Millimeter wave radar point cloud outlier detection method and system based on angular spectrum analysis

Abstract

The application discloses a kind of millimeter wave radar point cloud outlier detection method and system based on angular spectrum analysis, method includes: obtaining millimeter wave radar antenna array, the echo scattering point cloud obtained by angle estimation;L2 norm normalization is carried out to pointing vector;Local neighborhood is constructed based on K nearest neighbor search, and local angular spectrum second order statistical matrix is constructed by outer product statistical average;Eigenvalue decomposition is carried out to matrix, and main peak energy proportion is extracted and angular spectrum instability index is defined;The angular distance between vectors in local neighborhood is counted, and the minimum included angle and included angle standard deviation are extracted, and angle weighted score is synthesized by weighting;According to the statistical average of angular distance set, adaptive angle threshold is generated, and comprehensive outlier degree score is generated;According to comprehensive outlier degree score, outlier is removed.The method can effectively suppress multipath effect and random noise under complex background, significantly improve the consistency of point cloud geometric structure, meet the demand of high-precision point cloud data for automatic driving and intelligent transportation.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar point cloud processing technology, specifically to a method and system for detecting outliers in millimeter-wave radar point clouds based on angular spectrum analysis. Background Technology

[0002] Millimeter-wave radar, with its all-weather perception capability, high motion sensitivity and strong robustness, has become a core sensor in the fields of autonomous driving, intelligent transportation and unmanned systems. However, due to the limitations of radar antenna aperture and complex electromagnetic scattering environment, the raw point cloud data generally faces severe challenges such as sparse point cloud, limited spatial resolution and outlier interference introduced by multipath reflection.

[0003] Existing point cloud outlier detection methods, such as density-based spatial clustering or statistical distribution-based filtering, mostly employ geometric criteria based on spatial Euclidean distance or local spatial density distribution. While these methods are effective in handling regularly distributed noise, they have significant limitations in millimeter-wave radar applications. Firstly, the spatial distance criteria used in existing technologies suffer from failure mechanisms in multipath scenarios. Since the ranging error of millimeter-wave radar is typically much larger than the angular measurement error, and the spatial topological relationship of coordinate points is easily distorted at long detection distances, algorithms relying solely on geometric distance struggle to maintain stability. More importantly, in scenarios with strong multipath interference, false points often accompany real targets, appearing highly close to, overlapping with, or even exhibiting similar geometric density characteristics in three-dimensional space. This makes it impossible for traditional geometric criteria to physically distinguish noise with similar spatial coordinates from target edges, leading to a large number of missed multipath interference points. Secondly, the judgment logic of existing algorithms is essentially based on the spatial coordinate information of the point cloud in the Cartesian coordinate system, which is a superficial geometric discrimination that ignores the essential physical characteristics of the echo signal.

[0004] Furthermore, the point cloud density of millimeter-wave radar varies nonlinearly with the detection distance and the target's reflective cross-section. Existing algorithms typically use a globally unified judgment threshold, which makes it easy to miss detections in sparse areas and over-detect in noise-dominated areas, making it difficult to adapt to complex and ever-changing dynamic conditions. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of existing technologies, specifically the poor local density adaptability, low multipath noise recognition rate, and susceptibility to false positives or false negatives caused by over-reliance on spatial geometric distribution in sparse or low-density scenarios. Furthermore, when distance dimension measurements are unreliable, this invention utilizes angular spectrum instability analysis and an adaptive threshold mechanism to achieve accurate identification and removal of outliers. To solve these problems, this invention adopts the following technical solution:

[0006] First, this invention proposes a method for detecting outliers in millimeter-wave radar point clouds based on angular spectrum analysis, comprising the following steps:

[0007] Obtain the echo scattering point cloud obtained by angle estimation from a millimeter-wave radar antenna array, wherein each scattering point in the point cloud contains an echo scattering pointing vector;

[0008] The echo scattering pointing vector of each scattering point is normalized to eliminate the interference of ranging error and scene scale on the angle domain analysis.

