WO2024060771A1

WO2024060771A1 - Millimeter wave radar data clustering method based on multi-frame doppler velocity dimension expansion

Info

Publication number: WO2024060771A1
Application number: PCT/CN2023/104570
Authority: WO
Inventors: 王帅; 李再兴; 孙浩
Original assignee: 深圳安智杰科技有限公司
Priority date: 2022-09-21
Filing date: 2023-06-30
Publication date: 2024-03-28
Also published as: CN115236627B; CN115236627A

Abstract

A millimeter wave radar data clustering method based on multi-frame Doppler velocity dimension expansion. The method comprises: S1. preprocessing inputted millimeter wave radar frame data, and performing three-dimensional coordinate normalization on the preprocessed frame data; S2. invoking an OPTICS method to obtain a normalized data clustering result; S3. extracting a clustering data label from the normalized data clustering result, and matching the multi-frame clustering data with present-frame data to obtain a present-frame data cluster. The method introduces a Doppler velocity dimension, and can effectively distinguish between targets with similar position dimension distances but which have relative motion, thereby solving the problem of not being able to distinguish between short-distance targets when clustering is dependent purely on position information. The method further introduces a multi-frame delayed data method for a small targets prone to missed detection, where, if a single-frame missed detection probability is p f, the missed detection probability of such a target can be reduced to p f N, thereby improving the performance of the clustering method.

Description

A millimeter-wave radar data clustering method based on multi-frame Doppler velocity expansion

Technical field

The invention belongs to the technical field of signal and information processing, and more specifically, relates to a millimeter wave radar data clustering method based on multi-frame Doppler velocity expansion.

Background technique

Millimeter-wave radar is one of the important sensors included in automotive ADAS. Because the carrier wave has the characteristics of high frequency and short wavelength, it can reduce the divergence of the emitted electromagnetic wave beam and improve anti-interference. And because of the large Doppler frequency shift, it can achieve a relatively large High speed measurement accuracy. Millimeter wave radar realizes distance, angle and speed detection by analyzing the difference in characteristics of the transmitted wave and the reflected wave of the object. When the environment has many interfering targets other than effective targets, such as obstacles, buildings, etc., or the detection target is in a strong reflection obstacle When nearby, the detection data will contain a large amount of invalid data and even cover the target information, causing detection performance to decrease.

The traditional processing method is to convert the distance and angle information of the radar data into a two-dimensional Cartesian coordinate system with the horizontal distance as the x-axis and the longitudinal distance as the y-axis, and use clustering methods such as K-means, DBSCAN, and OPTICS to The Euclidean distance between data points is used as an evaluation index to divide the data in the two-dimensional position dimension. However, it is more difficult to use traditional clustering methods to achieve accurate clustering of target echo data in the following situations: 1. Large-size target echo data is discontinuous; 2. There are multiple targets within the detection range and the distance between targets is small or The distance between the target and obstacles is small; 3. The number of echoes from small-sized targets is small. For case 1, large-sized targets such as short-distance trucks are prone to have many echoes in the head and tail parts of the target, but no echo in the middle part. The clustering method based on two-dimensional position information can easily divide them. Generate redundancy for multiple targets or introduce noise in clustering; for case 2, clustering based on two-dimensional position information cannot achieve effective distinction between targets or between targets and obstacles, resulting in target loss or obstacle echo data being normalized. as the target; in case 3, due to the small number of target echoes, it will be judged as noise with a high probability, causing the target to be lost. The above-mentioned target loss or target redundancy due to data clustering will greatly reduce the reliability of the detection system.

Contents of the invention

In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a millimeter wave radar data clustering method based on multi-frame Doppler velocity expansion, aiming to solve the problems caused by the data clustering methods in the existing technology when targeting small targets. Target loss or target redundancy thereby reducing detection reliability.

The present invention provides a millimeter wave radar data clustering method based on multi-frame Doppler velocity expansion, which includes the following steps:

S1: preprocessing the input millimeter wave radar frame data and normalizing the three-dimensional coordinates of the preprocessed frame data;

S2: Obtain the normalized data clustering results by calling the OPTICS method;

S3: Extract the clustering data label from the normalized data clustering result, and match the multi-frame clustering data with the current frame data to obtain the current frame data clustering.

