Training method for a machine learning model and method for recognizing navigation objects

Abstract

The invention relates to a training method (100) for a machine learning model (50) for recognizing navigation objects (42) in order to provide the recognized navigation objects (42) for an orientation during navigation of an at least partially autonomous robot (61) or vehicle (62), comprising the following steps: – providing (101) at least one scene representation (40) of at least one traffic scene (41), – providing (102) predefined line representations (503) of at least a selection of navigation objects (42) which are provided in the at least one traffic scene (41), – training (103) the machine learning model (50) on the basis of the at least one provided scene representation (40) and the provided predefined line representations (503) in order to output line representations (503) of those navigation objects (42) which are recognized in the at least one traffic scene (41), wherein the line representations (503) represent the navigation objects (42) in each case in a discretized point representation (303, 313) with a set of points (302).

Description

[0001]R.408353 - 1 - Description Title Training method for a machine learning model and method for recognizing navigation objects The present invention relates to a training method as defined in the preamble of claim 1. Furthermore, the invention relates to a method, a computer program, a device, a storage medium, and a machine learning model. State of the art In order to effectively orient oneself in the environment, for example on highways, during autonomous driving, it is often necessary to identify the relevant navigation objects. Such objects could be road markings, which enable understanding of traffic regulations and facilitate autonomous vehicle navigation to stay in lane. Another important component is curbs, as they limit the drivable area. Traffic light poles are also important objects,which can be used for vehicle localization. It is often of great importance to detect these objects in three-dimensional form and to ensure compatibility with different resolutions, as this can reduce the technical effort and associated costs for deploying vehicles or robots with cameras of different resolutions. A state-of-the-art method was presented in the publication "Zhan Qu et al. Focus on Local: Detection Lane Marker from Bottom Up via Key Point." This method enables the detection of lane markings in image space. Despite supporting multi-resolutions, the method can only detect vertical lane markings and provides only 2D pixel information about the position of lane markings. The decoding mechanism disclosed therein also assumesthat all detected lane markings traverse the entire image R.408353 - 2 -. However, this is not the case if non-contiguous lane markings are present at road intersections. Disclosure of the Invention The subject matter of the invention is a training method having the features of claim 1, a method having the features of claim 5, a computer program having the features of claim 8, a device having the features of claim 9, a computer-readable storage medium having the features of claim 10, and a machine learning model having the features of claim 11. Further features and details of the invention emerge from the respective subclaims, the description, and the drawings. Features and details described in connection with the training method according to the invention naturally also apply in connection with the method according to the invention, the computer program according to the invention,the device according to the invention as well as the computer-readable storage medium according to the invention and the machine learning model according to the invention, and vice versa, so that with regard to the disclosure of the individual aspects of the invention, reference is always made to each other. The subject matter of the invention is, in particular, a training method for a machine learning model for recognizing navigation objects in order to provide the recognized navigation objects for orientation during navigation of an at least partially autonomous robot or vehicle. For this purpose, for example, a navigation device for autonomous driving can be provided on the robot or vehicle in order to automatically control the movement of the robot or vehicle based on the recognized navigation objects and, for example, to carry out roadway recognition. The training method can comprise the following steps:which are preferably carried out consecutively and / or repeatedly: - Providing at least one scene representation of at least one traffic scene, wherein the respective scene representation can be implemented, for example, as a camera image. A plurality of scene representations can also be provided in order to enable reliable training. - Providing predefined line representations of at least a selection of navigation objects provided in the at least one traffic scene. The R.408353 - 3 - line representations can, for example, have been prepared manually by viewing the scene representation. In this way, training data can be provided for training. - Training the machine learning model based on this training data, i.e., the at least one provided scene representation and the provided predefined line representations,to output line representations of those