Method and apparatus for training student network of teacher-student training architecture

Abstract

The invention relates to a method and apparatus for training a student network of a teacher-student training architecture. A method for training a student network of a teacher-student training architecture having a teacher network (200) and a student network (202) for point cloud registration to map an environment for navigating a technical system.

Description

Technical Field

[0001] This invention relates to a method for training a student network with a teacher-student training architecture to perform point cloud registration for mapping an environment for navigating a technical system. This invention also relates to an apparatus for training a student network with a teacher-student training architecture to perform point cloud registration for mapping an environment for navigating a technical system. Furthermore, this invention relates to a method for performing point cloud registration for mapping an environment for navigating a technical system. Finally, this invention relates to a computer program having program code. Finally, this invention relates to a computer-readable data carrier having program code. Background Technology

[0002] For example, map matching or registration methods used in mapping are employed in the fields of autonomous driving and robot navigation. These map registration methods create maps of the environment. Autonomous vehicles and / or robots can then navigate based on such maps, for example.

[0003] Here, map matching or registration methods are used to unify sensor data used for mapping. Thus, for example, sensors may observe the environment from different viewpoints and / or angles, producing different sensor data, but these sensor data map the same real environment. To unify these different sensor data, map matching or registration methods are used to unify the sensor data or transform it to a common reference frame, thereby achieving a unified mapping of the environment. The goal of such map matching (also known as scan-matching) is to determine an alignment transformation between different input data, which can come from any sensor and / or from different viewpoints. This transformation is often the most important first step in creating a merged map from multiple partially overlapping sensor data (or representations derived from them).

[0004] For example, vehicle cameras can detect lane markings from different perspectives while driving. The detected camera data can be matched via scan matching, particularly pixel matching. Similar actions apply to point cloud data detected by lidar and / or radar sensors, which can be associated with a common reference frame using map matching or registration methods.

[0005] The field of map matching or registration methods has a large body of literature, which describes not only classical analysis schemes but also increasingly machine learning (ML) schemes. In recent years, within the ML-based approach, several methods have emerged to overcome the necessity of (manual) assignment of ground truth labels, which is associated with significant labeling costs and consequently time consumption.

[0006] The paper "Self-supervised geometric perception" by H. Yang et al., in Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition, pp. 14350–14361, 2021, is inspired by the student-teacher architecture and specifically proposes the SGP algorithm. Here, the student is a trainable feature matcher that outputs the predicted correspondence between the input data to be matched or unified. The teacher's RANSAC part is a learning-free robust solver that only estimates the rigid transformation between the input data to be matched or unified. The predicted transformation is then used as a pseudo-label to supervise the student, thereby iteratively improving the pseudo-label again. According to this scheme, the student is thus trained across multiple epochs using the same pseudo-labels until the teacher generates new labels.

[0007] The authors Q. Liu et al.'s paper, "Extend your own correspondences: Unsupervised distant point cloud registration by progressive distance extension" (arXiv preprint arXiv:2403.03532, 2024), extends the scheme proposed by H. Yang et al. to datasets from the automotive industry by progressively increasing the spacing between point clouds and spatially filtering correspondences that are near the vehicle itself. Similar to the case of H. Yang et al., a learning-free robust solver is used here as the teacher, but SC2-PCR is chosen.

[0008] The authors M. El Banani and J. Johnson's "Bootstrap your own correspondences" in Proceedings of the IEEE / CVF International Conference on Computer Vision, pp. 6433–6442, 2021, utilizes the fact that images and point clouds are coupled in RGB-D data and uses pseudo-labels from a randomly initialized image feature extractor to train a point cloud feature extractor.

[0009] Therefore, various known methods in the image domain demonstrate the efficient use of the model's own predictions in order to improve training without a large amount of labeled or identifiable training data.

[0010] Even though some solutions are already known from the prior art, there is still potential for improvement. Therefore, the object of the present invention is to describe methods for further improvement and / or improved devices.

