Method, computer program and data carrier for determining position of object relative to detection device

Through RGB-D cameras and machine learning technology, the position position of small carriers is identified and calculated, which solves the problem of position position recognition of small and medium-sized carriers in vehicle production, and achieves high-precision and robust position positioning, which is suitable for complex logistics environments.

CN118901088APending Publication Date: 2024-11-05BAYERISCHE MOTOREN WERKE AG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380028823.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-03-23
Filing Date
2023-03-17
Publication Date
2024-11-05

AI Technical Summary

Technical Problem

In vehicle production and logistics, it is difficult to accurately identify and locate the position of small carriers with flat surfaces in automated environments, especially in poor light conditions and complex scenarios, traditional two-dimensional computer vision and three-dimensional matching technologies are difficult to practical.

Method used

The RGB-D camera collects the depth information and image information of the object, and the marking points in the three-dimensional model are identified and called through the object recognition device to perform credibility checks and calculations. Combined with machine learning and artificial intelligence technology, the two-dimensional sum of the object is calculated. Six-dimensional pose, using depth information and neural network to identify the flat surface and rotational pose of an object.

Benefits of technology

It realizes high-precision pose recognition and positioning of small carriers in complex environments, improves robustness and applicability, can adapt to different types and complex combinations of small carriers, and reduces equipment costs and maintenance needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118901088A_ABST
    Figure CN118901088A_ABST
Patent Text Reader

Abstract

The invention relates to a method for determining the position of an object having at least one flat surface relative to a detection device, comprising the following steps: (S1) capturing at least one record of the object by means of a camera, the record comprising depth information and image information; (S2) identifying the object in the image information; (S3) invoking a model of the object, the model comprising at least one marking point (KP), the own pose of which is predefined with respect to the at least one flat surface; (S4) checking the plausibility of the at least one marking point (KP) in the record on the object; (S5) calculating a two-dimensional pose of the at least one flat surface by means of the at least one identified marking point (KP) in the recording; (S6) selecting a point set from the recorded depth information; and (S7) calculating the pose of the object by matching the set of points to the calculated pose. The invention also relates to a computer program and a data carrier.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for determining the pose of an object having at least one planar surface relative to a detection device according to claim 1. Furthermore, the invention relates to a computer program according to claim 10. Finally, the invention relates to an electronically readable data carrier according to claim 11. Background Art

[0002] In logistics, for example in vehicle production, transport carriers can be used to transport goods or parts to be manufactured. The transport carriers can be designed as small transport carriers, such as transport boxes. These transport boxes are transported to and / or remain at the production site, for example by a conveyor device, where they are used, for example, for vehicle production.

[0003] In this context, automation of production (e.g. with the aid of robots) is increasingly being developed, so that entire production lines or logistics systems are being operated automatically. It is advantageous to know the position (Lage), i.e. the location and orientation, of the objects used, such as carriers, in space, e.g. relative to the robot, as accurately as possible.

[0004] US20210004984A1 shows a method for training a 6D position estimation network based on iterative deep learning. In addition, US11182924B1 shows a system for estimating a three-dimensional posture and for determining one or more biochemical performance parameters of at least one person in a scene. Summary of the invention

[0005] The object of the present invention is to provide a method, a computer program and a data carrier by means of which objects having at least one flat surface, such as small transport vehicles, can be particularly advantageously identified and their pose and thus their position and orientation or rotation relative to a logistics device can thus be determined in a particularly precise manner.

[0006] According to the invention, this object is achieved by the subject matter of the independent claims. Advantageous embodiments and further developments of the invention are given in the dependent claims as well as in the description and the drawings.

[0007] A first aspect of the invention relates to a method for determining the position of an object having at least one flat surface (e.g. a small carrier having a flat bottom surface) relative to a logistics device, which may include, for example, a transport device and / or a robot. In addition to or as an alternative to the bottom surface, the side surface of the small carrier may also be used in the method.