[0009] Within the normalized vector space, a local neighborhood is constructed for each center point based on K-nearest neighbor search to ensure the stability of local feature statistics; a second-order statistical matrix of the local angular spectrum is constructed by outer product statistical averaging.

[0010] Eigenvalue decomposition is performed on the second-order statistical matrix of the local angular spectrum to extract the energy proportion of the main peak, and an angular spectrum instability index is defined.

[0011] The angular distances between vectors within a local neighborhood are statistically analyzed, the minimum included angle and the standard deviation of the included angle are extracted, and then weighted and synthesized into an angle-weighted score.

[0012] The judgment boundary is adjusted in real time based on the statistical average value of the angular distance set to generate an adaptive angle threshold. Combined with the angular spectrum instability index, angle weighted score and angle anomaly ratio, a comprehensive outlier score is generated to identify and remove outliers, and an optimized millimeter-wave radar point cloud is output.

[0013] Preferably, the echo scattering pointing vector of each scattering point is normalized, specifically by: normalizing the echo scattering pointing vector. L2 norm normalization is performed to obtain the unit vector in angular spectral space. The calculation formula is as follows:

[0014] .

[0015] Preferably, a local neighborhood is constructed for each center point based on K-nearest neighbor search, where K ranges from 15 to 30.

[0016] Preferably, the angular spectral instability index is obtained through the following steps:

[0017] Combine the unit vectors of the center point and its local neighborhood of K points A second-order statistical matrix of the local angular spectrum is constructed by the statistical average of the outer product. :

[0018] ,

[0019] in, Represents the conjugate transpose of a vector;

[0020] For matrix Eigenvalue decomposition is performed to obtain a set of eigenvalues ​​that reflect the degree of concentration and dispersion of energy distribution. And extract the energy percentage of the main peak. :

[0021] ,

[0022] Among them, the largest eigenvalue The energy intensity in the direction of the main beam is represented by the local total energy;

[0023] Define the angular spectrum instability index .

[0024] Preferably, the angle-weighted score is obtained through the following steps:

[0025] Calculate the angular distance between any two unit vectors in the local neighborhood and construct the set of included angles. any of the included angle elements Defined as:

[0026] ,

[0027] From the set Extract the smallest included angle and the standard deviation reflecting the breadth of distribution. Therefore, a weighted composite score is calculated. :

[0028] ,

[0029] in, is a discrete weighting coefficient, with a value ranging from 0.5 to 2.0, used to adjust the algorithm's sensitivity to angular dispersion.

[0030] Preferably, the judgment boundary is adjusted in real time based on the statistical average of the angular distance set to generate an adaptive angle threshold, specifically including the following steps:

[0031] According to the set of included angles arithmetic mean Adjust the judgment boundary in real time and calculate the local adaptive angle threshold. :

[0032] ,

[0033] in, As an environmental regulator;

[0034] Statistical sets The value exceeds the threshold. The sample proportion was used to obtain the angle anomaly proportion. And finally merge to generate a comprehensive outlier score. :

[0035] ,

[0036] in, , , These are non-negative coefficients, used to control the contribution weights of angular spectrum instability, geometric deviation, and abnormal dispersion ratio to the final outlier determination.

[0037] Preferably, The value ranges from 1.5 to 3.0 in a sparse point cloud environment. Take the smaller value; in dense point cloud environments, Take the larger value.

[0038] Meanwhile, this invention proposes a millimeter-wave radar point cloud outlier detection system based on angular spectrum analysis, comprising:

[0039] The data acquisition module is used to acquire the echo scattering point cloud obtained by angle estimation of the millimeter-wave radar antenna array. Each scattering point in the point cloud contains an echo scattering pointing vector estimated from its echo signal.

[0040] The normalization module is used to normalize the L2 norm of the echo scattering pointing vector at each scattering point;

[0041] The neighborhood construction module is used to construct a local neighborhood for each center point in the normalized vector space based on K-nearest neighbor search.

