imaging of the animal and preferably range imaging I 3D imaging from above the animal, that takes advantage of the plethora of data acquired in an imaging system.

[0015] A first embodiment therefore relates to a method for determining at least one body parameter of a bovine subject, the method comprising the steps of acquiring at least one range image of at least one bovine subject, processing the at least one range image to generate at least one 3D image of a preselected bovine subject, calculating the body parameter of said preselected bovine subject by modelling the at least one 3D image against a trained model employing artificial intelligence (Al). The at least one body parameter is preferably selected from the group of: the weight, the predicted slaughter (carcass) weight, the conformation class, the predicted carcass conformation class, the fat cover class, and the predicted carcass fat cover class. I.e. body parameters that relate to the shape and topology of the animal, in particular the back of the animal.

[0016] The present inventors have realized that many body parameters of a bovine subject correlate with the 3D shape, i.e. topology and / or topography, of the back of the bovine subject. I.e. parameters such as age of the bovine subject, lactation, etc., can be left out of the estimation. This makes it possible to asses important body parameters of a bovine subject based on range images alone, acquired from above. This makes it also possible to install the presently disclosed system in a stable, a cowshed, or the like and acquire images of the bovine subjects from above, for example while the bovine subjects are eating or while they pass through a lock, for example on their way to or returning from milking.

[0017] In the presently disclosed approach it has been shown that all it requires is a reference model, possibly for the specific breed of the bovine subject and / or for the specific individual animal. This further allows for the continuous monitoring of the well-being of the bovine subjects on a day-to-day basis and even multiple times every day.

[0018] The present disclosure further relates to a system for determining at least one body parameter of a bovine subject, comprising an imaging system for acquiring at least one range image of at least one bovine subject, and a processing unit configured for executing the presently disclosed method. The at least one body parameter is preferably selected from the group of: the weight, the predicted slaughter (carcass) weight, the conformation class, the predicted carcass conformation class, the fat cover class, and the predicted carcass fat cover class.

[0019] The presently disclosed system and method for determining at least one body parameter of a bovine subject can completely eliminate the need for individually and subjectively handling each animal in the process of monitoring bovine subjects, which enables the possibility of frequent monitoring of each individual animal in a large herd. For example the bovine subject may be walking through a lock I narrow aisle connecting a shed resting area with a feeding and / or milking area. The aisle may be so narrow that it only allow passages of one animal at the time. One or more images of the back of the bovine subject may then be acquired by one or more cameras placed above the aisle. The camera(s) may be of any type of camera providing 3D information, such as, but not limited to, a time of flight (ToF) camera, stereo camera, structured light camera, light-field camera, or a combination thereof. The imaging system may comprise a 2D camera and a depth sensor. The imaging system may be configured for acquiring topographic images.

[0020] As also disclosed in WO 2017 / 001538 by the same applicant, such cameras can also be installed over the feeding area and used for acquiring range images of the bovine subjects. The individual bovine subject can be identified by means of the range images, such that each bovine subject also can be monitored with regard to determination of the various body parameters disclosed herein.

[0021] The possibility that the bovine subject may be walking while acquiring an image allows for a frequent monitoring of the bovine subject, as this can happen without interfering with the daily life of the bovine subjects. This frequent monitoring of the bovine subject, further allows the farmer to gather data for statistics on the various body parameters of bovine subjects in a herd.

[0022] The presently disclosed methods can advantageously be computer implemented. The present disclosure also relates to a system for determining at least one body parameter of a bovine subject, comprising a computer-readable storage device for storing instructions that, when executed by a processor, performs the methods disclosed herein.

[0023] The presently disclosed methods may be implemented as one or more separate computer programs but may also be implemented as a plugin to an existing service running on a device or online.

[0024] The present disclosure therefore also relates to a computer program having instructions which when executed by a computing device or system causes the computing device or system to perform the determination at least one body parameter of a bovine subject as disclosed herein. Computer program in this context shall be construed broadly and include e.g. programs to be run on a PC or software designed to run on smartphones, tablet computers or other mobile devices. Computer programs and mobile applications include software that is free and software that has to be bought, and also include software that is distributed over distribution software platforms or running on a server.

[0025] Description of the drawings

[0026] The present disclosure will in the following be described in greater detail with reference to the drawings. The drawings are exemplary and are intended to illustrate some of the features of the present method and system and are not to be construed as limiting to the present disclosure.

[0027] Fig. 1 illustrates the contour lines of the back of a bovine subject corresponding to height drops relative to the spine level.

[0028] Fig. 2 is a schematic view of the canalizing guidance of animals through a passage I lock I narrow aisle from one shed area to another, on the way to milking or on the way back from milking.

[0029] Fig. 3 illustrates one embodiment of the presently disclosed approach of calculating weight of a cow based on range imaging.

[0030] Fig. 4A shows a 3D image of the back of a cow.

[0031] Fig. 4B shows an infrared 2D image of the back of the cow in fig 4A.

[0032] Fig. 5 shows and example of the comparison between test and validation when training convolutional neural network to predict weight of cows based on range imaging.

[0033] Fig. 6 illustrates the training approach for training a neural network

[0034] Figs. 7A-B shows two example of instance segmentation of cows passing through a passage.

[0035] Fig. 8 shows a comparison in robustness between the presently disclosed Al approach and the previously disclosed contour based approach for determining weight.

[0036] Figs. 9A-C shows the correlation between certified carcass slaughter weight and predicted weight based on range imaging for three different breeds of cows. Fig. 10A shows a comparison between actual and predicted conformation class.

[0037] Fig. 10B shows a comparison between actual and predicted fat cover class.

[0038] Detailed description

[0039] A preferred embodiment relates to a method for determining at least one body parameter of a bovine subject, the method comprising the steps of acquiring at least one range image of at least one bovine subject, processing the at least one range image to generate at least one 3D image of a preselected bovine subject, calculating the body parameter of said preselected bovine subject by modelling the at least one 3D image against a trained model employing artificial intelligence (Al).

[0040] The preselected bovine subject can be selected already as part of the range image acquisition because there might only be a single bovine subject in the at least one acquired range image and this single bovine subject naturally becomes the preselected bovine subject. However, there might be several bovine subjects in the range image acquisition and part of the image processing is to select one of the imaged bovine subjects to become the preselected bovine subject. See also the description further below of the segmentation process.

[0041] In one embodiment at least one infrared (I R) 2D image is acquired of the bovine subject. Preferably the range image acquisition includes acquisition of the IR 2D image data.

[0042] The range images can be 3D images that are directly acquired from an imaging system or it can 2D image data comprising depth information, e.g. from one or more RGB 2D cameras combined with depth information from an infrared sensor, like a time-of-flight system, as known from the Kinect system for Windows. 2D image data combined with depth information (e.g. coming from the IR camera) can be processed to become 3D images. They can be processed directly in the imaging unit, possibly as part of the acquisition process, and / or they can be post-processed. At least one 3D image can then be processed to generate at least one point cloud representing the bovine subject. The at least one 3D image and / or the at least one point cloud can then be used as input to the modelling. Typically the IR camera is provided as supplier of depth data, but the IR camera may also provide a “normal” 2D IR image of the bovine subject. The inventors have realized that the 2D IR image data from the infrared camera can be used to improve the modelling. The 3D image data and the point clouds representing the bovine subjects are not necessarily “perfect” because the range image acquisition in a stable environment is not optimal conditions for image acquisition. Hence, the generated 3D image data and point clouds often contain incomplete or erroneous data for various parts of the bovine subject making it more difficult for the modelling to determine the body parameters. In that regard the inventors have realized that feeding the 2D IR data into the model provides better results, most likely because parts of the 3D digital model that lack data might be supplemented from the 2D IR data adding to the digital puzzle representing the bovine subject. Naturally the trained model is then preferably also trained on 3D image data I point clouds supplemented with 2D IR data. Hence, this additional 2D IR image data can be an extra modality fed into the Al model to enhance determination of the body parameters.