[0008] In order to be able to determine the pose of the object particularly advantageously, the method according to the invention comprises a plurality of steps:

[0009] In a first step, at least one recording of the object is acquired by means of a camera, the recording comprising depth information and image information. In a second step, the object is recognized in the image information by an object recognition device. In a third step, a (particularly three-dimensional) model of the object is called up based on the recognition in the second step, the model comprising at least one marking point, the intrinsic pose of which relative to the at least one planar surface is predetermined and therefore in particular fixed and / or known. In this case, the at least one marking point stored in the model is identified on the object in the recording and a marking point recognition is thus performed. In a fourth step, a plausibility check is performed for the previously identified marking point. In a fifth step, the two-dimensional pose or orientation of the at least one planar surface of the object is calculated by means of the at least one identified marking point in the recording, wherein the calculation is advantageously performed using at least two and in particular at least three marking points. In a sixth step, a point set is selected from the depth information of the recording, the point set being associated with the surface of the object or oriented in the at least one surface. Finally, in a seventh step, a six-dimensional pose of the object is calculated by adapting the point set or the surface defined by the point set to the calculated orientation, which pose describes the position and rotation or orientation of the object relative to the logistics device.

[0010] In the first step, i.e., acquisition, the recording itself can be generated by a camera, and additionally or alternatively, the data of the recording can be received from an electronic computing device that implements at least part of the method. The detection device includes a camera and can additionally have, for example, a holding device, on which the camera is fixed, for example, relative to the logistics robot. In this case, the holding device can be designed to be stationary or movable, for example.

[0011] The camera is in particular designed as an RGB-D camera and / or a stereo camera, so that the recording generated by the camera includes image information, i.e., for example, classic color information, as in conventional photography. In addition, the recording also includes depth information, which is generated, for example, by means of infrared light reflection and time-of-flight, and the depth information in the recording can be included in the recording, for example, as a point cloud of the environment. In order to be able to carry out the method, the object should be oriented relative to the camera in such a way that at least a part of the at least one planar surface appears in the recording.

[0012] The camera can be, for example, part of a logistics device or at least fixedly mounted relative to a corresponding part of the logistics device that receives objects, such as a conveyor belt. The orientation of the camera is therefore known. The object recognition in the second step of the method according to the invention is carried out in particular by an image recognition algorithm or an object recognition algorithm in the image information and is used to be able to prepare the correct model.

[0013] The model of the object can be stored as a two-dimensional model or a three-dimensional model, wherein the at least one marker describes at least one characteristic feature of the object. The model can be trained for a small carrier vehicle based on a neural network. Advantageously, the model has at least three markers and in particular at least four markers, since the two-dimensional pose can be uniquely determined. Thus, these markers can, for example, describe the corner points of the at least one flat surface.

[0014] For this purpose, a further algorithm is used, which can also be implemented by the object recognition device. The algorithm of the object recognition device is implemented in particular by a neural network.

[0015] In a fourth step, the at least one marking point on the object in the record is subjected to a plausibility check. In other words, a plausibility check is performed, wherein the identified marking points are checked based on logistics-specific post-processing rules and are combined into a marking point combination.

[0016] In the fifth step, calculation, the surface identified and recognized by means of the marking points is projected into the recorded image plane. This allows, for example, deformations due to an inclined position relative to the image plane to be detected, thereby making it possible, for example, to particularly advantageously determine, in particular project, the two-dimensional pose or orientation of the surface.

[0017] The recorded depth information is used in the selection of a point set, wherein corresponding points are used in the selection, the corresponding points being located in the flat surface of the object or being able to be associated with the flat surface at least by means of the recording. A surface model of a surface can be determined by means of the selected points, the surface model describing the at least one flat surface of the object. Finally, in a last step, the six-dimensional pose of the object and thus its position and orientation or rotation in space are calculated, wherein the selected point set is adapted to the calculated orientation or pose. Advantageously, the surface model determined by means of the point set is used here so that the surface model is adapted to the two-dimensional pose in the seventh step of the method.

[0018] In other words, in the method, an RGB image (ie a color image) and an associated depth image are generated as records. In addition, a learning-based approach is used for object recognition, which can also be used for two-dimensional marker point identification (English: "key point extraction"). Markers, for example, markers that are specific to or characterize a small carrier vehicle, can then be identified by domain transfer or domain adaptation.