[0011] This task is solved by the method according to the features of claim 1. This task is solved by the apparatus according to the features of claim 10. Summary of the Invention

[0012] According to the first aspect, a method is proposed for training a student network in a teacher-student training architecture for point cloud registration to map the environment for navigation of a technical system. This teacher-student training architecture comprises a teacher network (hereinafter referred to as the teacher) and a student network (hereinafter referred to as the student). The method includes the following steps:

[0013] - Provide a first point cloud for points of objects in the environment and a second point cloud for points of objects in the environment, wherein the points of the first point cloud and the second point cloud are registered to the objects from different viewpoints;

[0014] - Features are extracted from points in the first and second point clouds by the encoder of the teacher network.

[0015] - Features are extracted from points in the first and second point clouds, respectively, by the encoder of the student network;

[0016] - Estimate point pairs between extracted point features of the first and second point clouds, the point features being embedded into the embedding space by the decoder of the teacher network;

[0017] - Estimate point pairs between extracted point features of the enhanced first point cloud and the enhanced second point cloud, the point features being embedded into the embedding space by the decoder of the student network;

[0018] - The spacing estimated by the decoder is optimized by applying an optimization algorithm to remove outlier measurements from the estimated spacing and by subsequently applying an iterative nearest-point algorithm to maximize the consistency or rigid transformation between the extracted point features of the first and second point clouds.

[0019] - New point pairs are generated by applying the Nearest-Neighbour algorithm based on the consistency / rigid transformation between the extracted point features of the first and second point clouds;

[0020] - The loss function is optimized based on the pixel spacing estimated by the decoder of the student network and the newly generated pixel spacing to train the student network; and

[0021] - Provides a trained student network for point cloud registration to map the environment and navigate technical systems.

[0022] It is understood that the steps and other optional steps according to the invention do not necessarily have to be performed in the indicated order, but may be performed in a different order. Furthermore, other intermediate steps may be provided. Each step may also include one or more sub-steps without departing from the scope of the method according to the invention.

[0023] According to the second aspect, a device is proposed for training a student network in a teacher-student training architecture for point cloud registration to map the environment for navigating a technical system. The teacher-student training architecture includes a teacher network and a student network. The device includes evaluation and computational means configured to perform the following steps:

[0024] - Provide a first point cloud for points of objects in the environment and a second point cloud for points of objects in the environment, wherein the points of the first point cloud and the second point cloud are registered to the objects from different viewpoints;

[0025] - Features are extracted from points in the first and second point clouds by the encoder of the teacher network.

[0026] - Features are extracted from points in the first and second point clouds, which are respectively enhanced, by the encoder of the student network;

[0027] - Estimate point pairs between the first and second point clouds using the decoder of the teacher network;

[0028] - Estimate point pairs between extracted point features of the enhanced first point cloud and the enhanced second point cloud, the point features being embedded into the embedding space by the decoder of the student network;

[0029] - The spacing estimated by the decoder is optimized by applying an optimization algorithm to remove outlier measurements from the estimated spacing and by subsequently applying an iterative nearest-point algorithm to maximize the consistency or rigid transformation between the extracted points of the first and second point clouds.

[0030] - New point pairs are generated by applying the nearest neighbor algorithm based on the consistency / rigid transformation between the extracted point features of the first and second point clouds.

[0031] - The loss function is optimized based on the pixel spacing estimated by the decoder of the student network and the newly generated pixel spacing to train the student network; and

[0032] - Provides a trained student network for point cloud registration to map the environment and navigate technical systems.

[0033] The statements made for this method apply accordingly to this device. It is understood that, according to common linguistic practice, linguistic variations on the features described according to the method can be restated for this device without explicitly citing such a statement here.

[0034] A self-distillation scheme is proposed for learning point cloud registration in an unsupervised manner. In this scheme, each sample of a point cloud pair is fed to a teacher network, and augmented samples of the point cloud pair are fed to a student network. The teacher includes a (hyper)parameter-trainable feature extractor or encoder and a solver or decoder. The decoder, along with the encoder, is preferably part of a common neural network. For example, the teacher may also have a RANSAC part. The RANSAC part is preferably learning-free and robust. The solver preferably enforces Konsistenz consistency of correspondence between point pairs and optimizes the unsupervised in-point rate to eliminate the need for ground truth labels. Therefore, this scheme simplifies the training method compared to related methods and outperforms them in the case of multiple datasets of point cloud pairs. This scheme simplifies unsupervised point cloud registration by eliminating the need for pseudo-label validators. Furthermore, it eliminates the need for manually labeled bootstrap features and progressive datasets.