[0043] Hence, the at least one body parameter may be calculated by modelling the 3D image data, and / or the point cloud, and optionally IR 2D data, against a trained model employing artificial intelligence.

[0044] A 3D image representing the preselected bovine subject may comprise (x,y,z) data, wherein the (x,y) data corresponds to a 2D image of the preselected bovine subject, and the z-data corresponds to depth data. A fourth coordinate may be the intensity in the 2D image, preferably normalized intensity data, e.g. normalized to the maximum intensity in the image or to a predefined maximum intensity. An example of a 3D image representing a cow is shown in fig. 4A, where the color in the pixels represents the depth data, i.e. lighter colors means that the points are further away.

[0045] The IR 2D image data representing the preselected bovine subject may comprise (x,y) data, wherein the (x,y) data corresponds to an infrared 2D image of the preselected bovine subject, preferably with normalized IR intensity. E.g. normalized to the maximum intensity in the image or to a predefined maximum intensity. An example of a IR 2D image representing a cow is shown in fig. 4B. It is the same cow as in fig. 4A acquired with the case imaging unit, but here showing the infrared image. In fig. 4B the color in the pixels represents the intensity in the image and as seen in the image the markings on the cows back are visible in the 2D IR image, in contrast to the 3D image in fig. 4A, where only the topography of the cow’s back is visible.

[0046] Tests have shown that by including the 2D IR data as input to the model the prediction of the body parameters is improved.

[0047] The plurality of range images may be a time series of range images, and / or the sequence of 3D images of the preselected bovine subject may be a time series of 3D images. If a time series is used as input to the model, it is preferred that the model as well has been trained on time series of image data.

[0048] The sequence of range images 13D images may correspond to fixed number of range images 13D images acquired with a predefined duration between each image, such as at least 2, 3, 4, 5, 7, or at least 10 range images 13D images, preferably 3 range images 13D images, preferably acquired with a duration of less than 10, 8, 5, 4, 3, 2, or less than 1 second, preferably 1 second, between each range image 13D image.

[0049] The range images are preferably acquired with a high frequency, i.e. the range images may be acquired with a frame rate of at least 10, or 15 fps, preferably at least 20 frames per second (fps), more preferably at least 25 fps, or even at least 30, 40 or at least 50 fps.

[0050] The sequence of 3D images, typically selected from the acquired range images, may comprise the preselected bovine subject walking forwardly. Correspondingly it may be advantageous that the selected sequence of 3D images for use in the modelling comprise the preselected bovine subject walking forwardly over a minimum distance of at least 1 meter, or at least 2 m, preferably at least 3 m, or 4 m, more preferably at least 5 m, most preferably at least 7 m, such as between 5 and 10 m. Similarly the sequence of 3D images selected for modelling comprise the preselected bovine subject during a minimum period of at least 1 second, or at least 2 second, preferably at least 3 seconds, more preferably at least 5 seconds, most preferably at least 7 or 10 seconds, such as between 5 and 10 seconds.

[0051] Preferably the range images are acquired from above a passage where through a plurality of bovine subjects are passing through, such as passing through at least once daily, e.g. as illustrated in fig. 3. Correspondingly the step of identifying the preselected bovine subject by means of at least one ID reader located near one or both ends of the passage. Actually this identification may be part of selecting the preselected animal, because the time instance of identification readout provides a reference to the time series of range images such that a relevant sequence of 3D images can be generated for the identified bovine subject to become the preselected bovine subject. Hence, the sequence of 3D images of the preselected bovine subject preferably corresponds to a temporal (window) selection of the plurality of range images selected such that only an identified I preselected bovine subject is present in the sequence of 3D images.

[0052] The image processing step preferably comprise background removal to isolate the bovine subject from the surroundings of the range images. This background removal can also be part of the segmentation process, because for example known background elements can be segmented, identified and removed from the 3D image data.

[0053] The image processing preferably also comprise correction of perspective of the range images and / or the 3D images, in order to provide a good reference coordinate system for the 3D setup. Perspective correction may be required if the range images are acquired by a fixed imaging unit locates above a passage where through the animals move, and where multiple images are acquired while the animal walks through the passage. Correction of perspective can be provided by image processing methods known in the art. One example is to generate a point cloud based on each acquired range image 13D image as described herein, because typically each point position in the point cloud has its own set of cartesian coordinates which makes it relatively straightforward to correct for different perspectives in the image acquisition.

[0054] The image processing advantageously comprises the step of segmenting at least part of the acquired range images by means of for example instance segmentation in order to select a bovine subject among the segmented instances and generate said at least one sequence of 3D images of the preselected bovine subject. Instance segmentation can for example be provided by means of Mask-RCNN (Mask region-convolution neural network) which is an open source model which is a variant of a Deep Neural Network, and which can detect objects in an image and generate a high-quality segmentation mask for each instance. I.e. instance segmentation can provide a mask that describes, comprises and / or delimits the associated pixels for each bovine subject in the acquired range images. Other variants of instance segmentation neural networks exist, for example Cascade R-CNN and Mask2Former; these can also be applied here. In that regard the segmented bovine instance located closest to an ID reader and at the time of a corresponding ID reading can be associated with an identification of a segmented bovine subject, leading to a selection of the preselected bovine subject.

[0055] The presently disclosed approach may further comprise the step of generating point clouds of the preselected bovine subject based on the range images and the image processing, in particular the segmentation, e.g. point clouds can be generated based on the masks from the instance segmentation. The depth information can then turn the point clouds into real 3D point clouds with 3D coordinates in order to generate the at least one sequence of 3D images comprising point cloud coordinates representing the preselected bovine subject in the images, i.e. a sequence of 3D images with labelled point cloud data. And this can be combined with labelled 2D infrared data. Perspective correction can also be applied here. These data with several modalities can be input to the trained machine learning model, e.g. a neural network, e.g. a convolutional neural network, in order to determine the at least one body parameter of the bovine subject.

[0056] Hence, a sequence of 3D images, which can be used as input to the Al model, may comprise 3D point cloud data, that may represent movement of the preselected bovine subject in said sequence. And a sequence of 3D images used as input to the Al model may further comprise 2D IR data of the preselected bovine subject.

[0057] As stated in here a plurality of range images is typically acquired of the bovine subject, e.g. while it walks through a passage. As discussed herein this plurality of range images may be analyzed as a time series and fed into the model such that the movement of the bovine subject is part of the modelling, i.e. such that the modelling includes a temporal parameter. However, it may also use the plurality of range images as separate image data and perform the modelling on each or many of the acquired images such that a body parameter output is provided for a plurality of range images in order to utilize as much acquired data as possible. In that regard it becomes possibly to statistically analyse the plurality of resulting body parameters in order to evaluate the credibility of the body parameters and select the most credible and mainly use the median or average or the like to overall improve precision in the body parameter estimation.