[0019] For example, the two-dimensional pose is calculated with the coordinates u, v and Φ with the aid of image information, i.e., in particular in an RGB image, for example with the aid of projection geometry. The image information is in particular previously identified marking points and / or previously identified combinations of marking points. Additionally or alternatively, a two-dimensional grid is calculated in the area of ​​the two-dimensional pose of the surface with the aid of the image information in the record. The selection of a set of points is performed, for example, on a random basis within the two-dimensional grid (in particular calculated in step 5), the set of points corresponding to 3D depth points in particular based on depth information. Another feasible and robust solution for determining the 3D plane is, for example, a plane equation, which can be calculated, for example, with the aid of the so-called RANSAC algorithm. RANSAC is English and stands for "random sample consensus", meaning agreement with a random sample.

[0020] In order to calculate the pose of the object in the last step, a projection of a two-dimensional translation along the normal and a plane, in particular the image plane, can now be performed, wherein a projection of a two-dimensional rotation about the normal can also be performed. Thus, by combining translation and rotation, a complete six-dimensional pose or attitude is now obtained. "Six-dimensional" means that the three-dimensional space coordinates contain the position, i.e. the distance and direction of the object, for example relative to the origin of the coordinate system and in particular relative to the detection device, and the other three coordinates contain, for example, rotations relative to the three coordinate axes of a Cartesian coordinate system, so that the complete pose of the object, i.e. the position and orientation, can be described.

[0021] The invention is based on the recognition that conventional approaches, such as two-dimensional computer vision and three-dimensional matching, are difficult to implement under factory conditions, such as in vehicle production. For example, in a workshop, the light conditions for the camera may be poor, which may lead to sensor noise. In addition, a basic assumption is that clear and known objects should be "found" in the camera recordings and that essentially the same input (in terms of objects and recordings) is present in each run, but this ideal situation does not correspond to reality.

[0022] Advantageously, the camera can be designed as simply as possible and thus cost-effectively, and the method is particularly robust, which is achieved by the combination of conventional computing and machine learning methods and thus by means of artificial intelligence that can be implemented in the method according to the invention.

[0023] Therefore, the advantage of the method according to the invention is that the method can be optimized for example for small carriers as objects and can therefore be used advantageously in logistics. This particularly causes particularly high accuracy and / or robustness. Here, for example, a variety of different objects are used in vehicle production, and the objects can be accurately identified, so the method can be particularly advantageously extended to small carriers of different numbers and combinations and can therefore also be applied in the case of high product complexity and / or container complexity or object complexity. Here, in the method according to the invention, the problem can be advantageously decomposed into a learning-based scheme and a mathematical calculation of the posture. Therefore, another advantage of the method according to the invention is that the method is particularly advantageously extended and can therefore achieve an advantageous and / or low-maintenance implementation. Therefore, for example, a low-cost camera can be used. In addition, for example, the condition of the object (especially when it comes to small carriers) can be taken into account, or this only plays a secondary role. For example, the object may have a damaged part, color, sticker and / or the like and can still be identified despite this.

[0024] In an advantageous embodiment of the invention, before the orientation and / or the pose are calculated, as a step of the method, the at least one identified marker point is subjected to a plausibility check, in particular by a post-processing unit. In other words, in particular logistics-specific post-processing rules are defined. In particular, four marker points are used for determining or checking the plausibility. In order to ensure robustness, any number of marker point combinations (4 marker points) can be learned.

[0025] However, for industrial practice, a trade-off should be made between the necessary number of marker combinations and the cost for learning. Therefore, in particular, two marker combinations (ie, 8 markers) can be used.

[0026] If the marking points depict, for example, the corners of a rectangular, non-square surface of an object, then two marking points lying opposite each other on the short side have a shorter distance relative to each other than two marking points lying opposite each other on the long side of the rectangle. If it is now determined in the credibility check that the distance between two identified marking points (which should have a shorter distance) is longer than the originally longer distance, it follows that the consistency defined by the logistics-specific post-processing rules cannot be maintained. It is possible that two marking points (English: key points) are determined visually superimposed on each other, which two marking points may not actually have such a small distance from each other. Therefore, by means of the further method step, i.e. the credibility check, it is ensured in a particularly advantageous manner that errors in determining the pose of the object can be avoided. Such errors can lead to collisions with the robot.

