Methods for generating training datasets for training an evaluation algorithm, methods for training an evaluation algorithm, and methods for evaluating the alignment of two map datasets.

Abstract

The invention relates to a method for generating training datasets for training an evaluation algorithm, by means of which an alignment of two map datasets (302, 304) for determining navigation information for a mobile device (100) that is moving or is intended to move in an environment (120) can be evaluated, comprising for each of a plurality of training datasets: providing (310) two input feature datasets (312, 314), wherein the underlying map datasets each contain environmental information that has been acquired by means of a sensor of the mobile device; providing (340) a transformation dataset (332) that has been generated upon alignment of the two input feature datasets; providing (350) a reference transformation dataset (352) as ground truth; determining (360) a correlation dataset (362) based on the two input feature datasets and / or the transformation dataset;Determine (370) a goodness-of-fit measure (372) depending on the accuracy of a match between the transformation dataset and the reference transformation dataset; and provide (380) the training dataset (382), wherein the training dataset comprises the correlation dataset and the goodness-of-fit measure.

Description

[0001] The present invention relates to a method for generating training data sets for training an evaluation algorithm by means of which an alignment of two map data sets for determining navigation information for a mobile device that is moving or is intended to move in an environment can be evaluated, a method for training such an evaluation algorithm, a method for evaluating an alignment of two map data sets, a mobile device and a system and a computer program for carrying out the methods. Background of the invention

[0002] Mobile devices such as at least partially automated vehicles or robots typically move in an environment such as a home, a garden, a factory hall, on the street, in the air, or in water. One of the fundamental problems of such or other mobile devices is to navigate and, in particular, to orientate itself, i.e. to know what the environment looks like, in particular where obstacles or other objects are, and where it is (absolutely) located. For this purpose, the mobile device can be equipped with various sensors, such as cameras, lidar sensors, radar sensors, or even inertial sensors or GNSS sensors (generalization of GPS sensors for coarse positioning), with the help of which the environment and the movement of the mobile device can be recorded, for example, in two or three dimensions. Disclosure of the invention

[0003] According to the invention, a method for generating training data sets for training an evaluation algorithm, a method for training an evaluation algorithm, and a method for evaluating an alignment of two map data sets, a mobile device, and a system and a computer program for implementing the methods with the features of the independent patent claims are proposed. Advantageous embodiments are the subject of the dependent claims and the following description.

[0004] The invention generally relates to mobile devices that move, or at least can move, in an environment such as a road or a work area. Examples of such mobile devices (or mobile work devices) are, for example, robots and / or drones and / or partially or (fully) automated vehicles (on land, water or in the air). Robots that can be considered include household robots such as cleaning robots (e.g. in the form of vacuum and / or mop robots), floor or street cleaning devices, construction robots or lawnmower robots, as well as other so-called service robots, as well as at least partially automated vehicles, e.g. passenger transport vehicles or goods transport vehicles (including so-called industrial trucks, e.g. in warehouses, but generally also passenger cars and trucks), and also aircraft such as so-called drones or watercraft.

[0005] Such a mobile device comprises, in particular, a control or regulating unit and a drive unit for moving the mobile device, allowing the mobile device to be moved within its environment, e.g., along a movement path. Navigation information can be determined for this purpose, such as specific instructions regarding the direction in which the mobile device should travel in order to follow the movement path. These instructions can then be implemented via the control or regulating unit and the drive unit. In general, this can be referred to as the navigation of the mobile device.

[0006] In addition, a mobile device can have one or more sensors that can be used to record the environment or information in the environment, as well as possibly from the mobile device itself. As mentioned, these can be cameras, lidar sensors, radar sensors, ultrasonic sensors, or even inertial measurement units (or inertial sensors), as well as radodometry, which can be used to record the environment and the movement of the mobile device, for example, in two or three dimensions. Depending on the type of mobile device, other or additional sensors may also be provided.

[0007] One aspect of navigation is so-called mapping, as it is used, for example, in the creation of maps of the surrounding area. Such maps can also be used as sensors in a certain sense. In this context, scan matching or map alignment is also referred to. One goal here is to determine an approximate transformation between input data, which can originate from any sensor, such as lane markings detected by a vehicle camera or radar echoes. This transformation is usually a key first step in creating a consolidated map from multiple, partially overlapping sensor data (or a representation derived from them). Generally speaking, it involves aligning two map data sets.

