DE102021206625A1 - Computer-implemented method and system for supporting an imaging sensor installation and training method - Google Patents

Abstract

The invention relates to a computer-implemented method and system for supporting the installation of an imaging sensor (10), in particular an interior camera of a motor vehicle, with a calculation (S3) of a geometric displacement (GV) of the plurality of static-semantic scene elements (14a-h) of the first Data set (DS1) with the reference topology of the second data set (DS2) and an output (S4) of at least one piece of information (I) to support the installation of the imaging sensor (10) using the calculated geometric displacement (GV) of the plurality of static-semantic scene elements (14a-h). The invention also relates to a computer program and a computer-readable data carrier.

Description

The invention relates to a computer-implemented method for supporting the installation of an imaging sensor.

The invention further relates to a system for supporting an installation of an imaging sensor.

In addition, the invention relates to a computer-implemented method for providing a trained machine learning algorithm to support an installation of an imaging sensor.

More and more interior cameras are currently being used in motor vehicles. Likewise, stationary cameras are used to monitor public space, e.g. at traffic intersections, on or in buildings, etc., which are intended to capture a given scene with defined static scene elements. In this context, the orientation and positioning of the camera sensor or camera are important to ensure optimal capture of the scene.

US10820307B2 describes systems and processes for automatic camera installation guidance. A 3D point cloud data processing algorithm comprises several overarching stages. A 3D depth camera captures 3D image data of a shipping container. A container property evaluation application determines a representative container point cloud and converts the 3D imagery to 2D depth image data, compares the 2D depth image data to one or more template image data, performs segmentation to extract 3D point cloud features, determines exterior features of the shipping container and scores the exterior features using an exterior feature metric, determines interior features of the shipping container and scores the interior features using an interior feature metric, and generates an orientation adjustment instruction to indicate to an operator, the 3D depth camera in a second direction for use during a shipping container loading session, wherein the second direction is different than the first direction.

When the camera sensor is used in a motor vehicle, for example when people are being transported, it serves to provide various services for the driver and passengers, such as detecting assaults or safety-critical events by the driver and/or passengers. Likewise, the camera sensor can be used to detect tiredness on the part of the driver. The camera alignment must be done by the owner of the vehicle himself. In the simplest case, the position/installation can be simulated using instructions. Known methods can essentially be divided into two categories.

There are image feature-based methods that are based directly on image features and content. These can be previously placed markers or templates that need to be recognized. But this also includes recognizing and analyzing the image content. The features identified in this way can be compared with a reference model in order to issue appropriate instructions.

Machine learning methods can also be used in parallel. For this purpose, a data set with corresponding scenes of a camera is recorded or synthetically generated. A machine learning process is trained to calculate the current position of the camera or to issue the necessary instructions directly.

There are also position-based methods. In this context, position sensors or acceleration sensors can be used to calculate the position of the camera. This also includes geometric methods such as SLAM, which calculate a position of the camera in space.

One of the main problems is the variety of cars. Each type of car requires a different positioning of the camera, depending on the installation options, but also on the arrangement and size of the interior. Another problem is that a user usually does not know what the camera sees or what is relevant for the system. There is also the risk that the driver will want to hide certain areas of his interior from the camera.

It is therefore an object of the invention to provide an improved method and system for supporting installation of an imaging sensor.

The object is achieved with a computer-implemented method for supporting the installation of an imaging sensor having the features of patent claim 1.

The object is also achieved with a method for providing a trained algorithm by machine to support the installation of an imaging sensor with the features of patent claim 12 .

The object is also achieved with a system for supporting the installation of an imaging sensor having the features of patent claim 14 .

Furthermore, the object is achieved with a computer program having the features of patent claim 15 and a computer-readable data carrier having the features of patent claim 16.

Disclosure of Invention

The present invention creates a computer-implemented method for supporting the installation of an imaging sensor, in particular an interior camera of a motor vehicle.

