Method and system for processing radar reflections

By receiving the radar reflected signal and determining the energy in the predetermined angle area around the target angle, combining Doppler frequency and distance information to generate energy spectrum information, the problem of insufficient performance in radar data processing is solved, and the effect of object detection and classification is improved.

CN112240999BActive Publication Date: 2024-11-08APTIV TECHNOLOGIES AG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010690550.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-07-19
Filing Date
2020-07-17
Publication Date
2024-11-08
Estimated Expiration
2040-07-17

AI Technical Summary

Technical Problem

Existing radar data processing methods are difficult to provide more meaningful information, resulting in insufficient performance in object detection, classification and semantics or instance segmentation.

Method used

The computer hardware components receive radar reflected signals, determine radar reflected energy in a predetermined angle domain around the target angle, and combine Doppler frequency and distance information to generate energy spectrum information for improving object detection and classification performance. The method includes Fourier transforming the radar reflected signal, generating an angle spectrum, and processing feature vectors through a neural network to improve detection results.

Benefits of technology

The performance of object detection, classification and semantic segmentation is improved, and the processing effect of radar data is enhanced by considering the information of angle, distance and Doppler frequency regions, and more abundant feature information is provided for further processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112240999B_ABST
    Figure CN112240999B_ABST
Patent Text Reader

Abstract

The present disclosure relates to methods and systems for processing radar reflections. A computer-implemented method for processing radar reflections includes the following steps performed by computer hardware components: receiving radar reflections by at least one radar sensor; determining a target angle at which radar reflections associated with a potential target are received by at least one radar sensor; and determining the energy of the radar reflections received by at least one radar sensor within a predetermined angular region around the target angle.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to methods and systems for processing radar reflections. The energy of received radar reflections received within a predetermined angular region around a target angle may be determined and may be processed on a distributed computing system. Background Art

[0002] Radar is a commonly used sensor in automotive perception systems due to its low price compared to LiDAR and its high adaptability to environmental changes (e.g., weather or lighting conditions) compared to cameras.

[0003] The radar data may be preprocessed to form a point cloud of detections. Each point detection is characterized by some features, such as its RCS (reflection amplitude), rate of change of range (Doppler velocity), and spatial position. The method for obtaining these features represents a preprocessing pipeline typically implemented on a satellite sensor close to the radar sensor. The resulting features describing the point detections are sent to a central radar perception unit via a CAN (controller area network) bus, where tracker and classification tasks are applied to these features. The satellite sensors may be arranged on one or more first boards, and the central radar perception unit may be arranged on a second board. These features are then used to perform object detection, classification, semantic or instance segmentation based on a specific area of ​​coverage point detections.

[0004] However, according to conventional methods, if more meaningful radar data can be provided, the performance of object detection, classification, semantic or instance segmentation can also be improved.

[0005] Therefore, there is a need to provide systems and methods for enhanced radar data processing. Summary of the invention

[0006] The present disclosure provides a computer-implemented method, a computer system, and a non-transitory computer-readable medium. Implementations are presented in the specification and drawings.

[0007] In one aspect, the present disclosure relates to a computer-implemented method for processing radar reflections. The method includes the following steps performed by computer hardware components: receiving radar reflections by at least one radar sensor; determining a target angle at which radar reflections (in other words: radar data) related to a potential target are received by at least one radar sensor; and determining the energy of the radar reflections received by at least one radar sensor within a predetermined angular region around the target angle.

[0008] In other words, not only the energy of the radar signal received at the target angle (which is the angle of the expected potential target) is taken into account, but also the energy in a predetermined angular region (in other words: the region of the angle; in other words: the angular distance; in other words: the distance of the angle; in other words: the angular area; in other words: the area of ​​the angle; in other words: the angular interval; in other words: the (one-dimensional or two-dimensional) interval of the angle; in other words: the region of interest) is determined. By adding up (in other words: summing) all the energies of the radar signals in the predetermined angular region, the energy of the radar reflections received by the at least one radar sensor in the predetermined angular region around the target angle can be determined.