[0058] Hence, the body parameter may be based on calculating a plurality of said body parameters, for example based on a plurality of different range images acquired and analysed as described herein. The plurality of different range images that lay the foundation of determination of a single body parameter may for example be acquired each time the bovine subject is passing through a passage, for example over a predefined period such as one hour, two hours, 6 hours, 12 hours, 24 hours, two days, three days, one week, preferably over one day. This single body parameter can then based on statistics of said plurality of body parameters, preferably the median or average of said plurality of said body parameters. For example the weight of a cow can be determined as the median weight of the cow calculated over one day. If a cow is milked three times per day, the cow may pass through the passage six times a day. Each passage may result in for example three range images of the back of the cow. If each range image is good enough to be analysed, this may result in 18 calculations of the weight of the cow based on range imaging, and can therefore lay a good foundation for statistical analysis of the weight data and therefore improved precision in the daily, weekly, monthly and yearly monitoring of the cows.

[0059] The system may comprise one more ID readers for reading an identification number of a bovine subject entering and / or exiting the lock I passage.

[0060] Body parameters

[0061] The bovine subject body parameters that may be estimated with the presently disclosed approach is preferably selected from the group of: the weight, the predicted slaughter (carcass) weight, the conformation class, the predicted carcass conformation class, the fat cover class, and the predicted carcass fat cover class.

[0062] The weight is the present body weight of the bovine subject, i.e. like if the bovine subject were placed on a weight scale and manually weighed. However, if placing the bovine subject on a weight scale, the result would include the milk in the udder, which varies during the day and before and after milking. The advantage of using imaging acquired from above is that the udder is not part of the evaluation, i.e. it is the “true” body weight of the bovine subject that is monitored with the presently disclosed imaging approach. The reference data (for example in the form of manual weighing) should then advantageously be provided after milking, i.e. when the udder is substantially empty.

[0063] The predicted slaughter weight, i.e. carcass weight, is the weight of the carcass in the slaughter house, i.e. without for example head and blood. In the term “predicted” lies directly that the result of imaging of the alive bovine subject in a stable environment, can only be a prediction of a future slaughter weight. Bovine subjects typically end up in a slaughterhouse, where the weight of the carcass is registered. Feeding this data into an Al model makes it possible to accurately predict the carcass weight based on imaging of the cow while still alive, i.e. to determine the predicted slaughter weight. This provides data to the farmer in assisting on the decision on maintaining the bovine subject in the stable and / or sending it to the slaughterhouse, in order to increase profitability for the farmer.

[0064] The term “conformation” describes the shape, form, and type of an animal. Analyzing the conformation of a beef animal may help to predict the amount and quality of muscle, the structural soundness, the reproductive ability, and other traits that affect profitability. Conformation is typically divided into classes, i.e. conformation classes. For example Superior, Excellent, Very good, Good, Fair and Poor.

[0065] For example the conformation class “Superior” is characterized by very highly rounded round with double-muscled visibly separated seams, topside spreads very markedly over the symphysis (symphisis pelvis), very wide and very thick back, up to the shoulder, very rounded rump and very highly rounded shoulder.

[0066] Excellent: Round: very rounded, topside spreads markedly over the symphysis (symphisis pelvis), Back: wide and very thick, up to the shoulder, rump very rounded, shoulder: very rounded,

[0067] Very good: Round: rounded, topside spreads over the symphysis (symphisis pelvis), back: wide and thick, up to the shoulder, rump rounded, and shoulder: rounded

[0068] Good: Round: well-developed, topside and rump are slightly rounded, back: still thick but less wide at the shoulder, shoulder: fairly well-developed

[0069] Fair: Round: average development to lacking development, back: average thickness to lacking thickness, rump: straight profile, shoulder: average development to almost flat

[0070] Poor: Round: poorly developed, back: narrow with bones visible, shoulder: flat with bones visible At the slaughterhouse each carcass is provided with a conformation class. Feeding this data as reference data into an Al model makes it possible to accurately predict the carcass conformation class based on imaging of the cow while still alive, i.e. determine the predicted carcass confirmation class. However, it is also possible to assign a present confirmation class to the bovine subject being alive and imaged.

[0071] “Class of fat cover” is another classification of bovine subjects for estimating the amount of fat on the outside of the carcass and in the thoracic cavity. Fat cover is typically divided into five classes 1 , 2, 3, 4 and 5, with 1 being the leanest and 5 the fattest. At the slaughterhouse each carcass is provided with a fat cover class. Feeding this data as reference data into an Al model makes it possible to accurately predict the carcass fat cover class based on imaging of the cow while still alive, i.e. determine the predicted carcass fat cover class. However, it is also possible to assign a present fat cover class to the bovine subject being alive and imaged.

[0072] For example the fat cover class “1” is characterized by “No fat within the thoracic cavity”

[0073] 2: Slight; within the thoracic cavity the muscle is clearly visible between the ribs

[0074] 3: Average; within the thoracic cavity the muscle is still visible between the ribs

[0075] 4: High; the seams of fat on the round are prominent. Within the thoracic cavity the muscle between the ribs may be infiltrated with fat

[0076] 5: Very high; the round is almost completely covered with fat, so that the seams of fat are no longer clearly visible. Within the thoracic cavity the muscle between the ribs is infiltrated with fat

[0077] Passage

[0078] In one embodiment the bovine subject is walking during the image acquisition. The range images may hence be acquired while the bovine subject is moving, for example walking in a straight line. The ability to calculate the various body parameters of a moving animal is a great advantage as it allows monitoring of a large number of animals in a row, and hence this eliminates the need for individually handling the individual animal. When animals are walking from a shed resting area to a feeding and / or milking area, or the opposite direction, the bovine subjects may be passing through a narrow aisle thereby only allowing one animal to pass at a time. By placing an imaging system above this narrow aisle the bovine subjects will walk by the area below the camera which can then capture an image of each individual animal as they are passing through the aisle. A great advantage thereof is that the bovine subjects will not be overlapping the view of each other, seen from the 3D camera from above. Hence, the bovine subjects are passing the narrow aisle and passing through the frame area of the imaging unit as they are being canalized or lead from for example a rest area to a feeding and / or milking area or the opposite direction. Another advantage of the narrow aisle is that it is possible to gain a good control of the flow of animals.

[0079] An example of this process of leading the bovine subject trough a narrow aisle is shown in Figure 3. Here the thick lines 1 illustrates the boundaries of the areas of which the bovine subjects can move. The boundaries may be a hence or a wall or similar enclosure mechanism. Each animal is, for simplicity of the figure, illustrated by an ellipse in Figure 3. The arrow connected to each animal (ellipse in the figure) illustrates the direction of which the bovine subject is moving. The bovine subjects 2 to the left are the bovine subjects which have already passed the narrow aisle, the bovine subject 3 standing just below the 3D camera 4 is a bovine subject situated within the narrow aisle and currently being imaged by the 3D camera 4. The bovine subject behind 5 has just entered the narrow aisle and will be imaged as it soon passes through the area just below the 3D camera 4 hence, when animal 5 reaches the current position of animal 3. The bovine subjects 6 to the right all represent the bovine subjects still waiting to enter the narrow aisle to ultimately enter the area to the left of the aisle.

