Method for providing classification data, method for training classification system, control device, program and data memory

Abstract

The invention relates to a method for providing classification data for classifying a target object on the basis of detection points in a received radar signal. According to the invention, the method comprises the following steps executed by the control device: receiving radar data of a current sampling period of a detection point trajectory, the radar data comprising a detection point of the detection point trajectory assigned to the target object; extracting at least one predetermined input feature from a detection point in the radar data of the current sampling period; providing at least one input feature of the current sampling period to a feature generation model of the classification system; updating the internal state of the feature generation model based on the at least one input feature of the current sampling period; providing the at least one temporal feature to a classification model of a classification system; and outputting classification data through a classification model based on the at least one time feature.

Description

Technical Field

[0001] The present invention relates to a method for providing classification data for classifying target objects based on detection points in received radar signals, a method for training a classification system, a control device, a program, and a data storage device. Background Technology

[0002] The vehicle's environmental acquisition system includes sensor devices configured to collect data about the vehicle's surroundings. Further evaluation is then performed based on the data collected by the sensor devices, for example, to classify objects within the collected data.

[0003] Environmental data collection can be performed, for example, using radar. A sensor device emits radar signals and receives the radar signals reflected at a target object (the target). The target object can be identified as a detection point within the reflected radar signals. Radar-based sensor devices are used, for example, for obstacle recognition, where a target object is examined to determine if it constitutes an obstacle to a vehicle. If a target object cannot be passed by a vehicle from above or below, it can be classified as an obstacle. Target objects that can be passed from above may be, for example, manhole covers or lowered curbs. Objects that cannot be passed from above may be, for example, boundary stones. Target objects that cannot be passed from below may be, for example, bridges of relatively low height.

[0004] Compared to lidar signals acquired by lidar-based sensor devices, radar signals have relatively low lateral and longitudinal resolution. Therefore, in some cases, it can be difficult to determine whether a target object constitutes an obstacle. Consequently, additional evaluation is required in classification methods based on acquired data.

[0005] US 10 611 371 B2 describes a system and method for predicting vehicle lane changes using a structured recurrent neural network.

[0006] US 11 726 477 B2 describes a method and system for trajectory prediction using recurrent neural networks with inertial behavior unfolding.

[0007] US 2023 0184921 A1 describes an attitude determination system based on radar point clouds.

[0008] US 2022 0198808 A1 describes a method, apparatus, computer device, and storage medium for identifying obstacles.

[0009] US 11 594 011 B2 describes a deep learning-based feature acquisition method for LiDAR localization of autonomous vehicles.

[0010] DE 10 2018 217 533 A1 describes a method for classifying radar cross-section (RCS) signals and their interferometric patterns measured by radar. In this method, the history of measured RCS values ​​of a target is approximated using a linear function, and the derived features (such as average error, slope, etc.) are used for actual classification. The aim of the classification is to categorize measured targets into two classes: "can be driven over" and "obstacles," using the derived features.

[0011] The drawback of approximating the history of detection points (e.g., RCS) of target objects acquired from radar signals using linear functions is the loss of significant information. This can lead to reduced classification performance. For example, based on detection points, it may be impossible or only possible to distinguish between ground targets that can be driven over (e.g., manhole covers, beverage cans) and low-lying obstacles that cannot be driven over (e.g., European-style pallets). In the worst case, this could lead to vehicle malfunction (false braking or unresponsiveness). Summary of the Invention

[0012] The objective of this invention is to classify target objects directly based on measured radar signals without using processed features, with a classification accuracy sufficient to distinguish target objects with very similar heights.

[0013] The objective of this invention is particularly to directly determine the classification of detection points based on detection points acquired in radar signals without using processed features (e.g., features approximated by linear functions), with a classification accuracy sufficient to distinguish target objects with very similar heights.

[0014] The aforementioned objective is achieved through the teachings of claim 1 and the enclosing independent claims. Advantageous embodiments of the invention are derived in the dependent claims.

