Method for discriminating between containers in a sorting line, and associated installation

Abstract

The invention relates to a method (400) for discriminating between items according to a type of packaging chosen from rigid packaging and flexible packaging, comprising the steps of acquiring (S10) at least one digital image (NUM) of an item by way of an image capture device, computing (S20) statistical parameters (PARAM) representative of a plurality of points located on a surface of the item in the digital image (NUM), constructing (S30) a vector (VECT) having components defined by values of the statistical parameters (PARAM), on the basis of the vector (VECT), computing (S40) a probability (PROBA) of the item being an item with rigid packaging or of being an item with flexible packaging, comparing (S50) the probability (PROBA) with at least one threshold value and, in response to the comparison, generating a parameter (DET) representative of the type of packaging of the item.

Description

METHOD FOR DISCRIMINATING BETWEEN CONTAINERS IN A SORTING LINE AND ASSOCIATED INSTALLATION TECHNICAL FIELD OF THE INVENTION

[0001] The invention relates to a method of discriminating between articles circulating on an article sorting line, with the aim of adapting the handling of each of these articles to their respective containers. TECHNOLOGICAL BACKGROUND

[0002] The sorting of items such as postal parcels is carried out within sorting centers comprising sorting lines intended to carry out automated sorting of items, each handling a flow of items. The items are handled along the lines by means of collective conveyor systems but require, for example for individual picking or placement, to be picked individually during an operation called "pick and place", or "pick and place" in English terminology, implemented by means of a robotic manipulator arm, as described for example in document FR 3 019 068 B1. An example of a manipulator arm using suction cups as a device for gripping items is described in document FR 3 031 046 A1. Of course, in order to make sorting lines profitable, they must handle as large a flow of items as possible and the speed of execution of the "pick and place" operation is essential.

[0003] The function of the manipulator arm is to grasp the item, move it to a predetermined location, and then release it. This operation must be rapid yet reliable in order to limit handling errors. An example of handling error is "fly-out," which is an unexpected release of the item while it is being moved at high speed by the arm and can result in the item being ejected from the sorting line.

[0004] There is therefore a need to maximize the speed of “pick and place” operations while limiting handling errors.

[0005] The items to be sorted moving along a sorting line may be contained in different types of packaging, in particular flexible plastic packaging known as "polybags" and rigid packaging known as "boxes" such as cardboard boxes. The quality of the gripping of a box by a manipulator arm is better than the quality of gripping a polybag, so that the acceleration imposed on a box by a manipulator arm without causing a fly-out may be higher than that imposed on a polybag by the same manipulator arm. Thus, it is understood that for a given level of reliability of the manipulation of an item by the manipulator arm, it may impose a stronger acceleration and a higher speed on a box than that which it may impose on a polybag.

[0006] Based on this observation, the applicant proposes a sorting method with differentiated treatment between a box and a polybag: this involves making the manipulator arm work at a level closer to its maximum capacities, in terms of speed of execution of the movement of the article, when it manipulates a box than when it manipulates a polybag.

[0007] Such a process, including differentiated treatment between a box and a polybag, requires the ability to automatically distinguish the box from the polybag.

[0008] To achieve this aim, a first aspect of the invention is a method for discriminating between articles according to a type of packaging chosen from rigid packaging and flexible packaging, comprising the steps of acquiring at least one digital image of an article by means of a camera, calculating statistical parameters representative of a plurality of points located on a surface of the article in the digital image, constructing a vector having components defined by values ​​of the statistical parameters, on the basis of the vector, calculating a probability for the article to be a rigidly packaged article or to be a flexiblely packaged article, comparing the probability to at least one threshold value and, in response to the comparison, generating a parameter representative of the type of packaging of the article.

[0009] This process allows discrimination between items packaged in flexible packaging such as a polybag and items packaged in rigid packaging such as a box. Thus, integrated into a sorting procedure, this process allows, for example, the optimization of the throughput of a device for inserting items to be sorted into a sorting machine using a robotic handling arm.