Among them, parameter initialization is required before inputting frame data: setting the number of delayed frames N, OPTICS method parameters and data preprocessing parameters.

Data preprocessing parameters include: energy threshold Pmin, angle range of interest [θmin θmax] and distance range [rmin rmax]; millimeter wave radar frame data includes: polar radius r in polar coordinates of the observation point, angle θ, Doppler velocity v and energy I.

As an embodiment of the present invention, preprocessing includes: using the set data preprocessing parameters as data filtering conditions, judging the energy, angle and distance of the input points in sequence. If the conditions are met at the same time, they are retained, and if they are not met, they are deleted.

Furthermore, after preprocessing and before normalization, it also includes:

Frame data is pushed onto the stack: When the number of frames stored in the frame data stack is <N, the current frame data is stored; when the number of frames = N, the clustering algorithm is executed; when the number of frames > N, the last frame data of the stack is deleted and Store the current frame data to the top of the stack to maintain the number of frame data in the stack as always N;

Frame data stack judgment: This step mainly takes effect when the method starting cycle number is <N. Its function is to start the subsequent process only when the frame data stack is full, otherwise it will wait for new data input.

As an embodiment of the present invention, the three-dimensional coordinate normalization is specifically as follows:

After the frame data passes the judgment condition, the format of the data in the stack converted into a matrix is as follows:

The first line stores the x position coordinate, the second line stores the y position coordinate, and the third line stores the Doppler velocity. The superscript is the frame number to which the data belongs, and the value range is 1~N.

Among them, two different normalization methods are used for processing the position dimension and velocity dimension:

Positional dimension data normalization: using linear transformation method, based on the maximum value in the array with minimum value , according to the formula Transform the whole to the range of 0~1;

Speed dimension data normalization: obtaining speed through statistical methods , where the speed dimension normalization function expression is as follows: ;Formula is the attenuation coefficient, used to control the Euclidean distance between the moving target and the fixed target;

The normalized data is: .

Furthermore, step S2 specifically includes:

According to the formula Obtain the Euclidean distance between data as the basis for OPTICS algorithm clustering;

Call the OPTICS method to get the normalized data clustering results, and generate a structure for each cluster. The data label is the normalized matrix column number.

Furthermore, step S3 is specifically:

Index the original radar matrix data according to the normalized matrix column number, and store the three-dimensional coordinates of the data into clustering;

Compare the multi-frame clustering data with the information of each dimension of the current frame data. If they are completely consistent, the match is successful and the data in the cluster is retained. Otherwise, it is deleted, thereby completing the current frame data clustering.

The invention also provides a millimeter wave radar data clustering system based on multi-frame Doppler velocity expansion, including a multi-frame data processing module, an OPTICS-based three-dimensional data clustering module and a current frame clustering data recovery module connected in sequence. , the multi-frame data processing module is mainly used to implement millimeter wave radar multi-frame data preprocessing, data storage and shifting, Doppler velocity dimension nonlinear normalization, position dimension linear normalization and other operations; the OPTICS The three-dimensional data clustering module is mainly used to implement three-dimensional data clustering based on the OPTICS method; the current frame clustering data recovery module is used to match multi-frame clustering data with the current frame data, and output the clustering results to subsequent modules, Used for target matching or tracking.

Through the above technical solution conceived by the present invention, compared with the existing technology, because the present invention introduces the Doppler velocity dimension, the method proposed by the present invention can be used to effectively carry out targets whose position dimensions are close but have relative motion. The distinction solves the problem that clustering based solely on location information cannot distinguish close targets. At the same time, the present invention introduces a multi-frame delayed data method. For small targets that are prone to missed detection, if the single frame missed detection probability is , using the method proposed by the present invention can reduce the probability of missed detection of such targets to , thereby improving the performance of the clustering method.

Description of the drawings

FIG1 is a block diagram of the structural principles of a millimeter wave radar data clustering system provided by an embodiment of the present invention.

FIG2 is a flowchart of an implementation of a millimeter wave radar data clustering method provided in an embodiment of the present invention.