[0015] A first aspect of the invention relates to a method for providing classification data for classifying target objects based on detection points in a received radar signal. The classification data is configured to classify target objects depicted by detection points in the received radar signal. The classification includes assigning the target objects to at least one pre-given output category.

[0016] The method includes the following method steps performed by a control device.

[0017] The first step involves receiving radar data for the current sampling period. The radar data includes detection points assigned to the target object by the radar tracker. The radar data may be based on the received radar signal for the current sampling period. In other words, the radar data relates to the characteristics of the received radar signal for the current sampling period. The radar data describes the detection points present in the radar signal, which are assigned to the target object by the radar tracker. It can be specified that, in the corresponding sampling period, radar signals are transmitted and reflected radar signals are received via a sensor device. The reflected radar signals may contain detection points attributable to reflections on the target object. The relevant detection points can be identified by the radar tracker in the corresponding reflected radar signals and described in the radar data for the corresponding sampling period, which is provided to the control device.

[0018] The next step involves extracting at least one pre-defined input feature from the detection points of the radar signal in the current sampling period. In other words, the radar data in the current sampling period is evaluated by the control device, and at least one pre-defined input feature is extracted from it.

[0019] In another step, at least one input feature from the current sampling period is provided to the feature generation model of the classification system. The feature generation model has an internal state that represents at least one input feature from a previous sampling period of the detection point trajectory. The feature generation model is configured to output at least one pre-given temporal feature based on the internal state and at least one input feature from the current sampling period. In other words, it is specified that at least one input feature from a sampling period is provided to the feature generation model. Based on at least one input feature from a previous sampling period, the feature generation model has an internal state. Therefore, the internal state represents the history of at least one input feature from a previous sampling period of the detection point trajectory.

[0020] The internal state of the feature generation model is updated based on at least one input feature from the current sampling period. Based on the internal state and at least one input feature from the current sampling period, the feature generation model outputs at least one pre-given temporal feature. The temporal feature is a feature dependent on the internal state, and therefore can describe the history of at least one pre-given input feature.

[0021] At least one temporal feature is provided to the classification model of the classification system. The classification model is configured to output classification data regarding the assignment / correspondence between a target object and at least one pre-given output category, based on the at least one temporal feature. In other words, at least one pre-given output category is pre-defined. The classification model is configured to determine classification data that describes the assignment of the target object to at least one output category. The classification data may, for example, describe whether a target object assigned to at least one temporal feature is assigned to an output category.

[0022] Another step involves outputting categorized data based on at least one time feature using a classification model. The output can, for example, be provided to a vehicle's driver assistance system.

[0023] The advantage of this invention is that history is evaluated not through an algorithm (e.g., fitting a process to at least one input feature), but through at least one temporal feature. Therefore, more aspects of the process of at least one input feature can be considered, which is typically not possible in other methods according to the prior art.

[0024] An improved embodiment of the present invention specifies that the classification data includes corresponding probability values ​​regarding the probability of assigning a target object to at least one pre-given output category. In other words, the classification data includes information about the probability of assigning a target object to at least one pre-given output category. For example, the probability values ​​can be specified as describing the probability of assigning the target object to the output category of ground targets that can be driven over from above, and the probability of assigning the target object to the output category of ground targets that cannot be driven over from above. The probability values ​​can be normalized such that the sum of the probability values ​​for each output category is 1 or 100%. The advantage of this improved embodiment is that the reliability of the assignment can be estimated using probabilities.

[0025] An improved embodiment of the present invention specifies that the feature generation model is constructed as an artificial recurrent neural network (RNN). An RNN is a type of artificial neural network capable of processing temporal or sequential data, which achieves this by storing information from the past and using that information in future computations. Unlike feedforward neural networks where data propagates only in one direction, RNNs allow neurons to communicate in two directions, even within the same layer of the network. This architecture enables the network to store information about the input sequence and, based on this, predict future sequences. Therefore, an RNN has a type of "memory" called a background vector, which is generated from previous inputs in the sequence. For each new input, the network updates the background vector by combining the new input with the previous background vector. In this way, the network can keep track of information throughout the sequence and consider this information when outputting. The feature generation model is designed to output at least one temporal feature representing its internal state based on its internal state. The internal state of the feature generation model depends on the history of at least one input feature from previously sampled sequences. That is, the feature generation model processes a sequence of at least one input feature and generates a temporal feature that summarizes or generalizes the background of the input feature sequence.

