A multi-target detection method based on millimeter wave radar

Abstract

This invention belongs to the field of autonomous driving technology, specifically providing a multi-target detection method based on millimeter-wave radar, applicable to multi-target detection in close-range scenarios. Building upon existing MUSIC algorithms, this invention performs azimuth estimation on the range-Doppler spectrum, solving the matching problem between the range, velocity, and azimuth of radar measurement points, and effectively improving azimuth resolution. Furthermore, it innovatively proposes using the continuity of azimuth angles for the same target point while clutter points exhibit disordered azimuth angles to filter out clutter points, reducing the likelihood of clutter being identified as targets. Simultaneously, it employs a combined AP clustering and DBSCAN clustering algorithm to cluster the filtered target points, eliminating the influence of the number of target points, reducing the probability of missed target detection, and improving target detection accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and more specifically, to an improved multi-target detection method of angle estimation and point-trace aggregation, which can be used for millimeter-wave radar multi-target detection in close-range scenarios. Background Art

[0002] In recent years, the automotive industry's demand for active safety and intelligence has grown dramatically, leading to more and more companies and research institutions participating in the promotion and development of driverless technology. Currently, no vehicle can achieve fully autonomous driving, but some models have been able to complete highly autonomous driving behaviors in experiments, such as starting, accelerating, braking, collision avoidance, parking, etc. In driverless technology, sensors are used to perceive the surrounding environment, moving vehicles and pedestrians, and respond in real time to ensure that the vehicle can drive normally, safely and reliably on the road.

[0003] At present, unmanned vehicle manufacturers such as Google, Uber, and Baidu mainly use sensor systems that combine millimeter-wave radar, cameras, and lidar. Due to the high cost of lidar, semi-automatic driving L2 is still the main type of vehicles actually put into production. The typical configuration of semi-automatic driving is 1V1R, that is, a multi-sensor fusion system that combines cameras and millimeter-wave radar. After the CMOS sensor replaces the CCD sensor, the cost of the camera is greatly reduced, and the outline of the target can be detected and recognized. However, the camera has poor performance in ranging and speed measurement, and is easily affected by the environment, such as in clouds, fog, rain, snow, hail and other weather conditions. It may fail. Lidar is also susceptible to harsh environments, so a sensor that can work all day and all day and is not easily affected by the environment is needed, that is, millimeter-wave radar. Millimeter-wave radar has the advantages of short wavelength, wide bandwidth, and strong penetration ability. Therefore, it has extremely high resolution and strong anti-interference ability in ranging and speed measurement. It is a sensor that will be widely used in unmanned driving technology.

[0004] Unmanned driving technology emphasizes real-time performance, which means that it needs to perceive and respond to changes in the surrounding environment in real time. Therefore, the commonly used multi-sensor fusion system is mainly based on a distributed fusion system structure. In a centralized fusion system, each sensor transmits the original observation data to the fusion center, which leads to a sharp increase in data transmission and computing, and cannot meet the real-time requirements within a controllable cost. In a distributed fusion system, each sensor has its own processor, which processes the observation data locally and then sends it to the fusion center. In a distributed architecture, due to the characteristics of millimeter-wave radar, multiple points will be generated for the same target in actual detection. Millimeter-wave radar needs to use a clustering algorithm to cluster the observation points to generate observation targets, which will lead to target missed detection and false detection in the processing process of the millimeter-wave radar sensor; target false detection is due to the presence of interference such as clutter, which easily causes the radar sensor to mistake clutter points for targets; target missed detection is due to weak echo signals, resulting in too few target points being mistaken for clutter. However, the current detection methods related to millimeter-wave radars do not solve the problems of radar target false detection and missed detection, and are not suitable for actual scene requirements. Summary of the invention