The method includes receiving image data of an individual image recorded by the imaging sensor or a sequence of individual images by a machine learning algorithm, which classifies a scene contained in a respective individual image into a plurality of static-semantic scene elements and a first data set representing a plurality of determined classes spends

The method also includes comparing the first data set with a second data set representing a reference topology of the scene, in particular an occurrence of scene elements and/or their arrangement in the scene.

The method also includes calculating a geometric displacement of the plurality of static-semantic scene elements of the first data set using the reference topology of the second data set and outputting at least one piece of information to support the installation of the imaging sensor using the calculated geometric displacement of the plurality of static-semantic scene elements .

The present invention also provides a computer-implemented method for providing a trained machine learning algorithm to support the installation of an imaging sensor, in particular an interior camera of a motor vehicle.

The method includes receiving a first training data set of image data of an individual image recorded by an imaging sensor or a sequence of individual images.

The method also includes receiving a second training data set of classified image data, with a scene contained in a respective individual image being classified into a plurality of static-semantic scene elements.

The method also includes training the machine learning algorithm using an optimization algorithm, which calculates an extreme value of a loss function for the classification of static-semantic scene elements contained in the image data.

The present invention also creates a system for supporting the installation of an imaging sensor, in particular an interior camera of a motor vehicle.

The system includes means for receiving image data using a machine learning algorithm, which classifies a scene contained in a respective frame into a plurality of static-semantic scene elements and outputs a first data set representing the plurality of determined classes.

Furthermore, the system comprises means for comparing the first data set with a second data set representing a reference topology of the scene, in particular an occurrence of scene elements and/or their arrangement in the scene, and means for calculating a geometric displacement of the plurality of static-semantic scene elements of the first data set with the reference topology of the second data set.

The system further comprises means for outputting at least one item of information to support the installation of the imaging sensor using the calculated geometric displacement of the plurality of static-semantic scene elements.

The camera component contains the physical structure of the camera including the mounting inside a vehicle. The alignment and/or the position should be flexibly adjustable.

The Image component provides an image to be evaluated for visible content. The image can be supplied by a camera in live operation, extracted from a video or represent a single image. Color and grayscale images are applicable.

The semantic segmentation is a machine learning method that detects semantic-static scene elements. A neural convolution network can be used for this purpose. For example, the architecture can be a UNet, be VGG or ResNet. A data set with images and ground truth labels of the semantic segmentation is created and the network is trained to classify the semantic scene elements pixel by pixel. The network can be trained to minimize cross-entropy loss.

A semantic model is a further level of abstraction in segmentation and is used to make content assessable and comparable. In this step, segmentation groups of a class are combined into instances and mapped to individual nodes. These nodes contain information about the central coordinates in the image, size (number of associated pixels) and may only occur once per element. Other elements that cannot be evaluated or can only be evaluated with difficulty are filtered out, such as the background, which due to its diversity does not allow any further comparison. New semantic information is also generated here, such as assigning instances to a page in the vehicle. The segmentation has difficulties in distinguishing whether, for example, a seat is located on the left or right in a vehicle, since the imaging features cannot be distinguished and can only be evaluated from the overall context.

In the context of the present invention, a machine learning algorithm is trained to perform the function of the semantic segmentation as well as the semantic model in a single step. Alternatively, however, this can be carried out in two separate steps, for example as explained above.

The semantic reference model describes the expected abstraction from the segmentation of the static scene when the camera is optimally aligned.

This can be created manually or calculated using an optimally aligned camera.

In this step, the current semantic model is compared with the reference. The nodes that belong to each other are evaluated individually and as an overall construct. The comparison of individual nodes provides information about the existence and comparability (size, position and neighborhood) of the points. The more nodes, the more accurate the overall rating. For example, an affine transformation can be estimated via the respective positions of all points associated with one another.

Other forms of representation such as quaternions are also conceivable. This gives information about how the current model is shifted compared to the reference. All degrees of freedom such as translation, rotation and scaling are taken into account.

More points stabilize the statement, for example, with only one point, only the statement about the translation is possible. It is also possible to use the calculated translation to determine the error between the semantic reference model and the semantic model that is being shifted. This can be, for example, a mean square deviation and provides information about how well the two models fit together.

