Method for autonomously or semi-autonomously controlling a motor vehicle in consideration of external objects, motor vehicle and method for training a neural network

Abstract

The invention relates to a method for autonomously or semi-autonomously controlling a vehicle. By applying at least one neural network, edge end points (7), edge mid points (5) and local tangent slopes (9) or curve curvatures (14) are determined. The edge end points (7) mark predetermined end regions of an object, the edge mid points (5) mark the linear geometry of an object, by means of local information about tangent slopes and / or curve curvatures, a regression calculation can be carried out in order to determine the overall object structure as precisely as possible, and in turn an optimized longitudinal control and / or lateral control of the motor vehicle can be carried out by means of a control device.

Description

Technical Field

[0001] This invention relates to a method for autonomously or semi-autonomously controlling a motor vehicle in consideration of an external object (referring to an external object), comprising the following steps: Image data is provided by at least one sensor device, which describes at least one image of the vehicle environment and environmental objects. The sensor device is preferably part of the vehicle. The image data is input into at least one neural network trained to determine feature data relating to (or including) predetermined features (representations) regarding the basic geometric shape (reference shape) and / or color of the objects. This feature data is determined in a first stage, and in a second stage, the neural network determines one or more object structures based on the feature data. The object structures are input to (or provided to) the vehicle's control unit, and based on the provided object structures, longitudinal and / or lateral control of the vehicle is performed according to control commands from the control unit. Background Technology

[0002] The object recognition used here constitutes a branch of image processing that focuses on identifying individual objects in an image. In the context of object recognition in driver assistance systems, the term "object" refers to, for example, traffic islands and / or lane markings and / or lane lines and / or other vehicles. However, traditional methods, such as those using bounding boxes or image recognition, have limitations because they can only identify objects with a specific, fixed shape as a whole. Therefore, providing the correct association and / or identification and / or order of pixels or groups of pixels, especially when identifying continuous lines, constitutes a challenging task.

[0003] To ensure the advantageous construction of such object recognition systems, it is desirable to be able to determine objects of arbitrary shapes, or object structures, or object centerlines. Here, "object structure" refers to the composition of the visual appearance of objects consisting of basic shapes (e.g., line segments) and / or representations of such basic shapes or forms, such as polygonal shapes.

[0004] Object recognition systems specifically designed for lane identification typically only consider highway scenarios, where lane markings and road edges can be described by straight lines. However, such systems fail in urban scenarios, where objects such as road edges and / or traffic islands may have arbitrary geometries, such as the polyline shape described above. A polyline shape or polyline refers to a sequence of open or closed lines and / or arc segments and / or straight line segments, which together no longer constitute a single straight line.

[0005] The publication titled “A Keypoint-based Global Association Network for Lane Detection,” authored by Jinsheng Wang et al., was published on April 15, 2022, and is available at https: / / arxiv.org / abs / 2204.07335 in November 2023. It discloses the use of regression methods in lane detection.

[0006] The publication titled “Focus on Local: Detecting Lane Marker from Bottom Up via Key Point” by Zhan Qu et al. (published on May 28, 2021, and available at https: / / arxiv.org / abs / 2105.13680 in November 2023) discloses a two-stage method for using local information in neural network pre-identification: first, coarsely identifying intermediate points, and then acquiring more precise locations of these points. Summary of the Invention

[0007] The objective of this invention is to enable better identification of the object structure based on image data, thereby allowing for more reliable control of autonomous or semi-autonomous vehicles.

[0008] This task is accomplished by a method having the features of claim 1, a motor vehicle having the features of claim 6, and a method for training a neural network for the motor vehicle according to claim 9 or 10. Advantageous improvements of the invention are described by the dependent claims, the following description, and the accompanying drawings.