[0008] One example of this is scan matching of so-called point clouds from lidar sensors or point clouds from other sensors. In particular, the entire point cloud measured or recorded by the laser scanner or sensor is used here; a point cloud is a set of points in the environment that are determined, for example, using the laser scanner or lidar sensor. Each point can be assigned a distance to the mobile device or laser scanner, as well as an orientation relative to a reference orientation of the mobile device or laser scanner. The point cloud usually corresponds to one or more "point lines" along the contour of the objects in the field of view; however, the points may also only lie approximately on such a line.

[0009] This point cloud alone does not, however, allow the position and / or orientation of the mobile device in the environment to be determined. To do this, the point cloud or set of points is compared with a reference point cloud or reference set of points. A transformation can then be determined which best covers or aligns the point cloud (or set of points) with the reference point cloud (or reference set of points), i.e. aligns the two. This transformation then corresponds to a position and / or orientation of the mobile device when the point cloud is acquired relative to the reference point cloud or a coordinate system of the reference point cloud. If the reference point cloud is a map of the environment or at least part of it, the current position and / or orientation of the mobile device in the environment can be determined. The reference point cloud (or map of the environment) can, for example, be constantly or repeatedly scanned.repeatedly adding new point clouds or parts of them.

[0010] SLAM is also used in this context. SLAM (simultaneous localization and mapping) is a robotics technique in which a mobile device such as a robot can or must simultaneously create a map of its surroundings and estimate its spatial position within this map. It is thus used, for example, to detect obstacles and thus supports autonomous navigation.

[0011] Based on such a SLAM graph, a map of the environment (environment map) in which the mobile device is moving can be determined. With each new map data set containing information about the environment and / or the mobile device, which information is obtained from or was based on one or more sensors of the mobile device, the map (or SLAM graph) can be expanded or updated.

[0012] This involves attempting to match the two map data sets, ie an existing and a new map data set, at least within certain tolerances, in order to determine the movement or trajectory of the mobile device.

[0013] In principle, the mapping process on the one hand and the localization process on the other can also be decoupled from each other (i.e., the map is generated, played on the mobile device, and then localized there). Scan matching, or the alignment of two map data sets, can be used in both cases, i.e., combined or decoupled mapping and localization processes.

[0014] In general, an alignment of two map data sets can be used to determine navigation information, such as a map or a trajectory, for a mobile device that is moving or is intended to move in an environment. This can be particularly relevant in the field of autonomous driving. For this purpose, two map data sets, each with environmental information, can be provided. In this case, the environmental information for both of the two map data sets has been recorded from the mobile device and / or the environment using a sensor on the mobile device. However, the same sensor does not have to have been used for both map data sets. The two map data sets are provided as input feature data, or input feature data is determined based on the two map data sets.

[0015] The two map datasets can then be matched. Generally, there are various approaches, both classical and machine learning (ML)-based, as described, for example, in "C. Choy, W. Dong, and V. Koltun, "Deep global registration," in Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, pp. 2514-2523, 2020." Many of these approaches use two steps: In the first step, so-called feature descriptors are generated, as described, for example, in "C. Choy, J. Park, and V. Koltun, "Fully convolutional geometric features," in Proceedings of the IEEE / CVF International Conference on Computer Vision, pp. 8958-8966, 2019." Based on these feature descriptors, the matching transformation is then determined, often as a combination of classical and ML methods.

[0016] The alignment or matching of the two map datasets is therefore performed using a machine learning algorithm. A so-called convolutional neural network (CNN) is particularly considered as a machine learning algorithm. Output data is generated from the input feature data via intermediate feature data (feature maps) in one or more intermediate layers (so-called hidden layers). The output data includes information about a transformative relationship (or simply a transformation) between the two map datasets.

[0017] Typically, such a transformative relation or transformation comprises a translation, represented, for example, by a translation vector, and a rotation, represented, for example, by a rotation matrix (both, depending on the situation, in 2D or 3D). This is especially true in the case of so-called FCGF-based scan matching methods, in which a feature vector encoding the geometric properties of the point is specified for each point in the point cloud. Thus, one feature data set is obtained for each map data set (point cloud).

[0018] The alignment can be achieved, for example, by forming point-to-point correspondences based on the feature data sets and applying a correspondence-based point cloud registration method (e.g., RANSAC registration). Alternatively, point-to-point correspondences can be formed based on the feature data sets and DGR can be applied.

