WO2021191976A1

WO2021191976A1 - Weight estimation device, weight estimation method, and program

Info

Publication number: WO2021191976A1
Application number: PCT/JP2020/012751
Authority: WO
Inventors: 将徳畠中; 小春秋葉; 望美鈴木
Original assignee: Ｎｔｔテクノクロス株式会社
Priority date: 2020-03-23
Filing date: 2020-03-23
Publication date: 2021-09-30
Also published as: JPWO2021191976A1

Abstract

This weight estimation device according to one embodiment is characterized by comprising: an acquisition means which acquires three-dimensional point group data representing a depth value at each point of a weight estimation target; a creation means which creates a projection image in which the point group data acquired by the acquisition means is projected onto a two-dimensional plane; a calculation means which calculates a prescribed index value by using at least one among the point group data and the projection image; and an estimation means which uses the index value calculated by the calculation means to estimate the weight of the estimation target by using a pre-created estimation model.

Description

Weight Estimator, Weight Estimator Method and Program

The present invention relates to a weight estimation device, a weight estimation method and a program.

Livestock farmers have traditionally been able to keep track of the weight of the livestock they are raising. This is because, for example, in domestic animals such as pigs, the value as meat may decrease when the body weight exceeds a certain level.

By the way, the number of livestock raised by livestock farmers may exceed hundreds (or hundreds, hundreds, etc.). Therefore, when using a weight scale or the like, it takes time and effort to grasp the weight of livestock. On the other hand, there is known a technique of photographing livestock with a video camera and estimating the weight from the area of the livestock in the captured image (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 6-3181

However, due to various causes (for example, the posture of the livestock at the time of shooting, the distance between the camera and the livestock, etc.), the error between the estimated body weight and the true body weight was large in the above-mentioned conventional technique.

One embodiment of the present invention has been made in view of the above points, and an object of the present invention is to estimate the body weight with high accuracy.

In order to achieve the above object, the weight estimation device according to the embodiment includes an acquisition means for acquiring three-dimensional point cloud data representing a depth value at each point of the weight estimation target, and the acquisition means acquired by the acquisition means. A creating means for creating a projected image in which point cloud data is projected on a two-dimensional plane, a calculating means for calculating a predetermined index value using at least one of the point cloud data and the projected image, and the calculating means. It is characterized by having an estimation means for estimating the weight of the estimation target by a estimation model created in advance using the index value calculated by.

The weight can be estimated with high accuracy.

It is a figure which shows an example of the whole structure of the weight estimation system which concerns on this embodiment. It is a figure for demonstrating an example of a shooting condition. It is a figure for demonstrating an example of a point cloud data. It is a figure which shows an example of the functional structure of the weight estimation processing part which concerns on this embodiment. It is a flowchart which shows an example of the learning process which concerns on this Embodiment. It is a figure for demonstrating an example of a projected image. It is a figure for demonstrating an example of the search of a fit circle candidate. It is a figure for demonstrating an example of error calculation. It is a figure for demonstrating an example of area removal. It is a figure for demonstrating an example of calculation of a correction projection area. It is a figure for demonstrating an example of calculation of body width. It is a figure for demonstrating an example of calculation of body length. It is a flowchart which shows an example of the weight estimation process which concerns on this embodiment.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In this embodiment, a weight estimation system 1 capable of estimating the weight of a pig with high accuracy will be described. In the present embodiment, a pig is assumed as an example of livestock for which the weight is estimated, but the livestock is not limited to pigs, and is, for example, various livestock such as cows, horses, sheep, and chickens. May be good. Further, the body weight estimation target is not limited to livestock, but may be, for example, pet animals (so-called pets), wild animals, and the like.

<Overall configuration>
First, the overall configuration of the weight estimation system 1 according to the present embodiment will be described with reference to FIG. FIG. 1 is a diagram showing an example of the overall configuration of the weight estimation system 1 according to the present embodiment.

As shown in FIG. 1, the body weight estimation system 1 according to the present embodiment includes a body weight estimation device 10 for estimating body weight and a camera device 20 for photographing a pig P, which is a body weight estimation target. The weight estimation device 10 and the camera device 20 are communicably connected to each other by, for example, wirelessly, wiredly, or both.