[0009] The target angle may be an angle at which a potential target is expected. The target angle may be set relative to a coordinate system or relative to a vehicle in which the at least one radar sensor is provided. For example, the coordinate system may be defined by the at least one radar sensor. For example, a forward axis may be defined by the at least one radar sensor. For example, the at least one radar sensor may include a plurality of antennas or a plurality of radar sensors that may be arranged along a straight line, and a forward direction may be defined as being orthogonal to the straight line and pointing in a direction in which the at least one radar sensor transmits and receives signals; for example, if the at least one radar sensor is arranged in a vehicle, the forward direction may point to the outside of the vehicle (in other words: away from the interior of the vehicle).

[0010] It has been found that considering the energy of radar signals in angular regions provides useful information to improve object detection, classification, and semantic or instance segmentation.

[0011] According to another aspect, the radar reflected energy received by the at least one radar sensor within a predetermined angular region around the target angle may be radar reflected energy received by the at least one radar sensor within a predetermined angular region around the target angle and related to the distance of the potential target.

[0012] Differentiating angular regions according to the range (in other words: distance) of potential targets can enhance the information content of the computed energy spectrum, since targets at similar angles but at different distances can be distinguished.

[0013] According to another aspect, the energy of radar reflections received by at least one radar sensor within a predetermined angular region around the target angle may be the energy of radar reflections received by at least one radar sensor within a predetermined angular region around the target angle and associated with a predetermined distance region around a distance of the potential target. It has been found that considering not only the angular region but also the distance region may also improve the information for further processing.

[0014] According to another aspect, the radar reflected energy received by the at least one radar sensor within a predetermined angular region around the target angle may be radar reflected energy received by the at least one radar sensor within a predetermined angular region around the target angle and related to a Doppler frequency of the potential target.

[0015] Differentiating angular regions according to the Doppler frequency of potential targets (which is related to velocity) can enhance the information content of the calculated energy spectrum because targets at similar angles but with different velocities can be distinguished.

[0016] According to another aspect, the energy of radar reflections received by at least one radar sensor within a predetermined angular region around the target angle may be energy of radar reflections received by at least one radar sensor within a predetermined angular region around the target angle and related to a predetermined Doppler frequency region around the Doppler frequency of the potential target. It has been found that considering not only the angular region but also the Doppler frequency region (which corresponds to a velocity region around the velocity of the potential target) may also improve the information for further processing.

[0017] According to another aspect, the computer-implemented method may further include the step of determining a correlation coefficient between energy of radar reflections received by at least one radar sensor at the target angle and energy of related radar reflections received by at least one radar sensor within a predetermined angular region around the target angle.

[0018] The correlation coefficient between the energy of the radar reflection related to the target angle (in other words: the angle of the potential target) and the energy in the area around the target angle is taken into account to improve the results. This correlation can provide information about whether the energy in the area around the target angle is actually correlated with the energy of the target angle.

[0019] According to another aspect, corresponding energies may be determined for a plurality of distances of a potential target. According to another aspect, corresponding energies may be determined for a plurality of speeds of a potential target.

[0020] Determining energy for multiple ranges and multiple velocities may be arranged to determine energy for each range-velocity channel in the range-Doppler domain and may provide a complex data cube having dimensions relating to range, velocity and angle (i.e., the angle of a potential target around which an angular region is formed within which the energy is determined).

[0021] According to another aspect, the computer-implemented method may further include the step of providing the energies of the multiple distances for the potential targets and the multiple speeds for the potential targets as energy-related feature vectors to a neural network for further processing.

[0022] It has been found that neural networks provide effective further processing, such as object detection, object classification and / or semantic or instance segmentation.

[0023] According to another aspect, the computer-implemented method may further include the steps of: determining an additional feature vector based on the radar reflection; and providing the additional feature vector to the neural network.

[0024] The neural network can process both data related to the energy in the angular region (which can be provided in the form of a feature vector; in other words: it can be reformatted to be represented by a vector, which can be called a feature vector) and additional feature vectors (e.g., conventional feature vectors) obtained from other processing. It has been found that feature vectors according to various embodiments improve detection results compared to using only conventional feature vectors.

[0025] According to another aspect, the radar reflections may include channel data received by multiple antennas; and determining the energy of the radar reflections received by at least one radar sensor (e.g., multiple antennas) within a predetermined angular region around a target angle may be based on performing a Fourier transform on the channel data received by the multiple antennas.

[0026] Performing a Fourier transform (e.g., Fast Fourier Transform; FFT) on the last dimension of the complex data cube of range, Doppler and channel information can produce a spectrum corresponding to the angle. The channel information is related to the radar reflections received by different antennas in the antenna array and therefore includes time shifts in the received signal. Performing a Fourier transform on these time shifted signals can produce angle information.