Figure 3 is an on-site photo of test scene 1 provided by the embodiment of the present invention.

Figure 4 shows the three-dimensional clustering results of the data clustering method based on multi-frame Doppler velocity dimension expansion for test scenario 1 data provided by the embodiment of the present invention.

Figure 5 shows the two-dimensional clustering results of the data clustering method based on multi-frame Doppler velocity dimension expansion for test scenario 1 data provided by the embodiment of the present invention.

Figure 6 shows the test scenario one data provided by the embodiment of the present invention based on the two-dimensional location plane clustering result one based on OPTICS.

Figure 7 shows the test scenario one data provided by the embodiment of the present invention based on the two-dimensional position plane clustering result two based on OPTICS.

Figure 8 is an on-site photo of test scenario 2 provided by the embodiment of the present invention.

Figure 9 shows the three-dimensional clustering results of the data clustering method based on multi-frame Doppler velocity dimension expansion for the second test scenario data provided by the embodiment of the present invention.

FIG. 10 is a two-dimensional clustering result of the data clustering method based on multi-frame Doppler velocity expansion for the test scenario 2 data provided by an embodiment of the present invention.

Figure 11 shows the two-dimensional position plane clustering results of test scenario two data provided by the embodiment of the present invention based on OPTICS.

Figure 12 is a live photo of test scene three provided by the embodiment of the present invention.

Figure 13 shows the three-dimensional clustering results of the data clustering method based on multi-frame Doppler velocity dimension expansion for the third test scenario data provided by the embodiment of the present invention.

FIG. 14 is a two-dimensional clustering result of the data clustering method based on multi-frame Doppler velocity expansion for the test scenario three data provided by an embodiment of the present invention.

Figure 15 shows the two-dimensional position plane clustering results of test scenario three data provided by the embodiment of the present invention based on OPTICS.

Implementation

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention.

The present invention is suitable for millimeter-wave radar data clustering methods, can overcome the defects of traditional clustering methods in the process of processing millimeter-wave radar data, and proposes a high-performance millimeter-wave radar data based on multi-frame Doppler velocity expansion. Performance clustering methods achieve improved performance of clustering methods at the expense of a small increase in data processing volume.

The structure of the high-performance clustering method for millimeter-wave radar data based on multi-frame Doppler velocity expansion proposed by this invention is shown in Figure 1. It can be divided into three modules according to function: multi-frame data processing module, OPTICS-based three-dimensional Data clustering module and current frame clustering data recovery module. The multi-frame data processing module is mainly used to implement millimeter wave radar multi-frame data preprocessing, data storage and shifting, Doppler velocity dimension nonlinear normalization, position dimension linear normalization and other operations; OPTICS three-dimensional data clustering The module is mainly used to implement three-dimensional data clustering based on the OPTICS method; the current frame clustering data recovery module is used to match multi-frame clustering data with the current frame data, and output the clustering results to subsequent modules for target matching or tracking. .

As shown in Figure 2, the millimeter wave radar data clustering method provided by the present invention includes the following steps:

1. Multi-frame data processing

Step 1.1: Initialization of method parameters: Set the number of delayed frames N, OPTICS method parameters (core density M, neighborhood distance etc.), data preprocessing parameters (energy threshold Pmin, angle range of interest [θmin θmax], distance range [rmin rmax]);

Step 1.2: Frame data input: Input the millimeter wave radar data, including the polar diameter r, angle θ, Doppler velocity v and energy I in the polar coordinates of the observation point;

Step 1.3: Frame data preprocessing: This step mainly judges the input points in sequence based on the specified data filtering rules and the data preprocessing parameters set in step 1.1. If the conditions are met, they will be retained. If not, they will be deleted and the method operation will be reduced. quantity;

Step 1.4: Frame data is pushed onto the stack: When the number of frames stored in the frame data stack is <N, the current frame data is stored. When the number of frames is >N, the frame data at the end of the stack is deleted and the current frame data is stored on the top of the stack;

Step 1.5: Frame data stack judgment: This step mainly takes effect when the method starting cycle number is <N. The function is to start the subsequent process only when the frame data stack is full, otherwise wait for new data input.