Instructions for the user can be generated from the previously calculated transformation. This can be in the form of text, audio or images. Examples audio/text: "Move the camera to the left" or "Rotate the camera 20 degrees to the right". Visually, it is conceivable to place arrows or crosses on the current image in order to guide the user.

One idea of the present invention is therefore to enable a user to adjust a position of the camera sensor using the generated instruction.

Once the camera sensor has been realigned by the user using the function, the process starts again, i.e. the realigned camera sensor is again compared to the reference model to generate another function to adjust the position of the camera sensor.

The process is repeated until the camera sensor is correctly aligned. Furthermore, instead of a single image, for example, a sequence of single images can be used and evaluated, with the best single image from the sequence then being used. This can improve the robustness of the system, since unfavorable influences such as a low-lying sun dazzling the camera sensor and/or any covering of the sensor by objects that can appear in a single image are no longer significant.

The information to support the installation of the imaging sensor using the calculated geometric displacement of the plurality of static semantic scene elements thus minimizes a deviation between the first data set and the second data set.

The invention combines machine learning methods for semantic segmentation with model-based knowledge for detecting the camera position and describing the correction steps derived therefrom. The invention is designed for use in a semantic-static scene environment. Semantic-static means the fact that essential semantic scene elements are present in the same position in the camera image and do not change over time. This fact is present, for example, in a vehicle interior with semantic elements such as seats, footwell, doors, windows, steering wheel, and so on.

A semantic segmentation is trained with the help of machine learning methods to detect the semantic-static elements of the scene.

While changing light conditions and shadows change the image content and features significantly, this has no impact on the semantic meaning. For example, a seat will continue to be classified as a seat regardless of the presence of light spots or shadows in the image. Machine learning methods, in particular neural networks, have shown that they can achieve a high degree of robustness in the face of varying lighting conditions. For the reasons mentioned, the invention achieves a high level of robustness, in particular with regard to varying lighting conditions and objects or people in the camera image.

Another essential distinguishing feature is that, in contrast to a method that determines the position of the camera in space, here the visible area of the camera to be interpreted is evaluated. This has the advantage that a more optimal working range of downstream algorithms can be found.

The invention also shows an advantage of interpretability compared to existing machine learning methods, which provide "black box" detections without additional reasoning or interpretation. The comparison of a topological model with a reference model as proposed in this invention follows a model-based/probabilistic approach. This allows a statement to be made as to why a camera adjustment must be made and enables a probabilistic interpretation of the probability of a camera position being correctly detected.

The comparison between reference model and semantic segmentation also integrates the model knowledge that the semantics in this particular scene (e.g. vehicle interior for specific vehicle model) does not change over time. An End2End machine learning method, on the other hand, would implicitly have to map a scene model of all possible scenes (e.g. all possible vehicle interiors of all vehicle types). Further knowledge of the scene can be integrated, such as that a vehicle interior must contain two seats, two doors and two windows depending on the perspective and vehicle type.

The machine learning method is trained to detect the static-semantic scene elements in different camera positions. Training data with camera positions Adjustment notes are not required, especially since these are laborious to label. This reduces the effort involved in generating the data set.

The invention only uses features that are present in the vehicle (semantic scene elements such as seats, doors, steering wheel, etc.) for determining the camera position and determining the installation instructions. It is therefore not necessary for trained persons to attach additional markings. In addition, installation instructions are prepared in a form (e.g. "move camera to the right") that can be implemented by untrained personnel. As a result, a vehicle owner can take care of the subsequent installation himself.

Advantageous embodiments and developments result from the dependent claims and from the description with reference to the figures.

According to a preferred development, it is provided that the second data set representing the reference topology of the scene includes a predetermined plurality of classified static-semantic scene elements contained in a single image of a sequence of single images recorded by an imaging sensor. The static-semantic scene elements of the second data set are thus comparable with the static-semantic elements of the first data set.