[0035] The first and second point clouds are voxelized and passed to the teacher's feature extractor. This feature extractor predicts latent features for all points in the point clouds. The student obtains an expanded or enhanced version of the point cloud. For all points in the first point cloud, corresponding points are searched in the second point cloud where the spacing between the teacher's features should be minimized, resulting in estimated correspondents (Entsprechungen) or point pairs. Correspondent prediction is improved by applying RANSAC, followed by ICP algorithms. New, improved correspondences are then generated by searching for nearest neighbors in the coordinate space.

[0036] This method employs a so-called Mean-Teacher model, based on a teacher-student architecture where the teacher is the exponential mean (EMA) of the student's parameters, thereby promoting consistency in predictions used for semi-supervised learning. Furthermore, this eliminates the need for labeling training data.

[0037] This method preferably employs a self-distillation technique, which involves using a non-contrastive loss function, in particular, to optimize pseudo-class labels. The data augmentation used to train students preferably improves generalization. Other data augmentation of the input data preferably includes rotation, noise injection, and / or distortion of the input data.

[0038] This invention resides in the field of surveying or map matching. It is particularly applicable in the fields of autonomous driving and / or robot navigation and / or generally when navigating technical systems in an environment. The invention is applied in conjunction with and therefore in the fields of robot navigation and surveying. This invention facilitates the creation of maps in an optimized manner, which can in particular be used as sensors in technical systems for navigation in an environment.

[0039] This invention can be used to analyze (input) data from real or virtual (simulated) sensors. For example, such real or virtual sensors can detect or simulate environmental measurements in the form of sensor signals or sensor data. Sensor data can be pixel-based, especially digital, image and / or video data. This image and / or video data can be detected, for example, by a camera, thermal imaging camera, motion sensor, and / or video camera. Sensor data can also exist as point cloud data. This point cloud data can be detected, for example, by a radar sensor, lidar sensor, and / or ultrasonic sensor.

[0040] The sensor data, whether actually detected or synthesized, preferably includes data relevant to the vehicle and related to mapping, such as the positions of lane markings, traffic signs, pillars, and / or other objects in vehicle-related coordinates. Furthermore, this sensor data may have additional semantic information. Here, the actual sensor origin or the reference frame on which the detection is based is irrelevant.

[0041] Based on sensor signals, information about elements or objects in the sensor's environment, encoded by the sensor signals, can preferably be obtained. In other words, indirect measurements can be performed based on sensor signals used as direct measurements. Here, the invention is preferably used for classifying sensor data, identifying the presence of objects in sensor data, and / or performing semantic segmentation of sensor data, for example, for traffic signs. Furthermore, the invention can be used to determine one or more continuous values, i.e., for performing regression analysis, such as alignment transformation for given input data. Additionally, the invention can be used for controlling technical systems, particularly for calculating control signals for controlling technical systems such as computer-controlled machines, such as robotic systems, vehicles, home appliances, power tools, manufacturing machines, personal assistants, and / or access control systems. Furthermore, the invention can be used in systems for information transmission, such as monitoring systems or medical (imaging) systems. Here, the invention enables the creation of (uniform) maps that can be used by technical systems as a sensor alternative for navigation.

[0042] It should be mentioned that this method can also be applied to pixel-based images, rather than point clouds. The method steps can be summarized as follows:

[0043] - Provide a first image containing pixels of objects in the environment and a second image containing pixels of objects in the environment, wherein the pixels of the first image and the second image are registered from different viewpoints;

[0044] - Features are extracted from the pixels of the first and second images by the encoder of the teacher network;

[0045] -Features are extracted from the pixels of the first and second images, which are enhanced respectively, by the encoder of the student network;

[0046] - The pixel spacing between the extracted pixel features of the first and second images is estimated using the decoder of the teacher network;

[0047] - The pixel spacing between the extracted pixel features of the enhanced first image and the enhanced second image is estimated by the decoder of the student network;

[0048] - The pixel spacing estimated by the decoder is optimized by applying an optimization algorithm to remove outlier measurements from the estimated pixel spacing and by subsequently applying an iterative nearest-point algorithm to maximize the consistency between the extracted pixel features of the first and second images.