The camera device 20 is a digital camera, a smartphone, a tablet terminal, or the like used by a photographer who shoots a pig P. The photographer can obtain imaging data by photographing the pig P from above the pig P using the camera device 20. The camera device 20 may be a camera fixedly installed at a predetermined position.

Here, the camera device 20 is provided with a depth sensor, measures the depth of each point within the shooting range, and generates shooting data indicating the height of each of these points. Therefore, the shooting data generated by the camera device 20 is represented by a point cloud of (x, y, z), where each position in the shooting range is (x, y) and the depth value at each of these positions is z. NS. Hereinafter, the shooting data represented by the point cloud (x, y, z) will be referred to as “point cloud data”.

The body weight estimation device 10 is a computer or computer system that estimates the body weight of the pig P from the point cloud data generated by the camera device 20. The body weight estimation device 10 estimates the body weight of the pig P by a regression equation calculated using the learning data prepared in advance. The training data is data for calculating a regression equation. For example, it is represented by a set of a point cloud data generated by photographing a pig P whose body weight is known and a correct weight indicating the weight of the pig P. Will be done. The regression equation may be called a regression model or the like, and is an example of an estimation model for estimating the weight of the estimation target.

Here, the weight estimation device 10 includes a weight estimation processing unit 100 and a storage unit 110. The weight estimation processing unit 100 is realized by processing one or more programs installed in the weight estimation device 10 to be executed by a processor such as a CPU (Central Processing Unit). Further, the storage unit 110 can be realized by using an auxiliary storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), for example.

The body weight estimation device 10 uses the weight estimation processing unit 100 to calculate a regression equation for estimating the weight of the pig P using the training data prepared in advance, and the weight of the pig P is calculated by this regression equation. Execute the "weight estimation process" to estimate. Further, the storage unit 110 contains information necessary for executing the learning process and the body weight estimation process and various processing results (for example, learning data, a regression equation, point group data of the pig P to be weight-estimated, a body weight estimation result, and the like. ) Is memorized.

The overall configuration of the weight estimation system 1 shown in FIG. 1 is an example, and may be another configuration. For example, the weight estimation system 1 may include a plurality of camera devices 20. Further, a part or all of the weight estimation processing unit 100 may be realized by a function provided by a cloud service or the like.

<Shooting conditions>
Here, the shooting conditions when the photographer shoots the pig P using the camera device 20 will be described. When photographing the pig P, for example, the photographing screen 1000 shown in FIG. 2 is displayed on the display or the like of the camera device 20. A guide 1100 for adjusting the shooting position of the pig P is displayed on the shooting screen 1000. Further, the guide 1100 includes a guide 1110 indicating the center of the photographing range. Here, in the present embodiment, it is assumed that the origin of the point cloud data (that is, the position where (x, y, z) = (0,0,0) is) is the camera position of the camera device 20. Therefore, the position indicated by the guide 1110 is the origin (that is, (x, y) = (0,0)) in any xy plane of the point cloud data. The shooting direction is the positive direction of the z-axis, the downward direction of the shooting screen 1000 (that is, the head direction of the pig P) is the positive direction of the x-axis, and the right direction of the shooting screen 1000 (that is, the left side direction of the pig P). Is the positive direction of the y-axis.

The photographer moves the camera device 20 above the pig P, adjusts the contour of the pig P to match the frame indicated by the guide 1100, and then presses the shooting button 1200 to shoot the pig P. .. As a result, even when a plurality of pigs P are photographed, it is possible to obtain point cloud data in which the relative positional relationship between the origin and the point cloud representing the pig P in the xyz space is substantially the same.

<Point cloud data>
Next, the point cloud data 2000 generated by photographing a certain pig P with the camera device 20 will be described. As shown in FIG. 3, the point cloud data 2000 is a set of points (x, y, z) in the xyz space. As described above, with the camera position of the camera device 20 as the origin, the head direction of the pig P is the positive direction of the x-axis, the left side direction of the pig P is the positive direction of the y-axis, and the shooting direction of the camera device 20 is. It is the positive direction of the z-axis.