Step 1.6: Three-dimensional coordinate normalization: After the frame data passes the judgment condition, the format of the data in the stack converted into a matrix is as follows:

The first line stores the x position coordinates, the second line stores the y position coordinates, and the third line stores the Doppler velocity. The superscript is the frame number to which the data belongs, and the value range is 1~N. For the position dimension and velocity dimension, the present invention adopts two different normalization methods to process.

Step 1.6.1 Normalize position dimension data: use linear transformation method, based on the maximum value in the array with minimum value , press the formula to transform the whole into the range of 0~1;

Step 1.6.2 Speed dimension data normalization: In order to increase the three-dimensional Euclidean distance between the moving speed target and the fixed obstacle as much as possible to eliminate the influence of the fixed obstacle detection data, after analyzing a large amount of engineering data, it was found that for urban traffic scenes In millimeter-wave radar data, the proportion of echo data from fixed obstacles is much larger than that of other targets, and the Doppler velocity distribution of this type of data is relatively concentrated. This velocity can be obtained through statistical methods. . The expression of the speed dimension normalization function is as follows:

in the formula is the attenuation coefficient, used to control the Euclidean distance between the moving target and the fixed target.

The normalized data is:

2. Normalized Data Clustering

Step 2.1 The formula for calculating the Euclidean distance between data is as follows:

Step 2.2 Call the OPTICS method to obtain the normalized data clustering results, and generate a structure for each cluster. The data label is the normalized matrix column number.

3. Current frame clustering data recovery

Step 3.1 Extract cluster data labels: Index the original radar matrix data according to the normalized matrix column number, and store the three-dimensional coordinates of the data into clusters.

Step 3.2 matches the multi-frame clustering data with the current frame data. If the match is successful, the data in the cluster is retained, otherwise it is deleted, and the clustering of the current frame data is completed.

By introducing the Doppler velocity dimension, the present invention can effectively distinguish targets with close distances but relative motion in the position dimension by using the method proposed by the present invention, solving the problem that clustering based solely on position information cannot distinguish close targets. question. At the same time, a multi-frame delay data method is introduced. For small targets that are prone to missed detection, if the single frame missed detection probability is , using the method proposed by the present invention, the probability of missed detection of such targets can be reduced to , thereby improving the performance of the clustering method.

In order to further illustrate the millimeter-wave radar data clustering method provided by the embodiment of the present invention, the details are as follows with reference to specific examples:

The millimeter-wave radar data clustering method based on multi-frame delay and Doppler velocity expansion is used to cluster the 77GHz radar data.

The test scene is a one-way lane, with fixed discrete irregularly distributed obstacles on both sides, and small targets traveling in the opposite direction.

The specific process includes:

(1) Initialization of method parameters. Set the number of delayed frames to 3, core density to 3, and neighborhood distance =0.2, neighborhood distance =0.1: energy threshold Pmin=1e-3, focus angle range [-15° 15°], longitudinal distance range [0m 60m], lateral distance range [-20m 20m]), attenuation coefficient =0.3;

(2) Read 3 frames of data into the stack sequentially. The row index corresponds to the horizontal position, the longitudinal position, the Doppler velocity and the energy. The column index corresponds to the data label. The data is preprocessed according to the following data point filtering rules. Output multi-frame data matrix:

R1: Delete detection data with energy threshold lower than 1e-3;

R2: Delete detection data with angles outside [-15° 15°];

R3: Delete the detection data with longitudinal distance > 60m;

R4: Delete data whose lateral distance is outside [-20m 20m].

(3) Location data normalization. The maximum and minimum values of the horizontal position and longitudinal position are extracted respectively, and normalized to the [0 1] range; the speed data is normalized. Take the length of the velocity interval as 0.1m/s, traverse the Doppler velocity data, extract the most densely distributed interval and let it be , in this case =-12.91m/s, adopted;

(4) Normalized data clustering based on the OPTICS method. The parameter is M=3, =0.2， =0.1. Since this method is relatively mature, it will not be described in detail here.