[0010] According to additional non-limiting characteristics of the first aspect of the invention, considered individually or in any technically feasible combination:

[0011] - the steps of calculating statistical parameters, constructing a vector, calculating a probability and comparing the probability to at least one threshold value can be implemented by means of a digital computer;

[0012] - the step of acquiring at least one digital image of an item may further comprise the steps of acquiring a digital image of a scene comprising the item by means of the camera and of segmenting the digital image of the scene so as to isolate the item and obtain the digital image of the item;

[0013] - at the stage of calculating a probability, the components of the vector can be weighted by weighting coefficients determined in advance during a machine learning stage on the basis of annotated images;

[0014] - a depth map can be established, and the statistical parameters can be representative of a surface topography;

[0015] - the statistical parameters can be representative of deviations of points constituting the surface from a plane approximating this surface;

[0016] - components of the vector can be chosen from a mean, a median, a standard deviation of the deviations, a first, a second, a third and a fourth moment of a normalized histogram of the deviations, a value of a cumulative histogram of the deviations, and integrals of the cumulative histogram;

[0017] - the statistical parameters are representative of a shade or texture of the surface of the item in the digital image, the digital image being in color;

[0018] - the step of calculating statistical parameters may comprise a sub-step of determining a plurality of points distributed over the surface and a sub-step of defining, for each of the points, increasing surface areas, each centered on the point considered, the statistical parameters being calculated on the basis of these surfaces;

[0019] - components of the vector are chosen from an average of gray levels of pixels of one or more of the zones, and from respective averages of the color levels of pixels of one or more of the zones; and

[0020] - the method may combine a probability calculation based on statistical parameters that may be representative of deviations of points constituting the surface from a plane approximating a surface of the item in the image with a probability calculation based on a tint or texture of a surface of the item in the digital image.

[0021] The invention extends to a method for processing articles implementing the method for discriminating articles according to the invention, and further comprising the step of transmitting a control signal to an article processing device in response to the generation of the parameter representative of the type of packaging of the article.

[0022] This process allows for differentiated processing between items according to their packaging, and therefore to optimize processing speed according to the priority objectives (speed, reliability) set by the practitioner.

[0023] A second aspect of the invention relates to a system configured to implement the method of the first aspect of the invention within a sorting facility.

[0024] In order to achieve this aim, a second aspect of the invention is an article sorting installation configured to implement the article processing method according to the invention.

[0025] This installation allows for the efficient processing of a flow of items to be handled individually, with processing adapted to their respective packaging types, box or polybag. More specifically, for a given level of reliability, individual item processing can be accelerated when the item is packed in a box.

[0026] According to additional non-limiting characteristics of the second aspect of the invention, considered individually or in any technically feasible combination:

[0027] - the article processing device may be a robotic manipulator arm;

[0028] - the robotic manipulator arm can be equipped with a suction gripping device. BRIEF DESCRIPTION OF THE FIGURES

[0029] Other characteristics and advantages of the invention will emerge from the detailed description of the invention which follows with reference to the appended figures in which:

[0030] Represents a portion of an article processing facility;

[0031] La represents a normalized histogram obtained in a first mode of implementation;

[0032] Larepresents a cumulative histogram of the normalized histogram of the ;

[0033] Illustrates the principle of a process of discrimination between a box and a polybag;

[0034] Illustrates a first mode of implementation of the process.

[0035] Illustrates a second mode of implementation of the process.

[0036] Illustrates a third mode of implementation of the process.

[0037] Illustrates the second mode of implementation. DETAILED DESCRIPTION OF THE INVENTION

[0038] General principle

[0039] The principle of the solution proposed by the applicant, which consists of a method 400 for discriminating between one of the IT articles packed in a box and articles packed in a polybag, is put into context using the, which represents a portion of a FAC installation for processing articles comprising an SL sorting line equipped with a SYS system configured to implement the discrimination method.

