A 4D millimeter wave radar point cloud high-resolution imaging method

Abstract

The application provides a 4D millimeter wave radar point cloud high-resolution imaging method, which comprises the following steps: designing a suitable radar signal waveform to obtain radar sampling data; sequentially performing fast Fourier transform on the sampling data, and obtaining a distance-Doppler two-dimensional heat map through non-coherent accumulation; according to the relationship between radar echo power and target distance, adaptively improving a constant false alarm rate target detection algorithm to determine effective target points, which contain target distance and speed information; based on a two-dimensional virtual array formed according to the MIMO principle, jointly estimating the horizontal angle and the pitch angle of the target to obtain the accurate horizontal position and height information of the target; and removing noise and clustering targets through a density-based clustering algorithm to realize point cloud imaging. The 4D radar point cloud imaging method has the following advantages: the adaptive target detection algorithm is used to effectively extract target information, and to a certain extent, the interference of external environment, system hardware and other factors is inhibited; and the super-resolution algorithm and the multi-dimensional optimization search algorithm are used to realize high-resolution and high-precision angle estimation, and to reduce the operation complexity.

Description

Technical Field

[0001] This invention belongs to the field of automatic detection technology and relates to a high-resolution imaging method based on 4D millimeter-wave radar point clouds. Background Technology

[0002] With the development of intelligent driving technology, the requirements for vehicle environmental perception capabilities in autonomous driving systems are constantly increasing. Millimeter-wave radar, with its advantages of low cost, all-weather operation, and good environmental adaptability, has been widely used in Advanced Driver Assistance Systems (ADAS). However, traditional millimeter-wave radar suffers from poor angular resolution and very sparse point clouds, only able to output informational point targets, which is far from meeting the requirements of modern autonomous driving applications. Emerging high-resolution four-dimensional imaging radar (4D radar), as the main development direction for next-generation 77GHz automotive millimeter-wave radar, can provide multi-dimensional information such as range, horizontal angle, elevation angle, and velocity, thereby obtaining high-density point clouds and outputting target contour information. This is of great significance for subsequent research on signal processing algorithms such as clustering and target tracking. Currently, signal processing algorithms based on 4D radar are still immature, suffering from noise interference, low angular resolution, and low angular accuracy, affecting the imaging quality of millimeter-wave radar point clouds. Summary of the Invention

[0003] The high-resolution point cloud imaging algorithm proposed in this invention can obtain relatively clear contour information of the target and realize the reconstruction of the surrounding environment. Among them, the adaptive target detection algorithm can reduce noise interference to a certain extent and reduce the probability of missed targets; the angle joint estimation algorithm can obtain the accurate angle of the target and distinguish the target from two adjacent angles, effectively expanding the target point cloud.

[0004] To achieve the above objectives, this invention proposes a high-resolution imaging method based on 4D millimeter-wave radar point clouds, comprising:

[0005] Step 1: Use radar equipment to acquire raw sampling data, which contains information about target reflection;

[0006] Step 2: Perform fast time-dimensional fast Fourier transform and slow time-dimensional fast Fourier transform on the intermediate frequency signal in sequence. The fast time-dimensional FFT captures the instantaneous frequency change of the signal, and the slow time-dimensional FFT is used to capture the motion change of the target over time.

[0007] Step 3: Increase the signal-to-noise ratio through incoherent accumulation to obtain a range-Doppler heatmap, which shows the target's information in terms of range and velocity;

[0008] Step 4: Target detection is performed using an improved adaptive constant false alarm rate target detection algorithm to determine valid target points and obtain the target's distance and velocity information;

[0009] Step 5: Using the two-dimensional virtual array formed by the MIMO radar, perform joint angle estimation to obtain the target's horizontal and vertical angle information;

[0010] Step 6: Noise removal and target clustering are performed using a density-based clustering algorithm to form a 3D point cloud.

[0011] A further feature of step 4 is an adaptive constant false alarm rate (CFAR) target detection algorithm, which includes: adaptively improving the traditional CFAR algorithm by utilizing the relationship between radar received signal power and radar detection range, combined with a piecewise threshold coefficient method. Specifically, this includes:

[0012] The power equation for radar received echo signals:

[0013]

[0014]

[0015] Among them, P t It is the radar's output power, G TX / RX Let σ be the gain of the transmitting and receiving antennas, σ be the radar cross section (RCS) of the target, and λ be the signal wavelength.

[0016] Since the radar receives power through a square-law detector, the actual radar received power and the relative distance d between the target and the radar have the following relationship:

[0017]

[0018] The threshold coefficient of the adaptive CFAR detector is set based on the relationship between radar signal energy intensity and distance attenuation. Furthermore, a regional threshold coefficient setting method is used to assign different threshold coefficients to different regions, thereby generating high-density, high-quality spatial point clouds. The adaptive threshold coefficient function proposed in this invention is as follows:

[0019]

[0020] Where α=2N(P) fa -1 / 2N -1) is the initial threshold coefficient value.

