Method and device for preprocessing an input signal

Abstract

The invention relates to a computer-implemented method (100) for preprocessing an input signal for signal evaluation of the input signal, comprising the steps of: providing (110) the input signal; determining (120) one or more preprocessing modules (210-230) of a plurality of preprocessing modules (200) for preprocessing the input signal on the basis of the signal evaluation following the preprocessing, the plurality of preprocessing modules (200) comprising a data sub-range selector (210) for selecting a data sub-range of the input signal, an encoding unit (220) for transforming a data format of the input signal or of the data sub-range and / or a standard preprocessing module (230) for applying standardized preprocessing; preprocessing (130) the input signal by means of the one or more determined preprocessing modules (210-230) in order to provide a preprocessed input signal to the signal evaluation.

Description

[0001] Method and apparatus for preprocessing a single signal

[0002] The invention relates to a computer-implemented method and a device for preprocessing an input signal for signal evaluation of the input signal, in particular for preprocessing a radar input signal.

[0003] With the development of high-resolution sensors such as radar sensors, increasingly complex and extensive data sets are generated, particularly in four dimensions. Classical radar signal processing is reaching its limits when it comes to utilizing the full potential of the raw data.

[0004] Radar sensors are capable of precise 3D surface detection along with velocity estimation. More recently, sensors are being built with an increasing number of transmitters and receivers, which can generate more data points and thus also output the third spatial dimension (height), which is generated by calculating an additional angle.

[0005] This dramatically increases the amount of data to be processed. For example, tasks such as angular resolution and determining the reflection location become increasingly difficult. Due to the growing complexity of determining valid points, machine learning (ML) methods are now being used to generate better predictions from the radar input or raw radar signal. Classical signal processing methods are reaching their limits here, while artificial intelligence (AI) can recognize information, relationships, and patterns in data and thus provide solutions for the processing tasks.

[0006] At the same time, classical signal processing uses radar-specific knowledge in many steps. Real-world contexts can also be relevant to perception.

[0007] This global knowledge is only partially reflected in the radar data during signal processing, meaning that AI can only recognize and model such relationships within this portion. Furthermore, recognizing the knowledge contained in the data requires a vast amount of data, which can only be recorded or simulated at extremely high cost. Therefore, AI solutions in radar signal processing have room for improvement in terms of both quality and efficiency.

[0008] It is an object of the present invention to provide a computer-implemented method and a device that at least improve upon one or more of the aforementioned disadvantages. In particular, it is an object of the present invention to provide an efficient and precise reduction and / or preprocessing of input signals, especially raw input signals, in order to improve their subsequent evaluation or use.

[0009] According to one aspect, the task is solved by a computer-implemented method for preprocessing an input signal for subsequent signal evaluation. This method can also, or alternatively, be designed to compress the input signal data for greater efficiency in the subsequent signal evaluation.

[0010] The procedure includes the following steps:

[0011] - Providing the input signal;

[0012] - Determining one or more preprocessing modules from a plurality of preprocessing modules for preprocessing the input signal based on the signal evaluation following the preprocessing, wherein the plurality of preprocessing modules includes a data subrange selector for selecting a data subrange of the input signal, an encoding unit for transforming a data format of the input signal or the data subrange and / or a standard preprocessing module for applying a standardized preprocessing;

[0013] - Preprocessing the input signal using one or more specific preprocessing modules to provide a preprocessed input signal to the signal evaluation. Since the amount of data from sensors is constantly increasing, efficient and precise processing is required. If a standardized procedure with the same steps were always applied, the total processing time would increase proportionally to the amount of data. Particularly in vehicle applications, not all parts of the data set are needed; however, such standardized procedures do not differentiate between the individual data points within the data set. The inventors have recognized that, especially depending on the subsequent signal evaluation using different AI models, the input signal can be preprocessed in such a way that it can be preprocessed according to the requirements of this subsequent signal evaluation.This results in a reduction of the data volume, for example, by identifying only specific regions of interest, data with a measured value above a predetermined threshold, or similar elements within the input signal and forwarding them to the signal processing unit. This leads to a reduction in the overall data volume, enabling faster transmission and signal processing. Furthermore, the signal processing can be improved so that, for example, a cognitive model receives only the data relevant to that model. This, in turn, improves the signal-to-noise ratio.