navigation objects that are recognized in the at least one traffic scene. In other words, the machine learning model is trained to recognize the navigation objects in the at least one provided scene representation and to output them in the form of line representations. The line representations can be characterized in that they each represent the navigation objects in a discretized point representation with a set of points. The line representations can also each comprise further parameters, which, for example, provide an explicit representation of the start and end points of the set of points and / or an indication of an order for connecting the points. This makes it possible to reproduce the respective navigation object in the form of a line. A further improvement can be achievedif the respective point is defined by two-dimensional coordinates (instead of the conventional definition by only one coordinate). This enables the representation of the navigation objects as line objects of any shape and a completely generic representation of the line for multiple resolutions. The invention can have the advantage of enabling the correct recognition and decoding of various shapes of navigation objects and, in particular, road markings. This overcomes the limitations of the prior art for the recognition of only vertical road markings and also enables the recognition of lanes in 3D - i.e., in three-dimensional space - and their classification according to their type (e.g., road markings can be assigned class designations such as solid lane, stop line, or dashed lane, etc.). Furthermore, not only road markings,but also other navigation objects such as curbs and / or poles, especially traffic light poles, will be recognized in 3D and provide the required desired features for the object of interest (e.g., for curbs, the height of the curbs can also be provided). The invention has several unique main aspects. It provides an approach for the consistent development of the recognition of navigation objects—particularly in the form of line-like objects of arbitrary shapes. Furthermore, it can provide the possibility of obtaining the recognition of navigation objects, referred to as objects for short, directly in 3D, which reduces the complexity of line recognition approaches for recovering 3D information. Furthermore, it can provide a possibility of classifying the semantic meaning of the objects. This allows the application of a single neural network,to detect several different types of objects with line shapes. These types of objects can include, for example, road markings, curbs, masts, and the like. The invention also allows for a reduction in the effort and thus the cost of embedded hardware, as the approach is very efficient while retaining the functionality of more complex image processing approaches (semantic segmentation and / or depth estimation). Furthermore, each line shape of the objects can advantageously be detected in 3D with a single approach. Furthermore, it is possible to give each detected object a semantic meaning that is important for the navigation and localization of robots and / or vehicles and can be used for this purpose. It is also possible to use different scaling variants (e.g., only road marking detection,only pole or curb detection or any possible combination of these object detections). Furthermore, the solution according to the invention can be used with different camera resolutions, if necessary, while training only at a single resolution. This makes it a very efficient solution, particularly for autonomous systems for autonomous driving. Preferably, it can be provided that the line representations each comprise at least one of the following parameters: - A confidence and / or heat map for estimating a pixel position of the points, - A classification mask for the semantic classification of the respective represented navigation object, R.408353 - 5 - - An offset parameter of a cell of a heat map with offset information to specify an offset to a pixel position of the points of the point representation, - Depth information for the respective point of the point representation,which is specific to the distance of the represented navigation object from the vehicle and / or robot, - A class label for the respective point of the point representation to semantically classify the point. In contrast to known solutions that are not capable of recognizing arbitrary linear objects, the invention can overcome this limitation by introducing special parameterization and decoding schemes and extend 3D recognition with a semantic classification that allows the same machine learning model, preferably CNN (Convolutional Neural Network), to be used for the recognition of various navigation objects, preferably linear objects such as curbs, posts, and road markings. The provided predefined line representations can be specified as ground truth and / or a parameterization of the line representations,In particular, a specification of the parameters for training can be specified by the ground truth. Alternatively or additionally, the at least one provided scene representation can be implemented as an at least or exactly two-dimensional camera image. Furthermore, within the scope of the invention, it can be provided that each of the points comprises offset information for specifying the order for connecting