[0027] The angular spectrum may be determined by applying a Fourier transform (e.g., a Fast Fourier Transform; FFT) on the channel dimension of a complex data cube for a particular range and Doppler channel pair. This may be referred to as a beam vector, and may correspond to the composite response of an antenna array element (e.g., the third element of the beam vector may be the (complex-valued) amplitude measured at the third antenna element). Due to production inaccuracies, it may be desirable to apply calibration to the beam vector. Since the Fourier transform may require uniformly sampled data, in the case of a non-uniformly spaced array (depending on the array geometry), a transform or interpolation to a uniformly linearly sampled array (ULA) may be provided.

[0028] According to another aspect, a first portion of the computer-implemented method may be performed by a first portion of a hardware component, a second portion of the computer-implemented method may be performed by a second portion of the hardware component, and the first portion of the hardware component may be connected to the second portion of the hardware component via a communication structure.

[0029] Separating hardware components into a first portion and a second portion connected to each other via a communication structure may provide a distributed computing system (eg, a multi-board computing system).

[0030] According to another aspect, the computer-implemented method may further include the step of sending the intermediate data as input for the second part of the computer-implemented method from the first part of the hardware component to the second part of the hardware component via the communication structure.

[0031] To limit the communication load on the communication structure, the intermediate data may be data that has been partially processed and reduced in size. The intermediate data may be data that can be obtained at the first part of the hardware component using the computing power available at the first part of the hardware component. There may be a trade-off between the computing power of the first part of the hardware component and the communication bandwidth of the communication structure; a higher computing power of the first part of the hardware component may allow the execution of processing that results in the transmission of data that has been reduced in size (which may reduce the communication bandwidth required for the communication structure); likewise, a lower computing power of the first part of the hardware component may result in a higher communication bandwidth requirement for the communication structure.

[0032] According to another aspect, the computer-implemented method may further include the following steps: encoding intermediate data in a first part of the hardware component as input to a second part of the computer-implemented method to obtain compressed data of a reduced size that is smaller than the size of the intermediate data, sending the compressed data from the first part of the hardware component to the second part of the hardware component via a communication structure, and decoding the compressed data in the second part of the hardware component, and providing the decoded compressed data as input to the second part of the computer-implemented method.

[0033] In one aspect, the decoded compressed data can be similar or identical to the intermediate data. For example, the decoded compressed data and the intermediate data can be represented by vectors of the same length, and the norm of the difference between the vector representing the decoded compressed data and the vector representing the intermediate data can be below a predetermined threshold. A neural network can be trained to provide a high degree of similarity between the decoded compressed data and the intermediate data. In another aspect, the decoded compressed data can include semantically rich dense feature vectors corresponding to the intermediate data. The intermediate data can be sparser than the decoded compressed data. Increasing the data density in the decoded compressed data can enhance further processing.

[0034] The encoding of the intermediate data and the decoding of the compressed data may be performed by one or more neural networks, and these one or more neural networks may be trained end-to-end.

[0035] In another aspect, the present disclosure is directed to a computer system comprising a plurality of computer hardware components configured to perform many or all of the steps of the computer-implemented methods described herein.

[0036] The computer system may include multiple computer hardware components (e.g., a processing unit (in other words: a processor), at least one memory unit (which may also be referred to as a memory), and at least one non-transitory data storage, such as a hard drive or solid-state drive). It will be understood that additional computer hardware components may be provided in the computer system and used to perform the steps of the computer-implemented method. The non-transitory data storage and / or the memory unit may include a computer program for instructing a computer to perform multiple or all steps or aspects of the computer-implemented method described herein, for example using a processing unit and at least one memory unit. The computer system may include at least one radar sensor.

[0037] A first portion of the computer hardware assembly is disposed in the radar sensor unit and a second portion of the computer hardware assembly may be disposed in the central radar perception unit.

[0038] It has been found that processing a large amount of data in the radar sensor unit may allow only a small amount of data to have to be transferred to the central radar perception unit. In this way, processing that may reduce the amount of data may be calculated on the radar sensor unit, and the resulting data may then be sent to the central radar perception unit where further processing that may be computationally expensive may be performed.