[0040] The sorting line SL comprises a robotic manipulator arm MA equipped with a gripping device GH and a conveyor belt CB moving arriving articles in the direction Dir represented by an arrow. The gripping device GH may comprise suction cups. The articles may arrive in bulk in a work zone WZ, the manipulator arm being responsible for unstacking them, i.e. gripping them one by one to place them in order on the conveyor belt CB.

[0041] The SYS system consists of a CAM digital camera and a CALC computer functionally connected to the CAM digital camera and the MA manipulator arm.

[0042] The method 400, illustrated by the, consists in a step S10 of acquiring at least one digital image NUM of a scene encompassing the working area WZ of the manipulator arm MA by means of the digital camera CAM, in which one or more articles IT, intended to be manipulated individually by means of the manipulator arm, are visible. A plurality of points of the image are located on the surface SURF of an article of interest among the one or more articles IT.

[0043] In a step S20, a plurality of statistical parameters PARAM representative of a plurality of points located on the surface SURF of the article of interest are calculated by means of the digital calculator CALC.

[0044] At a step S30, the calculator CALC constructs at least one vector VECT whose components are respectively defined by one of the statistical parameters PARAM calculated at step S20.

[0045] In a step S40, the calculator CALC calculates, on the basis of the at least one vector VECT, a probability PROB associated with the fact that the article of interest is packaged in a box-type packaging or is packaged in a polybag-type packaging.

[0046] The number of components defining the vector is defined by the practitioner, according to a compromise between the necessary calculation time and a target reliability level, the increase in the number of components improving the reliability of the discrimination, but only marginally beyond a certain number, while the calculation times increase significantly with the number of components. However, the calculation time is a crucial data, considering that it is a real-time calculation and is integrated into a chain of calculations including, in addition to the calculation of the statistical parameters and the probability calculation mentioned above, the acquisition of a digital image, the segmentation of the objects visible in this digital image by a neural network so as to identify the visible articles, the determination of the order of grasping these articles, or the determination of the trajectory of the manipulator arm for grasping the article.Thus, for a manipulator arm having a cycle of the order of 2 seconds, we consider that a time of the order of ten milliseconds can be devoted to calculating the components of the vector VECT and the associated probability.

[0047] In a step S50, the probability PROB is compared to one or more predetermined threshold values ​​to allow the probability PROB to be associated with a type of packaging, box or polybag, or with a status of indeterminacy of the type of packaging due to too high a possibility of error with regard to this association. Step S50 results in the generation of a parameter DET, representative of the determination of the type of packaging of the article, box, polybag, or indeterminate.

[0048] In a step S60, in response to the comparison of step S50 and the parameter DET, the computer sends a COM command to the manipulator arm MA. The manipulator arm MA is configured to, in response to receiving the COM command, adjust its operational speed, i.e. the acceleration imposed on an item that it grasps and moves, the maximum movement speed of this item, or these two parameters simultaneously.

[0049] Thus, the SYS system is configured to adapt the handling speed of each item individually to its packaging type. In particular, when the item is packed in a box, the speed and acceleration can be maximum, while they will be relatively lower when the item is packed in a polybag. The result is an increase in the throughput of items handled by the manipulator arm for a given error rate.

[0050] Illustrates a particular example of application of the discrimination method according to the invention, which consists of a bin picking operation carried out by means of a manipulator arm. The applications of the method are however not limited to this particular example, considered for the purposes of explanation.

[0051] The principle described above is applicable to several types of digital images, and several types of statistical parameters can be used, as described more precisely in the two implementation modes taken as examples below.

[0052] First mode of implementation

[0053] The first implementation method is a method 500 for discriminating between an item packaged in a box and an item packaged in a polybag, illustrated by the. This method is based on the differences between the apparent geometries of a box and a polybag. While the polybag is by nature flexible and therefore has a surface of irregular topography, a box is by nature rigid, so that its faces have a regular topography, closer to a perfectly flat surface than a polybag. Thus, this first implementation method takes advantage of the "flat" aspect of the surfaces of a box to identify this type of packaging.

