A Crossing Target Detection Method Based on MIMO Millimeter Wave Radar

Abstract

The present invention discloses a method for detecting a crossing target based on a MIMO millimeter-wave radar, which is applied to the field of target detection in an obstacle environment. The method addresses the problem that the existing technology does not perform full-process detection when detecting a crossing target and does not consider the multipath interference caused by obstacles in the scene. The present invention establishes an echo model based on the electromagnetic propagation mechanism of the crossing target in the non-line-of-sight area behind the obstacle and in the line-of-sight area in front of the radar. The method proposes a method for detecting the entire movement process of the crossing target, firstly acquiring 3D point cloud information of the scene based on the radar echo data, then using a reflection boundary estimation algorithm based on a static point cloud with zero velocity in the scene to estimate the boundary of a strong reflective surface in the scene, then using a ghost removal algorithm based on a dynamic point cloud with non-zero velocity in the scene and the acquired reflection boundary information to eliminate ghosts in the dynamic point cloud, and finally tracking the crossing target based on the center of the dynamic point cloud after ghost removal.

Description

Technical Field

[0001] The present invention belongs to the technical field of millimeter wave radar target detection, and in particular relates to a technology for detecting a crossing target in an obstacle environment. Background Art

[0002] With the rapid development of autonomous driving technology, the safety of drivers and pedestrians is receiving increasing attention. Accurate detection of moving targets, such as pedestrians, can provide early warning information to autonomous vehicles or drivers, enabling them to take effective measures to avoid traffic accidents. However, in urban or other congested environments, targets are often obscured by obstacles, making traditional line-of-sight detection methods inapplicable. A specific example is the detection of targets crossing behind a vehicle. In this scenario, when using radar to detect targets, parked vehicles not only make it difficult to detect crossing targets in non-line-of-sight areas, but also make crossing targets in line-of-sight areas susceptible to multipath signal interference.

[0003] Domestic and foreign research institutions have also conducted relevant research on the above-mentioned target detection scenarios. In 2012, A. Bartsch et al. used density images and frequency images based on a 77 GHz millimeter-wave radar to perceive targets, but failed to achieve good detection results in the cross-target scenario (A. Bartsch, F. Fitzek, and R. H. Rashofer, “Pedestrian recognition using automotive radar sensors,” Advances in Radio Science, vol. 10, pp. 45–55, 2012.); in 2014, M. Heuer et al. used a pre-detection tracking algorithm and a particle filter algorithm based on a 24 GHz millimeter-wave radar to achieve detection and tracking of cross-targets (M. Heuer, A. Al-Hamadi, A. Rain, and M.-M. Meinecke, “Detection and tracking approach using an automotive radar to increase active pedestrian safety,” in 2014 IEEE Intelligent Vehicles. Symposium Proceedings, 2014, pp. 890–893.); in 2019, A. Palffy et al. achieved early detection of crossing targets behind the vehicle based on a fusion sensor of millimeter-wave radar and lidar (A. Palffy, JFP Kooij, and DMG avrila, “Occlusion aware sensor fusion for early crossing pedestrian detection,” in 2019 IEEE Intelligent Vehicles Symposium (IV), 2019, pp. 1768–1774.).

[0004] However, none of the aforementioned studies focused on what happens after a target crosses the non-line-of-sight (NLOS) zone. When a target crosses the NLOS zone and enters the line-of-sight zone, multipath signals generated by obstacles create ghosting. These ghosting signals can interfere with the detection of the actual target, leading to errors in the detection and tracking of the target. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes a crossing target detection method based on MIMO millimeter-wave radar, which can realize the effective detection and tracking of the entire movement process of crossing targets from the non-line-of-sight area to the line-of-sight area in an obstacle environment.