[0049] - By applying the nearest neighbor algorithm, a new pixel spacing is generated based on the consistency / transformation between the extracted pixel features of the first and second images;

[0050] - The loss function is optimized based on the pixel spacing estimated by the decoder of the student network and the newly generated pixel spacing to train the student network; and

[0051] - Provides a trained student network for image registration to map the environment and navigate technical systems.

[0052] In another aspect, it is proposed that the loss function has a contrastive loss function, wherein the loss function is optimized until a threshold is reached or until a predetermined number of training iterations are reached.

[0053] Currently, a contrastive loss is adapted to train the student network. Since ground truth labels are unavailable, positive (inlier) point pairs or predicted correspondences of the point spacing estimated by the teacher network are used. Correspondingly, negative (outlier) point pairs are ignored below a specific threshold. It has been shown that excluding the validator leads to better performance.

[0054] In another aspect, it is proposed that the student network and the teacher network have the same network architecture.

[0055] Therefore, a student-teacher architecture is used, where the teacher generates pseudo-labels during continuous runs to train the student. This scheme allows for continuous improvement of the pseudo-labels, eliminating the need to generate new pseudo-labels only after multiple full training runs. The teacher network is updated using the EMA of the student's parameters and thus uses the same architecture. Here, a robust solver is integrated into the teacher to improve its estimation. In the final step, a contrastive loss is applied, where positive pairs are determined by the estimated correspondence of the pair or the resulting distance between the teacher and the student.

[0056] In another aspect, it is proposed that the encoder of the teacher network is a parameter-trainable encoder, the parameters of which are trained or updated using an exponential moving average of the parameters of the encoder of the student network. In yet another aspect, it is proposed that the decoder of the teacher network is a parameter-trainable decoder, the parameters of which are trained or updated using an exponential moving average of the parameters of the decoder of the student network.

[0057] The statement "Update Teacher Network with EMA Using Student Parameters" describes a method for updating the parameters of a teacher network in a teacher-student architecture. This method uses the exponential moving average (EMA) of the student network's parameters to update the teacher network's parameters. In this architecture, the teacher network serves as a reference model, while the student networks are models trained to mimic the behavior of the teacher network. During training, the student network's parameters are continuously adapted to optimize performance. The teacher network is updated with EMA as follows: instead of directly and immediately taking over the teacher network's parameters from the student network's current parameters, a smoothed version of those parameters is computed.

[0058] Here, the exponential moving average (EMA) weights more strongly on more recent values ​​compared to older values, resulting in smoothing and stabilizing abrupt changes and fluctuations in the student parameters. Specifically, this means that the parameters of the teacher network are a continuously adapted, smoothed version of the student parameters. Because these methods mitigate rapid and potentially excessive changes in the student parameters, they result in more stable parameters in the teacher network. Consequently, the teacher network becomes more robust and less susceptible to noise and short-term fluctuations.

[0059] In another aspect, it is proposed that an enhanced first point cloud is generated by data enhancement from the first point cloud, and an enhanced second point cloud is generated by data enhancement from the second point cloud, wherein the data enhancement includes rotating, twisting, and / or translating the data points of the first point cloud and / or the second point cloud.

[0060] In the forward pass, for example, two point clouds are randomly rotated to produce an augmented point cloud, and to force the student network to become rotationally invariant. During distillation, it is important to overcome the bootstrapping phase, where a randomly initialized teacher may provide less useful pseudo-labels. Here, the teacher is also not yet trained to be rotationally invariant, and therefore receives weaker results in terms of augmented sampling. To circumvent this, augmentation is currently applied only to the student's input.

[0061] In another aspect, it is proposed that the optimization algorithm for removing outlier measurements has the RANSAC (Random Sample Consensus) algorithm, which finds a set of inliers in the estimated point pairs and ignores outlier measurements in the estimated point pairs.