[0042] The angular spectrum analysis module is used to construct a second-order statistical matrix of the local angular spectrum by the statistical average of the outer product of all unit pointing vectors in the local neighborhood, perform eigenvalue decomposition on the second-order statistical matrix of the local angular spectrum, extract the main peak energy ratio, and define the angular spectrum instability index.

[0043] The angle statistics module is used to calculate the angular distance between unit pointing vectors in a local neighborhood, extract the minimum included angle and the standard deviation of the included angle, and synthesize a weighted angle score.

[0044] An adaptive threshold module is used to adjust the judgment boundary in real time based on the statistical average of the angular distance set, generate an adaptive angle threshold, and count the proportion of samples in the angular distance set that exceed the adaptive angle threshold to obtain the angular anomaly ratio.

[0045] The scoring fusion module is used to combine the angular spectrum instability index, angle-weighted score, and angle anomaly ratio to generate a comprehensive outlier score.

[0046] The outlier removal module is used to identify and remove outliers based on the comprehensive outlier score, and output an optimized millimeter-wave radar point cloud.

[0047] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method described in the present invention.

[0048] Finally, the present invention proposes an electronic device comprising: a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the method as described in the present invention are implemented.

[0049] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0050] (1) This invention uses the eigenvalue decomposition of the second-order statistical matrix of the angular spectrum as the core criterion. By extracting the energy proportion of the main peak and defining the angular spectrum instability index, it characterizes the stability of the local point cloud from the underlying physical dimension of signal coherence. This feature uses the maximum eigenvalue to characterize the energy intensity in the direction of the main beam, and can accurately identify energy dispersion phenomena caused by multipath effects or random noise. Compared with traditional detection methods based on spatial distribution density or distance consistency, this invention effectively solves the problem of identifying outliers when they present a pseudo-clustered state in physical space, and significantly improves the accuracy of eliminating false scattering points in complex electromagnetic environments.

[0051] (2) This invention normalizes the echo scattering pointing vector using the second norm, mapping all scattering points onto a unit hypersphere in angular spectral space. This eliminates the amplitude fluctuations caused by the attenuation of target reflection intensity with distance and differences in target material during the preprocessing stage. This mechanism decouples the detection operator from energy intensity and geometric features, focusing only on the relative angular topological relationships between point clouds. This achieves insensitivity of detection performance to detection distance and signal gain, ensuring that the algorithm has high evaluation consistency and detection robustness across the entire detection range.

[0052] (3) This invention introduces a dynamic threshold mechanism based on the statistical characteristics of angle sets. It uses an environmental adjustment factor to correct the local adaptive angle threshold in real time and generates a comprehensive outlier score by combining the angle anomaly ratio. This feature allows the algorithm to spontaneously adjust the judgment boundary based on the angle consistency of the local point cloud. In sparse environments, it improves the recall rate for isolated noise by tightening the threshold, while in dense environments, it protects the geometric details of the real target by relaxing the boundary. This mechanism overcomes the limitations of fixed thresholds in non-uniform point cloud distributions, effectively balancing the stringency and fault tolerance of detection, and meeting the stringent requirements of high-precision perception systems for the consistency of point cloud geometry. Attached Figure Description

[0053] Figure 1 This invention relates to a schematic diagram of a method for detecting outliers in millimeter-wave radar point clouds based on angular spectrum analysis.

[0054] Figure 2 This is a schematic diagram illustrating the local angular spectral stability involved in the embodiment.

[0055] Figure 3 This is a schematic diagram of the local angle distribution involved in the embodiment.

[0056] Figure 4 This is a schematic diagram of the overall outlier score distribution involved in the example.

[0057] Figure 5 This is a schematic diagram illustrating outlier removal in an embodiment.

[0058] Figure 6 This is a schematic diagram illustrating outlier removal in a parking lot scenario, as illustrated in the example.

[0059] Figure 7 This is a schematic diagram illustrating outlier removal in an urban street scene as described in the example.