[0027] In another advantageous embodiment of the present invention, the geometry and / or confidence and / or consistency of the at least one identified marking point are determined in the credibility check. In other words, the geometry, distance, confidence and / or consistency are determined as logistics-specific post-processing rules, so that it is possible to simply check whether the identified marking points are meaningful. The confidence can be determined individually for each marking point (especially with the help of artificial intelligence) first. In the credibility check, each marking point is then compared with a minimum value. If, for example, for four assigned marking points (marking point combinations), at least one marking point is below a threshold value or the minimum value, the marking point combination is discarded. In industrial practice, neural networks can be well trained, so that it is rare to discard both marking point combinations. As a result, a credibility check can be performed particularly advantageously.

[0028] In another advantageous embodiment of the invention, a machine learning method, in particular deep learning (English: Deep Learning) and / or at least one convolutional neural network (English: Convolutional Neural Network) is used to recognize the object and / or to distinguish the at least one marking point. In other words, the recording is machine-processed mainly by artificial intelligence, that is, it is carried out in an automated manner so that, for example, an algorithm that can independently detect and process problems is used. A convolutional neural network includes one or more convolutional neural network layers followed by a pooling layer. These layers form a unit that can be repeated any number of times in principle. With sufficient repetitions, the domain of deep learning is reached, which is characterized by the fact that different neural network intermediate layers are arranged between the input and output layers. In this case, the object is advantageously recognized by a first convolutional neural network and the at least one marking point is distinguished, for example, by a second convolutional neural network different from the first neural network, which can in particular be a convolutional neural network (CNN) with key feature extraction. The recognition or the discrimination is carried out in a particularly advantageous and / or efficient manner by using machine learning methods. Furthermore, for example, previously generated, labeled records are suitable as training data, wherein in particular corresponding models of the corresponding objects can also be stored as input for the training. In this case, supervised learning is used in particular.

[0029] In another advantageous embodiment of the invention, a point set is selected from the depth information randomly and / or by means of a plane equation. In other words, the depth information is randomly selected, which is associated with the at least one flat surface and which, for example, consists of a point cloud including points located in the surface. Additionally or alternatively, the selection is performed by means of a plane equation. This results in the following advantages, for example: systematic errors, for example when calculating the pose, can be reduced.

[0030] In another advantageous embodiment of the invention, at least one of the method steps is carried out or at least realized by domain adaptation (English: Domain Transfer) from a (particularly human) body to an object with three-dimensional or six-dimensional pose determination. Additionally or alternatively, the recognition and / or identification of the object and / or the calculation of the orientation or pose are carried out by means of a pattern comparison (English: Template Matching) or template matching.

[0031] In the domain adaptation, in particular, a human pose estimation algorithm can be used, which generally processes image coordinates (2D) and thus processes the recorded image information. Therefore, in the domain adaptation, a known, trained machine learning mechanism (which is used, for example, to recognize the posture of a person in the recording) can be matched to the object. In pattern comparison, i.e., template matching, a small area of ​​the recording is matched (abgleichen), for example, in the image information or in a template image that may be contained in a model of the object. Generally, in traditional computer vision, a learning-based algorithm is not used in (2D or 3D) template matching, but a rigid "consistency" is used: template matching is a technique for finding a small part of an image that is consistent with a template image in digital image processing. For the method proposed here, a learning-based solution with a neural network / artificial intelligence is preferred. This enables migration, that is, the template does not have to be exactly 100% the same. This allows existing changes to be reliably reflected.

[0032] By using domain adaptation and / or pattern comparison, the method can be carried out in a particularly advantageous manner, wherein, for example, computing resources can be advantageously saved.