[0062] RANSAC (Random Sample Consensus) is an iterative algorithm used to estimate the set of parameters in a model from a dataset that may contain outliers. The main objective of RANSAC is to find a set of inliers (data points belonging to a specific model) from a given dataset and ignore outliers (data points that do not fit the model).

[0063] The Iterative Closest Point (ICP) algorithm is a widely used algorithm for registering point clouds. It is used to transform two point clouds to maximize their consistency or rigidity transformation. The ICP algorithm is frequently used in computer vision, robotics, and 3D modeling.

[0064] In another aspect, a method for point cloud registration to map the environment for navigation of a technical system is proposed. This method includes the following steps:

[0065] - Especially from the perspective of the technical system, the first point cloud of points detecting objects in the environment from the first viewpoint;

[0066] - Especially from the perspective of the technical system, a second point cloud is used to detect objects in the environment from a second field of view that is different from the first field of view;

[0067] - Register the first and second point clouds using the currently trained student network for mapping the environment; and

[0068] - Navigate the technical system in the mapped environment based on the registered first and second point clouds.

[0069] In another aspect, a computer program having program code is claimed for performing at least a portion of the method in one of its aspects when the computer program is executed on a computer. In other words, a computer program (product) is proposed that includes instructions, which, when executed by a computer, cause the computer to perform the method and / or the steps of the method in one of its aspects.

[0070] In another aspect, a computer-readable data carrier having program code of a computer program is proposed for performing at least a portion of the method in one of its aspects when the computer program is executed on a computer. In other words, the present invention relates to a computer-readable (storage) medium comprising instructions that, when executed by a computer, cause the computer to perform the method and / or the steps of the method in one of its aspects.

[0071] The described design and improvement schemes can be combined with each other arbitrarily.

[0072] Other possible designs, improvements, and implementations of the present invention include combinations of features of the present invention that are not explicitly mentioned in the preceding or following descriptions of embodiments. Attached Figure Description

[0073] The accompanying drawings should facilitate a further understanding of embodiments of the invention. The drawings illustrate embodiments and, in conjunction with the description, serve to explain the principles and concepts of the invention.

[0074] Other embodiments and many of the advantages described above arise in consideration of the accompanying drawings. The elements shown in the drawings are not necessarily depicted to scale.

[0075] Figure 1 A flowchart illustrating one embodiment of the present invention is shown.

[0076] Figure 2 A block diagram illustrating an embodiment of the present invention is shown.

[0077] In the accompanying drawings, unless otherwise stated, the same reference numerals denote the same or functionally identical elements, components, or assemblies. Detailed Implementation

[0078] Figure 1 A schematic flowchart illustrates a method for training a student network to perform point cloud registration for mapping the environment and navigating a technical system. This method also utilizes… Figure 2 Described.

[0079] The teacher-student training architecture has a teacher network 200 and a student network 202.

[0080] In any implementation, the method may be performed at least in part by device 100, which may include a plurality of components not shown in more detail, such as one or more providing means and / or at least one evaluation and computing means. It will be understood that the providing means may be constructed in conjunction with the evaluation and computing means, or may be different from the evaluation and computing means. Furthermore, device 100, which may be part of a system, may include storage means and / or output means and / or display means and / or input means.

[0081] The computer-implemented method includes at least the following steps:

[0082] In step S1, a first point cloud 204 is provided for points of objects in the environment, and a second point cloud 206 is provided for points of objects in the environment, wherein the points of the first point cloud and the second point cloud 204, 206 are registered from different viewpoints.

[0083] In step S2, features are extracted from points in the first point cloud and the second point clouds 204 and 206 by the encoder 208 of the teacher network 200. The encoder 208 of the teacher network 200 is a parameter-trainable encoder, the parameters of which are trained or updated by the exponential moving average of the parameters of the encoder 210 of the student network 202.

[0084] In an optional step, data augmentation is performed on the first point cloud and the second point cloud 204, 206, or the corresponding points, to generate a first augmented point cloud 204' and a second augmented point cloud 206'. Data augmentation may include rotating and / or distorting and / or translating the data points of the first point cloud and / or the second point cloud 204, 206.