[0039] In another aspect, the present disclosure relates to a non-transitory computer-readable medium including instructions for performing multiple or all steps or aspects of the computer-implemented methods described herein. The computer-readable medium may be configured as: an optical medium such as a compact disk (CD) or a digital versatile disk (DVD); a magnetic medium such as a hard disk drive (HDD); a solid-state drive (SSD); a read-only memory (ROM) such as a flash memory; etc. In addition, the computer-readable medium may be configured as a data storage unit accessible via a data connection such as an Internet connection. The computer-readable medium may be, for example, an online data repository or a cloud storage unit.

[0040] The present disclosure also relates to a computer program for instructing a computer to execute multiple or all steps or aspects of the computer-implemented method described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Exemplary embodiments and functions of the present disclosure are described herein in conjunction with the following drawings which are schematically shown:

[0042] Figure 1 is an exemplary angular spectrum of a particular range-Doppler channel according to various embodiments;

[0043] Figure 2 is a scheme of a network according to various implementations;

[0044] Figure 3 is a separation network according to various embodiments;

[0045] Figure 4 is a separation network with autoencoder functionality according to various embodiments;

[0046] Figure 5 is an illustration of a spectrum graph comparing methods according to various embodiments with conventional methods; and

[0047] Figure 6 is a flow chart illustrating a computer-implemented method for processing radar reflections according to various embodiments. DETAILED DESCRIPTION

[0048] For point detections processed by the radar, information about distance [m], angle [°], and range-velocity [m / s] is calculated. This information can be determined based on identified targets in a complex data cube (CDC, which can also be called a compressed data cube), which can be represented by a tensor of dimensions R×D×C, where R is the dimension of the distance information, D is the dimension of the distance-velocity information (e.g., Doppler information), and C is the dimension of the channel information (e.g., the number of antennas, based on which the angle information can be determined).

[0049] According to various embodiments, useful features can be extracted from this data by using neighboring beam vectors (in other words: beam vectors that are adjacent to the beam vector associated with the potential target), where the reflected energy from the object will also be partially visible. For each range channel (bin) in the CDC, multiple Doppler channels are stored in the CDC. In order to determine how much energy from the object of interest falls into each Doppler channel, a Fourier transform can be determined in the last dimension (i.e., in the channel dimension), which produces a spectrum corresponding to the angle, and the spectrum can be stored in a data cube of dimensions R×D×A, where A is a dimension related to the angle (the angle is, for example, the angle at which at least one radar sensor receives, for example, a radar reflection related to (or from) a potential target).

[0050] A region of interest (e.g., a predetermined energy bandwidth, e.g., a 3 dB bandwidth) may be defined in the angular spectrum around a peak belonging to the target of interest (in other words: an energy peak). Based on the region of interest (e.g., a predetermined energy bandwidth), the closest channel in the spectrum within the bandwidth may be determined. For example, for a 3 dB bandwidth, the closest channel in the spectrum may be determined such that at θ 3dB- With θ 3dB+The angle θ is determined in such a way that the energy reflected by the radar at each angle in the angle region between the two is within 3 dB of the energy at the target angle. 3dB- and θ 3dB+ θ 3dB- With θ 3dB+ The angle between θ and 3dB+ With θ 3dB- The energy between can be accumulated (in other words: added; in other words: summed) along the Doppler dimension to obtain the energy distribution on the Doppler channel. For example, the following formula can be used to determine the average energy in the area of ​​interest:

[0051]

[0052] Where r represents the range channel, θ represents the angle channel, d represents the velocity (or Doppler channel), and N θ Represents θ 3dB+ With θ 3dB- is the number of angular channels between, and S(r, d, θ) is the Fourier transform of the CDC along the channel axis (e.g., determined using a Fast Fourier Transform FFT). It will be appreciated that although the above formula illustrates the use of a one-dimensional sum accumulation (for a 2D radar), a two-dimensional sum (for a 3D radar) may also be used.

[0053] Figure 1 An exemplary angular spectrum 100 of a particular range-Doppler channel according to various embodiments is shown. A plurality of points 106 in the angular spectrum 100 illustrate the energy 104 of the reflected signal at angle 102. The target 108 may be determined, for example, by any frequency estimation method. This determination may be made prior to determining the target local energy. The energy associated with the angle in the region of interest, such as at point 110 (e.g., θ ) may then be averaged using the above formula. 3dB- ) and point 112 (e.g., θ 3dB+ )between.

