Point cloud clustering method and device based on millimeter wave radar indoor perception

Abstract

The invention relates to a point cloud clustering method and device based on millimeter wave radar indoor perception. The method comprises the following steps: acquiring point cloud data obtained based on millimeter wave radar detection in a physical space, and mapping the point cloud data to a preset graph space to obtain a mapping graph corresponding to the point cloud data; performing feature accumulation updating processing on the mapping graph according to different time levels to obtain feature graph layers of the mapping graph on different feature dimensions; and performing point cloud clustering processing on the plurality of feature layers to obtain a clustering result corresponding to the point cloud data, wherein the clustering result is used for representing the distribution condition of the point cloud data in each feature dimension so as to detect and track an entity object in the physical space. By adopting the method, multi-feature modeling and refined clustering recognition can be performed on the point cloud data in an indoor millimeter-wave radar sensing scene, and high-precision structural support is provided for stable detection and precise tracking of entity objects in a physical space.

Description

Technical Field

[0001] The present application relates to the field of radar perception technology, and in particular to a point cloud clustering method and device based on millimeter-wave radar indoor perception. Background Art

[0002] The field of radar perception involves using millimeter-wave radar to detect targets in indoor environments, enabling human presence perception and spatial activity analysis. Related point cloud clustering methods obtain point cloud data from radar echo signals and then cluster and identify this point cloud data using traditional clustering methods such as K-Means and DBSCAN. However, in indoor millimeter-wave radar perception scenarios, limited hardware resolution and signal susceptibility to interference result in sparse point cloud data and complex morphological distribution, making traditional clustering methods difficult to effectively address. Summary of the Invention

[0003] Based on this, it is necessary to provide a point cloud clustering method, device, computer equipment and computer-readable storage medium based on millimeter-wave radar indoor perception to address the above technical problems, which can be used to perform multi-feature modeling and refined clustering recognition on point cloud data in indoor millimeter-wave radar perception scenarios, and provide high-precision structural support for stable detection and accurate tracking of physical objects in physical space.

[0004] In a first aspect, the present application provides a point cloud clustering method based on millimeter-wave radar indoor perception, comprising: Acquire point cloud data obtained based on millimeter-wave radar detection in a physical space, and map the point cloud data into a preset graph space to obtain a mapping graph corresponding to the point cloud data; Perform feature accumulation and updating processing on the mapping graph according to different time levels to obtain feature layers of the mapping graph at different feature dimensions; Point cloud clustering processing is performed on multiple feature layers to obtain clustering results corresponding to the point cloud data, and the clustering results are used to characterize the distribution status of the point cloud data in various feature dimensions to detect and track physical objects in the physical space.

[0005] In a second aspect, the present application also provides a point cloud clustering device based on millimeter-wave radar indoor perception, comprising: A mapping module, configured to obtain point cloud data obtained based on millimeter-wave radar detection in a physical space, and map the point cloud data into a preset graph space to obtain a mapping graph corresponding to the point cloud data; An updating module is used to perform feature accumulation and updating processing on the map according to different time levels to obtain feature layers of the map at different feature dimensions; A clustering module is used to perform point cloud clustering processing on multiple feature layers to obtain clustering results corresponding to the point cloud data. The clustering results are used to characterize the distribution of the point cloud data in various feature dimensions to detect and track physical objects in the physical space.

[0006] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above steps when executing the computer program.

[0007] In a fourth aspect, the present application further provides a computer-readable storage medium on which a computer program is stored, and the computer program implements the above steps when executed by a processor.