[0019] Therefore, instead of source data, we will also refer to a transformation dataset. The source data or transformation dataset and / or the two underlying map datasets can then be provided for use in determining the navigation information.

[0020] Machine learning algorithms or CNNs that can be considered include those described in "C. Choy, W. Dong, and V. Koltun, "Deep global registration," in Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, pp. 2514-2523, 2020." or in "C. Choy, J. Park, and V. Koltun, "Fully convolutional geometric features," in Proceedings of the IEEE / CVF International Conference on Computer Vision, pp. 8958-8966, 2019."

[0021] However, it has now been shown that this alignment does not always work equally well. That is, such an alignment, which is ultimately a kind of estimate, can be good or bad. However, if the alignment is poor, navigation based on the transformation dataset or, possibly, the two map datasets would produce poor results.

[0022] Against this background, a method for evaluating such an alignment of two map datasets using an evaluation algorithm is proposed, as well as the training of such an evaluation algorithm and the generation of training data for it. Based on such an evaluation, it can then be decided, for example, whether the transformation dataset or the two underlying map datasets should be used to determine the navigation information.

[0023] This section will first explain how training data can be generated for this purpose, then how an evaluation algorithm can be trained or adapted based on such training data. Finally, an application of the evaluation algorithm will be explained.

[0024] The generation of training data or training data sets includes the following steps for each of a large number of training data sets: Two input feature data sets are provided, which correspond to two map data sets or which have been determined based on the two map data sets, as already mentioned above. The two map data sets each include environmental information, wherein the environmental information has each been acquired from the mobile device and / or the environment using a sensor of the mobile device, as also already mentioned.

[0025] Furthermore, a transformation data set is provided, which was generated by aligning the two input feature data sets, wherein the transformation data set includes information about a transformative relationship between the two input feature data sets, as also already mentioned. For example, the transformation data set and the two input feature data sets can be stored accordingly after performing an alignment.

[0026] Furthermore, a reference transformation dataset is provided as ground truth. Such a reference transformation dataset also contains information about a transformative relationship between the two input feature datasets. However, while the transformation dataset was obtained during the alignment process and is therefore ultimately an estimate, the reference transformation dataset specifies the actual transformative relationship between the two input feature datasets, or at least it is known that this transformative relationship corresponds sufficiently closely to reality. The reference transformation dataset can be acquired, for example, using a mobile device with additional high-precision sensors and then made available for later use.

[0027] The transformation data set and the reference transformation data set can each contain, for example, a rotation matrix and a translation vector.

[0028] The two input feature datasets, the transformation dataset and the reference transformation dataset, initially represent only raw data from which the training dataset is then created. This raw data can also contain optional, additional problem data or other information, such as the initial point clouds (map datasets) or method-specific information such as point-to-point correspondences and their weighting.

[0029] A correlation data set is then determined based on the two input feature data sets and / or the transformation data set. The correlation data set comprises information about a correlation between at least two of the following data sets: the two input feature data sets and the transformation data set.

[0030] The correlation dataset can, for example, be a data vector. In this case, the correlation dataset specifically indicates a relationship between the feature datasets and / or the transformation dataset. However, particular care must be taken to ensure that the correlation dataset is not determined based on the reference transformation dataset, since such a correlation dataset must also be determined later when applying the evaluation algorithm if there is no reference transformation dataset.

[0031] In general, the correlation data set can be determined or derived using a feature generation method selected by the user. In one embodiment, the correlation data set comprises information about a correlation between the two input feature data sets, wherein the correlation comprises distances between a point in one of the two input feature data sets and a point in the other of the two input feature data sets that is closest to this point. A point is understood here, in particular, to be a feature vector.

[0032] A concrete example is that for each point or feature vector in one of the two input feature data sets, the closest point or feature vector to this point in the other of the two input feature data sets is determined with respect to a suitable norm, e.g., the Euclidean norm. In this way, a set M of distances between the corresponding points or feature vectors is obtained. A simple example of feature generation and the resulting data vector (correlation data set) would then be, for example, (25% quantile of M50% quantile of M75% quantile of M).

[0033] Furthermore, the accuracy of the match between the transformation dataset and the reference transformation dataset is determined. Furthermore, a quality measure is determined depending on the accuracy of the match between the transformation dataset and the reference transformation dataset.