According to a further preferred development, it is provided that a respective scene element of the first data set of the plurality of static-semantic scene elements is assigned to a node, with the respective node containing information about central coordinates, in particular a number of associated pixels, of the assigned static-semantic scene element and/or has local neighborhoods to other nodes.

This makes it possible, among other things, to filter illogical situations in the event of faulty segmentation and to gain knowledge about the static semantic use scene. For example, it is expected that a left and right door may not be directly adjacent, but other elements such as seats must be in between.

According to a further preferred development, it is provided that an affine transformation is carried out between nodes of the first data set and nodes of the second data set.

In this step, the current semantic model is compared with the reference. The nodes that belong to each other are evaluated individually and as an overall construct. The comparison of individual nodes provides information about the existence and comparability of the points. The more nodes, the more accurate the overall rating. An affine transformation, for example, can be used to estimate the respective positions of all the points that are matched to one another.

According to a further preferred development, it is provided that the geometric shift has a translation, a rotation and/or a scaling of the plurality of static-semantic scene elements of the first data set with respect to the reference topology of the second data set. An optimal positioning and/or alignment of the camera sensor can thus be achieved in an advantageous manner.

According to a further preferred development, it is provided that the information to support the installation of the imaging sensor includes an instruction to change a spatial position of the imaging sensor with one to six degrees of freedom. Depending on the positioning requirement, the appropriate information is thus output to the user, including the corresponding degrees of freedom.

According to a further preferred development, it is provided that the at least one piece of information to support the installation of the imaging sensor is output to an app on a mobile device and/or an online service, in particular a cloud-based one.

The camera installation help can thus also be calculated on an external device such as in a cloud. For this purpose, images are uploaded to the cloud, which are to be evaluated with regard to a camera installation. If the installation is carried out via an app on the mobile device, there is the advantage that the user can carry out the installation efficiently and in a short time if the user is e.g. at the installation site and can therefore implement the instructions immediately.

Installation of the imaging sensor is complete when an actual position of the imaging sensor is within a predetermined range of a target position of the imaging sensor. In this way, the tolerance range within which the installation is considered to be complete can be specified in an advantageous manner.

According to a further preferred development, it is provided that the machine learning algorithm assigns unknown objects and/or people to a predefined class, in particular a background class. If these detectors detect an object or person, it can be assumed that this area cannot be evaluated or can only be evaluated to a limited extent.

It is also possible to generate further information in the event of a large model error or if only a few nodes are detected when there are many people in the vehicle or at the installation site at the same time. For example, if there is more than one person, an instruction can be “Please let passengers get out for the most stable installation routine possible”.

This is to prevent hallucinations of various non-visible objects. This can stabilize the entire system in difficult or unknown situations so that no false statements are made because the background is not evaluated.

According to a further preferred development, it is provided that the image data received by the imaging sensor of a recorded individual image or a sequence of individual images is two-dimensional, with the machine learning algorithm using semantic segmentation of image positions of the static-semantic scene elements and/or depth information of the image data using estimated values approximated.

In the current variant, a 2D planar plane is assumed. This is sufficient for the range of action of an interior camera. However, the displacements in space usually behave differently than assumed here and an error occurs. This error can be minimized with an additional 3D estimation.

According to a further preferred development, it is provided that the static-semantic scene elements have a vehicle seat, a footwell, a vehicle door, a vehicle window and/or a steering wheel of a motor vehicle senior A plurality of different scene elements of the vehicle can thus advantageously be detected.

The configurations and developments described can be combined with one another as desired.

Further possible configurations, developments and implementations of the invention also include combinations of features of the invention described above or below with regard to the exemplary embodiments that are not explicitly mentioned.

character list

The accompanying drawings are provided to provide a further understanding of embodiments of the invention. They illustrate embodiments and, together with the description, serve to explain principles and concepts of the invention.

Other embodiments and many of the foregoing advantages will become apparent by reference to the drawings. The illustrated elements of the drawings are not necessarily shown to scale with respect to one another.