[0008] The above-mentioned point cloud clustering method, device, computer equipment and computer-readable storage medium based on millimeter-wave radar indoor perception, first, according to the processing method of mapping the point cloud data obtained by the millimeter-wave radar into a unified graph space, thereby realizing the regular expression of the point cloud data at the structural level, which is convenient for subsequent local analysis and feature calibration on the graph structure; secondly, according to the cumulative update processing of the features of each element in the mapping graph at different time levels, a set of feature layers reflecting multiple update levels is generated, and a multi-angle feature view is constructed for the same frame of point cloud data, avoiding secondary search of the point cloud to reduce the computational consumption of the detection process; thirdly, according to the similarity of the data points in multiple feature layers, clustering processing is performed to obtain a clustering result for describing the distribution status of the point cloud data under multi-dimensional features; based on this, in the indoor millimeter-wave radar perception scenario, multi-feature modeling and refined clustering recognition can be realized based on the internal structure of a single frame of point cloud data, providing high-precision structural support for the stable detection and accurate tracking of physical objects in physical space. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0010] Figure 1 1 is a flow chart of a point cloud clustering method based on millimeter-wave radar indoor perception in one embodiment; Figure 2 The figure is a structural block diagram of a point cloud clustering device based on millimeter-wave radar indoor perception in one embodiment. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0012] In one embodiment, Figure 1 As shown, a point cloud clustering method based on millimeter-wave radar indoor sensing is provided. This embodiment uses the method applied to a server as an example. It is understood that the method can also be applied to a terminal, or to a system including a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps S101 to S103.

[0013] Step S101: acquiring point cloud data obtained based on millimeter-wave radar detection in a physical space, and mapping the point cloud data into a preset graph space to obtain a mapping graph corresponding to the point cloud data.

[0014] Among them, millimeter-wave radar refers to a sensing device that measures distance and direction based on the millimeter-wave frequency band (30 to 300GHz), such as a radar module installed on the ceiling for indoor personnel perception; physical space refers to an actual three-dimensional environmental area, such as a real scene of an indoor room or corridor.

[0015] The point cloud data represents a data set consisting of multiple spatial points detected by a millimeter-wave radar, such as a list of coordinate points converted from radar echoes.

[0016] Among them, the graph space represents a two-dimensional structured grid or graphic representation area used to planar map and display the original point cloud data; the mapping graph represents the graphical structure formed after projecting or transforming the point cloud data into the standard graph space according to preset rules.

[0017] For example, a millimeter-wave radar device first scans the physical environment to obtain a single frame of point cloud data generated at the current scanning frequency. This point cloud data consists of multiple coordinate points, each of which represents the specific location of the reflected echo signal detected by the millimeter-wave radar at a specific moment and in a specific spatial direction. These coordinate points also reflect the spatial contours of obstacles or moving targets in the physical space. Furthermore, this raw point cloud data undergoes a coordinate transformation, converting it from the local coordinate system of the millimeter-wave radar device into a uniformly defined graph space to form a mapping diagram with a unified structure and reference scale. This graph space is typically constructed as a regular two-dimensional grid structure, which is used to carry and record relevant information such as point cloud density and reflection intensity within each region of the physical space. During the coordinate transformation process, spatial coordinate normalization and voxelization can be used to classify different coordinate points in the point cloud data into specific location units in the graph space, thereby establishing a mapping relationship between point clouds and primitives in the graph space. Geometric connections between point clouds can also be established through adjacency relationships between primitives.

[0018] Step S102 : performing feature accumulation and updating processing on the map according to different time levels to obtain feature layers of the map at different feature dimensions.

[0019] Among them, the time level represents the time dimension considered when describing the degree of feature update or retention when performing feature accumulation and update processing on the mapping graph; for example, the greater the update degree or the smaller the retention degree, the smaller the time dimension; the smaller the update degree or the greater the retention degree, the larger the time dimension, so as to obtain the best signal-to-noise ratio through long-term accumulation, thereby improving the detection performance of weak moving targets.

[0020] Among them, the feature layers at different feature dimensions represent a multi-layer graph structure composed of different types of structural or attribute information extracted from the mapping graph, which is used to represent the state of point cloud data in different physical properties or statistical properties. For example, it includes layers corresponding to reflection intensity distribution, layers corresponding to point cloud density distribution, layers corresponding to spatial displacement distribution, etc., so as to introduce a cross-dimensional semantic extension mechanism in the two-dimensional graph structure, so that the originally flat image expression has multi-dimensional information carrying capacity.