[0009] The method for autonomously or semi-autonomously controlling a vehicle while taking external objects into account, as described in this invention, is based on the general method described at the beginning. The determination of the object structure by the at least one neural network includes: a) Divide the at least one image into image units (preferably all image units of the same size, and particularly preferably provided in a checkerboard pattern. For example, an image unit may consist of the same number of image rows and columns, that is, four rows and four columns of image points can be combined into one image unit, i.e., 4 x 4 = 16 image points. Other sizes are also possible. The image unit may also be the same as a pixel, i.e., an image point.); b) By applying the edge endpoint determiner of the at least one neural network to the feature data, image units with edge endpoints having an object structure are determined; c) By applying the edge midpoint determiner of the at least one neural network to the feature data, image units with edge midpoints of object structures are determined; Determining the object structure further includes: d) For each image cell with an edge midpoint, determine the tangent slope of the object structure within the relevant image cell by applying a tangent slope determiner to the feature data, and / or e) By applying a curve curvature determiner to the feature data, the curve curvature of the object structure within these relevant image units is determined; and f) Using the regression unit of a neural network, regression calculations are performed based on the determined tangent slope and / or curve curvature to determine the overall object structure.

[0010] Regarding the feature data, it should be noted that it can be determined by a so-called backbone network, wherein the backbone network is designed as or included in the at least one artificial neural network. The backbone network may, for example, include at least one residual neural network and / or densely connected convolutional network pre-trained on predetermined features regarding the basic geometric shape and / or color of an object. The at least one neural network may be additionally or alternatively trained on data describing the shape and / or structure of polygons. The feature data may be included or provided, for example, in the form of feature vectors. In other words, the at least one neural network can be configured to receive image data or visual data from at least one sensor device and determine and / or extract feature data therefrom. The at least one neural network may include a convolutional neural network (CNN) and / or a graph neural network (GNN). Additionally or alternatively, the at least one neural network may be designed as an encoder-decoder network. In summary, the at least one neural network can perform semantic segmentation.

[0011] A so-called feature map can be generated from this at least one neural network. This at least one neural network may include a large number of convolutional layers, thereby the feature map includes a receptive field (e.g., a capture 3D model). 3 to 200 With 200 image points or pixels, this receptive field can capture the entirety of objects such as traffic islands and / or road edges and / or lane markings.

[0012] In this invention, the at least one neural network post-connection / cascade has at least three, preferably four, different image processing functions, which can at least be grouped independently and process features in the feature map based on image points.

[0013] The at least one neural network, as a first image processing function, includes a so-called edge endpoint determiner, configured to determine and / or mark the edge endpoints or ends of an object's structure. Here, edge endpoints mark predetermined object end regions (specifically line ends) of the object. The feature data is subjected to binary classification by the edge endpoint determiner, where the binary classification may include: normalizing the feature vector of the feature data, and the normalized value, i.e., the classification value, exceeding a predetermined threshold represents the presence or presentation of an edge endpoint of the object structure. Edge endpoints can be identified by recognizing in the image data that an image region is part of a line type (e.g., a curb), and that from that edge endpoint, the remaining portion of the line extends in only one direction (rather than in two different directions). For example, an edge endpoint can therefore be identified and / or set at an ending lane line and / or lane marking. In other words, each image point in the image can be examined to determine whether and / or whether an edge endpoint is present.

[0014] An edge endpoint determiner can be used to generate a so-called seed point map, which includes or indicates the marked edge endpoints, or so-called seed points. For example, if image data depicts white lane arrows on a gray road background, then the two ends of the lane arrows (the tip and the opposite end) can be identified as edge endpoints. For this purpose, the edge endpoint determiner can be trained to recognize shapes that are between 20×20 cm and 75×75 cm in size and terminate or cease to extend in a certain spatial direction. Thus, the object terminates in this direction, and an edge endpoint can be set there. Therefore, one-dimensional points (edge ​​endpoints or seed points) are determined by means of so-called seed point detection. Thus, start and / or end points that can serve as edge endpoints can be learned.