Note that how many points a set of points is composed of one point cloud data depends on, for example, the resolution of the camera device 20 and the like.

<Functional configuration>
Next, the functional configuration of the body weight estimation processing unit 100 according to the present embodiment will be described with reference to FIG. FIG. 4 is a diagram showing an example of the functional configuration of the body weight estimation processing unit 100 according to the present embodiment.

As shown in FIG. 4, the weight estimation processing unit 100 according to the present embodiment includes the acquisition unit 101, the preprocessing unit 102, the image creation unit 103, the processing unit 104, the index value calculation unit 105, and regression. The formula calculation unit 106 and the weight estimation unit 107 are included.

The acquisition unit 101 acquires (reads) the learning data stored in the storage unit 110 in the learning process. As described above, the learning data is data represented by a set of a point cloud data generated by photographing a pig P whose body weight is known and a correct answer weight indicating the weight of the pig P. be.

Further, the acquisition unit 101 acquires the point cloud data generated by photographing the pig P to be weight-estimated from the storage unit 110 (or the camera device 20) in the weight estimation process.

The pre-processing unit 102 performs a predetermined pre-processing on the point cloud data acquired by the acquisition unit 101 in the learning process and the weight estimation process. Here, as preprocessing, for example, when points other than the tonp P portion are included in the point cloud data, such points are removed, or the point cloud of the tonp P portion is resampled and converted to a predetermined resolution. Resampling and processing can be mentioned. The pretreatment is a treatment for improving the accuracy of estimating the body weight of the pig, and does not necessarily have to be performed.

The image creation unit 103 creates a projected image in which the point cloud data after the preprocessing is projected onto the xy plane in the learning process and the weight estimation process.

In the learning process and the weight estimation process, the processing unit 104 performs processing for removing the head portion and the tail portion of the pig P from the point cloud data and the projected image after the preprocessing. This is because the head and tail parts of the pig move during shooting, and the head and tail parts are included or not included in the point cloud data, which adversely affects the weight estimation accuracy.

The index value calculation unit 105 calculates a predetermined index value by using the point cloud data after processing and the projected image after processing in the learning process and the weight estimation process. Here, in the present embodiment, the following four index values are calculated.

-Corrected projected area-Posture correction value-Shooting height correction value-Body length and body width Here, the corrected projected area is the area of the projected image corrected in consideration of the inclination (inclination) of the body surface of the pig P. It is a value. The posture correction value is a value representing the posture of the pig P (for example, raising or lowering the neck). The shooting height correction value is a value representing the distance from the camera device 20 to the pig P. The body length and the body width are values representing the body length and the body width of the pig P, respectively.

The regression equation calculation unit 106 calculates the regression equation using the index value calculated by the index value calculation unit 105 and the correct body weight in the learning process. Here, in the present embodiment, the regression equation is represented by the following equation (1).

Estimated weight = a ₁ x (corrected projected area) + a ₂ x (posture correction value) + a ₃ x (shooting height correction value) + a ₄ x (body length) / (body width) + b ... (1)
Here, a ₁ , a ₂ , a ₃ , a ₄ , and b are parameters of the regression equation. The regression equation calculation unit 106 calculates the regression equation _{by estimating these parameters a 1} , a ₂ , a ₃ , a ₄ , and b by a known method (for example, the least squares method).

In the present embodiment, the case where the above four index values are used will be described, but it is not necessary to use all of these four index values, and at least one index value may be used. _{In such a case, among the parameters a 1} , a ₂ , a ₃ , and a ₄ of the above equation (1), the parameter corresponding to the unused index value (that is, the parameter to be multiplied by the index value) is selected. It may be set to 0. For example, when the corrected projected area is not used, the regression equation shown in the above equation (1) _{is set to a 1} _{= 0, and the estimated weight = a 2} × (posture correction value) + a ₃ × (shooting height correction value). It may be + a ₄ × (body length) / (body width) + b. Similarly, for example, when the posture correction value and the shooting height correction value are not used, the regression equation shown in the above equation (1) _{is set to a 2} = 0, a ₃ _{= 0, and the estimated weight = a 1} ×. (Corrected projected area) + a ₄ × (body length) / (body width) + b. The same applies when no other index value is used.