[0034] Here, too, a norm can be determined between the transformation data set and the reference transformation data set, in particular one norm between the translation vector of the transformation data set and the translation vector of the reference transformation data set, and one norm between the rotation matrix of the transformation data set and the rotation matrix of the reference transformation data set. Both values ​​of the norms can then be added together. Here, too, the Euclidean or another norm can be used. If the result, i.e. the sum of the two values ​​of the norms, is less than a specified threshold, the accuracy can be assumed to be sufficient, for example; if the result is greater than the specified threshold, it can be assumed to be insufficiently accurate.

[0035] The quality measure can comprise one of several quality classes. For example, only two quality classes can be provided, e.g., "sufficiently accurate" and "insufficiently accurate," corresponding to the threshold example above. However, the use of more than two quality classes is also conceivable, e.g., three quality classes: "low inaccuracy," "medium inaccuracy," and "high inaccuracy." For this purpose, two threshold values ​​can be specified, for example.

[0036] Likewise, the quality measure can comprise a continuous or quasi-continuous value, e.g., between zero and one. For this purpose, the aforementioned result, i.e., the sum of the two norm values, can be normalized to a value between zero and one. Likewise, the error of the transformation in a specific norm can be used as a quality measure. In this case, it is not a multiclass problem, but a regression problem.

[0037] The training dataset is then provided, where the training dataset includes the correlation dataset and the quality measure.

[0038] To train the evaluation algorithm, a plurality of training data sets are provided, which were generated, for example, as explained above. Each training data set thus comprises a correlation data set and a quality measure. The evaluation algorithm is then adapted or trained based on the plurality of training data sets. This is done in such a way that the evaluation algorithm determines an evaluation result for a target correlation data set and a target transformation data set, which assigns a quality measure to the target transformation data set. The adapted evaluation algorithm is then provided.

[0039] A classifier or classification algorithm, in particular, can be considered as an evaluation algorithm, in which one of several classes or quality classes is determined as a quality measure. For example, a suitable classification approach can be selected based on given training data. A generic, suitable classifier accepts input data in the selected form and delivers a real number, a probability value for one of the classes, or a binary value (0, 1) as output. The classifier is trained on a set of labeled training data, i.e., the training data sets mentioned above. Concrete examples of possible classification mechanisms are support vector classifiers, logistic regression, gradient boosting classifiers, as well as neural networks or general machine learning algorithms. Training can be carried out depending on the type of evaluation algorithm or classification algorithm and, for example,This can involve adjusting the weights of the individual neurons in a neural network. For example, a so-called loss function can be used, which is then optimized.

[0040] A trained evaluation algorithm can then be used to evaluate an alignment of two map data sets. Two input feature data sets are provided that correspond to two map data sets or that have been determined based on the two map data sets and as the map data sets have already been described, i.e., the map data sets each include environmental information that has been captured by a sensor of the mobile device from the mobile device and / or the environment.

[0041] Furthermore, a transformation data set is provided which has been generated by an alignment of the two input feature data sets, wherein the transformation data set comprises information about a transformative relation between the two input feature data sets.

[0042] It should be noted here that, for example, the input feature data sets are determined from the received map data sets, to which the alignment is then applied, resulting in the transformation data set. In general, however, the alignment itself could also have taken place elsewhere and at a different time.

[0043] Based on the two input feature data sets and / or the transformation data set, a correlation data set is then determined, which comprises information about a correlation between at least two of the following data sets: the two input feature data sets, the transformation data set.

[0044] Based on the correlation data set and the transformation data set, an evaluation result is then determined using the evaluation algorithm, which assigns a quality measure to the transformation data set.

[0045] The evaluation result is then provided for use in determining the navigation information, in which navigation information is determined based on the two map data sets and / or the transformation data set depending on the evaluation result.

[0046] Determining the navigation information based on the two map data sets and / or the transformation data set as a function of the evaluation result comprises, in one embodiment, using the two map data sets and / or the transformation data when determining the navigation information if the quality measure according to the evaluation result corresponds to a predetermined quality criterion, and not using the two map data sets and / or the transformation data when determining the navigation information if the quality measure according to the evaluation result does not correspond to the predetermined quality criterion.

[0047] However, it can also be provided that the two map data sets and / or the transformation data are used to determine the navigation information with a weighting, wherein the weighting is determined as a function of the quality measure according to the evaluation result.

[0048] In summary, once training datasets have been created and a classifier or evaluation algorithm has been trained on them, it can be used in the map creation process chain. For example, after FCGF-based scan matching has been performed, feature generation can first be applied to the output data (two input feature datasets and a transformation dataset), and then the trained classifier or evaluation algorithm can be applied. The output of the classifier or evaluation algorithm can then be used to reduce the influence of inaccurately matched point cloud pairs.