[0015] The at least one neural network, as a second image processing function, includes an edge midpoint determiner, configured to determine and / or mark edge midpoints and / or keypoints and / or connection points of the object's object structure based on feature data. The edge midpoints mark the linear geometry / line-like geometry of the object, where the linear geometry is part of the object's object structure. The linear geometry may consist of straight lines and / or arcs and / or a mixture of both and / or polygonal faces.

[0016] The feature data is classified into a different binary category than the first one by the edge midpoint determiner. This second binary classification may include: normalizing the feature vector of the feature data. The normalized value, i.e. the classification value, exceeds a predetermined threshold, which indicates the existence of a linear geometry of the object structure.

[0017] Therefore, it is preferable to perform binary classification on the feature vectors separately, where the feature vectors can include parameterizable (numerical) attributes of the object's geometry and / or pattern / symbol in a vector manner. Various different features of this pattern can constitute different dimensions of the feature vector. Thus, subsequent binary classification can be simplified based on feature vectors because they greatly reduce the number of attributes to be classified (e.g., instead of considering the entire image, only a feature vector of N numbers needs to be considered, where N is less than the number of image points).

[0018] In the example of the lane arrows described above, for instance, each segment or line segment of the boundary line formed by the transition between the white lane arrow and the gray road background can be individually identified as an edge midpoint. Therefore, the edge midpoint determiner can be trained to identify segments of length between 20 cm and 75 cm that belong to one of the environments of continuous lines extending in two directions as edge midpoints.

[0019] Additionally or alternatively, it may be specified that non-maximum suppression (NMS) is used to identify and / or depict only edge endpoints and / or edge midpoints that have intensity or luminance values ​​exceeding a predetermined threshold.

[0020] As part of the method described according to the invention, as a further image processing function, the tangent slope of the object structure within the relevant image unit is determined by applying a tangent slope determiner to the feature data, and / or the curve curvature of the object structure within the relevant image unit is determined by applying a curve curvature determiner to the feature data, followed by regression. The tangent slope, also referred to as a "gradient," is a numerical value relating to a local property of the object structure. If the object structure represents, for example, a lane marking (or an instance of the object structure represents a lane marking), the tangent slope indicates whether the lane marking is straight or in which direction the straight line extends, or whether it might present a shape different from a straight trajectory. The tangent slope / gradient represents the first derivative of the object structure, abstractly speaking, the derivative of the object's geometric representation. The curve curvature represents its second derivative and is typically determined, for example, by manually applying an inscribed circle to the curve structure. The regression calculation may include applying the regression criteria or regression network of the at least one neural network to the determined data (tangent slope and / or curve curvature) output by the tangent slope determiner and / or curve curvature determiner. A curve curvature determiner can be a part of an artificial neural network trained to determine the curvature of a curve. Its training can be carried out using so-called labeled data, which includes lines with different curvatures and provides the true or actual curvature value (“Ground Truth”) generated by that curvature.

[0021] From the four image processing functions described above, we can determine whether edge endpoints and / or edge midpoints exist in an image unit, and especially, if edge midpoints exist, what tangent slope and / or curve curvature the object structure possesses there. The first two image processing functions (edge ​​endpoint determination and edge midpoint determination) can be performed independently, thus independently yielding the aforementioned information or estimates. Determining the tangent slope and determining the curve curvature can also be performed independently, thus independently yielding tangent slope information and curve curvature information.

[0022] According to an advantageous embodiment of the invention, multiple object structures may be identified during the determination process (based on at least one image). In this case, performing possible duplicate checks (“non-maximum suppression” methods, a method for suppressing non-maximums) is helpful. As mentioned above, advantageously, the object structure or instances of the object structure represent lane markings.