In the body weight estimation process, the body weight estimation unit 107 estimates the weight of the pig P by the regression equation shown in the above equation (1) using the index value calculated by the index value calculation unit 105. However, in the weight estimation process, the parameters a ₁ , a ₂ , a ₃ , a ₄ , and b estimated in the learning process are used.

<Learning process>
Hereinafter, the learning process according to the present embodiment will be described with reference to FIG. FIG. 5 is a flowchart showing an example of the learning process according to the present embodiment. Steps S101 to S105 of FIG. 5 are repeatedly executed for each learning data stored in the storage unit 110. For example, when N learning data are stored in the storage unit 110, steps S101 to S105 are repeatedly executed N times.

First, the acquisition unit 101 acquires one learning data from the learning data stored in the storage unit 110 (step S101).

Next, the preprocessing unit 102 performs a predetermined preprocessing on the point cloud data included in the learning data acquired in step S101 above (step S102). Here and thereafter, as preprocessing for the point cloud data, it is assumed that points other than the tonp P portion are removed and conversion to a predetermined resolution is performed by resampling the point cloud of the tonp P portion. The conversion to a predetermined resolution by resampling means that each point included in the point cloud data is resampled and a three-dimensional point cloud group having a predetermined predetermined interval (for example, the interval between adjacent points) is set. It means to obtain a point cloud (such as a point cloud that is 1.0 cm).

Next, the image creation unit 103 uses the point cloud data preprocessed in step S102 above (or, if no preprocessing is performed, the point cloud data included in the training data acquired in step S101 above). ) Is projected onto the xy plane to create a projected image (step S103). Here, an example of the projected image is shown in FIG. The projected image 3000 shown in FIG. 6 is an image obtained by projecting the point cloud data 2000 onto a certain xy plane, and includes the pig region 3100 on which the point cloud of the pig P portion of the point cloud data 2000 is projected. Here, the point 3110 is the origin of the xy plane, and the point at the position indicated by the guide 1110 in FIG. 2 is projected onto the xy plane.

Further, at this time, the image creation unit 103 may calculate a dividing line 3120 that divides the pig region 3100 in the projected image 3000 into left and right (left and right with respect to the traveling direction when the pig walks). Such a dividing line 3120 is when the x-coordinate value included in the pig region 3100 is x ₁ , x ₂ , ..., X _M, and the x-coordinate value is x _i (1 ≦ i ≦ M). Assuming that the y coordinate value included in the pig region 3100 can take y _{min, i} ≤ y ≤ y _{max, i} , for each i = 1, ..., M-1, (x _i , (y _{max). , I-} y _{min, i} ) / 2) and (x _{i + 1} , (y _{max, i + 1-} y _{min, i + 1} ) / 2) are calculated by connecting them with a straight line on the xy plane. Alternatively, a curve represented by a function that approximates the point cloud {(x _i , (y _{max, i-} y _{min, i} ) / 2) | i = 1, ..., M} may be the dividing line 3120. .. It should be noted that such a function can be obtained by, for example, a library such as polyfit.

Next, the processing unit 104 uses the point cloud data preprocessed in step S102 above (or, if no preprocessing is performed, the point cloud data included in the learning data acquired in step S101 above). And the projected image created in step S103 above, processing is performed to remove the head portion and the tail portion of the pig P (step S104). Here, the processing unit 104 performs the processing according to the following Steps 1 to 7. In the following, as an example, a case where the point cloud data 2000 shown in FIG. 3 and the projected image 3000 shown in FIG. 6 are processed will be described.

Step1: First, as shown in FIG. 7, the processing unit 104 sets j = 1, the preset window size is L> 0, and the search area R ₁₁ = {(x, y) | x ₁₁ ≦ x ≦ x ₁₁ + L} and R ₂₁ = {(x, y) | x ₂₁ −L ≦ x ≦ x ₂₁ } are set. Note that x ₁₁ ≧ 0 and x ₂₁ ≦ 0 are start points of the x-coordinate of the search region and are set in advance.