[0026] An improved embodiment of the present invention specifies that the classification model is constructed as a kinematic neural network (KNN). An KNN is a machine learning model constructed by mimicking the structure of the human brain. It consists of a pre-given number of interconnected nodes or neurons. The classification model is designed to output classification data regarding the assignment of a target object to at least one pre-given output category, based on at least one temporal feature. That is, the classification model receives one or more temporal features and creates predictions about the assignment attributes of the target object to a specific output category based on these temporal features. Using an KNN as the classification model allows for the creation of complex assignment functions that can capture the non-linear relationship between temporal features and output categories.

[0027] An improved embodiment of the present invention specifies the extraction of multiple input features from radar data of the current sampling period. It is stipulated that the input features are scaled in a feature generation model according to their respective scaling factors. In other words, the extraction involves determining multiple input features from the radar signal, such as the distance, velocity, direction, and size of a target or surface. These input features are identified and selected according to pre-given criteria and algorithms and provided to the feature generation model. The scaling can be achieved by applying scaling factors that are adjusted according to the characteristics and requirements of the feature generation model. Scaling factors may include, for example, normalization, or minimum-maximum scaling, which allows the feature values ​​of the input features to be transformed to a uniform range. By applying these scaling factors, differences in units of measurement and orders of magnitude of the input features are eliminated, and better comparability and composability of the data are ensured.

[0028] An improved embodiment of the present invention specifies that at least one input feature of the current sampling period includes the radar cross-section of the target object. The radar cross-section is a measure of the target object's ability to reflect and return radar waves. It is defined as the projected area of ​​the target object in the direction of the incident radar wave and depends on the geometric and material properties of the target object. By using the radar cross-section as an input feature in the current sampling period, more accurate and reliable acquisition of the target object can be achieved. The radar cross-section can, for example, be used to determine the target object's distance, size, and orientation.

[0029] An improved embodiment of the invention specifies that at least one input feature of the current sampling period includes the ground clearance of the detection point. The ground clearance of the detection point includes information about its spatial orientation and extent. By using height as an input feature, more accurate and reliable classification of detection points regarding drivability from above and / or below can be achieved. The ground surface can be determined by applying suitable algorithms and methods, such as digital surface modeling (DSM). The ground surface can, for example, describe the surface of the road ahead of the vehicle, which can be identified in radar data.

[0030] An improved embodiment of the present invention specifies that, according to the formula Calculate the height above the ground, where h is the height above the ground, γ is the elevation angle of the detection point, r is the radial distance of the detection point, and p is the vertical installation position of the radar device.

[0031] An improved embodiment of the present invention specifies that a separate classification unit of the classification system is used for each target object. In other words, it specifies that multiple target objects are classified in this method. Specifically, it specifies that a corresponding classification unit of the classification system is provided for each target object. Each classification unit has a corresponding feature generation model and a corresponding classification model, which are used individually for the corresponding target object.

[0032] An improved embodiment of the present invention specifies that the method includes outputting a control signal via a control device. In other words, the control device outputs a control signal based on classification data. For example, a control signal may be output when a target object is assigned to a specific output category. For instance, it may be specified that when a target object is classified as not to be driven over, a control signal is output to initiate a response, such as intervening in vehicle guidance / control or issuing a warning signal to the vehicle driver.

[0033] A second aspect of the invention relates to a method for training a classification system.

[0034] Methods for training classification systems include training at least one feature generation model, which is trained to output at least one pre-given temporal feature based on the internal state of the feature generation model and at least one input feature.

[0035] Furthermore, the method for training the classification system includes training a classification model that is trained to output classification data based on at least one temporal feature, assigning the target object to at least one pre-given output category. In other words, the described method for training the classification system comprises two main parts: training a feature generation model and training a classification model.