[0006] The technical solution adopted by the present invention is: a method for detecting crossing targets based on MIMO millimeter-wave radar, the application scenarios of which include: MIMO millimeter-wave radar, parked vehicles, and moving targets moving from the non-line-of-sight area in front of the parked vehicle to the line-of-sight area in front of the radar;

[0007] The detection method includes the following steps:

[0008] S1. Use MIMO millimeter-wave radar to transmit linear frequency modulation signals into the application scenario and receive echo signals;

[0009] S2. Preprocess the echo signal to obtain 3D point cloud information of the scene;

[0010] S3. Based on the static point cloud with zero velocity in the scene, the boundary of the strong reflective surface in the scene is estimated using a reflection boundary estimation algorithm, and the line segment parameter information describing the boundary of the reflective surface is extracted;

[0011] S4, based on the dynamic point cloud with non-zero speed in the scene and the line segment parameter information of the reflective surface boundary obtained in S3, ghost removal algorithm is used to eliminate ghosts in the dynamic point cloud;

[0012] S5. For the center of the dynamic point cloud after removing ghosts in S4, the crossing target is tracked based on the nearest neighbor data association algorithm and the Kalman filter tracking algorithm.

[0013] The present invention's beneficial effects include: The method accurately detects and tracks obscured, crossing targets throughout their entire motion from non-line-of-sight to line-of-sight areas. By processing static point clouds, scene information can be effectively acquired to assist in eliminating ghost interference signals in dynamic point clouds. Combining scene information extracted from static point clouds with a ghost removal algorithm effectively eliminates ghost interference. Field test results demonstrate that the method effectively detects and tracks crossing targets behind parked vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of electromagnetic propagation across target detection scenario;

[0015] Among them, (a) is a top view of the target in the non-line-of-sight area, (b) is a top view of the target in the line-of-sight area, (c) is a 3D image of the target in the non-line-of-sight area, and (d) is a 3D image of the target in the line-of-sight area;

[0016] Figure 2 This is the flow chart of the proposed crossing target detection method;

[0017] Figure 3 This is the flow chart of the reflection surface estimation algorithm;

[0018] Figure 4The radar point cloud distribution characteristics of the vehicle surface and the spatial neighborhood diagram of the improved DBSCAN algorithm;

[0019] (a) is the radar point cloud distribution feature map of the adjacent vehicle surface, and (b) is the spatial neighborhood diagram of the improved DBSCAN algorithm.

[0020] Figure 5 This is the flow chart of the ghost removal algorithm;

[0021] Figure 6 This is a diagram of the measured scene and the measured data processing results;

[0022] Among them, (a) is the experimental scene diagram, and (b) is the measured data processing result diagram. DETAILED DESCRIPTION

[0023] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.

[0024] The present invention proposes a method for detecting a crossing target based on MIMO millimeter wave radar. The detection scene is as follows: Figure 1 In the scenario, the MIMO millimeter-wave radar is located at point O, a vehicle is parked in front of the left side of the radar, and a target crosses from the front of the parked vehicle to the front of the radar. Figure 1 The non-line-of-sight region shown in (a) crosses into Figure 1 (b) The radar continuously emits electromagnetic waves into the scene to detect targets that cross the scene. When the target is in the non-line-of-sight area, such as Figure 1 As shown in (a), the electromagnetic propagation paths of the detected target mainly include the ground reflection path and the side vehicle diffraction path. Without considering the combined path, the target detection path from the radar to the target and then back to the radar is mainly as follows: Figure 1 (c) shows the ground reflection path P1: O→C→B→C→O, and the side vehicle diffraction path P2: O→A→B→A→O; (Due to the short wavelength of the millimeter wave radar, the side vehicle diffraction echo is weak, so the side vehicle diffraction path P2 is ignored). When the target is in the line of sight area, such as Figure 1 As shown in (b), the direct line of sight path can detect the target, but the sidecar reflection path also has strong reflected echoes, which strongly interfere with target detection (specifically, this multipath signal will cause ghosting in the subsequent radar signal processing, thereby affecting the detection and tracking of the real target). At this time, without considering the combined path, the electromagnetic propagation path from the radar to the target and then back to the radar is mainly as follows: Figure 1 (d) shows the direct view path P3: O→F→O, and the sidecar reflection path P4: O→E→F→E→O.