• 1 ein Ablaufdiagramm eines computerimplementierten Verfahrens zur Unterstützung einer Installation eines bildgebenden Sensors gemäß einer bevorzugten Ausführungsform der Erfindung;
• 2 ein Ablaufdiagramm des computerimplementierten Verfahrens zur Unterstützung der Installation des bildgebenden Sensors gemäß der bevorzugten Ausführungsform der Erfindung;
• 3 ein Ablaufdiagramm computerimplementierten Verfahrens zum Bereitstellen eines trainierten Algorithmus maschinellen Lernens zur Unterstützung der Installation des bildgebenden Sensors gemäß der bevorzugten Ausführungsform der Erfindung; und
• 4 eine schematische Darstellung eines Systems zur Unterstützung einer Installation eines bildgebenden Sensors gemäß der bevorzugten Ausführungsform der Erfindung.

Show it:

• 1 a flowchart of a computer-implemented method for supporting an installation of an imaging sensor according to a preferred embodiment of the invention;
• 2 FIG. 14 is a flowchart of the computer-implemented method for supporting the installation of the imaging sensor according to the preferred embodiment of the invention;
• 3 a flow chart of a computer-implemented method for providing a trained machine learning algorithm to support the installation of the imaging sensor according to the preferred embodiment of the invention; and
• 4 12 is a schematic representation of a system for supporting an installation of an imaging sensor according to the preferred embodiment of the invention.

In the figures of the drawings, the same reference symbols designate the same or functionally identical elements, parts or components, unless otherwise stated.

This in 1 The method shown is used to support the installation of an imaging sensor 10, in particular an interior camera of a motor vehicle.

The method includes receiving S1 image data BD of an individual image 12 recorded by the imaging sensor 10 or a sequence of individual images 12 by a machine learning algorithm A1, which divides a scene contained in a respective individual image 12 into a plurality of static-semantic scene elements 14a-h classified and outputs a first data record DS1 representing a plurality of determined classes.

The method also includes comparing S2 the first data set DS1 with a second data set DS2 representing a reference topology of the scene, in particular an occurrence of scene elements 14a-h and/or their arrangement in the scene.

In addition, the method includes calculating S3 a geometric shift GV of the plurality of static-semantic scene elements 14a-h of the first data set DS1 with the reference topology of the second data set DS2 and outputting S4 at least one piece of information I to support the installation of the imaging sensor 10 using the calculated geometric displacement GV of the plurality of static semantic scene elements 14a-h.

The second data set DS2 representing the reference topology of the scene comprises a predetermined plurality of classified static-semantic scene elements 14a-h scene contained in a single image 12 of a sequence of single images 12 recorded by an imaging sensor 10 .

A respective scene element 14a-h of the plurality of static-semantic scene elements 14a-h of the first data set is assigned to a node 18a-h. The respective node 18a-h has information I about central coordinates, in particular a number of associated pixels, of the associated static-semantic scene element 14a-h and/or about location-related adjacencies to other nodes 18a-h.

An affine transformation is carried out between nodes 18a-h of the first data record DS1 and nodes 18a-h of the second data record DS2.

The at least one piece of information I to support the installation of the imaging sensor 10 is preferably sent to an app 20 of a mobile device advised. Alternatively, the piece of information I can be output to an online service 22, in particular a cloud-based one.

The output by the app 20, a separate device and/or the online service 22 can, for example, take the form of visual feedback (e.g. display on the device or LEDs) or via a loudspeaker (e.g. instructions such as "Please turn the device down") .

Installation of the imaging sensor 10 is completed when an actual position P1 of the imaging sensor 10 is within a predetermined range of a target position P2 of the imaging sensor 10 . The machine learning algorithm A1 is also configured to assign unknown objects and/or persons to a predefined class, in particular a background class.

The image data BD of a recorded individual image 12 or a sequence of individual images 12 received by the imaging sensor 10 are two-dimensional. The machine learning algorithm A1 is also capable of approximating image positions of the static-semantic scene elements 14a-h via semantic segmentation and/or depth information of the image data BD via estimated values.

Alternatively or additionally, an installation position of the imaging sensor 10 can be approximated via a semantic segmentation.