[0060] Figure 8 This is a comparison chart of the performance indicators of the present invention and the prior art involved in the embodiments. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0062] Example 1: This example provides a method for outlier detection in millimeter-wave radar point clouds based on angular spectral analysis, as proposed in this invention. Through effective processing and analysis of millimeter-wave radar point clouds, it solves the problem of outlier detection in sparse, low-density scenes in existing technologies, and is particularly suitable for millimeter-wave radar point clouds characterized by ranging errors and where the angular scattering direction dominates the geometric structure. (Reference) Figure 1 This method includes the following implementation steps:

[0063] (a) Feature engineering and local neighborhood construction:

[0064] During millimeter-wave radar detection, ranging errors are significantly affected by the environment, while angular information exhibits stronger pointing consistency within a local range. Therefore, this embodiment first uses feature engineering to reduce interference from the range dimension.

[0065] Obtain the echo scattering point cloud obtained by angle estimation from a millimeter-wave radar antenna array, wherein each scattering point in the point cloud contains an echo scattering pointing vector. ;

[0066] Subsequently, for vectors L2 norm normalization is performed to obtain the unit vector in angular spectral space. :

[0067]

[0068] This normalization process can shield the scale interference of echo intensity and detection distance fluctuations on angular domain feature analysis. After obtaining the unit vector, a K-nearest neighbor search is used in the feature space to construct a local neighborhood for each center point, where the number of neighborhood points K is set to a range of 15 to 30 to provide robust local feature support.

[0069] (II) Calculation of angular spectrum instability index and stability modeling:

[0070] To quantify the physical stability of the local structure of a point cloud, this embodiment performs second-order statistical modeling on the distribution characteristics of unit vectors within the neighborhood. First, it combines the unit vectors of the center point and its K local neighbors. Construct the second-order statistical matrix of the local angular spectrum. :

[0071]

[0072] in, This represents the conjugate transpose of a vector. By examining this matrix... Perform eigenvalue decomposition to obtain the set of eigenvalues ​​that reflect the energy distribution characteristics. Based on this, the energy percentage of the main peak is calculated. :

[0073]

[0074] in, The largest eigenvalue represents the energy contribution in the main beam direction. Furthermore, the angular spectral instability index is defined. This index can objectively reflect the degree to which local scattering direction changes from a single stable main direction to a multi-directional stray state.

[0075] Figure 2 This is a schematic diagram of local angular spectral stability. The left figure shows the case where the main peak energy accounts for a high proportion and the eigenvalue distribution is extremely uneven in the stable state of the angular spectrum; the right figure shows the unstable state caused by multipath interference, etc., which is used to illustrate the principle of distinguishing the stability of scattering structures by eigenvalue distribution.

[0076] (III) Geometric Constraint Analysis and Angle-Weighted Score Extraction:

[0077] To supplement the constraints on the determination of angular spectrum instability from a geometric perspective, this embodiment introduces local angular statistical features. An angle set is constructed by calculating the angular distance between unit vectors of any two points in the local neighborhood. Any of the included angle elements is defined as:

[0078]

[0079] From the set Extract the smallest included angle and the standard deviation of the included angle, which reflects the breadth of distribution. This leads to the synthesis of angle-weighted scores. :

[0080]

[0081] Among them, the weighting coefficient The value ranges from 0.5 to 2.0 and is used to dynamically adjust the algorithm's sensitivity to local angular distribution diffusion phenomena, thereby distinguishing real edge points from random outlier noise.

[0082] Figure 3 The statistical distribution of the angle set calculated pairwise from the scattering direction unit vectors in the local neighborhood of the center point is shown, reflecting the consistency of the angular structure.

[0083] (iv) Integration of adaptive threshold mechanism and comprehensive scoring:

[0084] To address the uneven density distribution of point clouds in millimeter-wave radar, this embodiment employs a dynamic threshold adjustment criterion. Based on the included angle set... arithmetic mean Set a local adaptive angle threshold:

[0085]

[0086] Among them, the regulating factor The value ranges from 1.5 to 3.0. The value ranges from 1.5 to 3.0 in a sparse point cloud environment. Take the smaller value; in dense point cloud environments, Take the larger value.