[0005] The purpose of the present invention is to provide a millimeter wave radar target detection method based on azimuth continuity and an improved clustering algorithm, in view of the fact that the existing millimeter wave radar related detection methods cannot well solve the problem of radar target misdetection and missed detection. The present invention is based on the existing MUSIC algorithm, and azimuth estimation is performed on the range-Doppler spectrum, which not only solves the problem of matching the range velocity and azimuth of radar measurement points, but also effectively improves the azimuth resolution. In addition, it is innovatively proposed to filter out clutter points by utilizing the characteristics that the azimuth of the same target point is continuous while the azimuth of clutter points is disordered, thereby reducing the possibility of clutter being identified as a target. At the same time, an algorithm combining AP clustering and DBSCAN clustering is used to cluster the target points after filtering the clutter, which is no longer affected by the number of target points, reduces the probability of missed detection of the target, and improves the target detection accuracy and robustness.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] A multi-target detection method based on millimeter wave radar, characterized in that it includes the following steps:

[0008] S1: Analyze radar echo data and obtain range-Doppler spectrum through fixed clutter cancellation and range-Doppler FFT transformation;

[0009] S2: Take the modulus of the range-Doppler spectrum of the multi-receiving antenna and perform non-coherent accumulation, and perform two-dimensional CFAR detection to obtain the radar measurement point trace;

[0010] S3: For each radar measurement point (i, j), the azimuth of the radar measurement point is estimated using a multiple signal classification algorithm, and interference points are filtered out according to the continuity of the azimuth to obtain the target point;

[0011] S4: Normalize the distance, speed and azimuth of all target points, and then combine the AP clustering algorithm and the DBSCAN clustering algorithm to achieve point aggregation and obtain the final clustering target.

[0012] Furthermore, in step S3, the specific process of filtering out interference points according to the continuity of the azimuth angle is as follows:

[0013] Define an azimuth matrix with the same dimension as the distance-Doppler spectrum of the receiving antenna and initialize it to a matrix of all negative 1s, then fill the azimuth estimation values ​​of the radar measurement point (i, j) into the azimuth matrix in sequence;

[0014] Then the azimuth-rectangular window method is used to iterate the azimuth matrix: for any non-negative 1 measurement point in the azimuth matrix, the measurement point in the neighborhood of the measurement point is defined as the reference unit, and the comparison threshold eps is set. The absolute value of the azimuth difference between the reference unit and the measurement point is compared to see whether it is less than or equal to the comparison threshold eps. If there is a reference unit that meets the conditions in the field, the measurement point is determined to be the target point, otherwise it is an interference point and is filtered out;

[0015] Furthermore, the specific process of step S4 is as follows:

[0016] S4.1: For each target point (i, j), calculate the distance and speed of the target point, and normalize the distance, speed and azimuth of all target points; the normalized results of the distance, speed and azimuth of the target point (i, j) constitute its feature vector;

[0017] S4.2: Use the AP clustering algorithm to perform initial clustering on the feature vectors of all target points to obtain AP clustering targets;

[0018] S4.3: For each AP cluster target, perform denormalization processing to obtain the distance, speed and azimuth of the cluster target, and calculate the position (x, y) of the cluster target in the rectangular coordinate system based on the distance and azimuth;

[0019] S4.4: Set the minimum number of samples to 1 and the neighborhood to 0.6m, perform secondary clustering (DBSCAN) based on the positions of all cluster targets in the rectangular coordinate system, and obtain the final cluster target.

[0020] Furthermore, in step S1, a non-recursive secondary canceller is used for fixed clutter cancellation.

[0021] Furthermore, in step S2, the two-dimensional CFAR detection is specifically as follows: presetting a reference unit, a protection unit and a scale factor, and determining whether the unit to be detected is a radar target point trace:

[0022]

[0023] Where D is the power of the unit to be detected, α is the scale factor, β 2 is the interference power estimate, calculated from the frequency mean of the reference unit.