[0080] Imaging unit

[0081] The imaging unit comprises one or more cameras, preferably cameras adapted for range imaging, such as range cameras. There may also be several imaging units distributed along a bovine subject passage such that the bovine subjects can be imaged over a longer distance while moving. Examples of range imaging techniques applicable herein are time-of-flight, stereo triangulation, structured light, light-field imaging, e.g. 4D light-field imaging etc. The imaging unit may also be configured for acquiring topographic images. Each range camera may be provided with a depth sensor and a 2D camera, such as a RGB camera, for example as known from the Kinect system. The imaging unit may comprise at least one infrared (IR) sensor, such as an IR camera, preferably providing 2D IR data. Depth data can then be acquired by means of the IR signal.

[0082] The imaging unit is preferably configured for acquiring images over a time period with acquisition of multiple frames per second (fps), such that the output is more or more video sequences. E.g. at least 5, 10, or 15 fps, preferably at least 20 fps, most preferably at least 25 fps or even at least 30 fps, preferably both RGB, depth and IR, such high fps that video sequences comprises both 3D data and 2D IR data.

[0083] The time period for video acquisition may be triggered by detection of bovine subjects near the imaging unit, and video acquisition continues while one or more bovine subjects pass by. Subsequent image processing, e.g. by segmentation, can separate data such that video sequences of each bovine subject can be extracted. The presently disclosed approach is preferably configured such that a single bovine subject is imaged in a time period while passing the aisle I the imaging unit, of at least 1 second, preferably at least 2 seconds, more preferably at least 3 seconds, most preferably at least 5 seconds, or even 8 or 10 seconds, for example between 5 and 10 seconds. The bovine subject is preferably moving in the video sequence such that the movement dynamics of the single bovine subject is recorded in the video sequence.

[0084] Alternatively or supplemental the presently disclosed approach can be configured such that a single bovine subject is imaged over a predefined minimum distance to be certain that there is movement, preferably walking movement, in the images. E.g. a minimum distance of 2 m, or at least 3 m, preferably at least 5 m, most preferably at least 7 or even at least 10 m for example between 5 and 10 meters.

[0085] The imaging system typically includes at least one optical device I detector, e.g. a camera, and in order to keep the at least one optical device clean the device may have to eventually be cleaned and / or possibly protected from the dirt of the shed environment. One possible solution is to use a protective cover keeping the dirt away from the actual optical device. In the preferred embodiment the at least one optical device is protected by a protective cover to prevent dirt from covering the functioning parts of the optical device, directly. In a further embodiment the cover of the at least one optical device is cleaned using an automatized cleaning system such as an automatic windscreen wiper. Hence, if the optical device cover is getting too dirty for capturing the desired data the cover can get cleaned automatically by running the cleaning system such as an automatic windscreen wiper.

[0086] Another approach to keep the at least one optical device clean is for example by having the at least one optical device protected by a gate or a shutter only opening for a short period of time upon acquiring the data. Hence the optical device will only be exposed for a very short amount of time upon data collection by the at least one optical device. Thus, in one embodiment the at least one optical device is protected by a gate or a shutter only opening for a short period of time upon acquiring the data. Hence, the gate or shutter will open briefly for image acquisition, and will then close again to avoid the optical device getting dirty. In the latter case the optical device will contain both a protective cover and a gate / shutter, the gate / shutter hence preventing the cover of the optical device to get dirty too quickly. To optimize the time of which the optical device is exposed to collect data, hence to optimize the time of the shutter to open the shutter or gate of the system may include another feedback system, to determine when a bovine subject is standing in a proper position for data acquisition. This feedback system may be an independent system based on a sensor placed next to the 3D camera. The position of the bovine subjects may be determined from another optical device and / or sensor not covered by a shutter or a gate and this determines when a bovine subject is in a proper position for imaging. When this happens the shutter will open and the range images will be taken after which the shutter closes immediately. It may also be possible to acquire a series of images before the shutter closes.

[0087] In a further embodiment the at least one optical device includes a cleaning alert system sensing when the sensor needs to get cleaned and notify the user through said alert system such as a light turning on, a noise playing or a wireless signal transmitted to a computer. This is particularly useful if the device does not have an automatized cleaning system or if it is insufficient to fully clean the device. In this case the alert system will make the user aware that further cleaning of the device or the cover of the device is needed. The user might then manually clean the device or activate an automatized cleaning system. The cleaning process may then be automatic or manual. If the automated cleaning process is sufficient it may not be necessary to notify the user and this step might be left out as long as the system can efficiently self-clean.

[0088] Identification

[0089] It is an advantage for the further data analysis and monitoring of the individual animal to know the identity of the specific imaged animal. Identification may be manual, for example by reading the identification number of the ear tags of the bovine subject, or electronic identification means such as a transponder, radio frequency ID (RFID), by pattern recognition of the bovine subject, by a known, orders sequence of the bovine subjects or any other mean of identification. The bovine subject may be identified prior to the image acquisition or after the image acquisition.

[0090] If using subject specific reference it might be necessary to know the identity of the bovine subject in the image(s). But as shown in WO 2017 / 001538 it is possible to (uniquely) identify animals in a population of known animals based on images of the bovine subject’s back. Hence, the presently disclosed approach may further comprise the step of identifying the bovine subject based on said at least one 3D image.

[0091] Identification of the individual bovine subject may be manual, for example by reading the identification number of the ear tags of the animal, or electronic identification means such as radio frequency ID (RFID), by pattern recognition of the animal, by a known, orders sequence of the animals or any other mean of identification. An identification (ID) reader, for example for reading of (RFID) ear tags, is typically located near the aisle, i.e. at the entrance and / or at the exit of the aisle, where the animals are supposed to pass one by one. Note that an entrance may be an exit and vice versa, dependent on which way the animals pass. Hence, the animal may be identified prior to the image acquisition, during the image acquisition, and / or after the image acquisition.

[0092] As shown in WO 2017 / 001538 it is possible to (uniquely) identify animals in a population of known animals based on images of the animal’s back. Hence, the presently disclosed approach may further comprise the step of identifying the animal based on said at least one 3D image.

[0093] The system may also comprise ID tags. Such ID tags may be connected to animals to be identified. ID tags may be visual and / or electronic ID tags. Electronic ID tags may be electronic ear tags and / or electronic ID tags attached to an animal such as in a collar. A single animal may be marked with one or more ID tags such as at least one visual ID tag and / or at least one electronic ID tag. An example is at least one visual ear ID tag combined with at least one electronic ID tag in a collar. Another example is at least one visual ear ID tag combined with at least one electronic ear ID. Segmentation

[0094] A plurality of bovine subject may be passing through at the same time and the acquired 3D images may contain multiple animals. Hence, segmentation can advantageously be applied to isolate the individual bovine subject in the images, for example in the form of semantic segmentation to group pixels in a semantically meaningful way such that pixels belonging to different objects are grouped separately. However, preferably in the form of instance segmentation such that a label is assigned to each object in the images, such that instance segmentation can be provided of the individual bovine subject in the 3D images, for example in the sequence of 3D images. Segmentation is also a way of selecting a bovine subject in the images to become the preselected bovine subject.

[0095] Instance segmentation typically includes object detection of all objects in an image in order to classify individual objects and localize each object instance using for example a bounding box, and then segmenting of each instance in order to classify each pixel into a fixed set of categories without differentiating object instances. In the present approach it is most important to classify the individual bovine subjects such that a sequence of 3D images comprising more than one bovine subject is segmented to separate the individual bovine subjects in to separate instances.