[0014] Signal evaluation can include signal utilization and / or be signal processing.

[0015] Providing the input signal may involve receiving the input signal, in particular from a sensor and / or a transmitter-receiver.

[0016] The input signal can include physically measured and / or simulated measurements. For example, in the case of a radar sensor, the input signal can characterize a sampled environment.

[0017] The numerous preprocessing modules can be stored in advance, particularly in storage. The method can involve accessing these numerous preprocessing modules. For example, the method can be executed with decentralized preprocessing modules, such as those deployed in a cloud.

[0018] One or more specific preprocessing modules can be executed once and / or multiple times for preprocessing. Accordingly, the preprocessing can be varied and adapted to the subsequent signal evaluation.

[0019] The method can further include determining the sequence of several specific preprocessing modules for preprocessing the input signal based on the signal evaluation. The sequence can vary depending on the signal evaluation and / or be at least partially identical and / or at least partially different.

[0020] The data subrange selector can be configured to select a data subrange of the input signal or a pre-processed input signal based on intensity, pattern recognition, world knowledge, and / or signal evaluation. The intensity can be or encompass a data density (sparsity).

[0021] Using the data subrange selector, data subranges of the input signal, particularly the input signal in the form of a data cube, can be selected according to intensity using pattern recognition and / or world knowledge. This world knowledge can be specific to automated driving and driver assistance systems. Furthermore, world knowledge can be used, for example, to consider hardware requirements or restrictions, especially those of the radar sensor or other sensors, knowledge of special cases, and / or the weighting of scenarios and data. Selection using the data subrange selector can occur during inference, i.e., when using the sensor, or during the training of AI models. Particularly during training, this selection can lead to robust predictions by an underlying neural network and / or AI model.When selecting data subsets during inference, the neural network and / or the AI ​​model can support or replace signal processing in unclear data situations.

[0022] The selected data subset can be fed into a deep neural network and / or AI model. For radar signals, this can include an output for data points or objects within the corresponding spatial area. The neural network and / or AI model can be applied to multiple areas of interest. The neural network and / or AI model can be supported by classical signal processing, which is applied to areas that are clearly defined for it.

[0023] The encoding unit can be configured to transform the input signal, or a pre-processed input signal, into a data format usable by a further pre-processing module and / or signal evaluation. Latency and data density of the pre-processed input signal can negatively impact the efficiency of a computer algorithm or neural network, although data density can be increased by using multiple pre-processing modules. The further pre-processing module can be one that processes the pre-processed input signal after the encoding unit.

[0024] The encoding unit can include using or performing positional encoding for transformation. With positional encoding, data can be selected from the input signal, particularly the data cube, based on a heuristic. A threshold can be used for this purpose, for example, for the intensity of the data. Alternatively or additionally, a pattern recognition algorithm and / or world knowledge about a data origin, such as environmental and parameter properties of the sensor and its design, can be used. The encoding unit can add a positional vector to the selected data, which specifies the location within the input signal, particularly the data cube. The data format achieved using the encoding unit is well-suited, among other things, for training transformer architectures.The encoding unit can transform the input signal by selecting points or data using a heuristic or a threshold value for the intensity of the input signal, particularly the data cube. The points or data can be discovered using a peak-finding algorithm and stored as points (instead of as a range of values ​​within a tensor). This significantly reduces the data size and changes the data type. The points found in this way contain both coordinates of actual targets ("true positives") and false positives. They serve as candidates for the sensor's detection targets, especially the radar. A neural network, particularly a PointNet and / or an AI model, can be used to train on these points and output the validity of the point candidates and / or an object classification of the point candidates.