[0036] Training a feature generation model involves learning to output at least one pre-given temporal feature based on the model's internal state and at least one input feature. This means that the feature generation model should learn to generate an internal representation of the input features that is useful to a classification model. The feature generation model can be an RNN, trained to transform the time series of input features into an internal state that can then be used as input to a classification model.

[0037] Training the classification model involves learning to generate classification data based on at least one temporal feature, which involves assigning a target object to at least one pre-given output category. Therefore, the classification model is trained to make judgments about the output category of a target object based on the temporal features generated by the feature generation model.

[0038] In addition to training the feature generation model, the method also includes training a classification model that learns directly from temporal features to generate categorical data.

[0039] Overall, the goal of this method for training classification systems is to create models capable of assigning target objects to output categories based on input features. By training both the feature generation model and the classification model, it can be ensured that they generate appropriate internal representations of the data and are robust against different input features and output categories.

[0040] For application situations or scenarios that may occur in these methods but are not explicitly described herein, it can be stipulated that error messages and / or requests for user feedback and / or settings of default settings and / or pre-given initial states shall be output according to the corresponding methods.

[0041] A third aspect of the invention relates to a control device configured to perform a method according to a first aspect of the invention for providing classification data for classifying target objects based on received radar signals. Additionally or alternatively, the control device is configured to perform a method according to a second aspect of the invention for training a classification system.

[0042] To perform the described steps, a processor circuit may be provided, which has programming or software comprising program instructions that, when executed, cause the processor circuit to perform an embodiment of the method. For this purpose, the processor circuit may have at least one microprocessor and / or microcontroller. The program instructions may be stored in the processor circuit's data memory.

[0043] A fourth aspect of the invention relates to a program comprising program instructions that, when executed, cause processor circuitry to perform an implementation of one of these methods.

[0044] A fifth aspect of the invention relates to a data memory that includes program instructions that, when executed, cause processor circuitry to perform an implementation of one of these methods.

[0045] The present invention also includes improvements to the control device according to the invention, the computer program according to the invention, and the storage medium according to the invention, which have features as described in the section relating to improvements to the method. Therefore, the corresponding improvements to the control device according to the invention, the computer program according to the invention, and the storage medium according to the invention will not be repeated here.

[0046] The present invention also includes combinations of features of the described embodiments. Attached Figure Description

[0047] Embodiments of the present invention are described below. Therefore:

[0048] Figure 1 A schematic diagram of a vehicle with control devices is shown;

[0049] Figure 2 A schematic diagram showing the radar cross-section of different targets as a function of distance;

[0050] Figure 3 A schematic diagram of the classification system of the control device is shown;

[0051] Figure 4 A schematic diagram showing the results of the classification model;

[0052] Figure 5 A flowchart illustrating a method for providing classification data for classifying target objects based on received radar signals is shown.

[0053] Figure 6 A schematic diagram of a method for training a classification system is shown. Detailed Implementation

[0054] The embodiments described below are preferred embodiments of the present invention. In these embodiments, the components of the described embodiments are independent features of the present invention that should be considered separately, and these features also independently further constitute the present invention. Therefore, they should be considered as part of the present invention, whether individually or in combinations different from the illustrated combinations. Furthermore, the described embodiments can be supplemented by other features of the present invention already described.

[0055] In the accompanying drawings, elements with the same function are represented by the same reference numerals.

[0056] Figure 1 A schematic diagram of a vehicle with control devices is shown.

[0057] Vehicle 10 may include a radar device 12 configured to transmit radar signals 14 into the environment of vehicle 10 and receive reflected radar signals 14 during a sampling period. Radar device 12 may be configured to provide raw radar data 18 for the corresponding sampling period to a radar tracker 16. Radar tracker 16 may be configured to identify detection points 22 in the raw radar data 18 for the corresponding sampling period, which may be assigned to a target object 24. This assignment may describe the allocation of detection points 22 to a detection trajectory 40 of the target object 24. Radar tracker 16 may be configured to track the target object 24 by identifying corresponding detection points 22 in the corresponding sampling period. Radar tracker 16 may be configured to provide radar data 20 for the current sampling period to a control device 26, the radar data including detection points 22 and their allocation to the target object 24. Control device 26 is configured to receive radar data 20 and extract at least one pre-defined input feature 28 of the detection points 22 from the radar data 20 of the current sampling period. At least one pre-given input feature 28 may include, for example, the radar cross section of the detection point 22 of the target object 24, the ground clearance of the detection point 22, the model error of the elevation shaper, or other input features 28 in the feature list.