[0096] One example of instance segmentation is the known Mask-RCNN (Mask regionconvolution neural network, it is a variant of a Deep Neural Network and detects objects in an image and generates a high-quality segmentation mask for each instance) that can provide a mask that describes the associated pixels for each bovine subject in the sequence of 3D images. Other variants of instance segmentation neural networks exist, for example Cascade R-CNN and Mask2Former, that can also be applied here. Edge based image processing is also an approach to instance segmentation where edges in the image are detected to identify the outline of the animals.

[0097] Selecting a 3D image representing a bovine subject for further analysis with regard to determination of at least one body parameter depends on the quality of the acquired image data. See for example fig. 7A showing an image acquired from above a narrow passage where though bovine subjects pass. Two cows are visible in the image in fig. 7A, but only part of the back of the front cow and only the head and part of the shoulders of the back cow. The image has been subject of instance segmentation and part of the front cow is marked as a result of the instance segmentation, but the segmented data does not include the whole cow and cannot be used for further analysis. The cow in the back of the image in fig. 7A can neither be used. In contrast the example in fig. 7B shows an image from the same passage, but where only a single cow is visible and has been fully segmented.

[0098] Quality of the data may also relate to whether valid depth data are available for the pixels within the segmented instance, because otherwise it will be difficult to predict the body parameters.

[0099] Hence, the bovine subject selected among the segmented instance(s) is advantageously selected based on 1) extension of the segmented instance(s) in the corresponding image, for example the extension of the segmented instances must be within the image, and / or 2) fraction of valid depth data in the segmented instance. The fraction of valid depth data is for example at least 70%, preferably at least 80%, more preferably at least 90%, most preferably at least 95%, or even 99%.

[0100] Breed

[0101] As used herein the bovine subject may be a bovine subject, preferably including both cows and bulls, whether adult or newborn animals. Consequently the breed of the bovine subjects may be selected from the group of: The Jersey breed, the Holstein breed, the Holstein-Friesian cattle population, Holstein Swartbont cattle population, the Deutsche Holstein Schwarzbunt cattle population, the US Holstein cattle population, the Red and White Holstein breed, the Deutsche Holstein Schwarzbunt cattle population, the Danish Red population, the Finnish Ayrshire population, the Swedish Red and White population, the Danish Holstein population, the Swedish Red and White population and the Nordic Red population.

[0102] Al model

[0103] The Al model used in the presently disclosed approach may be selected from the group of: machine learning, neural network, such as recurrent neural network, and / or convolutional neural network, or any combination thereof. The preferred choice is the convolutional neural network (CNN) which is also used in the examples herein. The trained reference model may be specific to said preselected bovine subject, for example specific to the breed of said bovine subject. However, a common reference model across breeds may also be used, and as demonstrated herein also provides good results.

[0104] The trained reference model may be trained using a combination of self-supervised learning and semi-supervised learning, wherein self-supervised learning approach is provided to pre-train the Al model on unlabelled data, and then the pretrained Al model is fine-tuned on a smaller set of labelled data using semi-supervised learning

[0105] The trained reference model may be trained using machine learning, neural network, recurrent neural network, convolutional neural network, or any combination thereof.

[0106] Fig. 3 illustrates an example of a CNN approach for determining weight of a Holstein cow by means of range imaging, where the cow is found in the image by means of instance segmentation using Mask-RCNN. The cow specific features that is input to the model in 2) is a 3D image of the back of the cow with depth data (to the left, see also fig. 4A) and 2D IR image of the back of the cow (to the right, see also fig. 4B). The actual weight estimation is based on a convolutional neural network with two hidden layers.

[0107] Figs. 5 illustrates the training of the CNN model based on a reference dataset where bovine subjects of the same breed where range images are acquired and the bovine subjects are manually weighed, substantially concurrently, at least once per day, over a period of several weeks. 80% of the dataset were used for training, 10% for testing and 10% for validation. Fig. 5 illustrates the difference between prediction (testing) and validation and as seen from fig. 5 the model was successfully trained in the end.

[0108] Acquisition of reference datasets based on manual weighing of the bovine subjects is naturally quite tedious. Almost all milking cows at some point end up in the slaughter house, where the carcass is weighed and registered to the specific bovine subject to become a certified carcass weight. I.e. new certified slaughter weight data is continually generated - but it is not the same parameter as the manual weighing of the live bovine subjects. The present inventors have realized that certified carcass slaughter weight can be used as supplementary data to the trained reference model. One solution is to correlate the certified carcass slaughter weight with the predicted weight from range imaging obtained according to the presently disclosed approach. It has turned out that there is a linear correlation between the predicted weight based on range imaging and a subsequent certified carcass slaughter of the same bovine subject, in particular if there is a limited amount of days between the range image(s) has been acquired and the cow has been slaughtered. Linear correlations between the certified carcass slaughter weight (X-axis) and the weight of the live animal (Y-axis) are illustrated in fig. 9A for Holstein cattle, in fig. 9B for RDM and in fig. 9C for Jersey cattle. Hence, certified carcass slaughter weight for the specific bovine subject can be converted into weight of the live animal and fed into the model, as long as there are range images of the specific bovine subject that have been acquired shortly before slaughtering. This can increase the amount of reference data manyfold and thereby improve the accuracy of the Al model, preferably gradually improve, and optionally maintain the accuracy of the Al model.

[0109] Naturally the Al model for determining carcass slaughter weight of the bovine subject may be trained with certified carcass slaughter weight as outcome. However, what is more surprising is that the Al model for determining weight of the live bovine subject based on range imaging, may be trained with certified carcass slaughter weight as outcome.

[0110] Hence, the trained model for determining weight may be at least partly trained with certified weight as outcome, and / or the trained model for determining weight may be at least partly trained with certified carcass slaughter weight as outcome. In that regard training data comprising weight calculated based on range imaging vs. certified carcass slaughter weight is preferably based on data obtained with a duration less than a predefined number of days between acquisition of certified carcass slaughter weight and range image acquisition, such as less than 14 days, preferably less than 10 days, more preferably less than 7 days, even more preferably less than 5 days, most preferably less than 4 days, or even less than 3 or 2 days. I.e. it is important the training data using certified carcass slaughter weight is based on reliable data, i.e. based on data of the same bovine subject which has been imaged not too long before being slaughtered.

[0111] The Al models for determining predicted conformation class and predicted fat cover class of a bovine subject are preferably trained with certified data on conformation class and fat cover class, respectively. Such data are continuously generated at the slaughter house for the specific bovine subject and can be fed back into the model.

[0112] A validation of the prediction of conformation class is illustrated in fig. 10A with certified conformation class on the X-axis and predicted conformation class on the Y-axis for a number of bovine subjects. The RMSE is 0.51 , which is acceptable. The presently disclosed range imaging approach is therefore suitable for predicting the future carcass conformation class of the live bovine subject. However, there are a number of outliers and there is a tendency to overestimate the bovine subjects that end up with low fat cover class of 1 and underestimate the bovine subjects that end up with high fat cover class of 4.

[0113] A validation of the prediction of fat cover class is illustrated in fig. 10B with certified fat cover class on the X-axis and predicted fact cover class on the Y-axis for a number of bovine subjects. The fat cover class values are provided from the slaughter house as discrete integer values between 1 and 5; in this group of bovine subjects there are however no animals in the 5 category, because the fat cover class of 5 is rarely seen for milking cows. The predicted fat cover class obtained according to the presently disclosed approach can be provided as real numbers, as also indicated in fig. 10B. The RMSE in fig. 10B is 0.21, which is acceptable and usable in practice. The presently disclosed range imaging approach is therefore suitable for predicting the future fat cover class of the live bovine subject.