[0087] Statistical sets Exceeding the threshold The sample proportion is used to obtain the angular anomaly proportion. Finally, a comprehensive outlier score is generated through a weighted fusion method. :

[0088]

[0089] Among them, the weighting coefficient , , All are non-negative numbers, used to balance the contribution of angular spectrum stability, geometric deviation, and dispersion ratio to the final outlier determination.

[0090] Figure 4 This is a schematic diagram of the overall outlier score distribution in this embodiment. It shows the outlier score distribution after integrating the angular spectrum instability index, angular statistical features, and angular anomaly statistical results, illustrating the effectiveness of this method in distinguishing outliers.

[0091] (v) Outlier identification and anomaly removal:

[0092] This embodiment concludes with the comprehensive outlier score. The method identifies outliers based on a criterion generated from the global statistical characteristics of the point cloud. Points scoring above the criterion are considered outliers and removed. Through multi-dimensional joint constraints, this method effectively reduces the false detection rate in strong multipath environments and sparse point cloud regions, outputting millimeter-wave radar point clouds with higher geometric consistency and cleaner noise background. Figure 5 This is a schematic diagram of outlier removal according to an embodiment of the present invention. The diagram shows the outliers identified in the point cloud data and the target geometric structure after removal.

[0093] To further verify the effectiveness of the present invention, this embodiment provides applications of the method in different typical scenarios:

[0094] (1) Parking lot scenario: In the parking lot scenario, this method uses the second norm to normalize the difference in shielding distance, thereby achieving accurate removal of strong reflection interference. Figure 6 This paper presents a comparison of the original and processed point clouds of the same target, such as a human body and background vehicles, in a typical environment of strong metallic reflection in an underground parking lot. In the image, the left side represents the actual data acquisition environment, and the middle side shows the original 3D point cloud distribution before outlier removal. The horizontal, vertical, and height axes correspond to the horizontal position of the detection area, the vertical distance, and the target height, respectively. In the original point cloud, besides the high-density red point clusters representing the target in the central area, there are numerous scattered interference points with strong reflections generated by parking space metal fixtures, pillars, and ceiling pipes. These interference points appear as scattered blue and green point clusters deviating from the main target. This method first uses L2 normalization to mask the differences in echo energy at different distances, and then identifies these interference points through angular spectrum analysis. Although they have a certain intensity, their scattering direction deviates significantly from the main target direction. Comparing the results on the right after outlier removal, it is clear that the algorithm successfully identified and removed outliers distributed at the edge and far end of the target, significantly shielding the interference of environmental metal reflections on the target geometry, making the outline of the target point cluster purer and the boundary clearer, thereby greatly improving the reliability of subsequent target clustering and recognition.

[0095] (2) Urban street scene: In the urban street scene, the present invention effectively removes the ghosting and trailing interference of dynamic targets through angular spectrum coherence analysis. Figure 7 This paper demonstrates the point cloud processing performance of this invention in a typical open and dynamic environment of urban roads / streets. The left-hand real-world image shows a typical urban asphalt road surface and roadside trees. The middle sub-image shows the original point cloud before outlier removal. Due to the specular reflection caused by the smooth road surface and the scattering of microwaves by surrounding vegetation, a large number of multipath false points appear around moving targets such as pedestrians or vehicles. These points are close to the edges of the red core point clusters and appear as translucent blue and green impurities. Unlike the metallic noise points in parking lot scenes, this type of noise is closely connected to the target and moves synchronously with the target, making it easy for traditional algorithms to misidentify it as a target. To solve this problem, this invention identifies the energy phase of these false points as exhibiting obvious random instability by quantifying the angular spectral coherence within the local neighborhood. Comparing the right-hand image after outlier removal, it is clear that this algorithm accurately removes ghosting and trailing noise while preserving the point clusters and details of the real target. Experimental results demonstrate that this technology effectively solves the problems of blurred target outlines and interference from false targets in dynamic urban environments, providing reliable underlying perception data for subsequent high-precision trajectory association and obstacle avoidance decisions.