[0058] The control device 26 is configured to provide a classification system 30 for classifying the target object 24 using detection points 22. The classification system 30 has a feature generation model 32 and a classification model 36.

[0059] It has been specified that the control device 26 establishes a corresponding classification unit for the classification system 30 for each target object 24. In other words, the corresponding classification unit can be assigned to the detection point trajectory 40 of the target object 24. The control device 26 is configured to provide the input features 28 assigned to the target object 24 during the sampling period of the detection trajectory 40 to the feature generation model 32 of the classification system 30. The feature generation model 32 is configured to update its internal state based on at least one input feature 28 in the current sampling period. By continuously updating the internal state of the feature generation model 32 based on the input features 28 of the corresponding sampling period, the history of the input features 28 is represented by the internal state of the feature generation model 32.

[0060] It has been specified that the feature generation model 32 of the classification system 30 is configured to update its internal state upon receiving at least one input feature 28 of the current sampling period of the detection trajectory 40, and output at least one pre-given temporal feature 34 based on the internal state and the at least one input feature 28 of the current sampling period. The control device 26 is configured to provide at least one temporal feature 34 to the classification model 36. The classification model 36 of the classification system 30 is configured to output classification data 42 based on at least one temporal feature 34 regarding the assignment of the target object 24 to at least one pre-given output category 38. For example, it can be specified that three output categories 38 are pre-given. One output category 38 may depict a target object 24 that can be driven from below, another output category 38 may depict a target object 24 that can be driven from above, and yet another output category 38 may depict a target object 24 that is an obstacle. This target object is a target object 24 that cannot be driven from below and / or cannot be driven from above.

[0061] The classification data 42 indicates which output category 38 the target object 24 is assigned to and / or with what probability that the target object 24 is assigned to the relevant output category 38. The classification model 36 is configured to output classification data 42 related to the assignment of the target object 24. The classification data 42 can, for example, be provided to the driver assistance device 44 of the vehicle 10. The control device 26 can also be configured to output a control signal 46 based on the classification data 42. For example, the control signal 46 can be output when the target object 24 is assigned to a specific output category 38 with a specific probability. For example, it can be specified that the control signal 46 is output when the target object 24 is assigned to the output category 38 classifying the target object 24 as an obstacle with a probability higher than a certain threshold. This can thus control the driver assistance device to issue a warning signal.

[0062] Figure 2 A schematic diagram showing the radar cross-section of different targets as a function of distance;

[0063] The radar cross-section can be at least one input feature 28, and is determined for a corresponding sampling period. The target object 24 can be an aluminum can, a European-style pallet, or a car. It can be seen that the radar cross-section curve of the target object 24 exhibits characteristic changes with distance.

[0064] The methods used to date utilize linear function approximations to depict the observed RCS measurements of target object 24. Features derived from this (such as average error and slope) can be used to generate classifications regarding the feasibility of driving over it. Smaller errors, i.e., better linear approximations, indicate a higher probability of driving over the observed target object 24. This is attributed to the multipath propagation characteristics of electromagnetic waves. Exemplary RCS curves for a drivable target object 24 (aluminum can P1), a low stationary obstacle (European standard pallet P2), and a stationary obstacle (car P3) are shown below. A limitation of the methods used to date is that a significant amount of information is lost using linear function approximations. The remaining information is theoretically sufficient to distinguish the drivable target object 24 from obstacles, however, it can only distinguish the drivable target object 24 from low obstacles to a limited extent. By using an RNN as a feature generation model 32, more precise temporal structures in the radar signal can be identified, which can be used as the NN in the classification model 36 to generate classifications with a significantly lower misclassification rate.