_{Step2: Next, the processing unit 104 creates a circle S 11} that is a point on the boundary of the pig region 3100 _{and fits the point included in the search region R 11} (that is, a circle that approximates these point groups). presume. Similarly, the processing section 104 estimates a _{circle S 11} that is a point on the boundary of the pig region 3100 and fits the point included in the _{search region R 21.} Hereinafter, these circles will be referred to as "fit circle candidates". The fit circle candidate can be estimated by the least squares method. For example, refer to "https://myenigma.hatenablog.com/entry/2015/09/07/214600".

Step3: Subsequently, the processing unit 104 sets the search area for the projected image 3000 as _{j ← j + 1, x 1j} = x _{1 (j-1)} + d ₁ , x _2j = x _{2 (j-1)} −d _2. R _1j = {(x, y) | x _1j ≤ x ≤ x _1j + L} and R _2j = {(x, y) | x _2j- L ≤ x ≤ x _2j } are set. Here, d ₁ , d ₂ > 0 are preset values. That is, the processing unit 104 _{slides the search area R 1 (j-1)} _{by d 1 in the} x-axis direction to make the search area R _1j, and sets the search area R _{2 (j-1)} in the x-axis direction. Slide d ₂ to obtain the search area R _2j .

Step4: Next, the processing unit 104 estimates a _{circle S1j} that is a point on the boundary of the pig region 3100 and fits the point included in the _{search region R1j.} Similarly, the processing unit 104 estimates a _{circle S 11} that is a point on the boundary of the pig region 3100 and fits the point included in the _{search region R 2j.}

Steps 3 to 4 described above are repeated, for example, with the range of values that the x coordinate can take in the projected image 3000 as X ₁ ≤ x ≤ X ₂ until x ₂₁ −L ≤ X ₁ and X ₂ ≤ x _1j + L are satisfied. Will be executed. As a result, a set of fit circle candidates {S _1j } and {S _2j } can be obtained.

Step5: Then, the processing unit 104 from the set of fit circle candidate _{{S 1j}} with selecting fit circle _{S 1} for removing head portion, from the set of fit circle candidate _{{S 2j}} to select the fit circle S ₂ to remove the tail portion. Here, a method of selecting the fit circle S _{1 will be described.}

As shown in FIG. 8, with the center _{of the fit circle candidate S 1j} _{as O j} , a straight line passing through O _j and Q _m _{is drawn for each point Q m} on the boundary B of the pig region 3100, and the straight line and the fit circle are drawn. _{Let P mj} be the intersection with candidate S _1j . Further, the distance between _{P mj} and Q _m on the straight line _{is Δ mj} . Each point _{Q m (m = 1, ···} ) on the distance delta _mj boundary B on the calculated relative to the mean value of these distances delta _mj an error of fit circle candidate _{S 1j.} Then, based on the error of each fit circle candidate S _1j , one fit circle candidate is selected as the _{fit circle S 1} from the set of fit circle candidates {S _1j}. As a method for selecting fit circle S _1, for example, the error is smallest fit circle candidate may be selected as fit circle S _1, may be selected fit circle S ₁ from the degree of change of the error. As a method of selecting a fit circle from the degree of change in the error, for example, when plotting the x-coordinate of the center of each fit circle candidate and the error of the fit circle candidate, the error of the fit circle candidate having the minimum value is obtained. x-coordinate is the largest-fit circle candidate at medium (that is, an error takes a minimum value, and fit circle candidate on the most right side) may be selected as fit circle S _1. However, if the error is not present fit circle candidate takes a minimum value may be selected smallest-fit circle candidate errors as fit circle S _1. Although the same applies to the method of selecting the fit circle S _2, when selecting the fitting circle S ₂ from the degree of change of the error is smallest x-coordinate among the fit circle candidate error takes a minimum value fit circle candidate (i.e., an error takes a minimum value, and fit circle candidate on the leftmost) is selected as the fit circle S _2.