[0114] The robustness of the presently disclosed Al model approach for determining weight is in fig. 8 compared to a previously used contour model approach, where only contour data from the range images of bovine subjects were used to determine weight of the animals. The graph shows on the X-axis the prediction error in kilos in the contour LDA approach, and the Y-axis the prediction error in kilos with the presently disclosed CNN approach. If the two weight determination approaches were similar, the points would be scattered around (0,0). It can be seen that the contour LDA approach works but has much more outliers and wrong weight predictions. A marking shows two of the 3D images where the contour approach fails, but where the CNN approach is correct, and as seen from the images they contain noisy data. This type of noisy data disturbs the contour model, which relies on only few contour points from the back of the cow. In contrast the CNN approach can take account of such noisy data and still provide a correct weight of the cow, making the presently disclosed approach much more robust when employed in a real world environment, such as a stable.

[0115] A prediction error common to the two approaches are also illustrated with two examples, and as seen from two of the images in fig. 8, only part of the cow is imaged in the two images. Filtering out such images by means of for example instance segmentation and a mask outlier filter, as disclosed herein, may filter out such images from the determination of body parameters. And even if prediction errors occur, the statistics from several measurements per day can take account of such outliers.

[0116] Successful determination of present weight and prediction of carcass slaughter weight, conformation class and / or fat cover class provides a monitoring tool for the farmer to determine when to send the bovine subjects to the slaughter house to optimize both milking and meat revenue. The payment from the slaughter house to the farmer is depending on all three slaughter parameters, and the presently disclosed approach therefore also provides a tool for the farmer to provide an indication of whether data from the slaughter house is actually correct.

[0117] Examples

[0118] Example 1

[0119] In order to generate a reference of the cattle of the Jersey breed, the back of the individual animals of a herd of 102 Jersey cattle were imaged during four weeks using the system and method of the present disclosure. The animals of the herd had an average weight of 460 kg and had a weight range from 350 to 650 kg. During a period of 4 weeks, 1329 measurements of the backs of the animals of the herd were carried out, the number of measurements per animal varied between 3 and 15 measurements during the 4 weeks. Among the 1329 measurements 1149 of them were carried out in the morning and 180 was carried out in the afternoon at any day during the 4 weeks period. The images were acquired at the same time (within same minute) as a measurement of the animal weight was acquired using a weight scale. As with the present approach range images are acquired of the reference animals, and 3D images and optionally point clouds can be generated representing bovine subjects with known weight. Also sequences of range images can be acquired as part of the reference model generation, i.e. training data. After generation of the training data and a reference model it was therefore possible to test and cross-validate the presently disclosed approach.

[0120] The same approach was followed with one more herd of Jersey and three herds of Holstein and one herd of Red Dairy Cattle (RDM, short for “Rod Dansk Malkekvaeg”) herds. The result of applying the presently disclosed approach with a CNN model to determine weight and using instance segmentation with a mask outlier filter is summarised in the table below. The models used are specific for each breed. The number in parenthesis is the total number of animals that have been manually weighed for each breed.

[0121] Example 2

[0122] Data are applicable across different herds, the same approach as explained above in example 1 was repeated for a different herd of Jersey cattle at another farm. The second herd included 162 cattle of the Jersey breed with an average weight of 519 kg, i.e. on average the cows in herd 2 were approx. 60 kg heavier, and the results show that data are applicable across herds.

[0123] Example 3

[0124] In order to test whether data were applicable across different cattle breeds, a combined model across the three breeds was created based the data in example 1. The results are summarized below.

[0125] As seen from the table above, the prediction of Holstein and Jersey cows are a bit worse, whereas the prediction of RDM weight is slightly improved. In that regard it is notable that the RDM specific model has the least number of reference weights. The conclusion is that a combined breed model indeed is usable, and a breed specific model with a limited amount of reference data can be improved by including data from other breeds, whereas a breed specific model might be more accurate if enough reference data is available.

[0126] The model can only be slightly improved by including additional animal specific parameters, the variances decrease; the repeatability even slightly increases. It must be concluded that the approach disclosed herein extracted from 3D imaging of the back of the animal, is sufficient to accurately predict the weight of the animal if a specific reference model is available, in particular a breed specific reference model, but also a model spanning several breeds, i.e. animal specific parameters can be left out of the weight prediction making a practical implementation feasible in a real life environment.

[0127] Example 4

[0128] The trained reference models can be continuously improved by constantly feeding new verified data regarding one or more of the body parameters into the model, for example the carcass conformation class, the carcass fat cover class and the carcass slaughter weight, which typically is high quality data that is provided from the slaughter house. I.e. data that can be provided as output data in the training of the model.

[0129] Tests have shown that the presently disclosed approach become more robust when slaughter data is included in the training of the model. Robustness is important when implementing such imaging solutions in the real world.

[0130] Items

[0131] 1. A method for determining at least one body parameter of a bovine subject, the method comprising the steps of acquiring at least one range image of at least one bovine subject, processing the at least one range image to generate at least one 3D image of a preselected bovine subject, calculating the body parameter of said preselected bovine subject by modelling the at least 3D image against a trained model employing artificial intelligence (Al), wherein the body parameter is selected from the group of: the weight, the predicted slaughter (carcass) weight, the conformation class, the predicted carcass conformation class, the fat cover class, and the predicted carcass fat cover class.

[0132] 2. The method according to item 1 , wherein the bovine subject is a bovine subject of known breed.

[0133] 3. The method according to any of the preceding items, comprising the step of segmenting the at least one 3D image to generate at least one point cloud representing the preselected bovine subject.

[0134] 4. The method according to any of the preceding items, wherein a 3D image representing the preselected bovine subject comprise (x,y,z) data, wherein the (x,y) data corresponds to a 2D image of the preselected bovine subject, preferably with normalized intensity, and the z-data corresponds to depth data.

[0135] 5. The method according to any of the preceding items, wherein at least one infrared (IR) 2D image is acquired of the bovine subject.

[0136] 6. The method according to any of the preceding items 5, wherein the range image acquisition includes acquisition of the IR 2D image data.

[0137] 7. The method according to any of the preceding items 5-6, wherein at least one point cloud representing the preselected bovine subject is generated based on both 3D image data and IR 2D image data.

[0138] 8. The method according to any of the preceding items 5-7, wherein the at least one body parameter is calculated by modelling the 3D image and IR 2D data, representing the preselected bovine subject, against a trained model employing artificial intelligence.

[0139] 9. The method according to any of the preceding items 5-8, wherein IR 2D image data representing the preselected bovine subject comprise (x,y) data, wherein the (x,y) data corresponds to an infrared 2D image of the preselected bovine subject, preferably with normalized IR intensity.

[0140] 10. The method according to any of the preceding items, wherein the at least one range image and / or the at least one 2D IR image is acquired of the back of the bovine subject.

[0141] 11. The method according to any of the preceding items, wherein at least one sequence of range images are acquired of the at least one bovine subject to generate at least one sequence of 3D images of the preselected bovine subject.

[0142] 12. The method according to any of the preceding items 11 , wherein the sequence of range images 13D images correspond to a time series of range images 13D images.