[0054] According to various embodiments, a correlation coefficient between the peak 108 and neighboring channels (along the Doppler dimension, but in the range-Doppler-map domain) may be determined to weight the energy distribution.

[0055] According to various embodiments, energy may also be accumulated for adjacent range channels.

[0056] The energy distribution position along the abscissa may be corrected relative to the ego-motion of an observer (eg, a car) on which the radar sensor is mounted.

[0057] The energy distribution can be used as a (CDC) feature for object detection (eg, object classification).

[0058] The CDC features extracted by each point detection can be combined with the point features to classify the object with the help of an end-to-end trainable neural network. This can be called feature embedding.

[0059] Figure 2 A scheme 200 of a network according to various embodiments is shown. The CDC features that may be determined as described above may be arranged as an energy distribution 202 in the Doppler domain (e.g., arranged as a matrix with dimensions equal to the number of range channels times the number of Doppler channels). A 2D convolutional feature encoding 204 may be applied to the (weighted) energy distribution 202 at each point. After multiple convolution, pooling, and activation function pairs, a semantically rich feature map 206 with a fixed width and channel size may be created. The feature map 206 may undergo a reshaping process 208 to obtain a feature vector 210.

[0060] The additional point features 212 may be processed by a plurality of fully connected layers (e.g., a first fully connected layer 214 for obtaining a first processed feature vector 216, and a second fully connected layer 218 for obtaining a feature vector 220 having the same shape as the feature vector 210) to encode the point features 212 and make them have the same shape as the shaped CDC feature map (in other words: the feature vector 210).

[0061] CDC feature map 206 (in the form of feature vector 210 ) is masked with another feature vector 220 using element-wise multiplication 222 to obtain a masked feature vector 224 .

[0062] Each point can be processed individually using the same network with shared weights. In order to determine whether an object is within a region proposal, features from points within the region proposal need to be combined. This can be achieved using a commutative (in other words: symmetric) function 226, which can provide element-wise accumulation and ensure that the point arrangement is not changed. The region proposal can also be called an anchor.

[0063] Additional networks may be used for further processing of these combined features 228. It will be appreciated that this further processing 228 may be performed according to conventional processing and that various embodiments are not limited to a particular manner of further processing 228. Likewise, various embodiments are not limited to a particular source of region candidates and preprocessing of point detections of their combinations.

[0064] Normally, CDC data is not sent from satellite sensors to radar perception units due to bandwidth limitations.

[0065] Figure 3 FIG. 3 shows a separation network 300 according to various embodiments. Figure 2 The components of the illustrated network 200 are similar or identical such that the same reference numerals may be used and repeated descriptions may be omitted. The network 300 (i.e., a split network) may be split into a first portion of the network 300 implemented on one or more satellite sensors 302 and a second portion of the network 300 (e.g., the remainder of the network 300) implemented on a central perception unit 304. Training of the network may be performed in the combined network. After training, the network may be split. As indicated by arrow 306, conventional point features may be sent from the satellite sensor 302 to the central radar perception unit 304. As indicated by arrow 308, feature vectors 210 representing feature maps 206 may be sent from the satellite sensor 302 to the central radar perception unit 304.

[0066] To further reduce the communication overhead, an end-to-end trainable auto-encoding method can be provided to compress the information before sending it over the CAN (Controller Area Network) bus. The network can include two parts, one running on the satellite and the other running on the perception unit. On the satellite, the CDC feature vectors are encoded into feature vectors of smaller size. These vectors are then passed and decoded on the perception unit without losing much information.