[0021] A further feature, in step 5, is a high-precision, high-resolution joint angle estimation algorithm, which specifically includes:

[0022] First, the horizontal channel data acquired by the radar is extracted; then, according to the beamforming algorithm, the received horizontal channel signal data is weighted and summed to obtain the output signal, and then horizontal angle filtering is performed. The filtering strategy is as follows:

[0023] n peaks =P value(i) ≥max{P value}·(10^(-scale / 10))

[0024] Here, scale = 1; based on the obtained horizontal angle corresponding to the target point, the deterministic maximum likelihood (DML) super-resolution algorithm is used to estimate the initial value of the pitch angle; considering the complexity of the DML algorithm, the alternating projection algorithm (AP) is combined to transform the multidimensional search problem into multiple one-dimensional search problems, and finally estimate the accurate pitch angle of the target. The specific process of the pitch angle estimation algorithm includes: first, extracting the pitch dimension data of the horizontal angle corresponding to the target to obtain its covariance matrix; then, determining the initial value of the angle estimation using the DML algorithm; then, determining the iteration precision or the number of iterations, and iteratively solving using the AP algorithm to obtain the accurate pitch angle of the target.

[0025] This invention proposes a complete 4D radar high-resolution point cloud imaging algorithm, which achieves high-resolution and high-precision angle estimation, reduces some of the noise effects, generates a relatively dense point cloud image, and to a certain extent completes the restoration of the scene around the radar. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the point cloud high-resolution imaging algorithm based on 4D radar of the present invention.

[0027] Figure 2 This is a graph of the adaptive threshold coefficient function, which is an improvement on the traditional target detection algorithm of this invention.

[0028] Figure 3 This is a schematic diagram of a high-resolution joint angle estimation algorithm framework proposed in this invention;

[0029] Figure 4 This is a schematic diagram illustrating the specific process of the pitch angle estimation algorithm of this invention.

[0030] Figure 5 This is a flowchart illustrating the process of the density clustering algorithm of this invention.

[0031] Figure 6 This is a point cloud imaging result diagram according to an embodiment of the present invention. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0033] like Figures 1 to 6 As shown, a high-resolution imaging method based on 4D millimeter-wave radar point clouds according to a specific embodiment of the present invention includes the following steps:

[0034] See Figure 1 This invention proposes a complete high-resolution imaging algorithm based on 4D millimeter-wave radar point clouds, including the following steps:

[0035] S1: Use radar equipment to acquire raw sampling data, which contains information about target reflection;

[0036] S2: Perform fast time-dimensional fast Fourier transform and slow time-dimensional fast Fourier transform on the intermediate frequency signal in sequence. The fast time-dimensional FFT captures the instantaneous frequency change of the signal, and the slow time-dimensional FFT is used to capture the motion change of the target in time.

[0037] S3: Increase the signal-to-noise ratio through incoherent accumulation to obtain a range-Doppler heatmap, which shows the target's information in terms of range and velocity;

[0038] S4: Target detection is performed using an improved constant false alarm rate (CFAR) target detection algorithm to determine valid target points and obtain the target's distance and velocity information;

[0039] S5: Using a two-dimensional virtual array formed by MIMO radar, joint angle estimation is performed to obtain the target's horizontal and vertical angle information;

[0040] S6: Noise removal and target clustering are performed using a density-based clustering algorithm to preserve effective 3D point clouds and achieve point cloud imaging.

[0041] Specifically, step S4: This invention is based on a constant false alarm rate (CFAR) target detection algorithm with adaptive improvements. The mean-based CFAR algorithm calculates a local mean at each possible target location, then calculates a threshold based on the false alarm rate, and determines whether the detected signal strength exceeds the threshold to achieve target detection. If the detection unit is larger than the detection threshold, it is identified as a target; if the detection unit is smaller than the detection threshold, it is considered noise. When the threshold is set too low, clutter will be detected as a target, causing false alarms; when the threshold is set too high, the target may not be detected, causing missed alarms. Therefore, the selection of the threshold coefficient is crucial.

[0042] This invention adaptively allocates detection threshold coefficients based on the relationship between radar receiving antenna echo power and distance. Within the radar's primary detection area, it aims to present objects in that area as clearly as possible while minimizing the impact of interference signals. Due to inherent limitations of the radar system hardware, very high-energy interference can occur in areas very close to the radar; therefore, the threshold coefficient for these areas should be appropriately increased to suppress system noise caused by electronic components. For areas near the radar's primary detection area, air noise caused by radar signal attenuation due to atmospheric changes may occur; therefore, the threshold coefficient for these areas should be slightly larger than the threshold value for the primary detection area. This reduces the impact of atmospheric noise and the probability of missed target detection. The threshold coefficient should also be increased for areas outside the radar's primary detection range. (See also...) Figure 2 , is the image of the adaptive threshold coefficient function proposed in this invention.

[0043] In step S5, for the joint angle estimation method, please refer to [link / reference needed]. Figure 3 The present invention proposes a novel joint estimation algorithm framework, which specifically includes:

[0044] S7: Extract radar horizontal channel data;

[0045] S8: Obtain the horizontal angle based on the traditional beamforming algorithm (CBF);

[0046] S9: Based on the horizontal angle corresponding to the target point, the deterministic maximum likelihood (DML) super-resolution algorithm is used to estimate the pitch angle to obtain the initial value of the pitch angle;

[0047] S10: Solve using the alternating projection algorithm to obtain the final accurate value of the pitch angle.