(5) Current frame clustering data recovery. There are many methods available for this step, such as matching by label or strict matching by data. In this example, matching by data is used, that is, first extract the corresponding data points in the original multi-frame data matrix according to the clustering label, and use the exhaustive method to combine the two If the data is consistent, the current frame data clustering number will be assigned until the exhaustive search is completed, and the unmatched data will be the noise points of this frame.

The data clustering results in application examples of the millimeter wave radar data clustering method proposed by the present invention and the comparison results using the traditional OPTICS clustering method are shown in Figures 3 to 11.

Figure 3 is a photo of the on-site environment corresponding to test scene 1. There are irregularly distributed fixed obstacles on both sides of the road, and a larger vehicle appears on the right front. Figure 4 is the processing result of the data clustering method based on multi-frame Doppler velocity expansion proposed by the present invention. It can be seen from the figure that the method can effectively identify the vehicles appearing in the figure, and the clustering result is cluster 8 in the figure. Figure 5 shows the clustering result after restoring multi-frame data to this frame data. It can be seen from the figure that due to the large size of the vehicle, the echo data from the vehicle in this frame data is scattered. However, the method proposed by the present invention can Effectively cluster the data, and the clustering result is cluster 8 in the figure. Figures 6 and 7 show the results of clustering based on two-dimensional position data using OPTICS. The corresponding clustering data is labeled cluster 1. It can be seen from the two figures that clustering only single-frame data containing two-dimensional position variables cannot achieve ideal results. It can only identify part of the echo data or classify noise signals that are relatively close to the clustering results. The clustering results are all It is worse than the method proposed by the present invention. It can be seen that the proposed method can effectively solve the clustering problem in the case of discontinuous echo data of large-sized targets.

Figure 8 is a photo of the on-site environment corresponding to test scenario 2. There are irregularly distributed fixed obstacles on both sides of the road, and smaller motor vehicles appear near the obstacles in front. Figure 9 shows the processing results of the data clustering method based on multi-frame Doppler velocity dimension expansion. It can be seen from the figure that after the Doppler velocity dimension is introduced, due to the relative motion between the vehicle and the fixed obstacle, the vehicle and the obstacle are in Doppler Effective distinction can be made in the speed dimension. The vehicle cluster is labeled as cluster 2, and the adjacent obstacle cluster is labeled as cluster 5. Figure 10 shows the clustering result restored to the current frame. It can be seen from the figure that on the two-dimensional position plane, the vehicle echo signal and the obstacle signal are close to each other, but the proposed method can still effectively distinguish between the two. Vehicle clustering The label is cluster 2, and the adjacent obstacle cluster is labeled cluster 5. Figure 11 shows the clustering results based on OPTICS on the basis of two-dimensional position data. It can be seen that the method cannot divide vehicles and obstacles, but divides them into clusters labeled cluster 5. Through comparison, it can It can be seen that the proposed method can solve the clustering problem when the distance between the target and obstacles is small to a certain extent.

Figure 12 is a photo of the on-site environment corresponding to test scenario three. There are irregularly distributed fixed obstacles on both sides of the road, and small motor vehicles appear in front. Figure 13 shows the processing results of the data clustering method based on multi-frame Doppler velocity expansion. It can be seen from the figure that after introducing multi-frame data, small targets with a small number of echoes can be effectively clustered in three-dimensional space. The class label is cluster 2. Figure 14 shows the clustering result restored to the current frame. It can be seen from the figure that the number of echoes from the target in this frame data is only 1, but the proposed method can effectively distinguish it from noise points and form a single point cluster, labeled Cluster 2. Figure 15 shows the clustering result using the traditional OPTICS method. It can be seen from the figure that the target echo is divided into noise. According to the principle of the OPTICS method, it will be divided into noise when the core point density is greater than 1. If the core point is set If the point density is 1, all noise will be listed as valid targets, which will greatly increase the amount of subsequent calculations. Through comparison, it can be seen that the millimeter wave radar data clustering method proposed by the present invention can solve the clustering problem when the number of small target echoes is small.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements, etc., made within the spirit and principles of the present invention, All should be included in the protection scope of the present invention.