[0024] Furthermore, in step S3, the multiple signal classification algorithm is specifically:

[0025] S3.1: According to the radar measurement point (i, j), take the spectrum value of the i-th row and j-th column in the range-Doppler spectrum of the k-th receiving antenna as the response frequency r of the radar measurement point k ,

[0026] S3.2: Combine the response frequencies of the radar measurement points according to the order of receiving antennas to obtain the frequency vector of the radar measurement points: r = (r1, r2, ..., r K ), K represents the number of receiving antennas, k = 1, 2, ..., K;

[0027] S3.3: Perform correlation processing on the frequency vector r to obtain the covariance matrix: R = rr H ;

[0028] S3.4: Perform eigenvalue decomposition on the covariance matrix R to obtain multiple eigenvalues ​​λ1,λ2,λ3,...,λ K and the corresponding eigenvectors υ1,υ2,υ3,...,υ K ;

[0029] S3.5: Assuming that the radar measurement point trace contains only a single target, the eigenvector corresponding to the maximum eigenvalue is the target eigenvector, and the remaining N-1 eigenvectors are all noise eigenvectors. Therefore, the noise eigenvectors are sorted and combined in descending order of eigenvalue to form a noise subspace: represents the noise feature vector, k=1,2,...,K-1;

[0030] S3.6: Given the ratio of the receiving antenna spacing to the radar wavelength as d, and the angle of the target signal source relative to the radar as θ, the steering vector is: A = [1, e -j2πdsin(θ) ,...,e -j2π(K-1)dsin(θ) ];

[0031] S3.7: Construct spatial spectrum function:

[0032] S3.8: Traverse the angle θ and find the θ corresponding to the peak of the spatial spectrum function P(θ) as the azimuth angle estimate of the radar measurement point (i, j).

[0033] The beneficial effects of the present invention are:

[0034] The present invention provides a multi-target detection method based on millimeter wave radar, which directly estimates the azimuth angle using the MUSIC algorithm on the range-Doppler spectrum, and on this basis, uses the continuity of the azimuth angle to filter out clutter traces, and then uses an algorithm combining AP clustering and DBSCAN clustering to cluster the radar measurement traces after filtering out the clutter. The above three algorithms respectively solve the matching problem of the range speed and azimuth angle of the radar measurement traces, the false detection problem of the radar measurement traces, and the missed detection problem of the clustered targets, which not only improves the detection accuracy of the target, but also reduces the false detection rate and missed detection rate of the target. In summary, the present invention has the advantages of high radar detection accuracy and high detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the multi-target detection method based on millimeter wave radar in the present invention.

[0036] Figure 2 It is a schematic diagram of the rectangular window method in two-dimensional CFAR in the present invention.

[0037] Figure 3 It is a schematic diagram of the azimuth angle-rectangular window method in the present invention.

[0038] Figure 4 This is a trajectory diagram of single target detection in an embodiment of the present invention.

[0039] Figure 5 It is a trajectory diagram of dual target detection in an embodiment of the present invention.

[0040] Figure 6 Graph showing the trajectories of three target detections in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention is described in detail below based on the accompanying drawings and preferred embodiments to make the purpose and effect of the present invention clearer and more complete; wherein the abbreviations and key words are defined as follows:

[0042] FFT (Fast Fourier Transform),

[0043] CFAR (False Alarm Rate, constant false alarm detection),

[0044] MUSIC (Multiple Signal Classification),

[0045] DBSCAN (Density-Based Spatial Clustering of Applications with Noise),

[0046] AP (Affinity propagation, neighbor propagation clustering).

[0047] This embodiment provides a multi-target detection method based on millimeter wave radar, such as Figure 1 As shown, the specific steps include:

[0048] S1: Analyze the radar echo data, and obtain the range-Doppler spectrum through fixed clutter cancellation and range-Doppler FFT transformation, as follows:

[0049] S1.1: Analyze radar sample data into processable echo data: Convert binary sample data into decimal echo data;

[0050] S1.2: For the echo data of the receiving antenna, a non-recursive secondary canceller is used to eliminate the fixed ground clutter. The specific formula is as follows:

[0051] y t =x t -2x t-1 +x t-2

[0052] Among them, x t represents the sampling data of the t-th period chirp signal, t = 3, 4, ..., T, T is the number of chirp signals emitted by the radar within the pulse interval;

[0053] S1.3: Rearrange the radar echo data into an echo matrix according to the radar sampling period;