the points, which indicates an offset to a preceding and a following point of the point representation in order to recreate the line representing the respective navigation object by connecting the points. This has the advantage that the conventional limitation on the shape of the navigation objects is overcome. According to a further possibility, it can be provided that the machine learning model is trained - preferably end-to-end -to output the recognized navigation objects represented in the discretized point representation in the output tensor of the machine learning model. This enables an improved representation of the navigation objects. R.408353 - 6 - The invention also relates to a method for recognizing at least one navigation object of at least one traffic scene in order to use the at least one recognized navigation object for orientation during navigation of an at least partially autonomous robot or vehicle. The following steps can be performed, preferably sequentially and / or repeatedly: - Providing at least one scene representation of the at least one traffic scene, which results from camera capture and / or depicts the at least one navigation object, wherein the capture is or was preferably performed by at least one camera of the vehicle or robot,- Processing the at least one scene representation by a machine learning model, preferably by inputting the scene representation as an input tensor for the machine learning model, preferably in the form of a CNN; - Obtaining a line representation of the at least one navigation object based on the processing performed, wherein the obtained line representation represents the respective navigation object in a discretized point representation with a set of points, wherein the line representation can include further parameters that provide an explicit representation of start and end points of the set of points and an indication of an order for connecting the points. Furthermore, the respective point can be defined by two-dimensional coordinates. The method according to the invention offers the same advantages.as they have been described in detail with reference to a training method according to the invention. Furthermore, it can be provided that the machine learning model has been trained by a training method according to the invention, preferably to output the line representation by an output tensor of the machine learning model after processing the at least one scene representation. The discretized implementation of the point representation can relate to a line being represented by several individual points. The machine learning model can advantageously be designed as a CNN. R.408353 - 7 - Advantageously, the invention can providethat at least one of the following steps for decoding the at least one navigation object is carried out based on the obtained line representation: - determining an estimated pixel position based on at least one confidence or heat map from the obtained line representation, - determining a pixel position that is more precise than the estimated pixel position based on at least one offset information from the obtained line representation, - reproducing the start and end points of the point representation based on the explicit representation of the start and end points from the obtained line representation, - iterating a complete line based on the specification of the order for connecting the points from the obtained line representation, - reproducing a three-dimensional position of the points based on depth information from the obtained line representation,- Performing a reproduction of a semantic class of the navigation object based on a class label for the respective point from the obtained line representation. The invention also relates to a machine learning model that has been trained by a training method according to the invention. The invention also relates to a computer program, in particular a computer program product, comprising instructions that, when the computer program is executed by a computer, cause the computer to execute at least one of the methods according to the invention. Thus, the computer program according to the invention provides the same advantages as those described in detail with reference to a method according to the invention. The invention also relates to a data processing device that is configured to execute at least one of the methods according to the invention. For example, a computer can be provided as the device.which executes the computer program according to the invention. The computer can have at least one processor for executing the computer program. A non-volatile data memory can also be provided, in which the computer program R.408353 - 8 - is stored and from which the computer program can be read by the processor for execution. The invention can also provide a computer-readable storage medium which has the computer program according to the invention and / or comprises instructions which, when executed by a computer, cause the computer toto carry out at least one of the methods according to the invention. The storage medium is designed, for example, as a data storage device such as a hard disk and / or a non-volatile memory and / or a memory card. The storage medium can, for example, be integrated into the computer. Furthermore, the respective (training) method according to the