[0054] In this first embodiment, the digital camera CAM is a digital camera having three lenses: two lenses dedicated to stereoscopic vision and a third lens dedicated to the acquisition of a color image. Thus, the at least one digital image of step S10 consists of at least three images captured simultaneously, two intended for the formation of a depth map and the third for a two-dimensional color image.

[0055] The camera is configured to internally form, in a sub-step S10a of step S10, said depth map and to carry out a pixel-by-pixel alignment between the color image and the depth map. Such a depth map makes it possible to position each visible point of the image in a three-dimensional positioning reference frame. This embodiment uses a digital camera having three lenses, but the invention is not limited to this particular means, and this embodiment only requires a depth map and a color image aligned with each other, independently of the means for obtaining this map and this image.

[0056] In a sub-step S10b of step S10, the visible items are identified in the color image by segmentation using a neural network. By identification between the color image and the depth map on which it is aligned, a cloud of points located in a three-dimensional space is recovered for each of the items.

[0057] In a sub-step S10c of step S10, one of the visible articles is chosen, for example according to a criterion of order of entry of the articles determined in a conventional manner to optimize the processing of the flow of arriving articles. A visible face of the chosen article, represented by the surface SURF on the, is used for digital processing during the rest of the method. It can be an upper face of the article, a face having an orientation substantially parallel to a work plane on which the article rests, or the visible face of the article having the orientation closest to that of the work plane, or the face having the largest area among the visible faces of the article. Subsequently, the points of the cloud associated with this face of the chosen article are used by the CALC calculator.

[0058] Step S20 consists of the CALC calculator exploiting the positioning of the points constituting the face of the article in the depth map, corresponding to the surface SURF on the.

[0059] In a sub-step S20a of step S20, a first plane is defined so as to approximate the chosen face, each point of the chosen face being used in the determination of the parameters a, b and c determining the equation Eq. 1 of this first plane.

[0060] Eq. 1

[0061] The chosen face is formed by a cloud of points from the depth map, each of these points being located in space by three x coordinates i , y i and z i , where the index i identifies a point i, for example in an orthonormal frame. The parameters a, b and c are determined conventionally by reduction of the quadratic error E which can be expressed by the equation Eq. 2

[0062] Eq. 2

[0063] In a sub-step S20b of step S20, filtering is applied so as to eliminate the points considered to be aberrant, generated for example by the noise of the stereoscopic camera. The filtering may for example consist of eliminating a certain proportion of the points, 10% for example, the eliminated points being those whose distances to the foreground are the greatest. Another filtering could be the use of a threshold on the distance separating a point from the foreground, threshold beyond which a point is considered to be aberrant. This threshold is determined by the practitioner in order to adapt it to his equipment and to the algorithms used for example in the determination of the depth map.

[0064] Subsequently, the points considered are the points having passed the filtering step, that is to say all the points of the chosen surface with the exception of the points considered as aberrant.

[0065] In a sub-step S20c of step S20, step S20a is repeated in order to determine the parameters a', b' and c' of a second plane approximating the chosen face, represented by the surface SURF in the, with the difference that the eliminated outlier points are excluded from the calculation of the parameters a', b' and c'. By proceeding in this way, the second plane constitutes a better approximation of the chosen face than the first plane.

[0066] In a sub-step S20d of step S20, the calculator CALC calculates a plurality of characteristics based on the relative positions of the points of the chosen face with respect to the second plane defined by the parameters a', b' and c'.

[0067] More specifically, in sub-step step S20d, the calculator calculates a relative error or deviation for each of the points of the surface chosen on the basis of equation Eq. 3.