[0085] In step S3, features are extracted from the points of the first point cloud and the second point cloud 204' and 206', which are respectively enhanced, by the encoder 210 of the student network 202.

[0086] In step S4, point pairs 212 are estimated between the extracted point features of the first point cloud and the second point clouds 204, 206, said point features being embedded into an embedding space by the decoder 214 of the teacher network 200. The decoder 214 embeds the point features into a metric space. In this space, the point spacing 212 is calculated using Euclidean distance.

[0087] In step S5, point pairs 216 are estimated between the extracted point features of the enhanced first point cloud and the enhanced second point cloud 204', 206', and the point features are embedded into the embedding space by the decoder 218 of the student network 202.

[0088] In step S6, the point spacing 212 estimated by decoder 214 is optimized by applying optimization algorithm 220 to remove outlier measurements 222 from the estimated point pairs and by subsequently applying iterative nearest-point algorithm 224 to maximize consistency 225 through a rigid transformation between points in the first point cloud and the second point clouds 204, 206. The optimization algorithm 220 for removing outlier measurements 222 includes the RANSAC (Random Sample Consensus) algorithm, which finds a set of inliers 227 in the estimated point pairs 212 and ignores outlier measurements 222 in the estimated point pairs 212.

[0089] In step S7, new point pairs 226 are generated by applying the nearest neighbor algorithm 228 based on the consistency 225 or rigid transformation between the points of the first point cloud and the second point cloud 204, 206.

[0090] In step S8, a loss function 230 is optimized based on the pixel spacing 216 estimated by the decoder 218 of the student network 202 and the newly generated pixel spacing 226 for training the student network 202. The loss function 230 has a contrastive loss function, wherein the loss function 230 is optimized until a threshold is reached or until a predetermined number of training iterations are reached.

[0091] In step S9, a trained student network 202' is provided for point cloud registration to map the environment for navigation of the technical system.

[0092] Figure 2 The self-distillation (DiReg) process for registration is illustrated. Two point clouds 204, 206 are passed to teacher network 200, while student network 202 obtains expanded or enhanced point clouds 204', 206'. Networks 200 and 202 preferably predict geometric features for all points in their point cloud pairs 204, 206, 204', 206'. Here, correspondences are preferably collected by searching for nearest neighbors in the feature vectors of teacher network 200. Based on these correspondences, RANSAC preferably estimates a rigid transformation or consistency 225 to align the two point clouds 204 and 206. Next, nearest neighbors are searched in coordinate space to obtain improved correspondences for supervising student network 202. Preferably, a stopping gradient operator 232 is also used to clarify that backpropagation is not performed through teacher network 200.

Claims

1. A method for training a student network of a teacher-student training architecture for point cloud registration to map an environment for navigating a technical system, wherein the teacher-student training architecture has a teacher network (200) and a student network (202), wherein the method includes the steps of: - Provide (S1) a first point cloud (204) for points of objects in the environment and a second point cloud (206) for points of objects in the environment, wherein the points of the first point cloud and the second point cloud (204, 206) are registered to the objects from different viewpoints; - Features (S2) are extracted from points in the first point cloud and the second point cloud (204, 206) by the encoder (208) of the teacher network (200); - Features (S3) are extracted from the points of the first and second point clouds (204', 206') respectively enhanced by the encoder (210) of the student network (202); - Estimate (S4) the point pairs (212) between the extracted point features of the first point cloud and the second point cloud (204, 206), the point features being embedded into the feature space by the decoder (214) of the teacher network (200); - Estimate (S5) the point pairs (216) between the extracted point features of the enhanced first point cloud and the enhanced second point cloud (204', 206'), the point features being embedded into the feature space by the decoder (218) of the student network (202); - The point spacing (212) estimated by the decoder (214) is optimized (S6) by applying an optimization algorithm (220) to remove outlier measurements (222) from the estimated point pairs (212) and by subsequently applying an iterative nearest point algorithm (224) to maximize the consistency (225) between the points of the first point cloud and the second point cloud (204, 206); - By applying the nearest neighbor algorithm (228) based on the consistency (225) between points in the first point cloud and the second point cloud (204, 206), a new point pair (226) is generated (S7); - The loss function is optimized based on the pixel spacing estimated by the decoder of the student network and the newly generated pixel spacing to train the student network; as well as - Provide a trained student network for point cloud registration to map the environment for navigation of the technology system.