[0021] For example, in a mapping diagram formed based on single-frame point cloud data, the characteristic values ​​at each primitive position are locally updated or retained according to the settings of different time levels, thereby generating feature layers with different update degrees or retention degrees; the above-mentioned feature accumulation update processing does not involve comparison or merging between multiple frames of data, but only adjusts the point cloud attributes carried by the primitives through internal calculations for the current frame data. In the actual processing process, first of all, it is necessary to clarify the parameter settings of the time level. This parameter is used to express the balance between the update amplitude and the retention ratio of the characteristic values ​​contained in each primitive. For example, a smaller time level value indicates a faster response to the current feature change, that is, the characteristic value of the primitive is updated significantly, and the original state is less retained; while a larger time level value indicates a tendency to maintain the existing features, and the update amplitude is relatively small. On this basis, when processing the eigenvalues ​​in each primitive, we first calculate the difference between the new eigenvalue generated by projecting the current point cloud data onto the primitive and the original eigenvalue. Then, combined with the set time level parameters, we adjust the original eigenvalue according to specific weights, so that it gradually approaches the new state while maintaining the historical state.

[0022] The essence of this processing method is to refine the state of each primitive in the mapping image within a single frame, thereby generating a set of feature layers reflecting different update strategies; each feature layer represents the primitive state distribution of the same frame data under different feature response speed settings. Therefore, these feature layers have the same primitive structure but present different feature differences, which not only maintains the uniformity of the data source, but also improves the richness of the structural expression, and can be used as multiple references to distinguish the details of the point cloud structure in subsequent processing.

[0023] In step S103, point cloud clustering processing is performed on multiple feature layers to obtain clustering results corresponding to the point cloud data. The clustering results are used to characterize the distribution of the point cloud data in various feature dimensions to detect and track physical objects in the physical space.

[0024] Among them, the clustering result represents the classification set obtained by grouping the data points of multiple feature layers according to their similarities on the corresponding feature layers, which is used to reflect the areas with consistency in structure or attributes in the graph space.

[0025] Among them, the distribution of point cloud data in each feature dimension represents the statistical characteristics of the point cloud data in the feature layer, such as spatial distribution, density, and change trend, which is used to assist clustering judgment and entity recognition.

[0026] Among them, the entity object represents a target individual with integrity and consistency identified by point cloud clustering in the physical space, which is used for further identification, positioning or behavior tracking, such as a specific person or a specific object.

[0027] For example, based on the various feature information recorded in multiple feature layers, point cloud regions with consistency or similarity in spatial position, structural distribution, reflection intensity, and change trend are identified in the graph space and divided into several independent clustering units. Specifically, first, multiple feature layers need to be aligned according to the primitive positions to ensure that the same positions in each feature layer can be uniformly accessed and processed; then, at each primitive position, the corresponding feature value set in each feature layer is extracted to form a multidimensional feature vector describing the multidimensional state of the primitive position; after constructing the multidimensional feature vector set corresponding to all primitive positions, the distance measurement between primitives in the multidimensional feature space is analyzed based on the multidimensional feature vector set, and primitives that are close or highly consistent with each other are classified into one category, and the original point cloud regions corresponding to the primitives of this category are marked as the same clustering unit; at the same time, the boundaries between different clustering units need to be smoothed to avoid spatial segmentation errors caused by feature disturbances; in addition, to improve recognition accuracy, abnormal point clouds or edge points should be eliminated to enhance the performance of the clustering results in terms of spatial continuity and structural integrity. Based on this, the final clustering results not only reflect the structural division of point cloud data in the graph space, but can also be further mapped back to the physical space as the basis for identifying and tracking different physical objects.

[0028] In the above-mentioned point cloud clustering method based on indoor perception by millimeter-wave radar, first, according to the processing method of mapping the point cloud data acquired by the millimeter-wave radar into a unified graph space, the regular expression of the point cloud data at the structural level is realized, which is convenient for subsequent local analysis and feature calibration on the graph structure; secondly, according to the cumulative update processing of the features of each element in the mapping graph at different time levels, a set of feature layers reflecting various update degrees is generated, and a multi-angle feature view is constructed for the same frame of point cloud data, avoiding the secondary search of the point cloud and reducing the computational consumption of the detection process; thirdly, according to the similarity of the data points in multiple feature layers, clustering processing is performed to obtain the clustering results used to describe the distribution status of the point cloud data under multi-dimensional features; based on this, in the indoor millimeter-wave radar perception scenario, multi-feature modeling and refined clustering recognition can be realized on the basis of the internal structure of the single-frame point cloud data, providing high-precision structural support for the stable detection and accurate tracking of entity objects in the physical space.