[0068] Eq. 3

[0069] The gaps are defined as the distance from each of the points considered in the background, and can be considered as random variables in terms of probabilities. We can consider, for a box and a polybag, a histogram of the deviations , which is interpreted as follows: H(x)=y means that there are y points of the chosen surface which are associated with a relative error between x and x+0.5, if we work with a step of 0.5 (quantification of the values ​​of ). Figure 2 represents such a normalized histogram according to equation Eq. 4.

[0070] Eq. 4

[0071] Figure 3 represents the distribution function, or cumulative histogram , of the histogram , which can be calculated according to Eq. 5.

[0072] Eq. 5

[0073] Figures 2 and 3 illustrate that a polybag and a box have very different characteristics in terms of their deviations respective. Such statistical characteristics can therefore be used to determine the type of packaging, box or polybag, of any item. For this purpose, the characteristics illustrated by Figures 2 and 3 alone are insufficient, and a more extensive set of characteristics must be calculated for each point on the chosen surface.

[0074] At a sub-step S20e of step S20, the CALC calculator calculates a set of statistical characteristics based on the deviations , these characteristics constituting statistical parameters PARAM representative of the topography of the points located on the surface of the article of interest, and more specifically of the deviations of the points of the surface from a plane approximating this surface. A set of characteristics can be made up of the following ten characteristics: mean, median and standard deviation of the , first, second, third and fourth moment of the normalized histogram gaps , a value of the cumulative histogram , and the two integrals, respectively from 0 to 9 and from 10 to 19 of the cumulative histogram

[0075] The feature set defined above was determined as a compromise between the computation time required and the reliability of discrimination between box and polybag. In practice, other choices of features to constitute the claim set are possible, and their number could also be different. The number of features constituting the vector VECT is between 6 and 14, preferably between 8 and 12, more preferably between 9 and 11. It is 10 in the present example.

[0076] In step S30, the computer CALC constructs the vector VECT on the basis of the deviations , of the normalized histogram , and the cumulative histogram . More specifically, each of the 10 features of the feature set calculated in step S20 constitutes one of the components X m of the vector VECT, m being here a natural integer varying from 1 to 10.

[0077] In step S40, by means of a logistic regression calculation, the CALC calculator calculates the probability PROBA associated with the fact that the item of interest is packed in a box, according to the equation Eq. 6 Eq. 6 in which the components are each weighted by a weighting coefficient .

[0078] The PROBA probability is intended to be compared to threshold values ​​Thr1 and Thr2 to determine whether the item packaging is a box, a polybag, or should be considered undetermined, as detailed below. In a preparatory step, optimal values ​​of the weighting coefficients are determined for this purpose by machine learning using a database of sets of images of articles packaged in boxes or polybags, annotated so as to indicate the nature of the packaging of the articles, these images being of a nature similar to those obtained in steps S10 of this first embodiment, with a depth map and a color image used for the identification of the articles. The learning is designed to minimize the number of confusions between the types of article packaging, between box and polybag during the following step S50.

[0079] In step S50, the calculator CALC compares the probability PROBA to two predetermined threshold values ​​Thr1 and Thr2. If PROBA is smaller than or equal to the first threshold value Thr1, then the packaging of the item of interest is considered to be a polybag. If PROBA is larger than or equal to the second threshold value Thr2, then the packaging of the item of interest is considered to be a box. If PROBA is larger than Thr1 and smaller than Thr2, then the packaging of the item of interest is considered to be uncertain.

[0080] The first threshold value Thr1 and the second threshold value Thr2 can be respectively set to the values ​​(1 / 2 - incert_margin) and (1 / 2 + incert_margin), where incert_margin represents a parameter for adjusting the compromise to be established between the performance of the manipulation by the manipulator arm and the minimization of confusions between a box and a polybag. This parameter makes it possible to very simply adjust the performance / confusion compromises according to the intended application, which can be characterized by a flow of articles to be handled, an article picking dynamic, or the type of manipulator arm used, and in particular its article gripping mechanism.