[0025] Additionally or alternatively, the encoding unit can include a reduction of the input signal by creating columns or so-called "pillars" for transformation. For this, see Lang, Alex & Vora, Sourabh & Caesar, Holger & Zhou, Lubing & Yang, Jiong & Beijbom, Oscar. (2019). PointPillars: Fast Encoders for Object Detection From Point Clouds. 12689-12697. 10.1109 / CVPR.2019.01298. Here, a data tensor is sliced ​​into 1-dimensional vectors, which are then transformed into pillars, i.e., decorated singular data. In the case of raw radar data, these are latent vectors from the raw tensor, while in LiDAR object detection, they are vectors containing parts of a point cloud. This increases the data density. For example, the dimension of the virtual channels or that of the Doppler values ​​in the input signal, especially in the data cube, can be compressed into a pillar.The data then has two additional dimensions, longitudinal and lateral, plus the latent information within the respective pillars. An artificial neural network (e.g., a CNN) can then be trained on this data.

[0026] The previously described aspects of the encoding unit were described for the input signal, but these aspects are also applicable to the preprocessed input signal. The standard processing module can be configured for filtering, noise reduction, frequency analysis, amplitude and / or phase correction, and / or detection of predetermined characteristics of the input signal or a preprocessed input signal. The respective preprocessed input signal can be the input signal preprocessed by one or more of the specified preprocessing modules. The standard processing module can be configured to execute one or more preprocessing steps, which are, in particular, independent of the signal evaluation. Thus, these can be classic preprocessing steps that are executed independently of the subsequent course and / or further use of the data.

[0027] The numerous preprocessing modules can further include an additional unit for clustering and / or tracking. The clustering unit can be configured to group multiple data points of the input signal or a preprocessed input signal into a single object. Alternatively or additionally, the tracking unit can be configured to provide temporal information about data points, data groups, and / or the objects within the input signal or a preprocessed input signal, for example, in the case of tracking, the evolution of an object over time (i.e., the past). The respective preprocessed input signal can be the input signal that has been preprocessed by one or more of the specified preprocessing modules.

[0028] The input signal can include and / or consist of raw data. The input signal can be a simulated input signal of a simulation and / or a physically measured, in particular a real or actual, input signal.

[0029] Signal processing can involve evaluating and / or using the preprocessed input signal with a single AI model or a multitude of AI models, where the multiple AI models are trained for at least partially different applications of the input signal. The selection of one or more preprocessing modules can be based on one or more of the applications. The individual AI model can be trained for a specific application, and the input signal can be preprocessed accordingly. The input signal can be a radar input signal. The applications can be radar applications. The radar applications can include driving through a tunnel, which involves many strong reflections of the radar signal. For example, driving through a tunnel is relevant for a model that is supposed to detect mirror effects. Driving in a city involves different, more pronounced reflections compared to driving through a tunnel.Accordingly, various AI models can be provided for evaluating the radar signal, each specialized or trained for specific applications or areas of use. Through appropriate preprocessing of the input signal, the input signal can be made available to these AI models.

[0030] The signal evaluation can involve the AI ​​model or a multitude of AI models. The method can further include training the multitude of AI models based on the preprocessed input signal, whereby the input signal is preprocessed for one or more AI models and the preprocessed input signal is made available for training to the AI ​​model(s) for which the input signal was preprocessed. The AI ​​models can thus be specifically trained for at least partially different applications.

[0031] The procedure can further include performing signal evaluation based on the pre-processed input data signal.

[0032] The method may include providing the preprocessed input signal and / or the output of the AI ​​model(s) to a controller, in particular a vehicle. The controller may be configured to control the vehicle based on the output of the AI ​​model(s).