[0049] Two embodiments of the use of the classification result will be briefly explained below. During sorting, if the classifier outputs a high probability for a failed estimation or the label for a failed estimation, the corresponding point cloud pair can be excluded from further processing. The use of a weighting can be considered, for example, for all classifiers, not just a binary label (i.e. one of two classes). Here, an output can be transformed, for example, to probability values ​​in the interval [0; 1]. A weighting factor can then be calculated from these probability values. The higher the probability of a failed / inaccurate transformation estimate, the lower the weighting factor. In all further process steps (e.g. pose graph optimization), the influence of point cloud pairs can then be controlled according to the weighting factor..

[0050] Thus, by classifying the scan matching result, the influence of inaccurate transformation estimates in subsequent processing steps is reduced. This ultimately leads to higher map quality.

[0051] A system for data processing according to the invention or a computing unit, e.g. a control device or a control unit of a mobile device, or a server or other computer, is set up, in particular in terms of programming, to carry out a method according to the invention, e.g. in one of the described embodiments.

[0052] The invention also relates to a mobile device having such a data processing system or configured to receive navigation information determined as described above. The mobile device also has a sensor for detecting environmental information and is configured to navigate based on the navigation information.

[0053] The implementation of a method according to the invention in the form of a computer program or computer program product with program code for carrying out all method steps is also advantageous, since this entails particularly low costs, in particular if an executing control unit is also used for other tasks and is therefore already present. Finally, a machine-readable storage medium is provided with a computer program stored thereon, as described above. Suitable storage media or data carriers for providing the computer program are, in particular, magnetic, optical and electrical memories, such as hard disks, flash memories, EEPROMs, DVDs, and others. Downloading a program via computer networks (Internet, intranet, etc.) is also possible. Such a download can be wired or cable-based or wireless (e.g. via a WLAN network, a 3G, 4G, 5G or 6G connection, etc.).

[0054] Further advantages and embodiments of the invention will become apparent from the description and the accompanying drawings.

[0055] The invention is illustrated schematically in the drawing using an embodiment and is described below with reference to the drawing. Short description of the drawings Fig. 1 schematically shows a mobile device in an environment for explaining the invention. Fig. 2a, Fig. 2b schematically show adjustments of map data sets to explain the invention. Fig. 3 shows schematically a flow of a method in one embodiment. Fig. 4 shows schematically a sequence of a method in a further embodiment. Fig. 5 shows schematically a sequence of a method in a further embodiment. Embodiment(s) of the invention

[0056] In Fig. 1 schematically illustrates, by way of example, a mobile device 100 in an environment 120 to explain the invention. The environment 120 here includes, for example, a road 140 with lane markings 142. The mobile device 100 is, for example, a vehicle with a control or regulating unit 102 and a drive unit 104 (with wheels) for moving the vehicle 100, e.g., along a movement path 130, which here, for example, runs along the road 140 or a lane of the road.

[0057] Furthermore, the vehicle 100 has, for example, a sensor 106 embodied as a camera with a detection range. For clarity, the detection range is chosen to be relatively small here; in practice, however, the detection range can also be up to 180°, for example. Additional cameras and / or other sensors can also be provided. Using the sensor 106, the surroundings 120 can be detected, i.e., images of the surroundings or general environmental information can be generated or detected.

[0058] Furthermore, the vehicle 100 has a computing unit or a system 108 for data processing, e.g., a control unit, by means of which data can be exchanged with a higher-level system 110, e.g., via an indicated radio connection. In the system 110, for example, movement paths (or general navigation information) can be determined, which are then transmitted to the system 108 in the vehicle 100, which the vehicle is then to follow. However, it can also be provided that a movement path (or general navigation information) is determined in the system 108 itself or is received there in some other way. Instead of a movement path or the navigation information, the system 108 can, for example, also receive control information that has been determined based on a movement path or the navigation information, and according to which the control or regulating unit 102 can move the vehicle 100 via the drive unit 104 to follow a movement path.The movement path 130 is only indicated here as an example.

[0059] To determine the aforementioned navigation information, the images captured by camera 106—or map data sets in general, especially with other types of sensors—can be used. For example, an image can include or depict the lane markings or their positions, while other information in the image is ignored. In particular, a map data set can then be present as a point cloud, i.e., as a set of points that indicate the lane markings or their positions.