[0081] Step S50 results in the generation of a DET parameter, representative of the determination of the type of packaging of the article, box, polybag, or undetermined.

[0082] In step S60, in response to the comparison of step S50, the computer sends to the robotic manipulator arm MA a COM command representative of the result of the comparison and therefore of the operation of discrimination between polybag and box. The manipulator arm MA is configured to use a higher acceleration speed or maximum speed when it is considered that the article is packed in a box than when it is packed in a polybag.

[0083] For example, the manipulator arm may be configured to use 100% of its operational speed when the item is packed in a box and only 80% or less of its operational speed when the item is packed in a polybag.

[0084] When the packaging of the item is indeterminate, differentiated treatment is applied, for example, this may involve handling the item as if it were an item packaged in a polybag, for security, or using an auxiliary identification and / or sorting device configured to identify and / or isolate the item from other items to be sorted.

[0085] Second mode of implementation

[0086] The second implementation method is a method 600 for discriminating between an item packaged in a box and an item packaged in a polybag, illustrated by the. This method is based on the differences in shades and textures between a box and a polybag. Indeed, boxes are mainly made of cardboard with a particular color and texture due to their base material, which is corrugated cardboard. The applicant proposes to use these characteristics to identify items packaged in a box.

[0087] In this second embodiment, the digital camera CAM is a digital camera capable of producing color shots, for example in the form of RGB images, i.e. consisting of three channels, respectively red, green and blue. The at least one digital image of step S10 consists of at least one color digital image.

[0088] In a sub-step S10a' of step S10, the computer 40 uses this at least one digital color image to identify articles in the image by segmentation using a neural network.

[0089] In a sub-step S10b' of step S10, in the same way as in sub-step S10c of the first embodiment, one of the articles is chosen by the CALC calculator. One face of the chosen article corresponds to the surface SURF of the. For explanatory purposes, a substantially rectangular face will be considered, with four edges facing each other two by two, which represents the most frequent case, and has the advantage of applying to both polybags and boxes, the latter being generally parallelepiped.

[0090] Of course, other surfaces can be used. It is thus possible to consider the smallest oriented rectangle (its sides are not necessarily parallel to the axes of the NUM image) which encompasses the point cloud corresponding to the chosen article. This method applies equally well to a box as to a polybag.

[0091] In step S20, the computer extracts from the image of the chosen face a plurality of statistical parameters PARAM each representative of points located on the surface of the article of interest and their respective environments. More specifically, the statistical parameters used in this second mode of implementation are representative of the hues and textures in the respective local environments of the points considered. By local environments for each of the points is meant zones centered respectively on the points considered, sufficiently small so as not to overlap each other, and preferably of identical dimensions.

[0092] In a sub-step S20a' of step S20, a plurality of points P distributed over the image of the chosen face are determined. As illustrated by the, these may be, for example, 9 points comprising the barycenter of the surface SURF, the four midpoints of the four segments connecting the barycenter to the midpoints of the four edges of the surface SURF, and the four midpoints of the four segments connecting the barycenter to the four corners of the surface. It is practical to assimilate each of these points to one of the pixels forming the image of the surface SURF.

[0093] In a sub-step S20b' of step S20, for each of the 9 determined points P, a plurality of zones Z of increasing surfaces, each centered on the point with which they are respectively associated, are defined. Such zones have the advantage of highlighting, in addition to the shade, the texture of the material constituting the surface SURF, in the statistical parameters calculated in step S20c'. For example, for each point, it is possible to consider a first zone Z1 defined by a square with a side length of 3 pixels, a second zone Z2 defined by a square with a side length of 5 pixels, a third zone Z3 defined by a square with a side length of 11 pixels, and a fourth zone Z4 defined by a square with a side length of 21 pixels, each of the squares being centered on the point considered.