[0033] According to a second aspect, the task is solved by a computer-implemented method for training a computer intelligence model based on an input signal. The method comprises the following steps:

[0034] - Providing the input signal;

[0035] - Preprocessing the input signal and / or providing a preprocessed input signal using the method according to the first aspect;

[0036] - Creating training data and test data based on the pre-processed input signal, wherein the training data and the test data are at least partially, and in particular completely, different;

[0037] - Selecting a training data subset from the training data and a test data subset from the test data based on an application of the AI ​​model;

[0038] - Training the AI ​​model based on the training data segment;

[0039] - Evaluating the AI ​​model based on the test data excerpt.

[0040] The procedure described in the second aspect enables improved and more efficient training of the AI ​​model.

[0041] The training data and the test data can be configured so that they have no overlap. For example, a radar sensor can record a first and a second drive, providing a single input signal that characterizes both drives. The training data could be the data from the first drive, and the test data could be the data from the second drive. The training data and the test data can be the same type of data and / or belong to the same domain.

[0042] The training data segment and the test data segment can be selected using a data selector module.

[0043] The selection of the training and test data segments can be further based on world knowledge regarding the application of the AI ​​model. The quality of data, especially radar data, is highly dependent on and sensitive to the environment. The number, size, material, and positioning of objects affect the perceived radar signals. This applies to both simulated, artificial data and data from real-world recordings. For example, driving in an open, flat landscape with few surrounding objects offers less potential for error than an urban scenario or driving through a tunnel, where many strong reflections can occur.

[0044] Depending on the application or task of the AI ​​model, relevant data can be selected for training. For example, tunnel driving is relevant for an AI model that is supposed to detect mirror effects. The data selector can be used for this purpose, selecting training and / or test data based on world knowledge, in this case expert and subject-matter knowledge.

[0045] The method can further include determining and / or providing ground truth data in the preprocessed input signal, with the training and / or test data containing at least some of the ground truth data. Determining the ground truth data can, in particular, be based on world knowledge. Depending on the application or task of the AI ​​model, the focus can be on specific aspects of the data. Accordingly, specific labeling can be performed by experts. For example, if the AI ​​model is intended for resolving mirror effects, experts should distinguish between false positive results (ghost objects) and true positive results among the targets originating from a specific object. Optimizing angular accuracy in the case of field data ideally requires 3D bounding boxes or segmentation labels.

[0046] In the realm of radar raw data, very large volumes of data can be generated in a short time. These datasets often consist largely of small values ​​or exhibit low data density (sparsity). To efficiently train an AI module capable of processing radar raw data in real time, data selection is necessary. This can be achieved through an intelligent compression method (lossy or lossless), which typically reduces one or more dimensions of the input signal, particularly the radar data cube. A data cube is a multidimensional data vector containing information such as distance, angle, and relative velocity. Alternatively, a heuristic can be used to select elements of the input signal, especially the data cube, according to radar-specific criteria for use as input for the AI ​​model.

[0047] Furthermore, general knowledge, such as engineering expertise, can be incorporated into the design of the data format for the AI ​​module within the specific context of signal processing, particularly radar signal processing. For example, the resolution and discrimination capability of radar detections can be specifically tailored to a neural network of the AI ​​model through precise hardware and software parameterization, thereby adjusting its sensitivity to the setup. This general knowledge can also include expert knowledge of antenna design for signal processing, which can be further integrated into the AI ​​model design. This is particularly relevant when considering the discretization of angle and FFT data to make it kernel-readable and ensure a precise matching of the virtual antennas with the AI ​​input and output.

[0048] The task is solved according to a third aspect by a computer program product comprising instructions that cause a hardware component of a device to execute the procedure according to the first aspect and / or the second aspect when the computer program is loaded onto or executed by the hardware component. The computer program product may additionally or alternatively include instructions that, when the program is executed by the hardware component, cause it to execute the procedure according to the first aspect and / or the second aspect. The hardware component may be a processor.