[0067] Figure 4 FIG. 4 shows a separation network 400 with autoencoder functionality according to various embodiments. The various components of the separation network 400 with autoencoder are shown in FIG. Figure 2 The network 200 or Figure 3 The components of the network 300 shown are similar or identical, so that the same reference numerals may be used and repeated descriptions may be omitted. The satellite sensor 402 may be similar to Figure 3 4. The satellite sensor 302 of FIG. 402 may include an auto-encoding function, according to which the size of the feature vector 210 is reduced by one or more fully connected layers (e.g., the first fully connected layer 408 may reduce the size to obtain the feature vector 410, which may be further applied to the second fully connected layer 412 to obtain the feature vector 414). The size of the feature vector 414 may be smaller than the feature vector 210. As shown by arrow 406, the feature vector 414 (which may be referred to as intermediate data) may then be sent from the satellite sensor 402 to the central radar sensing unit 404. The central radar sensing unit 404 may be similar to Figure 3The central radar sensing unit 304 may include an automatic decoding function. In the central radar sensing unit 404, the received (encoded) feature vector 414 may be decoded, for example, using a fully connected layer 416 to obtain a decoded feature vector 418 (which may be close to or identical to the original feature vector 210 before encoding). For example, the norm of the difference between the original feature vector 210 and the decoded feature vector 418 may be below a predetermined threshold. For example, the size of the encoded feature vector 414 may be selected as follows: so that the predetermined threshold may be met (for example, if it is determined that the difference between the original feature vector 210 and the decoded feature vector 418 may be greater than a predetermined threshold, the size of the encoded feature vector 414 may be increased to allow more information to be encoded and result in a better similarity between the original feature vector 210 and the decoded feature vector 418).

[0068] The automatic encoding function of satellite sensor 402 and the automatic decoding function of central radar perception unit 404 can provide a feature vector 414 that is smaller in size than feature vector 210, which can reduce the communication load when feature vector 414 (instead of feature vector 210) is sent from satellite sensor 402 to central radar perception unit 404.

[0069] Figure 5 An illustration 500 of a spectrum diagram comparing methods according to various embodiments with conventional methods is shown. Figure 5 Spectrograms 502, 506 and 510 in the left part of FIG. 1 illustrate Doppler spectrograms obtained using conventional methods, and Figure 5 Spectral graphs 504, 508, and 512 in the right portion of exemplify spectrograms obtained using systems and methods according to various embodiments. Figure 5 The horizontal axis represents time and the vertical axis represents frequency.

[0070] The method according to various embodiments including feature encoding according to various embodiments shows better labeling, suppresses most of the clutter, and thus the classifier can work better than conventional methods. Table 1 provides performance results of the classifier using the method according to various embodiments.

[0071] Accuracy Recall f1 score Support bike 0.50 0.10 0.16 52 Car or van or SUV 0.96 0.94 0.95 3046 pedestrian 0.81 0.80 0.81 887 Average / Total 0.92 0.90 0.91 3985

[0072] Table 1

[0073] For the results of Table 1, bicycles, cars or vans or off-road vehicles and pedestrians are the categories into which the detected objects are classified. The precision indicates the ratio tp / (tp+fp), where tp is the number of true positives and fp is the number of false positives. Intuitively, precision is the ability of the classifier not to mark negative samples as positive. Recall is the ratio tp / (tp+fn), where tp is the number of true positives and fn is the number of false negatives. Intuitively, recall is the ability of the classifier to find all positive samples. The f1 score provides the harmonic mean of precision and recall. Support is the number of samples of the true response in this category. The rows marked "average / total" in Table 1 include the weighted average of precision, recall and f1 scores, where the weight is the support value.

[0074] Without using methods according to various embodiments, performance drops to an f1 score of 0.68.

[0075] Figure 6 A flowchart 600 illustrating a computer-implemented method for processing radar reflections according to various embodiments is shown. At 602, at least one radar sensor may receive a radar reflection. At 604, a target angle is determined at which the at least one radar sensor receives a radar reflection associated with a potential target. At 606, an energy of the radar reflection received by the at least one radar sensor within a predetermined angular region around the target angle is determined.

[0076] According to various embodiments, the radar reflected energy received by at least one radar sensor within a predetermined angular region around the target angle may be radar reflected energy received by at least one radar sensor within a predetermined angular region around the target angle and related to the distance of the potential target.

[0077] According to various embodiments, the radar reflected energy received by at least one radar sensor within a predetermined angular region around a target angle may be radar reflected energy received by at least one radar sensor within a predetermined angular region around a target angle and associated with a predetermined distance region around a distance to a potential target.

[0078] According to various embodiments, the computer-implemented method may further include the step of determining a correlation coefficient between energy of radar reflections received by the at least one radar sensor at the target angle and energy of radar reflections received by the at least one radar sensor within a predetermined angular region around the target angle.

[0079] According to various embodiments, respective energies are determined for a plurality of distances of a potential target. According to various embodiments, respective energies are determined for a plurality of speeds of a potential target.