Claims

[Corrected 28.09.2023 in accordance with Article 91]
A millimeter wave radar data clustering method based on multi-frame Doppler velocity expansion dimension, characterized in that it includes the following steps:

S1: Preprocess the input millimeter wave radar frame data, and normalize the three-dimensional coordinates of the preprocessed frame data;

After preprocessing and before normalization also includes:

Frame data is pushed onto the stack: When the number of frames stored in the frame data stack is <N, the current frame data is stored; when the number of frames = N, the clustering algorithm is executed; when the number of frames > N, the last frame data of the stack is deleted and Store the current frame data to the top of the stack to maintain the number of frame data in the stack as always N;

Frame data stack judgment: This step takes effect when the method starting cycle number is <N. Its function is to start the subsequent process only when the frame data stack is full, otherwise it will wait for new data input;

The three-dimensional coordinate normalization is specifically:

After the frame data passes the judgment condition, the format of the data in the stack converted into a matrix is as follows:

;

The first line stores the x position coordinate, the second line stores the y position coordinate, and the third line stores the Doppler velocity. The superscript is the frame number to which the data belongs, and the value range is 1~N;

S2: Obtain the normalized data clustering results by calling the OPTICS method;

S3: Extract the clustering data label from the normalized data clustering result, and match the multi-frame clustering data with the current frame data to obtain the current frame data clustering;

Step S3 is specifically as follows:

Index the original radar matrix data according to the normalized matrix column number, and store the three-dimensional coordinates of the data into clustering;

Compare the multi-frame clustering data with the information of each dimension of the current frame data. If they are completely consistent, the match is successful and the data in the cluster is retained. Otherwise, it is deleted, thereby completing the current frame data clustering.
[Correction 28.09.2023 under Rule 91]
The millimeter wave radar data clustering method according to claim 1, characterized in that before inputting frame data, parameter initialization is also required: setting the number of delayed frames N, OPTICS method parameters and data preprocessing parameters.
[Correction 28.09.2023 under Rule 91]
The millimeter wave radar data clustering method according to claim 2, wherein the data preprocessing parameters include: energy threshold Pmin, angle range of interest [θmin θmax] and distance range [rmin rmax];

The millimeter-wave radar frame data includes: polar diameter r, angle θ, Doppler velocity v and energy I in polar coordinates of the observation point.
[Correction 28.09.2023 under Rule 91]
The millimeter wave radar data clustering method according to claim 2, wherein the preprocessing includes: using the set data preprocessing parameters as data filtering conditions, and sequentially judging the energy, angle and distance of the input points. , if the conditions are met at the same time, it is retained, and if it is not met, it is deleted.
[Correction 28.09.2023 under Rule 91]
The millimeter wave radar data clustering method according to claim 1, characterized in that two different normalization methods are used for processing the position dimension and the velocity dimension:

Positional dimension data normalization: using linear transformation method, based on the maximum value in the array with minimum value , according to the formula Transform the whole to the range of 0~1;

Speed dimension data normalization: obtaining speed through statistical methods , where the speed dimension normalization function expression is as follows: ;Formula is the attenuation coefficient, used to control the Euclidean distance between the moving target and the fixed target;

The normalized data is: .
[Correction 28.09.2023 under Rule 91]
The millimeter wave radar data clustering method according to claim 1, wherein step S2 specifically includes:

According to the formula Obtain the Euclidean distance between data as the basis for OPTICS algorithm clustering;

Call the OPTICS method to obtain the normalized data clustering results, and generate a structure for each cluster. The data label is the normalized matrix column number.
[Correction 28.09.2023 under Rule 91]
A millimeter wave radar data clustering system that implements the millimeter wave radar data clustering method described in any one of claims 1 to 6 based on multi-frame Doppler velocity dimensionality expansion, characterized in that it includes sequentially connected multi-frame data processing module, OPTICS-based three-dimensional data clustering module and current frame clustering data recovery module,

The multi-frame data processing module is used to implement millimeter wave radar multi-frame data preprocessing, data storage and shifting, Doppler velocity dimension nonlinear normalization, and position dimension linear normalization operations;

The OPTICS three-dimensional data clustering module is used to implement three-dimensional data clustering based on the OPTICS method;

The current frame clustering data recovery module is used to match multi-frame clustering data with the current frame data, and output the clustering results to subsequent modules for target matching or tracking.