[0023] The motor vehicle according to the invention includes a control device for realizing autonomous or semi-autonomous control of the motor vehicle (this is achieved by the control device driving corresponding actuators for performing longitudinal and / or lateral control, i.e., actuators for the engine, brakes, steering mechanism, etc., with control commands). The motor vehicle includes a sensor device that provides or transmits digital image data about objects in the motor vehicle environment. The motor vehicle also includes at least one neural network having a backbone unit designed to determine feature data from the digital image data relating to predetermined features about the basic geometric shape and / or color of the objects. The neural network further includes: - Edge endpoint determiner, which is used to determine the image units of the edge endpoints of the object structure of the object; - An edge midpoint determiner, used to determine the midpoint of an image unit with an object structure; and - Tangent slope determiner, which is used to determine the tangent slope of an object structure within an image unit having an edge midpoint; and / or - Curve curvature determiner, which is used to determine the curve curvature of an object structure within an image unit having edge midpoints.

[0024] In addition, the neural network includes a regression unit for performing regression calculations based on the determined tangent slope and / or curve curvature to determine the overall object structure. The neural network is designed to transmit the identified object structure (or data describing it) (possibly internally) to the control device, which is then designed to output control commands to the actuator according to its configuration.

[0025] The motor vehicle is used internally to perform the methods described in this invention. An alternative could be to outsource units such as neural networks to external devices (“edge computing,” which outsources computational tasks to local computing devices as the vehicle passes by).

[0026] Advantageously, the control device possesses so-called computer vision capabilities. Therefore, these capabilities are integrated into the control device. In particular, the neural network can be part of the control device, resulting in high efficiency due to its compactness.

[0027] According to an advantageous implementation, the tangent slope determiner and the curve curvature determiner are designed to operate independently of each other.

[0028] The method according to the invention for training a neural network (and / or a neural network used in the method) for a motor vehicle of the type according to the invention comprises, in order to train a tangent slope determiner, adding digital information about local tangent slopes to an image of the lane, and inputting said digital information (i.e., the so-called "ground truth," basic facts, or basic information) into the tangent slope determiner. Thus, the tangent slope determiner can be a part of an artificial neural network trained to determine tangent slopes. Its training can be performed using so-called labeled data, which, on the one hand, includes lines with different curvatures, and respectively gives the true or actual tangent slope ("Ground Truth") generated by that curvature.

[0029] The method for training a neural network for a motor vehicle according to the present invention, or a method for training a neural network used in the method, alternatively or additionally includes, in order to train the curve curvature determiner, having an image of the lane provided with digital information about the local curve curvature, and inputting said digital information (also as “ground truth”) into the curve curvature determiner.

[0030] One advantageous implementation specifies that the determination of the object's structure is trained via backpropagation, using training data with annotations (labels) of that object structure. For example, the label for a line or polyline can be set as follows: its line position and / or line width and / or line length and / or line direction and / or line color and / or line type are known. Additionally or alternatively, as described above, one or more points can be marked on a line, wherein the percentage of the line's arc length up to each marked point is known in advance as a label.

[0031] Therefore, the at least one neural network is trained using supervised learning. During backpropagation, the input (e.g., image data converted into an input vector) is propagated through the at least one neural network. The output produced by the at least one neural network is compared with the desired output. The difference between these two values ​​is considered the error of the neural network, which is now backpropagated from the output layer to the input layer. During this process, the weights of the neuron connections in the at least one neural network are changed according to their influence on the error. This ensures that when the input is reapplied, the output approximates the desired output. The at least one neural network can be corrected through backpropagation.

[0032] Therefore, the parameters of the at least one neural network can be optimized or improved. With such improved parameters, in the application phase, the at least one neural network is suitable for determining a meaningful output vector (output) from an input vector (input) that is different from the initially learned input vector of the training cases.

[0033] For application cases or scenarios that may occur in this method but are not explicitly described herein, it may be specified that an error message and / or a prompt requiring user feedback be output according to this method, and / or default settings and / or a predetermined initial state be set.