2 FIG. 12 shows a flow chart of the computer-implemented method to support the installation of the imaging sensor according to the preferred embodiment of the invention.

The geometric displacement GV can have a translation, a rotation and/or a scaling of the plurality of static-semantic scene elements 14a-h of the first data set DS1 for the reference topology of the second data set DS2. In the present representation, there is a rotation of the plurality of static-semantic scene elements 14a-h of the first data set DS1 in relation to the reference topology of the second data set DS2.

The information I supporting the installation of the imaging sensor 10 includes an instruction for changing a spatial position P of the imaging sensor 10 with one to six degrees of freedom.

The static-semantic scene elements 14a-h include a vehicle seat, a footwell, a vehicle door, a vehicle window and a steering wheel of a motor vehicle.

3 12 shows a flow chart of a computer-implemented method for providing a trained machine learning algorithm to support the installation of the imaging sensor according to the preferred embodiment of the invention.

The method includes receiving a first training data set of image data BD of an individual image 12 or a sequence of individual images 12 recorded by an imaging sensor 10.

The method also includes receiving S2' a second training data set TD2 of classified image data BD, a scene contained in a respective individual image 12 being classified into a plurality of static-semantic scene elements 14a-h.

The method also includes training S3` of the machine learning algorithm A1 using an optimization algorithm which calculates an extreme value of a loss function for the classification of static-semantic scene elements 14a-h contained in the image data BD.

If a restricted view and/or an unknown object is detected by the imaging sensor 10, a loss in the optimization algorithm A2 is adjusted in such a way that the machine learning algorithm A1 outputs a predefined class or a plurality of predefined classes.

4 FIG. 12 shows a schematic representation of a system for supporting an installation of an imaging sensor according to the preferred embodiment of the invention.

The system 1 for supporting an installation of an imaging sensor 10, in particular an interior camera of a motor vehicle, comprises means 24 for receiving image data BD using a machine learning algorithm A1, which divides a scene contained in a respective individual image 12 into a plurality of static-semantic scene elements 14a -h classified and outputs a first data record DS1 representing the plurality of determined classes.

The system also includes means 26 for comparing the first data set DS1 with a reference topology of the scene, in particular an occurrence of scene elements 14a-h and/or their arrangement in the scene, representing the second data set DS2.

In addition, the system comprises means 28 for calculating a geometric displacement GV of the plurality of static-semantic scene elements 14a-h of the first data set DS1 with the reference topology of the second data set DS2 and means 30 for outputting at least one piece of information I to support the installation of the imaging sensor 10 using the calculated geometric displacement GV of the plurality of static semantic scene elements 14a-h.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• US 10820307 B2 [0005]

Claims (16)