[0060] In Fig. 2a and Fig. Figure 2b shows two such map data sets or images for illustrative purposes. These map data sets or images can be generated, for example, by the vehicle's camera. Fig. 1. Both map data sets 210, 220 each comprise a point cloud, the lane markings or their positions, such as in Fig. 1 shown.

[0061] As part of the vehicle's navigation, the two map data sets 210 and 220 are now to be aligned, i.e., matched, in order to find a transformative relation (or transformation) that maps one map data set to the other map data set. Such an alignment is indicated by 200a and 200b, respectively. This transformative relation then represents a movement of the vehicle between the points in time at which the two map data sets or images were acquired. This applies in particular to the case in which multiple observations originate from one vehicle. In the case where mapping and localization are decoupled, it may be the same or a similar observation made by multiple vehicles at very different points in time.

[0062] It should be noted here that the two images or map data sets are shown only as examples and for illustrative purposes. Instead of two images, an image and an existing map can also be aligned, with the map itself being based on images. After alignment, the new image can then be used to expand the map. In this case, one of the map data sets would be the map, while the other would be a (new) image.

[0063] As already mentioned, such an alignment can be good or less good. The alignment 200a in Fig. 2a - here the point clouds or map data sets are very well superimposed - is a good alignment, the alignment 200b in Fig. 2b, on the other hand - here there is an offset between the point clouds or map data sets - is a less good alignment.

[0064] If a transformative relation (or transformation) obtained during alignment 200b—generally a transformation dataset—is used for navigation, this can lead to problems, and navigation may be poor. One way to evaluate the quality of such an alignment is presented below.

[0065] In Fig. Figure 3 schematically illustrates a flow of a method in one embodiment, specifically for generating training data sets. In a step 300, two map data sets 302, 304, each containing environmental information, are acquired, each using a sensor of the mobile device, from the mobile device and / or the environment.

[0066] In a step 310, the two map data sets are provided as input feature data sets 312, 314. It is also conceivable that the input feature data are determined based on the two map data sets. As already mentioned above, feature descriptors can be extracted here, for example. The input feature data sets represent, for example, input data for a machine learning algorithm for alignment.

[0067] In a step 320, the two map data sets are then matched. This is done, for example, using a machine learning algorithm 330, e.g., a CNN. In the process, a transformation data set 332 is obtained, which includes information about a transformative relationship between the two map data sets or input feature data sets. The transformation data set 332 is then also provided in step 340. Furthermore, in step 350, a reference transformation data set 352 is provided as ground truth.

[0068] Based on the two input feature data sets and / or the transformation data set, a correlation data set 362 is then determined in step 360, wherein the correlation data set comprises information about a correlation between at least two of the following data sets: the two input feature data sets, the transformation data set. For example, the correlation can comprise distances between a point of one of the two input feature data sets and a point of the other of the two input feature data sets that is closest to that point.

[0069] Furthermore, in step 370, a quality measure 372 is determined depending on the accuracy of a match between the transformation data set 332 and the reference transformation data set 352. As mentioned, such a quality measure can, for example, comprise one of several quality classes, or a continuous or quasi-continuous value.

[0070] In step 380, a training data set 382 is then provided, wherein the training data set comprises the correlation data set 362 and the quality measure 372.

[0071] These steps can be repeated or, if necessary, carried out at least partially in parallel several times in order to obtain a large number of training data sets.

[0072] It should be mentioned that the alignment was also carried out independently and, for example, as part of an earlier navigation of the mobile device, so that ultimately only steps 310 and 340 are required insofar as they provide the two input feature data sets and the transformation data set.

[0073] In Fig. 4 schematically shows a sequence of a method in a further embodiment, namely for training the evaluation algorithm by means of which an alignment of two map data sets is evaluated.

[0074] For this purpose, step 400, which e.g. as in step 380 according to Fig. 3 mentioned training data sets are provided in a plurality. Thus, each training data set 380 comprises a correlation data set 362 and a quality measure 372.

[0075] In a step 410, the evaluation algorithm 412 is then adapted or trained based on the plurality of training data sets, specifically such that the evaluation algorithm determines an evaluation result for a target correlation data set and a target transformation data set, which assigns a quality measure to the target transformation data set. In step 420, the evaluation algorithm adapted or trained in this way is then provided.