[0065] The concept of this invention is to extract multiple input features 28 from detection point 22. These include, in addition to radar cross section (RCS), ground clearance (calculated from elevation angle, range, and the installation position of the radar sensor of radar device 12) and model error of the elevation beamformer. Other attributes can also be added to the feature list. Feature generation model 32 processes these input features 28 and updates its internal state in the process. Here, the internal state is a compressed representation of all the input features 28 observed so far. The output of feature generation model 32 is a processed temporal feature 34. This output depends on both the input features 28 and the internal state. The processing of input features 28 in feature generation model 32 is determined by the weight values ​​of each neuron, which are determined during offline learning. Temporal feature 34 is also the input signal to classification model 36. This network serves as the actual classifier, whereby temporal feature 34 describes the probability of output class 38.

[0066] Similar to feature generation model 32, the weights of each neuron used in classification model 36 are also determined during offline learning. Since detection point 22 itself has no historical information, an existing Radar Detection Tracker (RDT) is used to establish temporal correlations between detection points 22 from different radar measurement cycles. Here, a classification unit of classification system 30 (feature generation model 32 plus classification model 36) is established for each detection point trajectory 40.

[0067] The input feature 28 used to classify the detection point trajectory 40 is extracted from the detection point 22, which is also used in the update step of the radar tracker 16.

[0068] Figure 3 A schematic diagram of the classification system of the control device is shown.

[0069] The diagram illustrates a feature generation model 32 for the classification system 30, to which input features 28 corresponding to a sampling period can be provided. The feature generation model 32 can update its internal state each time it receives input features 28. Furthermore, upon receiving the corresponding input features 28, the feature generation model 32 can generate at least one pre-given temporal feature 34 and output it to the classification model 36.

[0070] After receiving at least one temporal feature 34 of the corresponding sampling period, classification model 36 can determine classification data 42 regarding assigning target object 24 to output category 38. For example, it can be specified that corresponding probability values ​​are determined, which describe the probability that target object 24 belongs to one of the output categories 38.

[0071] Figure 4 A schematic diagram showing the results of the classification model.

[0072] The described method was tested and verified in experiments.

[0073] The experimental setup is as follows: Objective: Classify output categories 38 "Driven over" (C1) and "Obstacle" (C2), ignoring output category 38 "Driven under". Output category 38 "Obstacle" (C2) includes stationary traffic participants, infrastructure, and "low obstacles". Compared to existing techniques, this method exhibits significantly better classification performance and a significantly lower misclassification rate. The results are shown in the confusion matrix.

[0074] Figure 5 This diagram illustrates a method for providing classification data based on received radar signals to classify target objects.

[0075] This method can be achieved through, for example, Figure 1 The control device 26 shown is executed.

[0076] The first step S1 of the method may include receiving radar data 20 for the current sampling period of the detection point trajectory 40 assigned to the target object 24. In other words, the control device 26 receives radar data 20, which describes, for example, detection points 22 acquired in the current sampling period and assigned by the radar tracker 16 to the detection point trajectory 40, which is assigned to the target object 24.

[0077] The second step S2 of the method may include extracting at least one pre-given input feature 28 from the radar data 20 of the current sampling period.

[0078] The third step S3 may include providing at least one input feature 28 of the current sampling period to a classification unit assigned to a specific target object 24 by a feature generation model 32 of the classification system 30. The feature generation model 32 may have an internal state characterizing at least one input feature 28 of a previous sampling period of the detection point trajectory 40. The feature generation model 32 may be configured to output at least one pre-given temporal feature based on this internal state and at least one input feature 28 of the current sampling period.

[0079] The fourth step S4 may include updating the internal state of the feature generation model 32 based on at least one input feature 28 of the current sampling period.

[0080] The fifth step S5 may include determining at least one time feature through the feature generation model 32.