[0094] In a sub-step S20c' of step S20, the calculator CALC calculates, for each of the determined points P, a set of statistical parameters PARAM based on hue information associated with pluralities of increasing surfaces each centered on a point of the chosen face. For example, a set of 10 statistical parameters PARAM may consist of the average of the gray levels (average of the values ​​of the three red, green and blue channels) of the pixels of the first zone, the respective averages of the red, green and blue levels of the pixels of the second zone, the respective averages of the red, green and blue levels of the pixels of the third zone, and the respective averages of the red, green and blue levels of the pixels of the fourth zone, the first, second, third and fourth zones being of increasing surfaces.For the reasons explained in the first implementation mode, the number of statistical parameters PARAM, intended for the construction of the vector VECT, is between 6 and 14, preferably between 8 and 12, more preferably between 9 and 11. It is 10 in the present example.

[0095] At step S30, the CALC calculator constructs for each of the 9 determined points a vector VECT whose components X m are respectively defined by one of the statistical parameters PARAM associated with the point considered, m being here a natural integer varying from 1 to 10.

[0096] In step S40, the calculator CALC calculates for each of the 9 points P considered an elementary probability PROBA_n, n representing an integer ranging from 1 to 9, associated with the fact that the article of interest is packed in a box. Each of the probabilities PROBA_n is calculated by logistic regression according to the same procedure as in step S40 of the first mode of implementation, with the difference that the weighting coefficients are obtained by means of a base of annotated digital images which are adapted to the second mode of implementation, of similar natures to those obtained in steps S10 of this second mode of implementation, with color images used for identifying the articles and characterizing the shades and textures associated with the points of one of their faces. The learning is designed to minimize the number of confusions between the types of article packaging, between box and polybag during the following step S50. The calculator CALC calculates the probability PROBA associated with the fact that the article of interest is packaged in a box by the arithmetic mean of the nine elementary probabilities PROBA_n.

[0097] Steps S50 and S60 of the second implementation mode are the same as those of the first implementation mode.

[0098] Third mode of implementation

[0099] The two implementation modes described above can be used independently. It is also possible to improve the robustness of the discrimination method by combining these two modes, according to the method 700 illustrated by the. The main interest is to limit false positives, that is to say to falsely consider that an article is packed in a box.

[0100] In this combination, methods 500 and 600 are implemented in parallel from their steps S10 to their respective steps S40. The digital color image NUM of the second embodiment can advantageously be obtained by means of the third lens of the digital camera employed in step S10 of the first embodiment.

[0101] Following steps S40, the two methods result in two respective PROBA probabilities, Pr1 and Pr2, distinct from each other. These two probabilities are then aggregated during an AGG aggregation step, and then steps S50 and S60 are applied to the result of this aggregation, which are the same for both implementation modes.

[0102] The manipulator arm is thus controlled in a more reliable manner, with a reduced level of confusion between the box and polybag packaging types compared to the first implementation mode and the second implementation mode.

[0103] The choice of the aggregation operator is dictated by operational choices: if the practitioner wishes, for example, to minimize discrimination errors, then he can choose to apply a minimization operator MIN to the probabilities Pr1 and Pr2, which results in the smallest probability between Pr1 and Pr2, or he can choose to calculate the product of the probabilities Pr1 and Pr2. Conversely, if the practitioner wishes to obtain a discrimination result on a maximum number of items, he can apply a more lax aggregation operator such as a maximization operator Max to the probabilities Pr1 and Pr2, which results in the largest probability between Pr1 and Pr2, or he can calculate the expression Pr1+Pr2 – Pr1*Pr2. The operators and calculations mentioned in this paragraph represent only a few examples among all the possibilities available to the practitioner, from which he will choose according to his objectives.

[0104] Of course, the invention is not limited to the embodiments described above and variant embodiments can be made without departing from the scope of the invention as defined by the claims.