[0025] like Figure 2As shown, the processing flow of the method of the present invention includes the following steps:

[0026] Step 1: Echo signal modeling

[0027] Assume that the parameters of the linear frequency modulation signal transmitted by the radar are: carrier frequency f0, frequency modulation slope μ, pulse width T, signal bandwidth B, and signal amplitude A0. Then, the radar transmission signal is:

[0028] s(t)=A0exp(j2πf0t+jπμt 2 )u(t)

[0029] in, t is time, u(t) is a rectangular function:

[0030]

[0031] When the transmitted signal is a wide-bandwidth signal, extended targets such as the human body and obstacles in the scene can be regarded as a collection of many scattering points. Therefore, the delay-based echo model can be expressed as:

[0032]

[0033] Where s(t-τ) represents the signal after s(t) is delayed by τ, l represents the lth path of the detection target, and i represents the i-th scattering point of an extended target in the scene. represent the backscattering coefficient and echo delay of the target respectively, are the backscatter coefficient and echo delay of the parked vehicle, respectively, and ξ is the noise. Therefore, in the above formula, the first part represents the echo signal of the crossing target, and the second part represents the echo signal of the parked vehicle.

[0034] Based on the above electromagnetic propagation analysis, when the target is located in the non-line-of-sight area in front of the parked vehicle, the radar detects the target through path P1. At this time, l∈{l1}, the echo model can be rewritten as:

[0035]

[0036] When the target is in the line-of-sight area in front of the radar, the propagation path of the electromagnetic wave can be P3, P4. At this time, l∈{l3,l4}, and the echo model can be rewritten as:

[0037]

[0038] When the echo signal passes through the mixing and filtering operation, an intermediate frequency signal can be obtained, which is recorded as y(t).

[0039] Furthermore, assuming that the virtual array of the MIMO radar system is a Q-root uniform linear array, after digitally sampling the intermediate frequency signals of all receiving antennas, the echo data matrix can be obtained:

[0040]

[0041] Where n represents the sampling point, p represents the p-th chirp signal in a frame (p=1,...,P), and q represents the q-th receiving antenna (q=1,...,Q). p,q (n) represents the echo sampling signal of the p-th chirp signal of the q-th antenna in a frame.

[0042] Step 2: Data preprocessing to generate radar 3D point cloud

[0043] Data preprocessing mainly includes 2D-FFT, constant false alarm detection, arrival angle estimation, and point cloud generation.

[0044] First, a fast Fourier transform is performed along the slow time dimension to complete pulse compression. Then, a fast Fourier transform is performed along the fast time dimension to obtain a range-Doppler spectrum. Then, a constant false alarm detection algorithm is used to extract strong scattering points in the scene that pass the detection threshold. The distance and velocity information of the scattering points that pass the detection threshold are further calculated.

[0045] The arrival angle of the point target detected after CFAR detection is estimated along the antenna dimension to obtain the angle of the scattering point in the radar coordinate system. Based on the acquired distance and angle information, the two-dimensional spatial position of the scattering point is calculated through coordinate transformation. Combined with the scattering point's velocity information, the radar's 3D point cloud information is finally obtained.

[0046] Step 3: Based on the static point cloud with zero velocity in the scene, the reflection boundary estimation algorithm is used to estimate the boundary of the strong reflective surface in the scene and extract related parameters.

[0047] like Figure 3 As shown, specifically including the following:

[0048] First, the improved DBSCAN algorithm is used to cluster the static point cloud. When electromagnetic waves propagate to the surface of the adjacent car at a certain angle, only the discontinuous parts of the car, such as the corners, door handles, and rearview mirrors, can strongly reflect the electromagnetic signal. Figure 4(a) shows a set of point cloud distribution characteristics of the side of the adjacent vehicle in a measured environment; although the point cloud distribution representing the side of the vehicle is discontinuous, they have a strong linear distribution feature in spatial position, so the traditional DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is improved to adapt to the point cloud distribution characteristics of the vehicle. Compared with the traditional DBSCAN algorithm, in addition to the density parameter ε and the radius parameter Minpts, the improved DBSCAN algorithm adds ρ, θ, △θ parameters to adjust the spatial neighborhood of the cluster to adapt to the distribution characteristics of the static point cloud generated by the reflection boundary of the adjacent vehicle; where ρ is the desired fan length, θ is the angle between the desired fan normal and the positive x-axis, and △θ is the deviation of the desired angle; Figure 4 (b) shows the spatial neighborhood of the improved DBSCAN algorithm; in the specific implementation of the algorithm, it is only necessary to incorporate the point cloud in the spatial neighborhood composed of the above three parameters into the spatial neighborhood of the traditional DBSCAN algorithm.