[0096] Figure 8 This invention demonstrates a point cloud detection method based on angular spectrum analysis, exhibiting excellent performance in precision, recall, and overall robustness. As shown in the figure, the method achieves a precision of 94.0%, significantly higher than existing comparative algorithms such as LOF, DBSCAN, and SOR, showcasing its advantages in identifying interference points and removing noise. The recall rate reaches 91.0%, effectively solving the problem of missed detections in sparse regions by traditional algorithms. Furthermore, a dynamic threshold mechanism enables the algorithm to adaptively adjust to changes in point cloud distribution, thereby enhancing its ability to detect isolated noise points. In addition, the overall robustness index, F1-Score, is 92.5%, far exceeding other comparative algorithms, indicating that this invention achieves a good balance between precision and recall. Through L2 normalization and the angular spectrum instability index, the algorithm maintains high stability in complex environments, such as multipath interference and target occlusion, significantly improving point cloud quality and detection consistency. This experiment uses a self-collected millimeter-wave radar point cloud dataset. The acquisition device is an IWR6843ISK millimeter-wave radar antenna array, covering 1000 frames of typical working conditions such as parking lots, urban blocks, and intersections.

[0097] Example 2: This example provides a millimeter-wave radar point cloud outlier detection system based on angular spectrum analysis, including:

[0098] The data acquisition module is used to acquire the echo scattering point cloud obtained by angle estimation of the millimeter-wave radar antenna array. Each scattering point in the point cloud contains an echo scattering pointing vector estimated from its echo signal.

[0099] The normalization module is used to normalize the L2 norm of the echo scattering pointing vector at each scattering point;

[0100] The neighborhood construction module is used to construct a local neighborhood for each center point in the normalized vector space based on K-nearest neighbor search.

[0101] The angular spectrum analysis module is used to construct a second-order statistical matrix of the local angular spectrum by the statistical average of the outer product of all unit pointing vectors in the local neighborhood, perform eigenvalue decomposition on the second-order statistical matrix of the local angular spectrum, extract the main peak energy ratio, and define the angular spectrum instability index.

[0102] The angle statistics module is used to calculate the angular distance between unit pointing vectors in a local neighborhood, extract the minimum included angle and the standard deviation of the included angle, and synthesize a weighted angle score.

[0103] An adaptive threshold module is used to adjust the judgment boundary in real time based on the statistical average of the angular distance set, generate an adaptive angle threshold, and count the proportion of samples in the angular distance set that exceed the adaptive angle threshold to obtain the angular anomaly ratio.

[0104] The scoring fusion module is used to combine the angular spectrum instability index, angle-weighted score, and angle anomaly ratio to generate a comprehensive outlier score.

[0105] The outlier removal module is used to identify and remove outliers based on the comprehensive outlier score, and output an optimized millimeter-wave radar point cloud.

[0106] Example 3: This example proposes a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method described in this invention.

[0107] Example 4: This example proposes an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method as described in this invention are implemented.

[0108] It should be noted that the processing flow of embodiments 2-4 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.

[0109] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0110] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for detecting outliers in millimeter-wave radar point clouds based on angular spectrum analysis, characterized in that, Includes the following steps: Obtain the echo scattering point cloud obtained by angle estimation from a millimeter-wave radar antenna array, wherein each scattering point in the point cloud contains an echo scattering pointing vector; Normalize the echo scattering pointing vector for each scattering point; In the normalized vector space, a local neighborhood is constructed for each center point based on K-nearest neighbor search, and a second-order statistical matrix of the local angular spectrum is constructed by outer product statistical averaging. Eigenvalue decomposition is performed on the second-order statistical matrix of the local angular spectrum to extract the energy proportion of the main peak, and an angular spectrum instability index is defined. The angular distances between vectors within a local neighborhood are statistically analyzed, the minimum included angle and the standard deviation of the included angle are extracted, and then weighted and synthesized into an angle-weighted score. The judgment boundary is adjusted in real time based on the statistical average value of the angular distance set to generate an adaptive angle threshold. Combined with the angular spectrum instability index, angle weighted score and angle anomaly ratio, a comprehensive outlier score is generated to identify and remove outliers, and an optimized millimeter-wave radar point cloud is output.