Computer-implemented method for supporting the installation of an imaging sensor (10), in particular an interior camera of a motor vehicle, with the steps: Receiving (S1) of image data (BD) of an individual image (12) recorded by the imaging sensor (10) or of a sequence of individual images (12) by a machine learning algorithm (A1), which classifies a scene contained in a respective individual image (12) into a plurality of static-semantic scene elements (14a-h) and outputs a first data set (DS1) representing a plurality of determined classes; Comparing (S2) the first data set (DS1) with a second data set (DS2) representing a reference topology of the scene, in particular an occurrence of scene elements (14a-h) and/or their arrangement in the scene; calculating (S3) a geometric displacement (GV) of the plurality of static-semantic scene elements (14a-h) of the first dataset (DS1) with the reference topology of the second dataset (DS2); and Outputting (S4) at least one piece of information (I) to support the installation of the imaging sensor (10) using the calculated geometric displacement (GV) of the plurality of static-semantic scene elements (14a-h). Computer-implemented method claim 1 , wherein the second data set (DS2) representing the reference topology of the scene contains a predetermined plurality of scene, classified static-semantic scene elements (14a- h) includes. Computer-implemented method claim 1 or 2 , wherein a respective scene element (14a-h) of the plurality of static-semantic scene elements (14a-h) of the first data set (DS1) is assigned to a node (18a-h), the respective node (18a-h) containing information (I) via central coordinates, in particular a number of associated pixels, of the associated static-semantic scene element (14a-h) and/or via locally-related adjacencies to other nodes (18a-h). Computer-implemented method claim 3 , An affine transformation being carried out between nodes (18a-h) of the first data set (DS1) and nodes (18a-h) of the second data set (DS2). Computer-implemented method according to one of the preceding claims, wherein the geometric shift (GV) is a translation, a rotation and/or a scaling of the plurality of static-semantic scene elements (14a-h) of the first data set (DS1) to the reference topology of the second data set (DS2) having. Computer-implemented method according to one of the preceding claims, wherein the information (I) to support the installation of the imaging sensor (10) comprises an instruction for changing a spatial position (P) of the imaging sensor (10) with one to six degrees of freedom. Computer-implemented method according to one of the preceding claims, wherein the at least one piece of information (I) to support the installation of the imaging sensor (10) to an app (20) of a mobile device and / or an, in particular cloud-based, online service (22) is issued. Computer-implemented method according to one of the preceding claims, wherein the installation of the imaging sensor (10) is complete when an actual position (P1) of the imaging sensor (10) is within a predetermined range of a target position (P2) of the imaging sensor (10 ) located. Computer-implemented method according to one of the preceding claims, wherein the machine learning algorithm (A1) assigns unknown objects and/or persons to a predetermined class, in particular a background class. Computer-implemented method according to one of the preceding claims, wherein the image data (BD) received by the imaging sensor (10) of a recorded individual image (12) or a sequence of individual images (12) are two-dimensional, the machine learning algorithm (A1) using a semantic Segmentation approximates image positions of the static-semantic scene elements (14a-h) and/or depth information of the image data (BD) via estimated values. Computer-implemented method according to one of the preceding claims, wherein the static-semantic scene elements (14a-h) have a vehicle seat, a footwell, a vehicle door, a vehicle window and/or a steering wheel of a motor vehicle. Computer-implemented method for providing a trained algorithm maschi nellen learning (A1) to support an installation of an imaging sensor (10), in particular an interior camera of a motor vehicle, with the steps: receiving a first training data set of image data (BD) of an imaging sensor (10) recorded individual image (12) or a sequence of frames (12); receiving (S2`) a second training data set (TD2) of classified image data (BD), a scene contained in a respective frame (12) being classified into a plurality of static-semantic scene elements (14a-h); and training (S3`) of the machine learning algorithm (A1) by an optimization algorithm which calculates an extreme value of a loss function for the classification of static-semantic scene elements (14a-h) contained in the image data (BD). Computer-implemented method claim 12 , wherein upon detection of a restricted view and/or an unknown object by the imaging sensor (10), a loss of the optimization algorithm (A2) is adapted in such a way that the machine learning algorithm (A1) outputs a predefined class or a plurality of predefined classes. System (1) for supporting the installation of an imaging sensor (10), in particular an interior camera of a motor vehicle, comprising: Means (24) for receiving image data (BD) by a machine learning algorithm (A1) which classifies a scene contained in a respective frame (12) into a plurality of static-semantic scene elements (14a-h) and a plurality of determined classes outputs the representative first data set (DS1); means (26) for comparing the first data set (DS1) with a second data set (DS2) representing a reference topology of the scene, in particular an occurrence of scene elements (14a-h) and/or their arrangement in the scene; means (28) for calculating a geometric displacement (GV) of the plurality of static-semantic scene elements (14a-h) of the first data set (DS1) with the reference topology of the second data set (DS2); and Means (30) for outputting at least one piece of information (I) to support the installation of the imaging sensor (10) using the calculated geometric displacement (GV) of the plurality of static-semantic scene elements (14a-h). Computer program with program code to implement the method according to one of Claims 1 until 11 to be performed when the computer program is run on a computer. Computer-readable data carrier with program code of a computer program for the method according to one of Claims 1 until 11 to be performed when the computer program is run on a computer

Images