[0049] After clustering the static point cloud, the Hough transform algorithm is used to perform linear parameter fitting on each cluster obtained by clustering. Assume that after the point cloud in a cluster is processed by the Hough transform algorithm, the polar coordinate parameter information extracted is Then in the Cartesian coordinate system (two-dimensional space coordinate system), the line corresponding to the parameter information is y=kx+b, and the line parameters (k, b) are The relationship is:

[0050]

[0051] Then, using the boundary point cloud coordinates of each cluster as a reference, the coordinates of the two ends of the fitting line are determined, and the length of the line segment is calculated. Finally, based on the length information of the real car, a length reference benchmark is set to eliminate invalid line segments that are too long or too short.

[0052] From the above, we can obtain the line segment information used to characterize the boundary of the reflecting surface, including the slope-intercept parameter (k, b) (where k represents the slope of the line and b represents the intersection of the line and the y-axis), the length of the line segment, and the port coordinate information.

[0053] Step 4: Based on the dynamic point cloud with non-zero speed in the scene and the scene information obtained from step 3, ghost removal algorithm is used to remove ghosts in the dynamic point cloud.

[0054] like Figure 5 As shown in the figure, the method specifically includes: firstly, clustering the dynamic point cloud using the traditional DBSCAN algorithm, and then extracting the center coordinates of each cluster to represent the cluster. The center coordinates of a cluster are recorded as (xc ,y c ):

[0055]

[0056] Among them, N is the number of point clouds in the cluster, (x j ,y j ) is the jth point cloud in the cluster. Subsequently, the obtained cluster center is further processed.

[0057] The geometric characteristics of ghost distribution indicate that the positions of ghosts caused by sidecar reflections (double-pass single reflections) and real targets are mirror-symmetrical about the reflection boundary. Therefore, after obtaining the planar distribution of the reflection boundary within the scene, we can search for cluster center pairs that are mirror-symmetrical about the line segment representing the reflection boundary based on the position information, thus achieving associative matching of cluster centers.

[0058] Assuming that the continuous reflection boundary in the scene obtained by step 3 is y=kx+b, then the center point (x c ,y c )The mirror image position about the boundary is (x mir ,y mir ):

[0059]

[0060] Due to the measurement error of radar, the ghost image may not be exactly located at the mirror image position of the real target with respect to the reflecting surface. Therefore, when calculating the mirror image symmetry point (x0, y0) of a cluster midpoint with respect to a certain reflection boundary, mir ,y mir ), set a mirror position (x mir ,y mir ) as the center and a search area with a radius of △R. The cluster centers falling within this area are matched with the original cluster center point (x0, y0). △R is an empirical value and can be selected as 0.5m. All cluster centers are traversed to complete the above cluster center matching operation.

[0061] Ghost detection and removal are performed on the matched cluster centers. The identification criterion is: in a set of matching point pairs, cluster centers that are relatively far from the radar are marked as ghosts. After ghost detection, cluster centers marked as ghosts are removed; unmarked and unmatched cluster centers remain.

[0062] Step 5: Based on the cluster center after ghost removal in step 4 (the center of the dynamic point cloud after clustering), the nearest neighbor data association algorithm and the Kalman filter tracking algorithm are used to track the crossing target.

[0063] The state of the Kalman filter consists of the lateral position, longitudinal position, lateral velocity, and longitudinal velocity of the target, represented by x:

[0064] x=[x,y,v x ,v y ] T

[0065] in,[·] T Indicates transpose.