Step 6: Then, the processing unit 104 removes the head portion and the tail portion in the pig region 3100 by using _{the fit circles S 1} and S _{2 obtained in the above Step 5, respectively.}

For example, when the fit circle S ₁ shown in FIG. 9 is obtained, the machined portion 104 _{selects the straight line parallel to the y-axis and in contact with the fit circle S 1} with the larger x-coordinate value as T _1. as deletes the region x-coordinate value is greater than the straight line T ₁ in pigs region 3100. As a result, the head portion in the pig region 3100 is removed.

Further, for example, if the fit circle S ₂ shown in FIG. 9 is obtained, the processing unit 104 is parallel to the y-axis, and, among the line tangent to fit circle S _2, a straight line towards the x-coordinate value is smaller As T ₂ , the region of the pig region 3100 whose x-coordinate value is smaller than that _{of the straight line T 2 is deleted.} As a result, the tail portion in the pig region 3100 is removed.

In the above, parallel to the y-axis, and, among the line tangent to fit circle S _1, but the straight line towards the x-coordinate value is larger and the T _1, not limited thereto, for example, parallel to the y-axis and, to a straight line passing through the center of the fit circle _{S 1} may be _{T 1,} the tangential fit circle _{S 1} at the intersection of the dividing lines 3120 and fit circle _{S 1} may be _{T 1.} Similarly, for example, parallel to the y-axis, and, to a straight line passing through the center of the fitting circle S ₂ may be T _2, the tangential fit circle S ₂ at the intersection of the dividing lines 3120 and fit circle S ₂ T _It may be 2.

Also, instead of deleting the region with a linear T _1, for example, among the pigs region 3100, a large x-coordinate value of the center of the fit circle S _1, and each on the circumference of the fit circle S ₁ of The area in which the x-coordinate value is larger than the point may be deleted. Similarly removed, for example, among the pigs region 3100, the x-coordinate value is smaller than the center of the fitting circle S _2, and an area in the range x-coordinate value is smaller than the points on the circumference of the fitting circle S ₂ You may.

Further, in step S104 above, both the head portion and the tail portion in the pig region 3100 were removed, but for example, instead of removing both, only one of them may be removed. Further, when the area of the area to be removed (deleted) is equal to or less than a predetermined threshold value, the area may not be removed.

Step 7: Finally, the processed portion 104 is projected onto the head portion and the tail portion when creating the point cloud (that is, the projected image 3000) corresponding to the portions (head portion and tail portion) removed in Step 6 above. The point cloud) is removed from the point cloud data 2000.

Return to Fig. 5. Following step S104, the index value calculation unit 105 calculates the corrected projected area, the posture correction value, the shooting height correction value, and the body length and body width as index values (step S105). The calculation method of each index value will be described below.

≪Corrected projected area≫
When the projected image 3000 is created from the point cloud data 2000 in step S103 above, the inclination of the body surface of the pig P is lost. Therefore, the area of the pig region 3100 becomes smaller than the area of the body surface of the pig P corresponding to the pig region 3100. Therefore, the corrected projected area obtained by correcting the area of the pig region 3100 in consideration of the inclination of the body surface of the pig P is calculated.

When resampling is performed in step S103 so that the distance between adjacent points is 1.0 cm, the distance between the pixels of the projected image 3000 is also 1.0 cm. Therefore, as shown in FIG. 10, four pixels at positions (x, y) = (4,6), (5,6), (4,5), (5,5) are in the pig area 3100. When included, the area of the area V ₁ (this area represents a "unit area") ^{surrounded by these four pixels is 1.0 cm 2} . Similarly, the area of the unit areas V ₂ to V ₅ is 1.0 cm ² . The area of the pig area 3100 is calculated by the total area of the unit areas included in the pig area 3100.

At this time, for example, among the points included in the point cloud data 2000, (x, y) = (4,1), (4,2), (4,3), (4,4), (4,5 ), The z coordinate value of the point corresponding to (4, 6) is Z ₁ = 1.0, and (x, y) = (5, 1), (5, 2), (5, 3), It is assumed that the z coordinate values of the points corresponding to (5,4), (5,5), and (5,6) are Z ₂ = 1.5. In this case, the inclination of the body surface of the pig P is lost in _{the unit regions V 1} to V _5.