Claims

Method for sorting articles (IT) including a differentiated treatment of the articles by means of a method (400; 500; 600; 700) for discriminating articles (IT) between articles according to a type of packaging chosen from rigid packaging and flexible packaging, comprising the steps of:- acquiring (S10) at least one digital image (NUM) of an article (IT) by means of a camera (CAM);- calculating (S20) statistical parameters (PARAM) representative of topography characteristics of a plurality of points located on a surface (SURF) of the article (IT) in the digital image (NUM) and / or representative of hues and textures in local environments of these points;- constructing (S30) a vector (VECT) having components (X m) defined by values ​​of the statistical parameters (PARAM);- based on the vector (VECT), calculating (S40) a probability (PROBA) for the article to be a rigidly packaged article or to be a flexiblely packaged article; and- comparing (S50) the probability (PROBA) to at least one threshold value and, in response to the comparison, generating a parameter (DET) representative of the type of packaging of the article, rigid packaging or flexible packaging, wherein the steps of calculating statistical parameters, constructing a vector, calculating a probability and comparing the probability to at least one threshold value are implemented by means of a digital computer (CALC). Method (400; 500; 600; 700) according to claim 1, the step (S10) of acquiring at least one digital image (NUM) of an article (IT) further comprising the steps of:- acquiring a digital image of a scene comprising the article by means of the camera; and- segmenting (S10b; S10a') the digital image of the scene so as to isolate the article and obtain the digital image of the article. Method (400; 500; 600; 700) according to any one of claims 1 to 2, wherein, in step (S40) of calculating a probability (PROBA), the components of the vector are weighted by weighting coefficients ( ) determined in advance during a machine learning step on the basis of images of articles, annotated so as to indicate a type of packaging of the articles of these images, rigid or flexible, the machine learning step being designed so as to minimize a number of confusions between the types of article packaging during the comparison step (S50). Method (500) according to any one of claims 1 to 3, in which a depth map is established (S10a), and in which the statistical parameters (PARAM) are representative of a topography of the surface (SURF), and in which the statistical parameters (PARAM) are representative of deviations ( ) of points constituting the surface (SURF) to a plane approximating this surface (SURF). The method (500) of claim 4, wherein components (X m ) of the vector are chosen from a mean, a median, a standard deviation of the deviations ( ), a first, a second, a third and a fourth moment of a normalized histogram ( ) deviations ( ), a value ( ) of a cumulative histogram ( ) deviations ( ), and integrals of the cumulative histogram. Method (600) according to any one of claims 1 to 3, in which the statistical parameters (PARAM) are representative of a shade or a texture of the surface (SURF) of the article (IT) in the digital image (NUM), the digital image being in color, and in which the step of calculating (S20) the statistical parameters (PARAM) comprises a step (S20a') of determining a plurality of points (P) distributed over the surface (SURF) and a step (S20b') of defining, for each of the points (P), zones (Z, Z1, Z2, Z3, Z4) of increasing surfaces, each centered on the point considered, the statistical parameters (PARAM) being calculated on the basis of these surfaces. The method (600) of claim 6, wherein components (X m) of the vector (VECT) are chosen from an average of gray levels of pixels of one or more of the zones (Z1), and from respective averages of the color levels of pixels of one or more of the zones (Z2, Z3, Z4). Method (700) combining the method (500) according to any one of claims 4 to 5 and the method (600) according to any one of claims 6 to 7, wherein the step (S50) of comparing the probability (PROBA) with at least one threshold value applies to an aggregation of the probability obtained by the method (500) according to any one of claims 4 to 5 and the probability obtained by the method (600) according to any one of claims 6 to 7. Method (400; 500; 600; 700) according to any one of the preceding claims 1 to 8, and further comprising the step of transmitting a control signal (COM) to an article processing device in response to the generation of the parameter (DET) representative of the type of packaging of the article. Article sorting installation (FAC) configured to implement the method according to claim 9. Installation (FAC) for sorting articles according to claim 10, the article processing device being a robotic manipulator arm (MA). Installation (FAC) for sorting articles according to claim 11, the robotic manipulator arm (MA) being equipped with a suction gripping device (GH).

Images