[0066] Considering the target motion as a uniform motion model, the state transfer matrix is:

[0067]

[0068] The measurement matrix is:

[0069]

[0070] Furthermore, the Kalman filter formula is used for prediction and update. The Kalman equations are as follows:

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] Where t represents the current time, t-1 represents the previous time, K represents the Kalman filter gain, P represents the state estimation error covariance matrix, z represents the measurement value vector, H represents the measurement matrix, Q represents the process noise covariance matrix, and R represents the measurement noise covariance matrix. They represent the predicted value of the state quantity at time t and the predicted value of the state estimation error covariance matrix, respectively, t 、P t , K t They represent the updated value of the state quantity at time t, the updated value of the state estimation error covariance matrix, and the updated value of the Kalman filter gain respectively.

[0077] Figure 6 (a) shows the experimental scenario adopted, Figure 6 (b) shows the processing results of the method of the present invention. Figure 6In (b), the right-hand reflection boundary segment of the car is the algorithm's estimated result, while the remaining three boundaries are the completed results. This shows that the boundary information of the reflecting surface of the parked vehicle to the left of the radar can be effectively estimated; ghost images caused by the reflected signal from the adjacent vehicle are effectively eliminated, effectively suppressing the interference of ghost images on the tracking of the real target; and the target can be correctly detected and tracked throughout its entire movement from the non-line-of-sight area in front of the parked vehicle to the line-of-sight area in front of the radar. Field experiments have verified the feasibility and effectiveness of this invention.

[0078] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A method for detecting a crossing target based on MIMO millimeter wave radar, characterized in that: Application scenarios include: MIMO millimeter-wave radar, parked vehicles, and moving targets from the non-line-of-sight area in front of the parked vehicle to the line-of-sight area in front of the radar; The detection method includes the following steps: S1. Use MIMO millimeter-wave radar to transmit linear frequency modulation signals into the application scene and receive echo signals; S2. Preprocess the echo signal to obtain 3D point cloud information of the scene; S3. Based on the static point cloud with zero velocity in the scene, a reflection boundary estimation algorithm is used to estimate the boundary of the strong reflective surface in the scene, and line segment parameter information describing the boundary of the reflective surface is extracted. Step S3 specifically includes the following sub-steps: S31, using the improved DBSCAN algorithm to cluster the static point cloud; Step S31 specifically includes incorporating the points in the spatial neighborhood of the traditional DBSCAN algorithm into the ρ , θ ,Δ θ The point cloud in the spatial neighborhood composed of three parameters; among them, ρ is the desired fan length, θ The angle between the desired fan normal and the positive x-axis, Δ θ is the deviation from the desired angle; S32, using the Hough transform algorithm to perform linear parameter fitting on each cluster obtained by clustering in step S31; S33, using the boundary point cloud coordinates of each cluster as a reference, determining the coordinates of the two ends of the fitting line and calculating the length of the line segment; S34, based on the length information of the real car, set a length reference benchmark, remove invalid line segments that are too long or too short; thereby obtaining line segment information used to characterize the boundary of the reflective surface; S4, based on the dynamic point cloud with non-zero speed in the scene and the line segment parameter information of the reflective surface boundary obtained in S3, ghost removal algorithm is used to remove ghosts in the dynamic point cloud; S5. For the center of the dynamic point cloud after removing ghosts in S4, the crossing target is tracked based on the nearest neighbor data association algorithm and the Kalman filter tracking algorithm.

2. The method for detecting a crossing target based on a MIMO millimeter wave radar according to claim 1, wherein: Step S4 specifically includes the following sub-steps: S41. Use the traditional DBSCAN algorithm to cluster the dynamic point cloud. After clustering, extract the center coordinates of each cluster to represent the cluster. S42, based on the position information of the line segment representing the reflection boundary, searching for a cluster center point pair that is mirror-symmetric about the line segment to achieve cluster center matching; S43, identifying and eliminating ghosts of the matched cluster centers.

Images