[0034] The present invention also includes a control device for a motor vehicle. The control device may have a data processing apparatus or a processor device configured to perform an embodiment of the method according to the invention. For this purpose, the processor device may include at least one microprocessor and / or at least one microcontroller and / or at least one FPGA (Field-Programmable Gate Array) and / or at least one DSP (Digital Signal Processor). As a microprocessor, a CPU (Central Processing Unit), GPU (Graphics Processing Unit), or NPU (Neural Processing Unit) may be used, respectively. Furthermore, the processor device may have program code configured to perform an embodiment of the method described in the present invention when executed by the processor device. The program code may be stored in the data memory of the processor device. The processor device may, for example, be based on at least one circuit board and / or at least one SoC (System-on-Chip).

[0035] This invention also includes improvements to the motor vehicle described herein, which have the features described in the improvements in conjunction with the method described herein. Therefore, corresponding improvements to the motor vehicle according to the invention will not be described again here.

[0036] The motor vehicle according to the present invention is preferably designed as an automobile, especially as a passenger car or commercial vehicle, or as a bus or motorcycle.

[0037] As another solution, the invention also includes a computer-readable storage medium comprising program code that, when executed by a computer or computer cluster, causes the latter to perform an embodiment of the method described herein. The storage medium may be provided at least partially as non-volatile data storage (e.g., as flash memory and / or as an SSD - solid-state drive) and / or at least partially as volatile data storage (e.g., as RAM - random access memory). The storage medium may be arranged within the computer or computer cluster. However, the storage medium may also operate on the Internet, for example, as a so-called app store server and / or cloud server. A processor circuit having, for example, at least one microprocessor may be provided by the computer or computer cluster. The program code may be provided as binary code and / or as assembly code and / or as source code in a programming language (e.g., C) and / or as a program script (e.g., Python).

[0038] The present invention also includes combinations of features of the described embodiments. Therefore, unless the embodiments are described as mutually exclusive, the present invention also includes implementations having combinations of features of multiple described embodiments. Attached Figure Description

[0039] Embodiments of the present invention are described below. In the accompanying drawings: Figure 1 A schematic diagram illustrating the architecture of at least one artificial neural network implementing an embodiment of the present invention is shown; Figure 2 A schematic diagram illustrating the determination of the slope of the curve tangent and the subsequent filtering is shown. Figure 3 A schematic diagram of a motor vehicle according to one embodiment of the present invention is shown. Detailed Implementation

[0040] The embodiments described below are preferred embodiments of the present invention. In these embodiments, the components described are individual, mutually independent features of the present invention, and these features further improve the present invention independently. Therefore, this disclosure should also include combinations of features other than those of the illustrated embodiments. Furthermore, the described embodiments may be supplemented by other features of the present invention already described.

[0041] In the accompanying drawings, the same reference numerals refer to elements that have the same function.

[0042] Figure 1A schematic diagram of the architecture of at least one artificial neural network 100 for implementing one embodiment of the present disclosure is shown. First, image data 1 can be provided, which describes at least one image of an environment including objects located therein, and this image data is received from at least one sensor device, for example, during vehicle movement.

[0043] To this end, it can be stipulated that sensor data from at least one sensor device (e.g., from a camera system and a radar system) are pre-weighted (multiplied) using their respective sensor-specific weighting factors, and a sum or sum of squares can be calculated from the thus weighted sensor data to produce a sum value. Sensor fusion can then be performed. If the sum value is greater than a threshold, image data 1 can be generated, indicating the presence of an object within the sensor's detection range. If the sum value is less than the threshold, image data 2 can be generated, indicating the absence of any object within the detection range. In this case, it can be stipulated that a message is output indicating that no object was detected. If an object is detected, the method can continue as follows: The image data 1 can then be input into at least one neural network, specifically into the backbone network 2 of that at least one neural network, which is trained to determine feature data, wherein the feature data includes predetermined features about the basic geometric shape and / or color of the object. This backbone network 2 may have a self-attention technique 11 and / or a feature pyramid network technique 12, and / or be designed to perform such techniques. This feature data can then be represented in a feature... Figure 3 superior.