[0029] In an exemplary embodiment, mapping the point cloud data into a preset graph space to obtain a mapping graph corresponding to the point cloud data includes steps S201 to S203.

[0030] Step S201 : determining the reference point positions in the image space and the scaling ratio of the image space corresponding to the point cloud data according to the geometric features of the image space.

[0031] Among them, the geometric features of the image space represent the spatial structural properties of the image space in terms of size, shape, resolution, etc., and are used to define the projection range and coordinate boundaries of point cloud data in the image space.

[0032] Among them, the reference point represents a specific pixel position in the image space used to align with the reference origin in the radar coordinate system, and is used to establish a spatial mapping relationship between the image space and the radar coordinate system. For example, if the center point of the image space is set as the reference point, the origin of the radar coordinate system will correspond to the reference point during mapping.

[0033] Among them, the radar coordinate system represents the local coordinate reference system used by the millimeter-wave radar itself in space, which is used to describe the spatial position relationship of the point cloud data detected by the radar from its own perspective.

[0034] The scaling ratio indicates the linear conversion ratio of the unit coordinate value between the two coordinate systems during the mapping process from the radar coordinate system to the image space. It is used to ensure consistent representation of data at different scales. For example, if a physical length unit of 1 meter corresponds to 10 pixel units in the image space, the scaling ratio is 10.

[0035] For example, in order to accurately map point cloud data to the image space, it is necessary to determine the reference points and scaling ratios applicable to the mapping process based on the geometric characteristics of the image space in terms of size, shape, resolution, etc., so as to establish a conversion relationship between each point in the image space and the corresponding point cloud data in the physical space. Specifically, the reference point refers to a specific position in the image space, which is used as the corresponding position of the origin or reference point in the radar coordinate system; the setting of the reference point can be based on the shape of the image space, and special points with significant reference effects such as the center point are selected to achieve spatial alignment between the radar coordinate system and the image space coordinate system, thereby ensuring that the position of the point cloud data presented in the image space is consistent with its structure in the physical space. Furthermore, the scaling ratio is used to describe the transformation ratio relationship between the unit length of the two coordinate systems to solve the scale inconsistency problem between the point cloud data and the image space; the setting of the scaling ratio can be unified according to the resolution setting of the image space and the actual range of the point cloud data, so that each unit length of the point cloud data can correctly express its spatial extensibility in the image space, thereby avoiding distortion of the element density due to scaling imbalance.

[0036] In step S202 , the corresponding point cloud coordinates of each point data in the point cloud data in the preset radar coordinate system are converted into pixel coordinates in the image space in combination with the reference point position and the scaling ratio.

[0037] For example, based on the scaling ratio and reference point position, the position of the center point of the image is first calculated according to the preset image space size. This center point is the reference point position and serves as the origin of the image space coordinate system, so that the radar coordinate system centered on the radar in physical space can be aligned with the geometric center in the image space. Subsequently, a coordinate mapping operation is performed on each point data in the point cloud data. Specifically, the horizontal point cloud coordinates of the point cloud data on the horizontal axis in the radar coordinate system are multiplied by the scaling ratio of the image space on the corresponding horizontal axis, thereby achieving unit conversion and obtaining the scaling result of the horizontal point cloud coordinates on the corresponding horizontal axis in the image space. Furthermore, this scaling result is added to the center point coordinates in the image space to obtain the horizontal pixel coordinates corresponding to the horizontal point cloud coordinates in the image space. The mapping process also uses the same method in the vertical coordinate axis direction to obtain the vertical pixel coordinates corresponding to the vertical point cloud coordinates in the image space. Based on this, the complete two-dimensional pixel position is finally generated according to the horizontal pixel coordinates and vertical pixel coordinates corresponding to each point cloud coordinate in the image space. To ensure that the pixel position is an integer, the calculation result needs to be rounded to meet the discrete requirements of the graph structure.