[0081] The sixth step S6 may include providing at least one temporal feature to the classification model 36 of the classification system 30, wherein the classification model 36 may be configured to output classification data 42 on assigning the target object 24 to at least one pre-given output category 38 based on at least one temporal feature.

[0082] Step S7 may include outputting classification data 42 based on at least one time feature through classification model 36.

[0083] This invention uses a recurrent neural network (RNN) to approximate the history of measured radar signals 14 of a target object 24 and identify patterns in the sequence. The input features 28 in this RNN are features in a so-called detection list. To assign detection points 22 belonging to the same target object 24 from radar measurements at different time points, a known radar detection tracker (RDT) is used. The RNN processes this information as a time series, stores the relevant information as an internal state, and outputs temporal features at its output. These temporal features can then be used as input signals by another neural network (NN) to determine the prediction of the output category 38 for the measured detection points. The output category 38 can be defined as "driveable from above," "obstacle," and "driveable from below." Each output category 38 receives a probability value at each prediction step, where the sum of the probability values ​​is always 100%.

[0084] Advantageously, compared to existing technologies, the necessity of determining the processed input features 28 is eliminated, thereby reducing the manual workload for development engineers. Historical approximations are more accurate, thus allowing for the identification of more precise structures in the radar signal 14. Consequently, classification performance is higher; that is, detection points can be classified with a lower error rate. This, in turn, may contribute to a reduction in potential functional failures in the vehicle 10, thereby improving the driver's system experience.

[0085] This method can be applied to other radar classification tasks that are also based on time series, such as pedestrian and cyclist classification. The described method is able to process the time series of radar data 20 at the detection point level and generate classifications for obstacle categories such as "can be driven over", "obstacle", or "can be driven under". A classification probability can be calculated for each category. The sum of all probabilities should be 100%. This method is implemented by using a recurrent neural network (RNN) as the feature generation model 32 and subsequently using a neural network (NN) as the classification model 36.

[0086] During the classification process for individual detection points, the following steps are performed:

[0087] 1. Extract input features 28 from detection points 22 used in the update step of radar tracker 16.

[0088] 2. Calculate the secondary feature "height above ground", where h is the height above ground, γ is the elevation angle of detection point 22, r is the radial distance of detection point 22, and p is the vertical installation position of radar device 12.

[0089] 3. Scale all input features 28. The scaling factor was also determined based on the training data during the offline process.

[0090] 4. Temporal features are generated by processing the input features 28 in the feature generation model 32.

[0091] 5. Classification data is generated by processing temporal features 34 in classification model 36, which includes the probability of target object 24 being assigned to output class 38.

[0092] Figure 6 A schematic diagram of a method for training a classification system is shown.

[0093] This method may include a method for training the feature generation model 32 and a method for training the classification model 36.

[0094] The first step T1 of the method for training classification system 30 may include initializing all variables and hyperparameters required for training feature generation model 32 and classification model 36. This may include the number of training epochs, learning rate, batch size, and other parameters.

[0095] The second step T2 may include training the feature generation model 32. This may include providing the feature generation model 32 with training input features and training time features, using this data to update the weights of the RNN using gradient descent, and repeating this process for a specific number of rounds or until the feature generation model 32 converges.

[0096] The third step T3 may include training the classification model 36. This may include providing the classification model 36 with temporal features generated by the feature generation model 32 and the corresponding output class 38, using this data to update the weights of the classification model 36 using gradient descent, and repeating this process for a specific number of rounds or until the classification model 36 converges.

[0097] The fourth step T4 may include testing the classification system 30 based on new data. This may include providing input features 28 to the feature generation model 32, generating temporal features 34 by the feature generation model 32, providing these temporal features 34 to the classification model 36, and outputting predictions about whether the target object 24 belongs to different output categories 38.

[0098] Finally, the training and testing process of classification system 30 can be repeated until appropriate accuracy and performance are achieved.

[0099] Overall, this example demonstrates how to use a recurrent neural network to achieve the classification of radar detection points as "can be driven over" and "can be driven under" using sequence pattern recognition.