2. The method according to claim 1, characterized in that, Normalize the echo scattering pointing vector at each scattering point, specifically by: normalizing the echo scattering pointing vector. L2 norm normalization is performed to obtain the unit vector in angular spectral space. The calculation formula is as follows: 。 3. The method according to claim 1, characterized in that, A local neighborhood is constructed for each center point based on K-nearest neighbor search, where the value of K ranges from 15 to 30.

4. The method according to claim 1, characterized in that, The angular spectrum instability index is obtained through the following steps: Combine the unit vectors of the center point and its local neighborhood of K points A second-order statistical matrix of the local angular spectrum is constructed by the statistical average of the outer product. : ， in, Represents the conjugate transpose of a vector; For matrix Eigenvalue decomposition is performed to obtain a set of eigenvalues ​​that reflect the degree of concentration and dispersion of energy distribution. And extract the energy percentage of the main peak. : ， Among them, the largest eigenvalue The energy intensity in the direction of the main beam is represented by the local total energy; Define the angular spectrum instability index .

5. The method according to claim 1, characterized in that, The angle-weighted score is obtained through the following steps: Calculate the angular distance between any two unit vectors in the local neighborhood and construct the set of included angles. any of the included angle elements Defined as: ， From the set Extract the smallest included angle and the standard deviation reflecting the breadth of distribution. Therefore, a weighted composite score is calculated. : ， in, is a discrete weighting coefficient, with a value ranging from 0.5 to 2.0, used to adjust the algorithm's sensitivity to angular dispersion.

6. The method according to claim 1, characterized in that, The decision boundary is adjusted in real time based on the statistical average of the angular distance set to generate an adaptive angle threshold, which specifically includes the following steps: According to the set of included angles arithmetic mean Adjust the judgment boundary in real time and calculate the local adaptive angle threshold. : ， in, As an environmental regulator; Statistical sets The value exceeds the threshold. The sample proportion was used to obtain the angle anomaly proportion. And finally merge to generate a comprehensive outlier score. : ， in, , , These are non-negative coefficients, used to control the contribution weights of angular spectrum instability, geometric deviation, and abnormal dispersion ratio to the final outlier determination.

7. The method according to claim 6, characterized in that, The value ranges from 1.5 to 3.0 in a sparse point cloud environment. Take the smaller value; in dense point cloud environments, Take the larger value.

8. A millimeter-wave radar point cloud outlier detection system based on angular spectrum analysis, characterized in that, include: The data acquisition module is used to acquire the echo scattering point cloud obtained by angle estimation of the millimeter-wave radar antenna array. Each scattering point in the point cloud contains an echo scattering pointing vector estimated from its echo signal. The normalization module is used to normalize the L2 norm of the echo scattering pointing vector at each scattering point; The neighborhood construction module is used to construct a local neighborhood for each center point in the normalized vector space based on K-nearest neighbor search. The angular spectrum analysis module is used to construct a second-order statistical matrix of the local angular spectrum by the statistical average of the outer product of all unit pointing vectors in the local neighborhood, perform eigenvalue decomposition on the second-order statistical matrix of the local angular spectrum, extract the main peak energy ratio, and define the angular spectrum instability index. The angle statistics module is used to calculate the angular distance between unit pointing vectors in a local neighborhood, extract the minimum included angle and the standard deviation of the included angle, and synthesize a weighted angle score. An adaptive threshold module is used to adjust the judgment boundary in real time based on the statistical average of the angular distance set, generate an adaptive angle threshold, and count the proportion of samples in the angular distance set that exceed the adaptive angle threshold to obtain the angular anomaly ratio. The scoring fusion module is used to combine the angular spectrum instability index, angle-weighted score, and angle anomaly ratio to generate a comprehensive outlier score. The outlier removal module is used to identify and remove outliers based on the comprehensive outlier score, and output an optimized millimeter-wave radar point cloud.

9. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: A processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method as described in any one of claims 1 to 7 are implemented.

Images