[0054] S1.4: Perform range dimension FFT transformation on the echo matrix to obtain a range FFT matrix;

[0055] S1.5: Perform Doppler dimension FFT transformation on the range FFT matrix to obtain the range-Doppler spectrum;

[0056] S2: Take the modulus of the range-Doppler spectrum of the multi-receiving antenna and perform non-coherent accumulation, and perform two-dimensional CFAR detection to obtain the radar measurement point trace, as follows:

[0057] S2.1: taking the modulus of the range-Doppler spectra of multiple receiving antennas and accumulating them to obtain a non-coherently accumulated range-Doppler spectrum;

[0058] S2.2: A two-dimensional CFAR detection is performed on the range-Doppler spectrum after non-coherent accumulation using a rectangular window method to obtain all radar measurement points; the two-dimensional CFAR detection is as follows: Figure 2 As shown, specifically: preset reference unit, protection unit and scale factor to determine whether the unit to be detected is a radar target point trace:

[0059]

[0060] Where D is the power of the unit to be detected, α is the scale factor, β 2 is the interference power estimate, calculated from the frequency mean of the reference unit;

[0061] S3: For each radar measurement point (i, j), the azimuth of the radar measurement point is estimated using a multiple signal classification algorithm, and interference points are filtered out according to the continuity of the azimuth; the details are as follows:

[0062] S3.1: According to the radar measurement point (i, j), take the spectrum value of the i-th row and j-th column in the range-Doppler spectrum of the k-th receiving antenna as the response frequency r of the radar measurement point k ,

[0063] S3.2: Combine the response frequencies of the radar measurement points according to the order of receiving antennas to obtain the frequency vector of the radar measurement points: r = (r1, r2, ..., r K ), K represents the number of receiving antennas, k = 1, 2, ..., K;

[0064] S3.3: Perform correlation processing on the frequency vector r to obtain the covariance matrix: R = rr H ;

[0065] S3.4: Perform eigenvalue decomposition on the covariance matrix R to obtain multiple eigenvalues ​​λ1,λ2,λ3,...,λ K and the corresponding eigenvectors υ1,υ2,υ3,...,υ K ;

[0066] S3.5: Assuming that the radar measurement point trace contains only a single target, the eigenvector corresponding to the maximum eigenvalue is the target eigenvector, and the remaining N-1 eigenvectors are all noise eigenvectors. Therefore, the noise eigenvectors are sorted and combined in descending order of eigenvalue to form a noise subspace: represents the noise feature vector, k=1,2,...,K-1;

[0067] S3.6: Given the ratio of the receiving antenna spacing to the radar wavelength as d, and the angle of the target signal source relative to the radar as θ, the steering vector is: A = [1, e -j2πdsin(θ) ,...,e -j2π(K-1)dsin(θ)];

[0068] S3.7: Construct spatial spectrum function:

[0069] S3.8: traverse the angle θ and find the θ corresponding to the peak of the spatial spectrum function P(θ) as the azimuth angle estimate of the radar measurement point (i, j);

[0070] S3.9: define an azimuth matrix of the same dimension as the distance-Doppler spectrum of the receiving antenna and initialize it to a matrix of all negative 1s, and then fill the azimuth estimation values ​​of the radar measurement point (i, j) in the azimuth matrix in sequence;

[0071] S3.10: Iterate the azimuth matrix through the azimuth-rectangular window method to filter out interference points; the azimuth-rectangular window method is as follows Figure 3 As shown, specifically: for any non-negative 1 measurement point trace in the azimuth matrix, define the measurement point trace in the neighborhood of the measurement point trace as a reference unit, set a comparison threshold eps (in this embodiment, the comparison threshold eps is 3°), compare whether the absolute value of the azimuth difference between the reference unit and the measurement point trace is less than or equal to eps, if there is a reference unit that meets the conditions in the field, then determine that the measurement point trace is a target point trace, otherwise it is an interference point trace and is filtered out;

[0072] S4: Normalize the distance, speed and azimuth of all target points, and then combine the AP clustering algorithm and the DBSCAN clustering algorithm to achieve point condensation, as follows:

[0073] S4.1: For each target point (i, j), calculate the distance and speed of the target point, and normalize the distance, speed and azimuth of all target points; the normalized results of the distance, speed and azimuth of the target point (i, j) constitute its feature vector;

[0074] S4.2: Use the AP clustering algorithm to perform initial clustering on the feature vectors of all target points to obtain AP clustering targets;

[0075] S4.3: For each AP cluster target, perform denormalization processing to obtain the distance, speed and azimuth of the cluster target, and calculate the position (x, y) of the cluster target in the rectangular coordinate system based on the distance and azimuth;

[0076] S4.4: Set the minimum number of samples to 1 and the neighborhood to 0.6m, perform secondary clustering (DBSCAN) based on the positions of all cluster targets in the rectangular coordinate system, and obtain the final cluster target.

[0077] The beneficial effects of the present invention are further illustrated below in conjunction with tests:

[0078] 1. Test conditions:

[0079] In this embodiment, the millimeter-wave radar used is the NXP MR3003 millimeter-wave radar, and the millimeter-wave radar host computer is used for data sampling; the transmitting antenna gain used is 15dBm and the receiving antenna gain is 14dBm; the carrier frequency used by the millimeter-wave radar is 76500MHz, the maximum detection distance is 20m, and the maximum detection speed is 25km / h.

[0080] 2. Test content:

[0081] In this embodiment, a self-made data set is used to sample radar data through the host computer platform. The data set evaluation indicators are three evaluation indicators: target detection rate, missed detection rate, and false detection rate. At the same time, the present invention uses the document "Research on Vehicle-mounted Millimeter-wave Radar Target Detection Technology" as a comparative example to compare the detection results. In this comparative example, the detection probability of a single target is 91%, the detection probability of a double target is 69%, and the detection probability of a triple target is 51%, and the number of samples is 100. Therefore, this embodiment uses a single target, a double target, and a triple target for testing, and the number of samples remains the same as 100. The target detection probability statistics table is shown in the following table:

[0082] Table 1: Statistical table of radar detection rate, missed detection rate and false detection rate in this embodiment

[0083]

[0084] Table 2: Comparison of detection rate indicators of the present invention and the comparative example

[0085]

[0086] In the comparative example, the target azimuth information is obtained by using angle-dimensional FFT transformation. In comparison, the present invention uses an improved MU SIC algorithm to calculate the azimuth, so that the azimuth estimation has high resolution. At the same time, Figure 3 As shown in the figure, the continuity of the sampling azimuth angle is used to filter out clutter, which can greatly reduce the probability of false detection of the target; and, by combining the AP clustering algorithm and the DBSCAN secondary clustering algorithm, the probability of missed detection of the target can be effectively reduced; finally, the present invention has a higher target detection rate, a lower missed detection rate and a lower false detection rate, as shown in Table 1; more specifically, compared with the indicators of the comparative example, the correct detection rate of the target in the present invention is increased by 3% in single target detection, 27% in multi-target detection, and 40% in triple target detection, that is, the present invention can detect the position and speed of the target more accurately, and then obtain the radar detection trajectory. The radar detection trajectory diagrams of the single target, multi-target and triple targets in this embodiment are shown in FIG. Figure 4 , Figure 5 , Figure 6 shown.