[0048] For the detailed process of the pitch angle estimation algorithm, please refer to [link / reference]. Figure 4 ,include:

[0049] S11: Obtain the covariance matrix by receiving data;

[0050] S12: The initial value of the angle estimate is determined by the DML algorithm;

[0051] S13: Determine the iteration precision or the number of iterations;

[0052] S14: Solve using the AP algorithm and determine whether the iterative accuracy has been achieved;

[0053] S15: Stop iterating if the iteration accuracy is met; otherwise, repeat step S14.

[0054] The noise removal and target clustering method described in step S6: Due to the sparse and uneven characteristics of millimeter-wave radar point clouds, and the presence of noise and clutter, this invention employs the DBSCAN algorithm for target clustering and noise removal. Please refer to [link to relevant documentation]. Figure 5 The DBSCAN algorithm steps include:

[0055] S16: Mark all points as unobserved points;

[0056] S17: Randomly select an unobserved point and mark it as an observed point P;

[0057] S18: If the ε-neighborhood of point P contains at least MinPts points;

[0058] S19: Create a new cluster C and add point P to C. Let N be the set of objects in the ε-neighborhood of p.

[0059] S20: For all points P' in N, they are marked as observed points if the ε-neighborhood of point P' contains at least MinPts points;

[0060] S21: Add these objects to N, and if P' is not yet a member of any cluster, add P' to C;

[0061] S22: Otherwise, mark point P as a noise point;

[0062] S23: Until all points are marked as observed points.

[0063] See Figure 6 This embodiment is a radar point cloud image of a real scene obtained by the 4D millimeter-wave radar high-resolution point cloud imaging method proposed in this invention.

Claims

1. A high-resolution imaging method based on 4D millimeter-wave radar point clouds, characterized in that, Includes the following steps: Step 1: Use radar equipment to acquire raw sampling data, which contains information about target reflection; Step 2: Perform fast time-dimensional fast Fourier transform and slow time-dimensional fast Fourier transform on the intermediate frequency signal in sequence. The fast time-dimensional FFT captures the instantaneous frequency change of the signal, and the slow time-dimensional FFT is used to capture the motion change of the target over time. Step 3: Increase the signal-to-noise ratio through incoherent accumulation to obtain a range-Doppler heatmap, which shows the target's information in terms of range and velocity; Step 4: Target detection is performed using an improved adaptive constant false alarm rate target detection algorithm to determine valid target points and obtain the target's distance and velocity information; Step 5: Using the two-dimensional virtual array formed by the MIMO radar, perform joint angle estimation to obtain the target's horizontal and vertical angle information; Step 6: Noise removal and target clustering are performed using a density-based clustering algorithm to form a 3D point cloud; Step 4 includes an adaptive constant false alarm rate (CFAR) target detection algorithm, comprising: By utilizing the relationship between radar received signal power and radar detection range, and combining it with the piecewise threshold coefficient method, an adaptive improvement is made to the traditional constant false alarm rate target detection (CFAR) algorithm, specifically including: The power equation for radar received echo signals: in, This refers to the radar's output power; GTX and GRX represent the transmit and receive antenna gains, respectively. The radar cross section (RCS) of the target. It is the signal wavelength; Since the radar receives power through a square-law detector, the actual radar received power and the relative distance d between the target and the radar have the following relationship: The threshold coefficient of the adaptive CFAR detector is set based on the relationship between radar signal energy intensity and distance attenuation. Furthermore, a regional threshold coefficient setting method is used to assign different threshold coefficients to different regions, thereby generating high-density, high-quality spatial point clouds. The proposed adaptive threshold coefficient function is as follows: in, This is the initial threshold coefficient value.

2. The high-resolution imaging method based on 4D millimeter-wave radar point clouds according to claim 1, characterized in that, Step 5, a high-precision, high-resolution joint angle estimation algorithm, specifically includes: First, the horizontal channel data acquired by the radar is extracted; then, according to the beamforming algorithm, the received horizontal channel signal data is weighted and summed to obtain the output signal, and then horizontal angle filtering is performed. The filtering strategy is as follows: in, Based on the obtained horizontal angle corresponding to the target point, the deterministic maximum likelihood (DML) super-resolution algorithm is used to estimate the initial value of the pitch angle. Considering the complexity of the DML algorithm, the alternating projection algorithm (AP) is combined to transform the multidimensional search problem into multiple one-dimensional search problems, and finally the accurate pitch angle of the target is estimated. The specific process of the pitch angle estimation algorithm includes: first, extracting the pitch dimension data of the horizontal angle corresponding to the target to obtain its covariance matrix; then, determining the initial value of the angle estimation by the DML algorithm; then, determining the iteration precision or the number of iterations, and iteratively solving by the AP algorithm to obtain the accurate pitch angle of the target.

Images