[0049] According to a fourth aspect, the problem is solved by a device for preprocessing an input signal for signal evaluation, comprising a hardware component, in particular a processor, and a memory. The computer program product, as described in the third aspect, is stored in the memory, and the hardware component is configured to execute the computer program product. The device may further include a sensor, in particular a radar sensor, for scanning an environment and outputting the input signal based on the scanned environment.

[0050] The task is solved according to a fifth aspect by a vehicle comprising the device according to the fourth aspect.

[0051] Features described in relation to the procedure according to the first aspect can be described as features of the second to fifth aspects and vice versa.

[0052] Preferred embodiments are explained by way of example with reference to the accompanying figures. These show:

[0053] Fig. 1 shows a schematic representation of a computer-implemented method for preprocessing an input signal;

[0054] Fig. 2 shows a schematic representation of the preprocessing of the input signal using a multitude of preprocessing modules;

[0055] Fig. 3 shows a schematic representation of a computer-implemented method for training a computer model;

[0056] Fig. 4 shows a schematic representation of a device for preprocessing an input signal; and

[0057] Fig. 5 shows a schematic representation of a vehicle with such a device.

[0058] In the figures, identical or essentially functionally equivalent or similar elements are designated with the same reference symbols.

[0059] Fig. 1 shows a computer-implemented method 100 for preprocessing an input signal for signal evaluation. The method 100 can be stored as a computer program on memory and executed by means of a hardware component, in particular a processor. The input signal here is a radar input signal provided by a radar sensor, in particular a vehicle 600. For this purpose, the radar sensor can scan the area around the vehicle 600 and provide the corresponding radar input signal.

[0060] Method 100 comprises providing 110 the radar input signal, for example, by means of the radar sensor. Method 100 further comprises determining 120 one or more preprocessing modules 210-240 or a plurality of preprocessing modules 200 for preprocessing the radar input signal based on the signal evaluation following the preprocessing. Accordingly, several, in particular different, preprocessing modules 210-240 are provided, each of which processes or preprocesses the input signal or the input into the respective processing module 210-240 in one or more predetermined ways. Preprocessing is necessary for many types of signals in order to improve signal evaluation.

[0061] The radar input signal is therefore preprocessed by one or more of the preprocessing modules 210-240. The sequence and selection of the preprocessing modules 210-240 for preprocessing the radar input signal depends on how the subsequent signal evaluation is performed. This provides flexible and demand-based preprocessing of the radar input signal. Since certain data from the radar input signal are not needed for every signal evaluation, or may even slow it down or degrade it, the large number of preprocessing modules 200 allows the radar input signal to be appropriately preprocessed, resulting in improved signal evaluation.

[0062] The multitude of preprocessing modules 200 comprises a data subrange selector 210 for selecting a data subrange of the input signal and an encoding unit 220 for transforming a data format of the input signal or the data subrange. Furthermore, the multitude of preprocessing modules 200 includes a standard preprocessing module 230 for applying standardized preprocessing, as well as an additional unit 240 for clustering to combine multiple data points of the input signal or a preprocessed input signal into an object. Consequently, for example, a specific data subrange of the input signal can be selected using the data subrange selector 210 and made available for further signal evaluation and / or preprocessing. When evaluating objects in the radar input signal, the data that can be assigned to an object can therefore be selected.

[0063] The method 100 further comprises preprocessing 130 of the radar input signal by means of one or more specific preprocessing modules 210-240 to provide a preprocessed radar input signal to the signal evaluation.

[0064] As schematically shown in Fig. 2, the input signal is provided as input to the multiple preprocessing modules 200. Based on the further course of the signal evaluation, one or more of the preprocessing modules 210-240 are used to preprocess the input signal. One or more of the preprocessing modules 210-240 can be used multiple times for preprocessing.

[0065] The pre-processed input signal is provided as output in Fig. 2 to a plurality of trained AI models 300. The plurality of AI models 300 comprises a first AI model 310, a second AI model 320, and a third AI model 330. Each of the AI ​​models 310-330 has been trained for a different application. For example, the first AI model 310 can be trained to evaluate reflections of the radar signal in a tunnel, the second AI model 320 to evaluate the radar signal in an urban environment, and the third AI model to evaluate predefined object classes.