[0033] In another advantageous embodiment of the invention, the two-dimensional pose of the surface is calculated by projecting geometry. Additionally or alternatively, a two-dimensional grid is inserted in the area of ​​the pose. In other words, a perspective view of a three-dimensional object is displayed in a two-dimensional plane in the recorded image information. Therefore, when calculating the pose or posture, the object is located in the projection plane or projection space, whereby the pose can be calculated particularly efficiently and / or effectively and therefore with high precision. By inserting a two-dimensional grid, the calculation in the seventh step of the method can be carried out in a particularly advantageous manner, for example, because, for example, the point set of the surface can be particularly advantageously matched to the calculated orientation or posture. The advantage resulting from this is that the method can particularly advantageously determine the pose of the object.

[0034] In another advantageous embodiment of the invention, the orientation of the camera relative to the logistics device is predetermined or known when the recording is acquired. The logistics device can, for example, include a robot with a mechanical arm, a linear axis and an in particular mobile transport robot, on which the camera or the detection device is mounted. The camera can, for example, be fixedly positioned on a ceiling above the logistics device. Additionally or alternatively, the method is used in vehicle production. This results in the advantage that, on the one hand, the position of the object can be determined particularly accurately, and, on the other hand, the logistics in vehicle production can be improved.

[0035] A second aspect of the invention comprises a computer program which can be loaded into an electronic computing device of an installation and comprises program modules so that when the computer program is executed in the electronic computing device or, for example, a control device, the steps of the method according to the first aspect of the invention are implemented.

[0036] Advantages and advantageous embodiments of the first aspect of the present invention may be regarded as advantages and advantageous embodiments of the second aspect of the present invention, and vice versa.

[0037] A third aspect of the invention relates to an electronically readable data carrier. The electronically readable data carrier comprises electronically readable control information stored thereon, the electronically readable control information comprising at least one computer program according to the second aspect of the invention and being designed such that the electronically readable control information can implement the method proposed herein according to the first aspect of the invention when the data carrier is used in an electronic computing device.

[0038] Here, the advantages and advantageous embodiments of the third aspect of the present invention can be regarded not only as the advantages and advantageous embodiments of the second aspect of the present invention, but also as the advantages and advantageous embodiments of the first aspect of the present invention, and vice versa.

[0039] Further features of the invention are apparent from the claims, the drawings and the description of the drawings. The features and feature combinations mentioned above in the description and the features and feature combinations mentioned below in the description of the drawings and / or shown individually in the drawings can be used not only in the respectively given combination but also in other combinations or individually. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will now be described in detail with reference to preferred embodiments and the accompanying drawings.

[0041] Figure 1 A schematic flow chart showing a method for determining a position of an object relative to a logistics device; and

[0042] Figure 2 The recording by means of a camera shows Figure 1 The order of the method steps of the method. DETAILED DESCRIPTION

[0043] Generally, in the production of motor vehicles, small carriers KLT are used in corresponding logistics areas or in logistics devices, such as sorting robots, in which components are transported, for example. These small carriers can be regarded as objects, and it is important to know the position and orientation of these small carriers relative to the logistics device and therefore the posture. As a result, for example, the automation of robots, for example, can be achieved in a particularly advantageous manner. Here, a method should be proposed below, which should solve the following task: robust object recognition and positioning can be achieved. It should be determined here what type of small carrier KLT the object is, because different types of small carriers KLT can be used in vehicle manufacturing or logistics devices. It should be positioned in a particularly advantageous manner and therefore with particularly high precision, that is, the six-dimensional posture or posture of the object and therefore the translation and rotation of the object should be fully determined.

[0044] The method proposed here is Figure 1 The schematic flow chart of the invention is shown in FIG. 1 and includes the following steps, which enable the position and posture of an object having at least one flat surface relative to a logistics device to be determined:

[0045] In a first step S1 of the method, at least one recording of an object comprising depth information and image information is acquired by means of a camera, which can be part of a logistics device, for example.

[0046] In a second step S2 , objects in the image information are recognized by an object recognition device, which may in particular execute an object recognition algorithm (eg based on machine learning).

[0047] In a third step S3, a (particularly three-dimensional) model of the object is called up based on object recognition so that the correct model can be assigned to the recognized object, the model comprising at least one marking point KP whose own pose (Eigenlage), i.e., the pose, relative to the at least one flat surface is fixed or known and is therefore predetermined.