[0100] List of reference numerals in the attached diagram:

[0101] 10 vehicles

[0102] 12 radar devices

[0103] 14 radar signals

[0104] 16 radar trackers

[0105] 18 Radar Raw Data

[0106] 20 radar data

[0107] 22 testing sites

[0108] 24 target objects

[0109] 26 control devices

[0110] 28 Input Features

[0111] 30-classification system

[0112] 32 Feature Generation Model

[0113] 34 Time Features

[0114] 36-class classification model

[0115] 38 output categories

[0116] 40 detection point trajectories

[0117] 42-category data

[0118] 44 Driver Assistance Devices

[0119] 46 control signals

[0120] C1, C2 Output Categories

[0121] P1-P3 variation curve

[0122] Steps S1-S6

[0123] Steps T1-T4.

Claims

1. A method for providing classification data (42) for classifying target objects (24) based on detection points in received radar signals (14), Its features are, The method includes the following steps performed by the control device (26): - Receive radar data (20) for the current sampling period, the radar data including detection points of the detection point trajectory (40) assigned to the target object (24); - Extract at least one pre-given input feature (28) from the detection points in the radar data (20) of the current sampling period; - Provide at least one input feature (28) of the current sampling period to the feature generation model (32) of the classification system (30), wherein the feature generation model (32) has an internal state that represents at least one input feature (28) from the previous sampling period of the detection point trajectory (40), and the feature generation model is configured to output at least one pre-given temporal feature based on the internal state and at least one input feature (28) of the current sampling period; - Update the internal state of the feature generation model (32) based on at least one input feature (28) of the current sampling period; - Determine at least one time feature using the feature generation model (32); - Provide at least one temporal feature to the classification model (36) of the classification system (30), wherein the classification model (36) is configured to output classification data (42) based on at least one temporal feature regarding the assignment of the target object (24) to at least one pre-given output category (38); and - Based on at least one time feature, the classification model (36) outputs classified data (42).

2. The method according to claim 1, Its features are, The classification data (42) includes corresponding probability values ​​regarding the probability of assigning the target object (24) to a pre-given output category (38).

3. The method according to claim 1 or 2, Its features are, The feature generation model (32) is designed as an artificial recurrent neural network.

4. The method according to any one of the preceding claims, Its features are, The classification model (36) was designed as an artificial neural network.

5. The method according to any one of the preceding claims, Its features are, Extract multiple input features (28) of the current sampling period from the radar data (20) of the current sampling period. The input features (28) are scaled in the feature generation model (32) according to their respective scaling factors.

6. The method according to any one of the preceding claims, Its features are, At least one input feature (28) of the current sampling period includes the radar cross section of the detection point.

7. The method according to any one of the preceding claims, Its features are, At least one input feature (28) of the current sampling period includes the ground height of the detection point (22).

8. The method according to any one of the preceding claims, Its features are, According to the formula Calculate the height above the ground, where h is the height above the ground, γ is the elevation angle of the detection point (22), r is the radial distance of the detection point (22), and p is the vertical installation position of the radar device (12).

9. The method according to any one of the preceding claims, Its features are, Use a separate classification unit of the classification system (30) for each target object (24).

10. The method according to any one of the preceding claims, Its features are, The method includes outputting a control signal (46) via a control device (26).

11. A method for training a classification system (30), the method comprising: - Train a feature generation model (32), which is trained to output at least one pre-given temporal feature (34) based on the internal state of the feature generation model (32) and at least one input feature (28). - Train a classification model (36), which is trained to output classification data (42) based on at least one temporal feature (34) regarding the assignment of the target object (24) with at least one pre-given output category (38).

12. A control device (26) configured to perform the method according to any one of claims 1 to 10 and / or the method according to claim 11.

13. A program product comprising program instructions that, when executed, cause a processor circuit to perform an embodiment of the method according to any one of claims 1 to 10 and / or an embodiment of the method according to claim 11.

14. A data storage device comprising program instructions that, when executed, cause processor circuitry to perform an embodiment of the method according to any one of claims 1 to 10 and / or an embodiment of the method according to claim 11.

Images