[0066] For each of these applications, at least a different data range is of interest. To enable efficient evaluation of the input signal, the input signal can be preprocessed using the numerous preprocessing modules 200 to suit one or more of the KL models 310-330. Consequently, multiple preprocessed radar input signals can also be provided based on a single radar input signal using the numerous preprocessing modules 200. The output of the KL model(s) 310-330 can then be used further, for example, for controlling the vehicle 600.

[0067] Fig. 3 shows a schematic representation of a computer-implemented method 400 for training a Kl model, such as one of the Kl models 310-330. The method 400 can be stored in memory as a computer program product and be executable by means of a hardware component, in particular a processor.

[0068] Method 400 comprises providing 410 of the input signal. Furthermore, method 400 comprises preprocessing 420 of the input signal and / or providing a preprocessed input signal using method 100. Consequently, providing 410 can be omitted if the preprocessed input signal is already provided. The inventors have recognized that preprocessing the input signal can improve the training of a computer model.

[0069] Procedure 400 further includes the creation of training data and test data based on the pre-processed input signal. The training data and the test data are at least partially, and in particular completely, different.

[0070] Furthermore, the procedure 400 includes the selection 440 of a training data subset from the training data and a test data subset from the test data based on an application of the AI ​​model. The application can be, as described above for radar signals, the evaluation of reflections in a tunnel. Accordingly, data relevant for evaluating reflections are selected from the training and test data.

[0071] The method 400 further comprises training 450 of the AI ​​model based on the training data segment and evaluating 460 of the AI ​​model based on the test data segment. Evaluation 460 allows the performance of the AI ​​model to be checked and corrected accordingly. Fig. 4 shows a schematic representation of a device 500 for preprocessing an input signal with a hardware component, in particular a processor 510 and a memory 520. At least one of the aforementioned computer program products is stored on the memory 520, and the processor 510 is configured to execute the computer program product.

[0072] Fig. 5 shows a schematic representation of a vehicle 600 with such a device 500. By preprocessing the input signal, in particular the radar input signal, data processing of the input signal can be improved.

[0073] Reference mark

[0074] Method for preprocessing an input signal

[0075] Providing the input signal

[0076] Determine one or more preprocessing modules

[0077] Preprocessing the input signal

[0078] Numerous preprocessing modules

[0079] Data subrange selector

[0080] Encoding unit

[0081] Standard preprocessing module

[0082] Additional unit

[0083] Variety of AI models, first AI model, second AI model, third AI model

[0084] Methods for training an AI model

[0085] Providing the input signal

[0086] Preprocessing the input signal and / or providing a preprocessed input signal

[0087] Creating training and test data

[0088] Selection of a training data segment from the training data and a test data segment

[0089] Training the AI ​​model

[0090] Evaluating the AI ​​model

[0091] Device for preprocessing an input signal

[0092] Hardware component

[0093] memory

[0094] vehicle

Claims

Patent claims 1. Computer-implemented method (100) for preprocessing an input signal for signal evaluation of the input signal, comprising the steps: Providing (1 10) the input signal; Identifying (120) one or more preprocessing modules (210-230) of a plurality of preprocessing modules (200) for preprocessing the input signal based on the signal evaluation following the preprocessing, wherein the plurality of preprocessing modules (200) comprises a data subrange selector (210) for selecting a data subrange of the input signal, an encoding unit (220) for transforming a data format of the input signal or the data subrange and / or a standard preprocessing module (230) for applying a standardized preprocessing; Preprocessing (130) of the input signal by means of one or more specific preprocessing modules (210-230) to provide a preprocessed input signal to the signal evaluation.

2. Method (100) according to claim 1, wherein the one or more specified preprocessing modules (210-230) are executed once and / or multiple times for preprocessing.