[0076] In Fig. 5 schematically shows a sequence of a method in a further embodiment, namely for evaluating an alignment of two map data sets by means of an evaluation algorithm, which, for example, as with respect to Fig. 4 explains how it has been adapted or trained.

[0077] For this purpose, in step 500, two map data sets 502, 504 are acquired, each with environmental information, by means of a sensor of the mobile device from the mobile device and / or the environment.

[0078] In a step 510, the two map data sets are provided as input feature data sets 512, 514. It is also conceivable that the input feature data are determined based on the two map data sets. As already mentioned above, feature descriptors can be extracted here, for example. The input feature data sets represent, for example, input data for a machine learning algorithm for alignment.

[0079] In step 520, the two map data sets are then matched. This is done, for example, using a machine learning algorithm 530, such as a CNN. In the process, a transformation data set 532 is obtained, which includes information about a transformative relationship between the two map data sets or input feature data sets. The transformation data set 532 is then also provided in step 540.

[0080] Based on the two input feature data sets and / or the transformation data set, a correlation data set 562 is then determined in step 560, wherein the correlation data set comprises information about a correlation between at least two of the following data sets: the two input feature data sets, the transformation data set. For example, the correlation can comprise distances between a point of one of the two input feature data sets and a point of the other of the two input feature data sets that is closest to that point.

[0081] Furthermore, then, step 570, based on the correlation data set 562 and on the transformation data set 532, and using an evaluation algorithm, e.g. 412 according to Fig.4, an evaluation result 574 is determined that assigns a quality measure 572 to the transformation data set. The quality measure 572 can be of the same type as the quality measure 372 with which the evaluation algorithm 412 was trained. For example, the quality measure can be determined first, which is then provided as the evaluation result 574 in step 580.

[0082] In a step 590, based on the map data sets and / or the transformation data set, and depending on the evaluation result 574, navigation information 592 for the navigation of the mobile device can then be determined, so that the mobile device can navigate. This can be the case, for example, if the quality measure according to the evaluation result corresponds to a predefined quality criterion. If the quality measure does not correspond to the predefined quality criterion, for example, if it indicates a poor fit or transformation estimate, the determination of the navigation information can be omitted, at least based on the current map data sets and / or the current transformation data set. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited non-patent literature

[0000] C. Choy, W. Dong, and V. Koltun, “Deep global registration,” in Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, pp. 2514-2523, 2020

[0015] C. Choy, J. Park, and V. Koltun, “Fully convolutional geometric features,” in Proceedings of the IEEE / CVF International Conference on Computer Vision, pp. 8958-8966, 2019 [0015, 0020] “C. Choy, W. Dong, and V. Koltun, “Deep global registration,” in Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, pp. 2514-2523, 2020

[0020]