[0048] In this case, at least one of the marking points KP on the object is identified in the recording, wherein an identification device can be used. The identification device can be designed similarly to the object recognition device and execute an algorithm (in particular based on machine learning) which can constitute an extension of the object recognition device.

[0049] In a fourth step S4 of the method, the identified marking points are subjected to a plausibility check based on logistics-specific post-processing rules. In addition, the marking points KP are combined into marking point combinations.

[0050] In a fifth step S5 , the two-dimensional orientation or pose of the surface in the recording and thus in the image plane is calculated with the aid of the at least one identified marking point KP, wherein advantageously at least three and in particular four marking points KP are identified.

[0051] In a sixth step S6 of the method, a point set is selected from the recorded depth information, which point set is assigned to at least one plane of the surface of the object, so that a surface model describing the surface can be derived from the point set.

[0052] Finally, in a seventh step S7 of the method, a (particularly six-dimensional) pose of the object is calculated by adapting the point set or a surface model derived therefrom to the calculated pose or orientation, which pose describes the position and orientation or rotation of the object relative to the logistics device.

[0053] The advantages of the proposed method are, for example, that a particularly robust recognition can be achieved through the combination of a neural network, a calculated scheme and machine learning. Thus, for example, problems that arise when using small carriers as objects can be dealt with. For example, there may be a particularly high complexity because a large number of different small carriers KLT may be used in a logistics device. In addition, as a sensor for acquisition, i.e., a camera, only a simple RGB-D camera can be used, which enables low-maintenance operation, which can also be expanded well. In addition, for example, favorable light conditions are not present everywhere on a logistics device, so sensor noise may occur, which can also be compensated particularly advantageously by the proposed method. In addition, the condition (Beschaffenheit) of the small carriers, for example, whether they have different wear phenomena, such as sticking labels, scratches, dirt, oil stains, broken edges, other damage, wrong colors, etc., plays a secondary role, because this can also be eliminated particularly advantageously by combining depth information with neural network posture determination. The advantages of the proposed method over traditional computer vision with template matching are embodied here, in which difficulties are expected with different wear phenomena.

[0054] It is advantageous here that, before calculating the orientation and / or the pose, as a step of the method, a plausibility check is performed on the at least one identified marker point KP (e.g. by a post-processing device). In particular, a plurality of marker points and marker point combinations are taken into account in the plausibility check. It is advantageous here that the geometry and / or the convergence and / or the consistency of the at least one identified marker point KP is determined for the plausibility check.

[0055] Figure 2 The recording with the aid of a camera shows how the individual steps of the method are carried out. Figure 2FIG. 1 shows the recorded image information in which two different small carriers KLT are shown. In this case, the respective small carrier KLT has at least one flat surface which is Figure 2 The bottom of the corresponding small carrier KLT is formed in the middle.

[0056] exist Figure 2 As can be seen in part b) of the figure, the object recognition device used in step S2 of the method has already set a bounding box (Bounding Box, BB) around at least one of the small carriers KLT. Here, for example, the larger small carrier KLT on the left should be further considered. Each method step can be performed for each identified small carrier.

[0057] After object recognition with the aid of the bounding box BB, step S3 is executed and the model including the marker points KP is called. These marker points are now identified by machine learning methods and as in Figure 2 These points are marked accordingly as can be seen in part c) of the figure.

[0058] A plausibility check can now be performed, in which, for example, it is detected that the identification of the at least one marking point KP, in particular the plurality of marking points KP (English: Key Point), is successful, whereby these marking points are further introduced into the Figure 2 Now, if Figure 2 As shown in part e) of the figure, the two-dimensional pose of the surface is calculated, which is represented by the axes shown in the coordinate system KOS.

[0059] Now, point sampling can be performed, wherein a set of points is selected that represents the surface of the small launch vehicle. The selection of the points can be based on randomness and lie within a grid that is already given in the two-dimensional pose, for example. This is done in Figure 2 In part f) of the figure, this is represented by the marked surface area FB.