3. Method (100) according to claim 1 or 2, further comprising: Determining a sequence of the several specified preprocessing modules (210-230) for preprocessing the input signal based on the signal evaluation.

4. Method (100) according to any one of the preceding claims, wherein the data subrange selector (210) is configured for selecting a data subrange of the input signal or a pre-processed input signal based on intensity, pattern recognition, world knowledge and / or signal evaluation, wherein the encoding unit (220) is configured for transforming the input signal or a pre-processed input signal into a data subrange intended for further processing. The preprocessing module and / or the signal evaluation is configured in a data format usable by the signal processing module, wherein the standard processing module (230) is configured for filtering, noise suppression, frequency analysis, amplitude and / or phase correction and / or detection of predetermined characteristics of the input signal or a preprocessed input signal, wherein the respective preprocessed input signal is the input signal preprocessed by means of one or more of the specified preprocessing modules.

5. Method (100) according to one of the preceding claims, wherein the plurality of preprocessing modules (200) further comprises an additional unit (240) for clustering and / or tracking, wherein the additional unit (240) for clustering is configured to combine several data points of the input signal or a preprocessed input signal into an object, and / or wherein the additional unit (240) for tracking is configured to provide temporal information of data points, data groups and / or the objects in the input signal or in a preprocessed input signal, wherein the respective preprocessed input signal is the input signal preprocessed by means of one or more of the specified preprocessing modules.

6. Method (100) according to one of the preceding claims, wherein the signal evaluation comprises evaluating and / or using the preprocessed input signal by means of a KL model or a plurality of KL models (300), wherein the plurality of KL models (300) are trained for at least partially different applications of the input signal, wherein the determination (120) of one or more preprocessing modules is based on one or more of the applications.

7. Method (100) according to any of the preceding claims, wherein the input signal is a radar input signal, and / or wherein the applications are radar applications.

8. Method (100) according to any of the preceding claims, wherein the signal evaluation comprises the Kl model (310-330) or the plurality of Kl models (300), the method (100) further comprising: Training the AI ​​model (310-330) or the plurality of AI models (440) based on the preprocessed input signal, wherein the input signal is preprocessed for the AI ​​model(s) (310-330) and the preprocessed input signal is provided to the AI ​​model(s) (310-330) for training for which the input signal has been preprocessed.

9. Computer-implemented method (400) for training a AI model based on an input signal, comprising the steps: Providing (410) the input signal; Preprocessing (420) the input signal and / or providing a preprocessed input signal by means of the method (100) according to any one of claims 1 to 8; Creating (430) training data and test data based on the pre-processed input signal, wherein the training data and the test data are at least partially, and in particular completely, different; Selections (440) of a training data subset from the training data and a test data subset from the test data based on an application of the Kl model; Training (450) the KL model based on the training data extract; Evaluating (460) the Kl model based on the test data sample.

10. Method (400) according to claim 9, wherein the selection (440) of the training data portion and the test data portion is further based on world knowledge regarding the application of the AI ​​model.

11. Method (400) according to claim 9 or 10, further comprising: Determining ground truth data in the preprocessed input signal, where the training data and / or the test data at least partially include the ground truth data, the determination of ground truth data is still based in particular on world knowledge.

12. Computer program product comprising instructions that cause a hardware component of a device to execute the method (100) according to any one of claims 1 to 8 and / or the method (200) according to any one of claims 9 to 11 when the computer program is loaded onto or executed by the hardware component.

13. Device (500) for preprocessing an input signal for signal evaluation of the input signal, comprising a hardware component, in particular a processor (510), and a memory (520), wherein the computer program product according to claim 12 is stored on the memory (520) and the hardware component (510) is configured for executing the computer program product.

14. Device (500) according to claim 13, further comprising: a sensor, in particular a radar sensor for scanning an environment and for outputting the input signal based on the scanned environment.

15. Vehicle (600) comprising the device (500) according to claim 13 or 14.

Images