Therefore, as shown in FIG. 10, 1.0 / sinθ ≒ 1.12 corrects the x-axis direction of the side of the unit region _V 1 ~ _{V 5} by (i.e., multiplied by 1.0 / sin [theta), corrected The rear unit area is W ₁ to W ₅ . Then, instead of the area of the unit area _V 1 ~ _{V 5,} using the area of the corrected unit area _W 1 ~ _{W 5,} calculates the area of the pig region 3100. This area is the corrected projection area. In the example shown in FIG. 10, the side in the x-axis direction of the unit region is corrected, but the side in the y-axis direction may be corrected, and both the side in the x-axis direction and the side in the y-axis direction are corrected. May be done.

≪Posture correction value≫
Since the posture of the pig P is different between the case where the pig P is raising the neck and the case where the pig P is lowering the neck, the accuracy of weight estimation may be affected. Therefore, the posture correction value is calculated as an index value.

Of the points included in the point cloud data 2000, let f be a polynomial function that approximates the points (x, y, z) corresponding to the points (x, y) on the dividing line 3120 described in FIG. That is, let f be a polynomial function such that z≈f (x, y), (where (x, y) is a point on the dividing line 3120). Such a function f can be obtained by, for example, a library such as polyfit.

Then, a predetermined coefficient of the polynomial function f (for example, a quadratic coefficient of x, a quadratic coefficient of y, a coefficient of xy, etc.) is used as the posture correction value.

In the above, the polynomial function f such that z = f (x, y) is obtained by using (x, y) as a point on the dividing line 320, but the present invention is not limited to this, and for example, a predetermined xz plane (for example, , Y = 0), a polynomial function f (that is, a polynomial function f such that z≈f (x)) that approximates the points (x, z) included in the point group data 2000 may be obtained. .. In this case as well, a predetermined coefficient of the polynomial function f (for example, a quadratic coefficient of x) may be used as the posture correction value.

≪Shooting height correction value≫
The distance from the camera device 20 to the pig P varies depending on the height of the photographer, the length of the hand, and the like, which may affect the accuracy of weight estimation. Therefore, the shooting height correction value is calculated as an index value.

Among the points included in the point cloud data 2000, the shooting height correction is performed for the linear distance between the point having the smallest z coordinate value at x = 0 and y = 0 and the origin (that is, the camera position of the camera device 20). It can be a value.

≪Body length and width≫
Since the body length and width differ depending on the pig P, it may affect the accuracy of weight estimation. Therefore, the body length and body width are calculated as index values.

As shown in FIG. 11, a straight line 3130 that passes through the point 3110 (that is, the origin of the xy plane) and ends at a point on the boundary of the pig region 3100 is calculated and is orthogonal to the dividing line 3120. Then, the length of this straight line 3130 may be the body width.

Further, as shown in FIG. 12, among the points included in the point cloud data 2100 after removal in Step 7, the points (x, y, z) corresponding to the points (x, y) on the dividing line 3120 are defined. The function to be approximated may be f, and the length of the curve 2110 represented by this function f may be the body length.

Return to FIG. After the steps S101 to S105 are repeatedly executed for each learning data, the regression equation calculation unit 106 calculates the regression equation using the index value and the correct body weight corresponding to each learning data (step S106). That is, the regression equation calculation unit 106 _{initializes the parameters a 1} , a ₂ , a ₃ , a ₄ , and b to appropriate initial values, and then uses the index value corresponding to the training data for each training data. The estimated body weight is calculated by the above formula (1). Then, the regression equation calculation unit 106 calculates the difference between the estimated body weight and the correct body weight corresponding to the learning data, and then minimizes the sum of the squares of the differences calculated for each learning data. , Parameters a ₁ , a ₂ , a ₃ , a ₄ , b are estimated. As a result, a regression equation for estimating the body weight of pig P is calculated.

<Weight estimation process>
Hereinafter, the body weight estimation process according to the present embodiment will be described with reference to FIG. FIG. 13 is a flowchart showing an example of the weight estimation process according to the present embodiment. _{As the parameters a 1} , a ₂ , a ₃ , a ₄ , and b of the regression equation shown in the above equation (1), those estimated by the above learning process are used.