[0087] The above description is only a specific implementation mode of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other alternative features that are equivalent or have similar purposes; all the disclosed features, or all the steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A multi-target detection method based on millimeter wave radar, characterized in that: The following steps are involved: S1: Analyze radar echo data and obtain range-Doppler spectrum through fixed clutter cancellation and range-Doppler FFT transformation; S2: Take the modulus of the range-Doppler spectrum of the multi-receiving antenna and perform non-coherent accumulation, and perform two-dimensional CFAR detection to obtain the radar measurement point trace; S3: For each radar measurement point (i, j), the azimuth of the radar measurement point is estimated using the multiple signal classification algorithm, and the interference point is filtered out according to the continuity of the azimuth to obtain the target point. The specific process of filtering out the interference point according to the continuity of the azimuth is: Define an azimuth matrix with the same dimension as the distance-Doppler spectrum of the receiving antenna and initialize it to a matrix of all negative 1s, then fill the azimuth estimation values ​​of the radar measurement point (i, j) into the azimuth matrix in sequence; Then the azimuth-rectangular window method is used to iterate the azimuth matrix: for any non-negative 1 measurement point in the azimuth matrix, the measurement point in the neighborhood of the measurement point is defined as the reference unit, and the comparison threshold eps is set. The absolute value of the azimuth difference between the reference unit and the measurement point is compared to see whether it is less than or equal to the comparison threshold eps. If there is a reference unit that meets the conditions in the field, the measurement point is determined to be the target point, otherwise it is an interference point and is filtered out; S4: Normalize the distance, speed and azimuth of all target points, and then combine the AP clustering algorithm and the DBSCAN clustering algorithm to achieve point aggregation and obtain the final clustering target.

2. The multi-target detection method based on millimeter wave radar according to claim 1, characterized in that: The specific process of step S4 is: S4.1: For each target point (i, j), calculate the distance and speed of the target point, and normalize the distance, speed and azimuth of all target points; the normalized results of the distance, speed and azimuth of the target point (i, j) constitute its feature vector; S4.2: Use the AP clustering algorithm to perform initial clustering on the feature vectors of all target points to obtain AP clustering targets; S4.3: For each AP cluster target, perform denormalization processing to obtain the distance, speed and azimuth of the cluster target, and calculate the position (x, y) of the cluster target in the rectangular coordinate system based on the distance and azimuth; S4.4: Set the minimum number of samples to 1 and the neighborhood to 0.6m, perform secondary clustering based on the positions of all cluster targets in the rectangular coordinate system, and obtain the final cluster target.

3. The multi-target detection method based on millimeter wave radar according to claim 1, characterized in that: In step S1, a non-recursive secondary canceller is used for fixed clutter cancellation.

4. The multi-target detection method based on millimeter wave radar according to claim 1, characterized in that: In step S2, the two-dimensional CFAR detection is specifically as follows: presetting a reference unit, a protection unit and a scale factor, and determining whether the unit to be detected is a radar target point trace: Where D is the power of the unit to be detected, α is the scale factor, β 2 is the interference power estimate, calculated from the frequency mean of the reference unit.

5. The multi-target detection method based on millimeter wave radar according to claim 1, characterized in that: In step S3, the multiple signal classification algorithm is specifically: S3.1: According to the radar measurement point (i, j), take the spectrum value of the i-th row and j-th column in the range-Doppler spectrum of the k-th receiving antenna as the response frequency r of the radar measurement point k , S3.2: Combine the response frequencies of the radar measurement points according to the order of receiving antennas to obtain the frequency vector of the radar measurement points: r = (r1, r2, ..., r K ), K represents the number of receiving antennas, k = 1, 2, ..., K; S3.3: Perform correlation processing on the frequency vector r to obtain the covariance matrix: R = rr H ; S3.4: Perform eigenvalue decomposition on the covariance matrix R to obtain multiple eigenvalues ​​λ1,λ2,λ3,...,λ K and the corresponding eigenvectors υ1,υ2,υ3,...,υ K ; S3.5: Assuming that the radar measurement point trace contains only a single target, the eigenvector corresponding to the maximum eigenvalue is the target eigenvector, and the remaining N-1 eigenvectors are all noise eigenvectors. Therefore, the noise eigenvectors are sorted and combined in descending order of eigenvalue to form a noise subspace: represents the noise feature vector, k=1,2,...,K-1; S3.6: Given the ratio of the receiving antenna spacing to the radar wavelength as d, and the angle of the target signal source relative to the radar as θ, the steering vector is: A = [1, e -j2πdsin(θ) ,...,e -j2π(K-1)dsin(θ) ]; S3.7: Construct spatial spectrum function: S3.8: Traverse the angle θ and find the θ corresponding to the peak of the spatial spectrum function P(θ) as the azimuth angle estimate of the radar measurement point (i, j).

Images