[0143] 13. The method according to any of the preceding items 11-12, wherein the sequence of range images 13D images correspond to fixed number of range images 13D images acquired with a predefined duration between each image, such as at least 2, 3, 4, 5, 7, or at least 10 range images 13D images, preferably 3 range images I 3D images, preferably acquired with a duration of less than 10, 8, 5, 4, 3, 2, or less than 1 second, preferably 1 second, between each range image 13D image.

[0144] 14. The method according to any of the preceding items, wherein the range images are acquired with a frame rate of at least 10 frames per second (fps), preferably at least 15 fps, more preferably at least 20 fps, most preferably at least 25 fps.

[0145] 15. The method according to any of the preceding items 11-14, wherein the sequence of 3D images comprises the preselected bovine subject walking forwardly.

[0146] 16. The method according to any of the preceding items 11-15, wherein the sequence of 3D images comprise the preselected bovine subject walking forwardly over a minimum distance of at least 3 meters.

[0147] 17. The method according to any of the preceding items 11-16, wherein the sequence of 3D images comprises the preselected bovine subject during a minimum period of at least 3 seconds, preferably at least 5 seconds.

[0148] 18. The method according to any of the preceding items 11-17, wherein the image processing comprises the step of generating at least one sequence of 2D I R images of the back of the preselected bovine subject corresponding to the at least one sequence of 3D images. The method according to any of the preceding items, wherein the body parameter is based on calculating a plurality of said body parameters. The method according to any of the preceding items 19, wherein the plurality of said calculated body parameters are based on a plurality of different range images. The method according to any of the preceding items 20, wherein the plurality of different range images are acquired while the bovine subject is passing through a passage, acquired over a predefined period such as one hour, two hours, 6 hours, 12 hours, 24 hours, two days, three days, one week, preferably over one day. The method according to any of the preceding items 19-21 , wherein the body parameter is based on statistics of said plurality of said body parameters, preferably the median or average of said plurality of said body parameters. The method according to any of the preceding items, wherein the at least one range image is acquired from above a passage where through a plurality of bovine subjects are passing through, such as passing through at least once daily. The method according to any of the preceding items 23, further comprising the step of identifying the at least one bovine subject and / or the preselected bovine subject by means of at least one ID reader located at one or both ends of the passage. The method according to any of the preceding items, wherein the image processing comprise background removal to isolate the bovine subject from the surroundings of the range image(s). The method according to any of the preceding items, wherein the image processing comprise correction of perspective of the range image(s) and / or the 3D image(s). The method according to any of the preceding items 3-26, wherein the image processing comprise correction of perspective of the point cloud. 28. The method according to any of the preceding items, wherein the image processing comprise the step of segmenting at least part of the acquired range images by means of instance segmentation in order to select a bovine subject among the segmented instances and generate said at least one 3D image of the preselected bovine subject.

[0149] 29. The method according to item 28, wherein instance segmentation is provided by means of Mask Region-based Convolutional Neural Network (Mask-RCNN), and / or by means of edge based image processing.

[0150] 30. The method according to any of the preceding items 28-29, wherein the segmented bovine instance located closest to an ID reader and a corresponding ID reading is associated with an identification of the preselected bovine subject.

[0151] 31. The method according to any of the preceding items 28-30, wherein the bovine subject selected among the segmented instances is selected based on 1) extension of the segmented instances in the corresponding image, and / or 2) fraction of valid depth data in the segmented instance.

[0152] 32. The method according to any of the preceding items 31 , wherein the extension of the segmented instances must be within the image.

[0153] 33. The method according to any of the preceding items 31-32, wherein the fraction of valid depth data must be at least 70%, preferably at least 80%, more preferably at least 90%, most preferably at least 95%, or even 99%.

[0154] 34. The method according to any of the preceding items 3-33, comprising the step of generating the point cloud of the preselected bovine subject based on the range image(s) and the image processing.

[0155] 35. The method according to any of the preceding items 10-34, wherein the at least one sequence of 3D images comprise point cloud coordinates representing the preselected bovine subject in the images.

[0156] 36. The method according to any of the preceding items, comprising the step of adding depth data to the point cloud data and perspective correcting the point cloud data based on range data / depth data in the range images.

[0157] 37. The method according to any of the preceding items 3-36, wherein the at least one 3D image used as input to the Al model further comprise 2D IR data of the preselected bovine subject.

[0158] 38. The method according to any of the preceding items, wherein the trained model comprises information of the topology of the back of bovine subjects versus the weight, for example in relation to the breed of said preselected bovine subject.

[0159] 39. The method according to any of the preceding items, wherein the trained model is at least partly trained with certified weight as outcome.

[0160] 40. The method according to any of the preceding items, wherein the trained model is at least partly trained with certified carcass slaughter weight as outcome.

[0161] 41. The method according to any of the preceding items 40, wherein training data comprising weight calculated based on range imaging vs. certified carcass slaughter weight is based on data obtained with a duration less than a predefined number of days between acquisition of certified carcass slaughter weight and range image acquisition, such as less than 14 days, preferably less than 10 days, more preferably less than 7 days, even more preferably less than 5 days, most preferably less than 4 days, or even less than 3 or 2 days.

[0162] 42. The method according to any of the preceding items, wherein the trained model is trained using a combination of self-supervised learning and semi-supervised learning, wherein self-supervised learning approach is provided to pretrain the Al model on unlabeled data, and then the pretrained Al model is fine-tuned on a smaller set of labeled data using semi-supervised learning.

[0163] 43. The method according to any of the preceding items, wherein the trained model is trained using machine learning, neural network, recurrent neural network, convolutional neural network, or any combination thereof.

[0164] 44. The method according to any of the preceding items, wherein the Al model is selected from the group of: machine learning, neural network, recurrent neural network, convolutional neural network, or any combination thereof. The method according to any of the preceding items, wherein the trained model is specific to said preselected bovine subject. The method according to any of the preceding items, comprising a step of identifying the at least one bovine subject and / or the preselected bovine subject based on at least one range image. The method according to any of the preceding items, comprising a step of identifying the at least one bovine subject and / or the preselected bovine subject based on RFID technology. The method according to any of the preceding items, wherein the breed of the bovine subject is selected from the group of: the Jersey breed, Friesian cattle population, Holstein Swartbont cattle population, the Deutsche Holstein Schwarzbunt cattle population, the US Holstein cattle population, the Red and White Holstein breed, the Deutsche Holstein Schwarzbunt cattle population, the Danish Red population (RDM), the Finnish Ayrshire population, the Swedish Red and White population, the Danish Holstein population, the Swedish Red and White population and the Nordic Red population. A system for determining at least one body parameter of a bovine subject, comprising: an imaging system for acquiring at least one range image of at least one bovine subject, and a processing unit configured for executing the method of any of the preceding items, wherein the at least one body parameter is selected from the group of: the weight, the predicted slaughter (carcass) weight, the conformation class, the predicted carcass conformation class, the fat cover class, and the predicted carcass fat cover class. The system according to item 49, configured to acquire said at least one range image while the bovine subject is standing in and / or walking through a lock and / or passage 51. The system according to any of the preceding items 49 to 50, configured such that said at least one range image is acquired from above the bovine subject thereby imaging said bovine subjects in a top-view.