First, the acquisition unit 101 acquires the point cloud data generated by photographing the pig P whose body weight is to be estimated from the storage unit 110 (or the camera device 20) (step S201).

Next, the preprocessing unit 102 performs a predetermined preprocessing on the point cloud data acquired in the above step S201 in the same manner as in step S102 of FIG. 5 (step S202).

Next, the image creation unit 103 is acquired in the point cloud data (or, if the preprocessing is not performed, in the above step S201) that was preprocessed in the above step S202, similarly to the step S103 in FIG. A projected image is created by projecting the point cloud data) onto the xy plane (step S203).

Next, the processing unit 104 was acquired in the point cloud data (or, if the preprocessing is not performed, in the above step S201) that was preprocessed in the above step S202, similarly to the step S104 in FIG. The point cloud data) and the projected image created in step S203 are processed to remove the head portion and the tail portion of the pig P (step S204).

Next, the index value calculation unit 105 calculates the corrected projected area, the posture correction value, the shooting height correction value, and the body length and body width as index values (step S205).

Then, the body weight estimation unit 107 estimates the weight of the pig P by the regression equation shown in the above equation (1) using the index value calculated in the above step S205 (step S206). From this, the weight of the pig P is estimated.

As described above, the body weight estimation system 1 according to the present embodiment removes the image region of the portion (that is, the head portion and the neck portion of the pig P) that is likely to adversely affect the body weight estimation, and then determines a predetermined index. Calculate the values and estimate the body weight by the regression equation using these index values. This makes it possible to estimate the weight of pigs with high accuracy.

The present invention is not limited to the above-described embodiment disclosed specifically, and various modifications and modifications, combinations with known techniques, and the like are possible without departing from the description of the claims. ..

1 Weight estimation system 10 Weight estimation device 20 Camera device 100 Weight estimation processing unit 101 Acquisition unit 102 Preprocessing unit 103 Image creation unit 104 Processing unit 105 Index value calculation unit 106 Regression formula calculation unit 107 Weight estimation unit 110 Storage unit

Claims

An acquisition means for acquiring three-dimensional point cloud data representing the depth value at each point of the weight estimation target, and
A creating means for creating a projected image by projecting the point cloud data acquired by the acquiring means on a two-dimensional plane, and
A calculation means for calculating a predetermined index value using at least one of the point cloud data and the projected image, and
An estimation means for estimating the weight of the estimation target by an estimation model created in advance using the index value calculated by the calculation means, and an estimation means.
A body weight estimator characterized by having.
The calculation means is
The body weight estimation device according to claim 1, wherein a value representing a body shape to be estimated is calculated as the index value.
The point cloud data is created by photographing the estimation target with a depth sensor.
The calculation means is
The weight estimation device according to claim 1 or 2, wherein the distance from the position where the estimation target is photographed to the estimation target is calculated as the index value.
The point cloud data is represented by (x, y, z), where the coordinate value of the point is (x, y) and the depth value is z.
The calculation means is
A polynomial function that approximates the depth value z corresponding to the coordinate values (x, y) of each point on a predetermined line or curve in the xy plane is calculated, and the predetermined coefficient of the polynomial represented by the calculated polynomial function is used as the index. The weight estimation device according to any one of claims 1 to 3, wherein the weight estimation device is calculated as a value.
The calculation means is
Any one of claims 1 to 4, wherein the area of the region representing the estimation target in the projected image is corrected by the depth value, and the corrected area is calculated as the index value. The weight estimation device described in.
An acquisition procedure for acquiring three-dimensional point cloud data representing the depth value at each point for weight estimation, and
A creation procedure for creating a projected image by projecting the point cloud data acquired in the acquisition procedure on a two-dimensional plane, and a procedure for creating the projection image.
A calculation procedure for calculating a predetermined index value using at least one of the point cloud data and the projected image, and
An estimation procedure for estimating the weight of the estimation target by an estimation model created in advance using the index value calculated in the calculation procedure, and an estimation procedure.
A weight estimation method characterized by a computer performing.
A program for making a computer function as each means in the weight estimation device according to any one of claims 1 to 5.