invention can also be designed as a computer-implemented method. Further advantages, features, and details of the invention emerge from the following description, in which exemplary embodiments of the invention are described in detail with reference to the drawings. The features mentioned in the claims and in the description can each be essential to the invention individually or in any combination. They show: Fig. 1 a schematic visualization of a method, a device,a storage medium and a computer program according to embodiments of the invention. Fig. 2 shows a schematic representation of a machine learning model according to embodiments of the invention. Fig. 3 shows a schematic representation of an output tensor of the machine learning model according to embodiments of the invention. Fig. 4 shows a schematic representation of a confidence and heat map feature map. Fig. 5 shows an example of the representation of start and end points. Fig. 6 shows a visualization of various line shapes. R.408353 - 9 - Fig. 7 shows an example decoding of the 3D line objects with their class labels. Fig. 8 shows an example of a 3D reconstruction of a line. Fig. 9 shows exemplary raw outputs of the neural network or convolutional neural network (CNN) before decoding. Fig. 1 shows a method 100, a device 10,a storage medium 15 as well as a computer program 20 and a machine learning model 50 according to embodiments of the invention are schematically shown. Fig. 1 illustrates, according to embodiments of the invention, a training method 100 for a machine learning model 50 for recognizing navigation objects 42 in order to provide the recognized navigation objects 42 for orientation during navigation of an at least partially autonomous robot 61 or vehicle 62. According to a first training step 101, at least one scene representation 40 of at least one traffic scene 41 can be provided. Furthermore, according to a second training step 102, predetermined line representations 503 of at least a selection of navigation objects 42 can be provided.wherein the navigation objects 42 are provided in the at least one traffic scene 41. According to a third training step 103, the training of the machine learning model 50 can be carried out on the basis of the at least one provided scene representation 40 and the provided predefined line representations 503 to output line representations 503 of those navigation objects 42 that are recognized in the at least one traffic scene 41. The line representations 503 can each represent the navigation objects 42 in a discretized point representation 303, 313 with a set of points 302. Furthermore, the line representations 503 can each comprise further parameters that provide an explicit representation of start and end points 501 of the set of points 302 and an indication of an order for connecting the points 302 in order to reproduce the respective navigation object 42 in the form of a line 301.wherein the respective point 302 is defined by two-dimensional coordinates x, y. R.408353 - 10 - Fig. 1 also shows a method 200 for recognizing at least one navigation object 42 of at least one traffic scene 41, in order to use the at least one recognized navigation object 42 for orientation during navigation of an at least partially autonomous robot 61 or vehicle 62. According to a first method step 201, provision of at least one scene representation 40 of the at least one traffic scene 41 is provided. Furthermore, according to a second method step 202, processing of the at least one scene representation 40 by a machine learning model 50 is possible. According to a third method step 203, a line representation 503 of the at least one navigation object 42 can be obtained based on the processing performed.whereby the respective navigation object 42 is represented in a discretized point representation 303, 313 with a set of points 302. In self-driving vehicle systems and / or in robotics, a complete 3D representation is often necessary to enable reliable and safe navigation of linear objects (road markings, curbs, posts). Embodiments of the invention enable this functionality for a camera-based system and preferably on embedded edge devices. An overview of the proposed approach according to embodiments of the invention is shown in Fig. 2. By way of example, a CNN 50 is used here, i.e., a convolutional neural network, which can have several layers of convolution operations, pooling, and activation functions. The CNN can receive an input image 40 and 3D lane markings, 3D posts, and 3D curbs (see lines 201, 202,203). It can also output the detailed semantic class of the detected object (e.g., in the case of road markings, it can output the type of road, such as solid road, dashed road, stop line, etc.). Fig. 2 schematically shows that the neural network 50 receives an input image 50 and can output the linear object detection and its position in 3D coordinates. Pole detection is illustrated by lines 201, curb detection by lines 202, and road marking detection by lines 203. The approach can be implemented in different scaling variants (e.g., only detection of road markings, only detection of poles,only detection of curbs) or in any possible combination of object detections. Further details of embodiments of the invention are described in more detail using the steps given below. First, a fully convolutional CNN can