[0165] 52. The system according to any of the preceding items 49 to 51 , configured to start acquisition of the range image(s) when triggered by at least one bovine subject approaching and / or entering a lock.

[0166] 53. The system according to any of the preceding items 49 to 52, wherein the imaging system is configured to acquire the range image(s) with a frame rate of at least 15 frames per second, preferably at least 25 frames second.

[0167] 54. The system according to any of the preceding items 49 to 53, comprising one or more ID readers for reading an identification number of a bovine subject entering and / or exiting the lock I passage.

[0168] 55. The system according to any of the preceding items 49 to 54, wherein the imaging unit comprise at least one RGB camera and a depth sensor, such as an infrared sensor for providing 2D infrared images.

Claims

Claims1. A method for determining at least one body parameter of a bovine subject, the method comprising the steps of acquiring at least one range image of the back of at least one bovine subject, processing the at least one range image to preselect a bovine subject and generate at least one 3D image of the preselected bovine subject, calculating the body parameter of said preselected bovine subject by modelling the at least one 3D image representing the preselected bovine subject, against a trained model employing artificial intelligence (Al), and wherein the at least one body parameter is the weight of the bovine subject.

2. The method according to claim 1, comprising acquiring at least one infrared (IR) 2D image of the back of the at least one bovine subject.

3. The method according to claim 2, wherein the range image acquisition includes acquisition of the IR 2D image data.

4. The method according to any one of the claims 2-3, wherein calculating the body parameter of said preselected bovine subject is by modelling the at least one 3D image representing the preselected bovine subject and the at least one IR 2D image representing the preselected bovine subject, against a trained model employing artificial intelligence (Al),5. The method according to any one of the preceding claims, wherein the body parameters are the weight of the bovine subject, and the predicted carcass slaughter weight of the bovine subject.

6. The method according to any one of the preceding claims, wherein the body parameters are furthermore the predicted carcass conformation class, and the predicted carcass fat cover class.

7. The method according to any one of the preceding claims, comprising the step of generating at least one point cloud representing the preselected bovine subject, and wherein the image processing comprises correction of perspective of the at least one 3D image, based on correction of perspective of the at least one pointcloud.

8. The method according to any of the preceding claims, wherein at least one sequence of range images are acquired of the at least one bovine subject to generate at least one sequence of 3D images of the preselected bovine subject.

9. The method according to claim 8, wherein the sequence of range images 13D images corresponds to fixed number of range images 13D images, preferably acquired with a predefined duration between each image, such as at least 2, 3, 4, 5, 7, or at least 10 range images 13D images, preferably 3 range images 13D images, preferably acquired with a duration of less than 10, 8, 5, 4, 3, 2, or less than 1 second, preferably 1 second, between each range image 13D image.

10. The method according to any one of the preceding claims, wherein a 3D image representing the preselected bovine subject comprises (x,y,z) data, wherein the (x,y) data corresponds to a 2D image of the preselected bovine subject, preferably with normalized intensity, and the z-data corresponds to depth data.

11. The method according to any one of the preceding claims 2-10, wherein IR 2D image data representing the preselected bovine subject comprises (x,y) data, wherein the (x,y) data corresponds to an infrared 2D image of the preselected bovine subject, preferably with normalized IR intensity.

12. The method according to any one of the preceding claims, wherein the body parameter is based on calculating a plurality of said body parameters, wherein the plurality of said calculated body parameters are based on a plurality of different range images, and wherein the plurality of different range images are acquired while the bovine subject is passing through a passage.

13. The method according to claim 12, wherein the range images are acquired over a predefined period of between 1 and 24 hours, such as over a day, where the bovine subject has passed through the passage several times.

14. The method according to any one of the preceding claims 12-13, wherein the body parameter is based on statistics of said plurality of said body parameters, preferablythe median or average of said plurality of said body parameters.

15. The method according to any one of the preceding claims, wherein the at least one range image is acquired from above a passage where through a plurality of bovine subjects are passing through, such as passing through at least once daily, and wherein the at least one bovine subject and / or the preselected bovine subject is identified by means of at least one ID reader located at one or both ends of the passage.

16. The method according to any one of the preceding claims, wherein the image processing comprises the step of segmenting at least part of the acquired range images by means of instance segmentation in order to select a bovine subject among the segmented instances and generate said at least one 3D image of the preselected bovine subject.

17. The method according to claim 16, wherein the segmented bovine instance located closest to an ID reader and a corresponding ID reading is associated with an identification of the preselected bovine subject.

18. The method according to any one of the preceding claims 16-17, wherein instance segmentation is provided by means of Mask Region-based Convolutional Neural Network (Mask-RCNN).

19. The method according to any one of the preceding claims 16-18, wherein the bovine subject selected among the segmented instances is selected based on 1) extension of the segmented instances in the corresponding image, and / or 2) a fraction of 90% of valid depth data in the segmented instance.

20. The method according to any one of the preceding claims, wherein the trained model is a convolutional neural network.

21. The method according to any one of the preceding claims, wherein the trained model is specific to a breed of said preselected bovine subject.

22. The method according to any one of the preceding claims, wherein the trained model is a common model applicable to several different breeds of bovine subjects.

23. The method according to any one of the preceding claims, wherein the trained model for determining weight of the bovine subject is at least partly trained with certified weight as outcome, and wherein the trained model for determining weight of the bovine subject is at least partly trained with certified carcass slaughter weight as outcome.

24. The method according to any one of the preceding claims, wherein training data for the Al model for determining weight of the bovine subject comprises weight calculated based on range imaging vs. certified carcass slaughter weight.

25. The method according to any one of the preceding claims, wherein the trained model is continuously updated, such as once per day, week or month, with certified data from slaughtered bovine subjects, such as certified carcass slaughter weight, certified carcass conformation class, and certified carcass fat cover class.

26. The method according to any one of the preceding claims, wherein updated training data is based on data obtained with a duration less than a predefined period, such as seven days, between acquisition of certified slaughter data and range image acquisition.

27. The method according to any one of the preceding claims, wherein the breed of the bovine subject is selected from the group of: The Jersey breed, the Holstein breed, the Holstein-Friesian cattle population, Holstein Swartzbunt cattle population, the Deutsche Holstein Schwartzbunt cattle population, the US Holstein cattle population, the Red and White Holstein breed, the Deutsche Holstein Schwarzbunt cattle population, the Danish Red population, the Finnish Ayrshire population, the Swedish Red and White population, the Danish Holstein population, the Swedish Red and White population and the Nordic Red population28. A system for determining at least one body parameter of a bovine subject, comprising: an imaging unit for acquiring at least one range image of at least one bovine subject, anda processing unit configured for executing the method of any one of the preceding claims, wherein the at least one body parameter is selected from the group of: the weight, the predicted carcass slaughter weight, the predicted carcass conformation class, and the predicted carcass fat cover class.

29. The system according to claim 28, configured to acquire said at least one range image while the bovine subject is standing in and / or walking through a lock and / or passage,30. The system according to any one of the preceding claims 28 to 29, configured such that said at least one range image is acquired from above the bovine subject thereby imaging said bovine subjects in a top-view,31. The system according to any one of the preceding claims 28 to 30, configured to start acquisition of the range image(s) when triggered by at least one bovine subject approaching and / or entering a lock.

32. The system according to any one of the preceding claims 28 to 31 , wherein the imaging unit comprises at least one RGB camera and a depth sensor in the form of an infrared sensor for providing 2D infrared images.

Images