[0060] Now this is recorded in Figure 2 Finally, as shown in FIG. h) and illustrated with the aid of the axes of the third spatial dimension of the coordinate system KOS, the six-dimensional pose of the object, i.e. the small carrier, is calculated by matching the point set to the calculated pose, wherein the pose describes the position and / or orientation of the object relative to the logistics device.

[0061] In steps S2 and S3 , machine learning methods are advantageously used, in which in particular deep learning is used, or at least one convolutional neural network is used for recognition and in particular a further convolutional neural network is used for discrimination, wherein the two networks can also be combined into one network.

[0062] Furthermore, in at least one of the method steps, a domain adaptation and thus a domain transfer is performed, for example, a neural network which was originally trained, for example, for recognizing the pose of a person in an image is adjusted. Furthermore, object recognition and / or identification of the at least one marker point and / or calculation of the orientation of the two-dimensional pose can be performed by means of a pattern comparison (i.e., template matching). When calculating the two-dimensional pose, projection geometry can be used and a two-dimensional grid can be advantageously inserted in the region of the pose.

[0063] The proposed method thus represents an advantageous, robust, hybrid 2.5D approach for 6D position estimation of small vehicle KLT in logistics with the aid of deep learning.

[0064] Reference numerals list

[0065] S1 First step

[0066] S2 Second step

[0067] S3 Step 3

[0068] S4 Step 4

[0069] S5 Step 5

[0070] S6 Step 6

[0071] S7 Step 7

[0072] KLT small launch vehicle

[0073] BB Bounding Box

[0074] FB area

[0075] KP Marker

[0076] KOS coordinate system

Claims

1. A method for determining a position of an object having at least one flat surface relative to a detection device, the method comprising the following steps: - (S1) acquiring at least one record of the object by means of a camera of a detection device, the record comprising depth information and image information; - (S2) identifying the object in the image information by an object recognition device; - (S3) based on the recognition, calling a model of the object, the model comprising at least one marking point (KP), the position of the marking point relative to the at least one planar surface being predetermined; - (S4) performing a credibility check on the at least one marking point (KP) on the object in the record; - (S5) calculating a two-dimensional pose of the at least one planar surface by means of at least one identified marker point (KP) in the recording; - (S6) selecting a point set from the recorded depth information, the point set being matched to the at least one flat surface of the object; and - (S7) calculating a six-dimensional pose of the object by matching the point set to the calculated two-dimensional pose, the pose describing the position and orientation of the object relative to the detection device.

2. The method according to claim 1, It is characterized in that Before calculating the orientation and / or the pose, as a further step of the method, the at least one identified marking point (KP) is subjected to a plausibility check.

3. The method according to claim 2, It is characterized in that In the plausibility check, the geometry and / or the plausibility and / or the consistency of the at least one identified marking point are determined.

4. The method according to any one of the preceding claims, It is characterized in that A machine learning method, in particular deep learning and / or at least one convolutional neural network, is used for recognizing the object and / or for identifying the at least one marking point (KP).

5. The method according to any one of the preceding claims, It is characterized in that The point set is selected randomly and / or with the aid of a plane equation.

6. The method according to any one of the preceding claims, It is characterized in that At least one of the method steps is carried out by domain adaptation of the 3D pose determination from a volume to an object and / or the recognition and / or the identification and / or the calculation of the orientation of the object is carried out based on a pattern comparison.

7. The method according to any one of the preceding claims, It is characterized in that The 2D pose is calculated by projecting the geometry and / or inserting a 2D grid in the region of the pose.

8. The method according to any one of the preceding claims, It is characterized in that When recording the recording, an orientation of the camera relative to the logistics device is predetermined and / or the method is used in vehicle construction.

9. A computer program, which can be directly loaded into a memory of an electronic computing device, and which has program modules so as to implement the steps of the method according to any one of claims 1 to 8 when the program is executed in the electronic computing device.

10. An electronically readable data carrier having electronically readable control information stored thereon, the electronically readable control information comprising at least one computer program according to claim 9 and being designed such that the electronically readable control information implements the method according to any one of claims 1 to 8 when the data carrier is used in an electronic computing device.