be trained according to a first step. For this purpose, each linear object can be represented as a set of points (see Fig. 3) with a binary classification mask for confidence 401 and / or heatmap 402 (see Fig. 4) and an additional offset (Δx, Δy) within the heatmap cell in order to achieve subpixel accuracy. Each point can additionally receive information about the depth and the class label. This information can be added in the channel dimension of the output tensor of the neural network. Furthermore, each point in the row can have the offsets to the previous (Δx, Δy) and the next (Δx,∆y) as well as to the previous point in the line and the next point in the line. This can be provided to connect all points in the correct order and reconstruct the complete line. Furthermore, the end point and the start point can have a connection to themselves to signal that the line begins / ends here (see Fig. 5, where the connections are represented by arrows). The described parameterization allows the lines to be represented as objects of arbitrary shape (Fig. 6). In a further step, a special decoding algorithm can be provided for the resulting output tensor in inference time to decode all lanes in 3D (Fig. 7,Fig. 8). Fig. 3 shows an example of a linear object in the CNN's output tensor. The line 301 can be represented as a set of connected points 302. Each grid cell in the image 303 corresponds to a pixel position in the output tensor. The points can be recognized as a binary classification. The cells 304 containing points 302 can be classified as 1.0, otherwise as 0.0 (remaining cells in the image 303). The output tensor is typically smaller than the original resolution, so that to restore the fine-grained point position, Δx and Δy offsets are also provided for each cell into which the line point falls, in order to restore the subpixel accuracy of the point position. R.408353 - 12 - In Figure 313 in Fig.3 the point connection is shown as an offset between the current point and Previous(∆x, ∆y) and the current point and Next(∆x,∆y). Each point 312 in the image 313 has this offset from its neighbors. This is used to obtain a complete line given the order of the connections between the points. Each point 312 can additionally receive depth information and a class label in the channel dimension of the output tensor. This is used to recover the positions of the points in 3D and their semantic meaning. Each line can be discretized as a set of points and a spatial pixel position grid used as a coordinate grid to locate them. The localization mechanism can be implemented as a binary classification problem (see Fig. 3). Each grid cell can be a pixel of the output tensor. The grid cells (pixels) containing line points can be classified as class - 1.otherwise, -0. The binary classification can be implemented as a confidence map 401 or a heatmap 402 (see Fig. 4). Fig. 4 shows two examples of the binary classification of the points in the output tensor. The confidence map is shown in representation 401. Pixels containing points are represented as 1.0 in the image, otherwise as 0. The heatmap is visible in representation 402. Each point is represented as 1.0, and a small Gaussian blur is applied near its vertex; otherwise, the point is represented as 0.0. This representation allows for increasing the amount of information for network training for positive examples and can, in some cases, be more robust than the confidence approach. In the case of the confidence map, the binary cross-entropy loss can be used as follows: It says here ^^ ^^ for ground truth labels, ^^ ^^for the CNN prediction and N for the number of samples. In the case of heatmap representation, a penalty can be applied that reduces the focal loss: R.408353 - 13 - It says here ^^ ^^ for the ground truth heatmap, ^^ ^^ for the predicted heatmap, ^^ and α are additional parameters (e.g., α=2, ^^ = 4 can be set) and N is the number of samples. The ^^ Heatmap for the binary ground truth mask ^^ ^^ can be created using a Gaussian kernel. If two Gaussian kernels of the same points overlap, the element-wise maximum can be taken. To calculate the current point with subpixel accuracy, the following can be applied: Where ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ - the position of the current point with subpixel accuracy, ^^ - the rough position of the point from the confidence / heatmap (presumably the cells in Fig. 3), ∆ ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ - offset to the cell in the x-direction (see Fig. 3), ∆ ^^^^ ^^ ^^ ^^ ^^ ^^ - offset to the cell in the y-direction (see also Fig. 3). This method can be used to calculate the next and previous points: as well as Where ^^^^ ^^ ^^ ^^, ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ - next and previous points, [∆ ^^^^ ^^ ^^ ^^∆ ^^^^ ^^ ^^] - offsets from the current point to the next point. Furthermore, [∆ ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^∆ ^^^^ ^^ ^^ ^^ ^^ ^^ ^^] are offsets between the current point and the previous point. The L1 loss can be used as a loss function for the offsets: (6) R.408353 - 14 - Here, ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^_ ^^ ^^, ^^^^ ^^ ^^ ^^_ ^^ ^^, ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^_ ^^ ^^ are the locations of the ground perception points, and the loss values ​​are divided by the number of points ^^ to normalize them. There are two ways to calculate the depth information: The depth ^^ can be represented as a disparity: (7) This parameterization enhances the importance of the nearest points. An l1 loss can be applied: Here is ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ the basic truth, ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ the CNN prediction of the disparity and ^^ ^^ ^^ ( ^^) the confidence ground truth mask. The second possibility is to directly predict the logarithmic depth as follows: ^^ ^^ ^^ - the true depth and log ( ^^ ^^ ^^ ) the CNN prediction of the log-depth This is ^^ ^^ ^^ ^^ ^^ ^^ ^^the basic truth ^^ ^^ ^^ ^^ ^^ ^^ ^^ CNN prediction of class labels. The total loss can be calculated as follows: R.408353 - 15 - A decoding algorithm can be provided during inference time for the resulting output tensor to decode all lanes in 3D (see Figs. 7 and 8). The algorithm is shown in further detail in Fig. 7. According to a first step 701, the coarse pixel position is recovered from the confidence or heatmap. In the case of confidence, a sigmoid activation can be applied to scale the logit values ​​between 0 and 1, and if the probability is greater than or equal to the defined threshold (e.g., 0.5), this cell can be considered the location of the lane point; otherwise, it cannot. In the case of a heatmap, the activations can be sigmoid-scaled, and then a 3x3 max-pooling operation can be applied, and the original heatmap can be compared with the max-pooled heatmap.Values ​​smaller than the "maxpooled" heatmap can be suppressed, so that only the peaks of the heatmap where the trace points are located are obtained. Then, for each peak, only the values ​​greater than or equal to the specified threshold (e.g., 0.5) can be determined. The result of step 1 can be a set of keypoints ( ^^1, ^^2,..., ^^ ^^) that indicate the rough location of the points of all lines. According to a second step 702, the refined position of each keypoint ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ and its neighboring nearest ^^^^ ^^ ^^ ^^ and ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ points can be obtained by adding appropriate offsets to the keypoints from the first step 701. According to a third step 703, a data association can be constructed between the current keypoint and its neighboring points. For each keypoint, the next ^^^^ ^^ ^^ ^^ and ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ points can be calculated.These calculated points can be used to find the nearest points in the ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ point set. A data association threshold can further be introduced in (∆x, ∆y). If the distance lies within the data association radius, this data association can be considered valid; otherwise, not. According to a fourth step 704, the reconstruction of the start and end points can be provided. The start and end points can be the ^^^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^ points for which one of the data association deltas to either the next or the previous point is zero (association with itself). According to a fifth step 705, for each start point, the entire lane can be iteratively reconstructed according to the association relationship, where all points associated with this start point and the following points are connected to the same line and considered as a global curve.The curve reconstruction can be terminated when either the endpoint is reached or when an incorrect assignment occurs. According to a sixth step 706, to recover the 3D position for each curve, the depth and the camera property can be used to project the 2D points of the curve into 3D space. The depth can be recovered in the case of disparity parameterization as follows: (12) and in the case of logarithmic parameterization: (13) where ^^ ^^ ^^ ^^ℎ. ^^ ^^ ^^ and ^^ ^^ ^^ ^^ℎ ^^ ^^ ^^ scalar values ​​of the expected minimum and maximum depth range in meters, and ^^ the given point index g. In the seventh step 707, to determine the semantic class, argmax can be applied to the class tensor, thus obtaining the label for each point on the curve: Here are ^^ ^^ ^^ ^^ ^^ ^^ ^^ and ^^ ^^ ^^ ^^ ^^ ^^ ^^ ^^argmax and softmax operations over channel dimensions. Figure 8 shows an example 3D reconstruction of a line 801 and the poles 802. The camera is labeled 810. Furthermore, the labels "Dashed Line" 805 and "Solid Line" 806 are indicated. R.408353 - 17 - Figure 9 shows example raw outputs 900 of the neural network / convolutional neural network (CNN) 50 before decoding. The CNN 50 can output the tensor 900 of the form [8 + number of classes, output height, output width]. For example, the first channel 901 corresponds to the confidence or heatmap, the next two 902 current (∆x, ∆y) offsets for the current point, the next two 903 offsets for the next point (∆x, ∆y), and the next two 904 offsets for the previous point (∆x, ∆y). One channel 905 can correspond to the depth, and the last channel 906 can correspond to the number of class labels for the classification.The above explanation of the embodiments describes the present invention exclusively by way of examples. Of course, individual features of the embodiments can be freely combined with one another, provided they are technically feasible, without departing from the scope of the present invention.