[0080] According to various embodiments, the computer-implemented method may further include the step of providing the energies of the multiple distances for the potential targets and the multiple speeds for the potential targets as energy-related feature vectors to a neural network for further processing.

[0081] According to various embodiments, the computer-implemented method may further include the steps of: determining an additional feature vector based on the radar reflection; and providing the additional feature vector to the neural network.

[0082] According to various embodiments, the radar reflections may include channel data received by multiple antennas, and determining the energy of the radar reflections received by at least one radar sensor within a predetermined angular region around the target angle may be based on performing a Fourier transform on the channel data received by the multiple antennas.

[0083] According to various embodiments, a first portion of the computer-implemented method is performed by a first portion of a hardware component, and a second portion of the computer-implemented method is performed by a second portion of the hardware component. The first portion of the hardware component may be connected to the second portion of the hardware component via a communication structure.

[0084] According to various embodiments, the computer-implemented method may further include the step of sending the intermediate data as input to the second part of the computer-implemented method from the first part of the hardware component to the second part of the hardware component via the communication structure.

[0085] According to various embodiments, the computer-implemented method may further include the following steps: encoding the intermediate data as input to the second part of the computer-implemented method in the first part of the hardware component to obtain compressed data of a reduced size that is smaller than the size of the intermediate data, sending the compressed data from the first part of the hardware component to the second part of the hardware component via the communication structure, and decoding the compressed data in the second part of the hardware component, and providing the decoded compressed data as input to the second part of the computer-implemented method.

[0086] Each of the above steps 602, 604, 606 and other steps may be performed by computer hardware components.

[0087] As described above, according to various embodiments, systems and methods for radar CDC data feature encoding for multi-board level solutions can be provided. According to various embodiments, systems and methods can be configured to generate new point features associated with CDC data of main beam vectors and adjacent beam vectors.

[0088] Radar CDC (Complex Data Cube) data contains useful information to improve detection and classification performance. After applying 2D FFT in range and Doppler direction and CFAR threshold (Constant False Alarm Rate Threshold), CDC contains discrete radar energy response as beam vectors. According to various embodiments, in order to use super-resolution methods, not only the beam vector of the peak is kept, but also the adjacent beam vectors. Therefore, more useful information can be extracted knowing the relationship between the main beam vector and the adjacent beam vectors.

[0089] Furthermore, according to various embodiments, a set of possible feature encodings may be provided to encode these features into small feature vectors for each point, thereby enabling the data to be transmitted over the CAN bus. The encodings may be learned in an end-to-end manner.

[0090] Reference numerals list

[0091] 100 Angle Spectrum

[0092] 102 Angle

[0093] 104 Energy

[0094] 106 points

[0095] 108 Target

[0096] 110 points

[0097] 112 points

[0098] 200 Network solutions according to various embodiments

[0099] 202 Energy Distribution

[0100] 204 2D Convolutional Feature Encoding

[0101] 206 Figures

[0102] 208 Plastic Processing

[0103] 210 Eigenvectors

[0104] 212 features

[0105] 214 First fully connected layer

[0106] 216 The first processed feature vector

[0107] 218 Second fully connected layer

[0108] 220 Eigenvectors

[0109] 222 Element-wise multiplication

[0110] 224 Masked eigenvectors

[0111] 226 Symmetric functions

[0112] 228 Further processing

[0113] 300 Separation Network According to Various Embodiments

[0114] 302 Satellite Sensors

[0115] 304 Central Perception Unit

[0116] 306 Arrow

[0117] 308 Arrow

[0118] 400 Separation network with autoencoder function according to various embodiments

[0119] 402 Satellite Sensor

[0120] 404 Central Radar Perception Unit

[0121] 406 Arrow

[0122] 408 First fully connected layer

[0123] 410 Eigenvectors

[0124] 412 Second fully connected layer

[0125] 414 Eigenvector

[0126] 416 Fully connected layer

[0127] 418 Decoded feature vector

[0128] 500 Spectrum Graph Example

[0129] 502 Doppler spectrum obtained using conventional methods

[0130] 504 Spectrum diagram obtained using systems and methods according to various embodiments

[0131] 506 Doppler spectrum obtained using traditional methods

[0132] 508 Spectrum diagram obtained using systems and methods according to various embodiments