[0044] In the first image processing step, the edge endpoints 7 of the object's structure can be determined by applying at least one neural network edge endpoint determiner 6 to the feature data, wherein the edge endpoints 7 respectively mark the predetermined object end regions of the object.

[0045] In the second image processing step, edge midpoints 5 of the object structure can be determined by applying at least one neural network edge midpoint determiner 4 to the feature data, wherein the edge midpoints 5 mark the linear geometry of the object, and the linear geometry is part of the object structure.

[0046] In a further first image processing step, the tangent slope 9 of the object structure for a single image unit or path point can be determined by applying at least one neural network tangent slope determiner 8 to the feature data.

[0047] In a further second image processing step, the curve curvature 14 of the object structure for a single image unit or path point can be determined by applying at least one curve curvature determiner 13 of a neural network to the feature data.

[0048] Units 4, 6, 8 and 13 can work independently of each other; however, units 8 and 13 are preferably based on the determination of the midpoint 5 of the edge of the object structure to avoid unnecessarily determining the tangent slope and / or curve curvature for points that do not belong to the object structure at all.

[0049] In the graph fusion unit 15 of the neural network, the output data 5, 7, 9, and 14 are fused together to obtain the object point graph. Unit 16 can perform regression in particular. Instead of performing narrow computation, the neural network applies regression criteria. Unit 16 of the neural network searches for polylines and outputs them in the results, thereby generating the actual graph as the object structure. Possible duplicate checks are performed in unit 17. Because actual line markings in, for example, digital images include multiple adjacent image points, there may be cases where multiple parallel lines are captured during image processing, but only one can actually be identified. Therefore, unit 17 outputs the final object structure. Thus, the object structure is determined by means of edge endpoints 7, edge midpoints 5, and further by means of tangent slopes 9 and / or curve curvatures 14.

[0050] Figure 2 A simplified diagram illustrating filtering of the output of a neural network is shown, such as that performed in unit 16.

[0051] The left side depicts image units, which can represent a single image point or a combination of multiple image points. The actual curve K of the object structure passes through several of these image units. Although curve K does not pass through image units 20a and 20f, it does pass through image units 20b, 20c, 20d, and 20e.

[0052] The tangent slope determiner determines the local slope, which is represented, for example, as a value between 0 and 1, where 0 represents the horizontal direction and 1 represents the vertical direction, with a difference of 90° (TT).

[0053] exist Figure 2 On the left side, the tangent slope is represented as Sb for image unit 20b, Sc for image unit 20c, Sd for image unit 20d, and Se for image unit 20e.

[0054] This was simplified after filtering, for example by applying a threshold criterion: an upward arrow was assigned when the slope was greater than 0.5, and a right or left arrow was assigned when the slope was less than 0.5. Therefore, in this example, the tangent slopes Sb, Sc, Sd, and Se are the new tangent slopes S'b, S'c, S'd, and S'e generated from the filtering.

[0055] Figure 3 The illustrated motor vehicle 200 is designed to achieve semi-autonomous or autonomous driving via a control device 210. For this purpose, the control device 210 sends control commands to the actuator 216, wherein... Figure 3 The actuator 216 in the diagram may symbolically represent multiple actuators, including those related to power equipment control, brakes, steering mechanisms, etc. The image sensor 214 (shown here as an optical camera) provides the control unit 210 with data about the environment. This camera specifically acquires images, which also show objects; in this example, the object may include lane markings 300.

[0056] In this example, the control device 210 of the motor vehicle 200 includes Figure 1 The neural network 100 in the example.

[0057] The vehicle can utilize the object structure provided by the neural network to employ computer vision functions, such as identifying lane 300 as described above, and thereby performing longitudinal and / or lateral control. By providing specific local information about tangent slope and / or curve curvature, such object structures can be captured with particular precision, parsed as polylines, and thus enabling reliable semi-autonomous or autonomous control of the vehicle.