Claims

R.408353 - 18 - Claims 1. Training method (100) for a machine learning model (50) for recognizing navigation objects (42), in order to provide the recognized navigation objects (42) for orientation during navigation of an at least partially autonomous robot (61) or vehicle (62), comprising the following steps: - providing (101) at least one scene representation (40) of at least one traffic scene (41), - providing (102) predetermined line representations (503) of at least one selection of navigation objects (42) that are provided in the at least one traffic scene (41), - training (103) the machine learning model (50) on the basis of the at least one provided scene representation (40) and the provided predetermined line representations (503) to output line representations (503) of those navigation objects (42) that are provided in the at least one traffic scene (41) be recognized,wherein the line representations (503) each represent the navigation objects (42) in a discretized point representation (303, 313) with a set of points (302), characterized in that the line representations (503) each comprise further parameters which provide an explicit representation of start and end points (501) of the set of points (302) and an indication of an order for connecting the points (302) in order to reproduce the respective navigation object (42) in the form of a line (301), wherein the respective point (302) is defined by two-dimensional coordinates (x, y). R.408353 - 19 - 2. Training method (100) according to claim 1, characterized in that the line representations (503) further each comprise at least one of the following parameters: - a confidence and / or heat map for estimating a pixel position of the points (302), - a classification mask for the semantic classification of the respective represented navigation object (42), - an offset parameter of a cell of a heat map with offset information in order to indicate an offset to a pixel position of the points (302) of the point representation (303, 313), - depth information for the respective point (302) of the point representation (303, 313), which is specific for a distance of the represented navigation object (42) to the vehicle (62) and / or robot (61), - a class label for the respective point (302) of the point representation (303, 313) in order to semantically classify the point (302). classify,wherein the provided predefined line representations (503) are specified as ground truth and / or a parameterization of the line representations (503), in particular a definition of the parameters, for the training (103) is specified by the ground truth, wherein the at least one provided scene representation (40) is embodied as an at least or exactly two-dimensional camera image.

3. Training method (100) according to claim 1 or 2, characterized in that each of the points (302) comprises offset information for specifying the order for connecting the points (302), which offset indicates an offset to a preceding and a following point (302) of the point representation (303, 313) in order to reconstruct the line (301) representing the respective navigation object (42) by connecting the points (302). R.408353 - 20 - 4. Training method (100) according to one of the preceding claims, characterized in that the machine learning model (50) is trained end-to-end to output the recognized navigation objects (42) represented in the discretized point representation (303, 313) in the output tensor of the machine learning model (50).

5. Method (200) for recognizing at least one navigation object (42) of at least one traffic scene (41) in order to use the at least one recognized navigation object (42) for orientation during navigation of an at least partially autonomous robot (61) or vehicle (62), comprising the following steps: - Providing (201) at least one scene representation (40) of the at least one traffic scene (41), which results from camera capture and depicts the at least one navigation object (42),- performing (202) processing of the at least one scene representation (40) by a machine learning model (50), - obtaining (203) a line representation (503) of the at least one navigation object (42) based on the processing performed, by which the respective navigation object (42) is represented in a discretized point representation (303, 313) with a set of points (302), characterized in that the line representation (503) comprises further parameters which provide an explicit representation of start and end points (501) of the set of points (302) and an indication of an order for connecting the points (302), wherein the respective point (302) is defined by two-dimensional coordinates (x, y). R.408353 - 21 - 6. Method (200) according to claim 5, characterized in that the following steps for decoding the at least one navigation object (42) are carried out based on the obtained line representation (503): - determining an estimated pixel position based on at least one confidence or heat map from the obtained line representation (503), - determining a pixel position that is more precise than the estimated pixel position based on at least one offset information from the obtained line representation (503), - performing a reproduction of the start and end points (501) of the point representation (303, 313) based on the explicit representation of the start and end points (501) from the obtained line representation (503), - performing an iterative reproduction of a complete line (301) based on the specification of the order for connecting the points (302) from the obtained line representation (503),- Performing a reproduction of a three-dimensional position of the points based on depth information from the obtained line representation (503), - Performing a reproduction of a semantic class of the navigation object (42) based on a class label for the respective point from the obtained line representation (503).

7. The method (200) according to claim 5 or 6, characterized in that the machine learning model (50) was trained by a training method (100) according to one of claims 1 to 4 in order to output the line representation (503) through an output tensor of the machine learning model (50) after processing the at least one scene representation (40).

8. A computer program (20) comprising instructions which, when the computer program (20) is executed by a computer (10), cause the computer (10) to execute the training method (100) according to one of claims 1 to 4 and / or the method (200) according to one of claims 5 to 7. R.408353 - 22 - 9. A data processing device (10) configured to execute the training method (100) according to any one of claims 1 to 4 and / or the method (200) according to any one of claims 5 to 7.

10. A computer-readable storage medium (15) comprising instructions which, when executed by a computer (10), cause the computer (10) to execute the steps of the training method (100) according to any one of claims 1 to 4 and / or the method (200) according to any one of claims 5 to 7.

11. A machine learning model (50) trained by a training method (100) according to any one of claims 1 to 4.

Images