[0133] 510 Doppler spectrum obtained using traditional methods

[0134] 512 Spectrum diagram obtained using systems and methods according to various embodiments

[0135] 600 Flowchart

[0136] 602 Methods and Steps

[0137] 604 Methods and Steps

[0138] 606 Methods and Steps

Claims

1. A computer-implemented method for processing radar reflections, the computer-implemented method comprising the following steps performed by computer hardware components: receiving radar reflections by at least one radar sensor; determining a target angle at which radar reflections associated with a potential target are received by the at least one radar sensor, wherein: The target angle corresponds to an energy peak in the angular spectrum; determining an energy of radar reflections received by the at least one radar sensor at the target angle; determining an energy of radar reflections received by the at least one radar sensor at a plurality of angles within a predetermined angular region around the target angle, wherein the predetermined angular region is determined by defining a predetermined energy bandwidth in the angular spectrum; determining a sum of an energy of radar reflections received by the at least one radar sensor at the target angle and an energy of radar reflections received by the at least one radar sensor at the plurality of angles within the predetermined angular region; and Based on the determination and detection and / or classification of the potential target, Wherein, the radar reflected energy received by the at least one radar sensor within a predetermined angular region around the target angle is radar reflected energy received by the at least one radar sensor within a predetermined angular region around the target angle and is related to a predetermined Doppler frequency region around the Doppler frequency of the potential target.

2. The computer-implemented method according to claim 1, in, The radar reflected energy received by the at least one radar sensor within the predetermined angular region around the target angle is the radar reflected energy received by the at least one radar sensor within the predetermined angular region around the target angle and is related to the distance of the potential target.

3. The computer-implemented method according to claim 1 or 2, in, The radar reflected energy received by the at least one radar sensor within the predetermined angular region around the target angle is radar reflected energy received by the at least one radar sensor within the predetermined angular region around the target angle and related to a predetermined distance zone around the distance of the potential target.

4. The computer-implemented method of claim 1 , further comprising the steps of: A correlation coefficient is determined between energy of radar reflections received by the at least one radar sensor at the target angle and energy of radar reflections received by the at least one radar sensor within the predetermined angular region around the target angle.

5. The computer-implemented method of claim 1, in, The corresponding energies are determined for multiple distances to potential targets.

6. The computer-implemented method of claim 1, in, The corresponding energies are determined for multiple velocities of potential targets.

7. The computer-implemented method of claim 1 , further comprising the steps of: The energies of the multiple distances to the potential targets and the multiple speeds to the potential targets are provided as energy-related feature vectors to a neural network for further processing.

8. The computer-implemented method of claim 7, further comprising the steps of: determining a further feature vector based on the radar reflections; and The further feature vector is provided to the neural network.

9. The computer-implemented method of claim 1, in, The radar reflections include channel data received by a plurality of antennas; and Wherein, determining the energy of radar reflections received by the at least one radar sensor within the predetermined angular region around the target angle is based on performing Fourier transform on the channel data received by the plurality of antennas.

10. The computer-implemented method of claim 1, in, A first portion of the computer-implemented method is performed by a first portion of the hardware components; wherein a second portion of the computer-implemented method is performed by a second portion of the hardware component; and Therein, the first part of the hardware component is connected to the second part of the hardware component via a communication structure.

11. The computer-implemented method of claim 10, further comprising the steps of: Intermediate data is sent as input to a second part of the computer-implemented method from a first part of the hardware component to a second part of the hardware component via the communication structure.

12. The computer-implemented method of claim 10, further comprising the steps of: encoding, in a first portion of the hardware component, intermediate data as input to a second portion of the computer-implemented method to obtain compressed data of reduced size that is less than a size of the intermediate data; sending the compressed data from the first portion of the hardware components to the second portion of the hardware components via the communication structure; as well as The compressed data is decoded in a second portion of the hardware component and the decoded compressed data is provided as input to a second portion of the computer-implemented method.

13. A computer system comprising a plurality of computer hardware components configured to perform the computer-implemented method according to any one of claims 1 to 12.

14. The computer system according to claim 13, in, A first portion of the computer hardware assembly is disposed in a radar sensor unit; and Wherein, the second part of the computer hardware component is arranged in the central radar sensing unit.

15. A non-transitory computer-readable medium comprising instructions for executing the computer-implemented method according to any one of claims 1 to 12.