[0058] In summary, these examples illustrate how it is possible to identify identifiable objects in a given camera image through polylines, where the neural network has different units (dedicated heads) that provide purely local geometric information.

Claims

1. A method for autonomously or semi-autonomously controlling a vehicle while taking into account external objects, comprising the following steps: - Image data (1) is provided by at least one sensor device, the image data describing at least one image of the vehicle environment and objects in the environment; - The image data (1) is input into at least one neural network, which determines feature data in a first stage, wherein the feature data relates to predetermined features about the basic geometric shape and / or color of the object; - In the second stage, the object structure of the object is determined by the neural network based on the feature data; - Provide the object structure to the vehicle's control unit; - Based on the control commands from the control unit, which are based on the provided object structure, perform longitudinal and / or lateral control of the vehicle; The object structure determination performed by the at least one neural network includes: a) Divide the at least one image into image units of preferably the same size. b) By applying the edge endpoint determiner (6) of the at least one neural network to the feature data, image units with edge endpoints (7) of object structure are determined; c) By applying the edge midpoint determiner (4) of the at least one neural network to the feature data, image units with edge midpoints (5) of object structure are determined; The determination of the object structure also includes: d) For each image unit with an edge midpoint, determine the tangent slope (9) of the object structure within that relevant image unit by applying the tangent slope determiner (8) to the feature data, and / or e) By applying the curve curvature determiner (13) to the feature data, the curve curvature (14) of the object structure within these relevant image units is determined; and f) Using the regression unit of the neural network, regression calculations are performed based on the determined tangent slope and / or curve curvature to determine the overall object structure.

2. The method according to claim 1, characterized in that, The tasks defined in steps b) and c) are performed independently of each other.

3. The method according to claim 1 or 2, characterized in that, The tasks defined in steps d) and e) are performed independently of each other.

4. The method according to any one of the preceding claims, characterized in that, When determining the structure of multiple objects based on at least one image, possible duplicates are checked.

5. The method according to any one of the preceding claims, characterized in that, Lane markings are represented by an object structure.

6. A motor vehicle having a control device (210) for achieving autonomous or semi-autonomous control and a sensor device (214) providing digital image data of objects in the environment of the motor vehicle, the motor vehicle also having at least one neural network (100) having a backbone unit (2) designed to determine feature data from the digital image data relating to predetermined features concerning the geometrical basic shape and / or color of the objects. in, This neural network also has: - Edge endpoint determiner (6), which is used to determine the image unit with the edge endpoint of the object structure of the object; - Edge midpoint determiner (4), which is used to determine the image unit with the edge midpoint of the object structure; as well as - Tangent slope determiner (8), which is used to determine the tangent slope (9) of the object structure within an image unit having an edge midpoint; and / or - Curve curvature determiner (13), which is used to determine the curve curvature (14) of an object structure within an image unit having an edge midpoint; and - Regression unit (16) is used to perform regression calculations based on the determined tangent slope and / or curve curvature, wherein the neural network is designed to transmit the identified object structure to the control device (210), which is designed to trigger longitudinal and / or lateral control of the motor vehicle based on the identified object structure.

7. The motor vehicle according to claim 6, characterized in that, The control device has computer vision capabilities.

8. The motor vehicle according to claim 6 or 7, characterized in that, The tangent slope determiner (8) and the curve curvature determiner (13) are designed to work independently of each other.

9. A method for training a neural network for a motor vehicle according to any one of claims 6 to 8, wherein, In order to train the tangent slope determiner (8), digital information about the local tangent slope is added to the image of the lane, and the digital information about the local tangent slope is input into the tangent slope determiner (8).

10. A method for training a neural network for a motor vehicle according to any one of claims 6 to 8, particularly the method according to claim 9, wherein, To train the curve curvature determiner (13), digital information about the local curve curvature is added to the image of the lane, and the digital information about the local curve curvature is input into the curve curvature determiner.

Images