2. The method according to claim 1, wherein the loss function (230) has a contrastive loss function, wherein the loss function (230) is optimized until a threshold is reached or until a predetermined number of training iterations are reached.

3. The method according to claim 1 or 2, wherein the student network (202) and the teacher network (200) have the same network architecture.

4. The method according to any one of the preceding claims, wherein the encoder (208) of the teacher network (200) is a parameter-trainable encoder, the parameters of which are trained or updated by an exponential moving average of the parameters of the encoder (210) of the student network (202); and wherein the decoder (214) of the teacher network (200) is a parameter-trainable decoder, the parameters of which are trained or updated by an exponential moving average of the parameters of the decoder (218) of the student network (202).

5. The method according to any one of the preceding claims, wherein an enhanced first point cloud (204') is generated by data enhancement from the first point cloud (204), and an enhanced second point cloud (206') is generated by data enhancement from the second point cloud (206), wherein the data enhancement includes rotating and / or twisting and / or translating data points of the first point cloud and / or the second point cloud.

6. The method according to any one of the preceding claims, wherein the optimization algorithm for removing outlier measurements has a RANSAC (Random Sample Consensus) algorithm and is configured to find a set of inliers in the estimated point pairs 212 and ignore outlier measurements in the estimated point pairs.

7. A method for performing point cloud registration to map the environment for navigating a technical system; the method comprising: -In particular, from the perspective of the technical system, the first point cloud (204) of the points of the objects in the environment detected from the first viewpoint; -In particular, from the perspective of the technical system, a second point cloud (206) detects points of objects in the environment from a second field of view different from the first field of view; - Using a student network trained according to any one of claims 1 to 6, register the first point cloud and the second point cloud for mapping the environment; and - Navigate the technical system in the mapped environment based on the registered first and second point clouds.

8. A computer program having program code for performing at least a portion of the method according to any one of claims 1 to 7 when the computer program is executed on a computer.

9. A computer-readable data carrier having program code of a computer program for performing at least a portion of the method according to any one of claims 1 to 7 when the computer program is executed on a computer.

10. An apparatus (100) for training a student network of a teacher-student training architecture for point cloud registration to map an environment for navigating a technical system, wherein the teacher-student training architecture has a teacher network and a student network, and wherein the apparatus (100) has an evaluation and computing device configured to perform the following steps: - Provide (S1) a first point cloud (204) for points of objects in the environment and a second point cloud (206) for points of objects in the environment, wherein the points of the first point cloud and the second point cloud (204, 206) are registered to the objects from different viewpoints; - Features (S2) are extracted from points in the first point cloud and the second point cloud (204, 206) by the encoder (208) of the teacher network (200); - Features (S3) are extracted from the points of the first and second point clouds (204', 206') respectively enhanced by the encoder (210) of the student network (202); - Estimate (S4) the point pairs (212) between the extracted point features of the first point cloud and the second point cloud (204, 206), the point features being embedded into the embedding space by the decoder (214) of the teacher network (200); - Estimate (S5) the point pairs (216) between the extracted point features of the enhanced first point cloud and the enhanced second point cloud (204', 206'), the point features being embedded into the embedding space by the decoder (218) of the student network (202); - The point spacing (212) estimated by the decoder (214) is optimized (S6) by applying an optimization algorithm (220) to remove outlier measurements (222) from the estimated point pairs (212) and by subsequently applying an iterative nearest point algorithm (224) to maximize the consistency (225) between the points of the first point cloud and the second point cloud (204, 206); - By applying the nearest neighbor algorithm (228) based on the consistency (225) between the extracted point features of the first point cloud and the second point cloud (204, 206), a new point pair (226) is generated (S7); - The loss function is optimized based on the pixel spacing estimated by the decoder of the student network and the newly generated pixel spacing to train the student network; as well as - Provide a trained student network for point cloud registration to map the environment for navigation of the technology system.

Images