Claims

[1] A method for generating training data sets for training an evaluation algorithm by means of which an alignment of two map data sets (302, 304) for determining navigation information for a mobile device (100) moving or intended to move in an environment (120) can be evaluated, comprising for each of a plurality of training data sets: Providing (310) two input feature data sets (312, 314) which correspond to two map data sets (302, 304) or which have been determined based on the two map data sets, wherein the two map data sets each comprise environmental information, wherein the environmental information has each been acquired by means of a sensor (106) of the mobile device (100) from the mobile device (100) and / or the environment (120); Providing (340) a transformation data set (332), wherein the transformation data set has been generated during an alignment of the two input feature data sets, and wherein the transformation data set comprises information about a transformative relation between the two input feature data sets; Providing (350) a reference transformation data set (352) as ground truth; Determining (360) a correlation data set (362) based on the two input feature data sets and / or the transformation data set, wherein the correlation data set comprises information about a correlation between at least two of the following data sets: the two input feature data sets, the transformation data set; Determining (370) a quality measure (372) as a function of an accuracy of a match between the transformation data set and the reference transformation data set; and Providing (380) the training data set (382), wherein the training data set comprises the correlation data set and the quality measure. [2] The method of claim 1, wherein the quality measure (372) comprises one of a plurality of quality classes, or wherein the quality measure comprises a continuous or quasi-continuous value. [3] Method according to claim 1 or 2, wherein the correlation data set comprises information about a correlation between the two input feature data sets, wherein the correlation comprises distances between a point of one of the two input feature data sets and a point of the other of the two input feature data sets that is closest to this point. [4] A method for training an evaluation algorithm by means of which an alignment of two map data sets (302, 304) for determining navigation information for a mobile device (100) that is moving or is to move in an environment (120) can be evaluated, comprising: Providing (400) a plurality of training data sets (382), each training data set comprising: - a correlation data set (362) which has been determined based on two input feature data sets and / or one transformation data set, and which comprises information about a correlation between at least two of the following data sets: the two input feature data sets, the transformation data set, wherein the two input feature data sets correspond to two map data sets or have been determined based on the two map data sets, wherein the two map data sets each comprise environmental information, wherein the environmental information has each been acquired by means of a sensor (106) of the mobile device (100) from the mobile device (100) and / or the environment (120), wherein the transformation data set has been generated upon alignment of the two input feature data sets, and wherein the transformation data set comprises information about a transformative relation between the two input feature data sets, and - a quality measure (372) indicating an accuracy of a match between the transformation data set and the reference transformation data set; Adapting (410) the evaluation algorithm (412) based on the plurality of training data sets such that the evaluation algorithm determines an evaluation result for a target correlation data set and a target transformation data set, which evaluation result assigns a quality measure to the target transformation data set; and Providing (420) the adapted evaluation algorithm. [5] A method according to claim 4, wherein the plurality of training data sets has been determined according to a method according to any one of claims 1 to 3. [6] A method for evaluating an alignment of two map data sets (502, 504) for determining navigation information for a mobile device (100) moving or to move in an environment (120), comprising: Providing (510) two input feature data sets (512, 514) which correspond to two map data sets or which have been determined based on the two map data sets, wherein the two map data sets each comprise environmental information, wherein the environmental information has each been acquired by means of a sensor (106) of the mobile device (100) from the mobile device (100) and / or the environment (120); Providing (540) a transformation data set (532), wherein the transformation data set has been generated during an alignment of the two input feature data sets, and wherein the transformation data set comprises information about a transformative relation between the two input feature data sets, Determining (560), based on the two input feature data sets and / or the transformation data set, a correlation data set (562), wherein the correlation data set comprises information about a correlation between at least two of the following data sets: the two input feature data sets, the transformation data set; Determining (570), based on the correlation data set and on the transformation data set, and using an evaluation algorithm (412), an evaluation result (574) that assigns a quality measure (572) to the transformation data set; and Providing the evaluation result for use in determining the navigation information, in which navigation information is determined based on the two map data sets and / or the transformation data set depending on the evaluation result. [7] The method according to claim 6, wherein determining the navigation information based on the two map data sets and / or the transformation data set, depending on the evaluation result, comprises: - Using the two map data sets and / or the transformation data when determining the navigation information if the quality measure according to the evaluation result corresponds to a predetermined quality criterion, and not using the two map data sets and / or the transformation data when determining the navigation information if the quality measure according to the evaluation result does not correspond to the predetermined quality criterion, or - Using the two map data sets and / or the transformation data set to determine the navigation information with a weighting, wherein the weighting is determined depending on the quality measure according to the evaluation result. [8] The method of claim 6 or 7, further comprising: Determining (590) the navigation information (592) based on the map data sets and / or the transformation data set as a function of the evaluation result, wherein the navigation information comprises in particular a map of the surroundings and / or a trajectory for the mobile device. [9] Method according to one of the preceding claims, wherein the sensor of the mobile device comprises one of the following sensors: a camera, a radar sensor, a lidar sensor, an ultrasonic sensor. [10] Method according to one of the preceding claims, wherein the two map data sets are in the form of point clouds or represent such. [11] A data processing system (108, 110) comprising means for carrying out the method according to any one of the preceding claims. [12] A mobile device (100) comprising a system (108) according to claim 11 and / or configured to receive navigation information determined according to a method according to claim 10, wherein the mobile device (100) has a sensor (106) for detecting environmental information and is configured to navigate based on the navigation information, preferably with a control or regulating unit and a drive unit for moving the mobile device according to the navigation information. [13] Mobile device (100) according to claim 12, which is designed as an at least partially automated moving vehicle, in particular as a passenger transport vehicle or as a goods transport vehicle, and / or as a robot, in particular as a household robot, e.g. a cleaning robot, a floor or street cleaning device or a lawnmower robot, and / or as a drone. [14] A computer program comprising instructions which, when executed by a computer, cause the program to carry out the method steps of a method according to any one of claims 1 to 10 when executed on the computer. [15] A computer-readable storage medium on which the computer program according to claim 13 is stored.

Images