[0038] Optionally, the mapping relationship between point cloud coordinates and pixel coordinates can refer to equations (1) and (2): (1) (2) in, and Respectively represent the horizontal pixel coordinates and vertical pixel coordinates of the i-th point in the point cloud data in the image space; and Respectively represent the scaling ratio corresponding to the horizontal coordinate axis and the scaling ratio corresponding to the vertical coordinate axis; and Respectively represent the horizontal point cloud coordinates and vertical point cloud coordinates of the i-th point in the point cloud data in the radar coordinate system; and They represent the horizontal pixel coordinates and vertical pixel coordinates corresponding to the reference point in the image space, that is, the horizontal pixel coordinates and vertical pixel coordinates corresponding to the center point of the image space; Represents a rounding function that converts scaled coordinate values ​​to the nearest integer corresponding to discrete positions in the image pixel.

[0039] The design of this type of mapping relationship can ensure the symmetry and scale uniformity of the radar detection area in the image space, so that the detection range centered on the radar can accurately cover the center of the image, and avoid the problem of element dislocation or boundary compression caused by scale mismatch during the spatial transformation process; through the above mapping operation, the original point cloud coordinates are uniformly converted into two-dimensional pixel coordinates in the image space, so that the point cloud data has an expression form compatible with the graph structure, while retaining the relative distribution pattern of the point cloud data in the physical space, so that the geometric relationship between points can be continued in the image space, avoiding the loss of structural information during the transformation process.

[0040] Step S203 , obtaining a mapping image corresponding to the point cloud data according to the pixel coordinates of each point data of the point cloud data in the image space.

[0041] For example, for each pixel coordinate position that has been projected by the point cloud, it is necessary to mark or assign a value to the corresponding primitive in the graph space; specifically, the assignment method can be set according to the attribute information of the point cloud data itself, for example, the value of each primitive is set to the reflection intensity or other original attributes of the corresponding point data in the point cloud data, so as to reflect the structural state of different primitives in the point cloud distribution. In the case where multiple point data are simultaneously mapped to the same primitive position, superposition processing or maximum value processing can be performed to ensure that the state presented by the primitive is representative. The construction of the entire mapping graph is a key process for structuring the continuously distributed spatial point cloud data into a discrete primitive structure, so that the original data has graph structure characteristics, so that it can adapt to subsequent operations such as layer construction, feature extraction and cluster analysis; in addition, the mapping graph not only retains the spatial distribution pattern of the point cloud data, but also reflects the changes in the physical properties of the point cloud data through the value transfer of the primitives, forming a data representation basis with both structural and attribute dimensions.

[0042] In this embodiment, first, the reference point positions and scaling ratios are set according to the geometric features of the image space, thereby achieving a unified correspondence between the point cloud data and the image space in the coordinate system and scale; secondly, coordinate transformation is performed according to the linear mapping relationship between the point cloud coordinates and the image space pixel coordinates, thereby ensuring that the pixel positions of the original point cloud data in the image space can accurately express their physical position relationships; thirdly, a point cloud mapping diagram with a graph structure is constructed according to the image space pixel coordinates of each point cloud data, which can effectively structure the original radar point cloud data into a standard graph form, thereby enhancing the adaptability and expression integrity of subsequent feature modeling and spatial clustering processing.

[0043] In an exemplary embodiment, mapping point cloud data into a preset graph space to obtain a mapping graph corresponding to the point cloud data includes steps S301 and S302.

[0044] In step S301, dynamic point cloud data and static point cloud data in the point cloud data are determined based on different frequency shift characteristics of the point cloud data. The different frequency shift characteristics reflect the different motion conditions of the physical object in the physical space relative to the millimeter wave radar.

[0045] The frequency shift feature represents the frequency offset value generated by the radial motion speed of the target relative to the radar in the millimeter-wave radar echo. It is used to distinguish points reflecting moving targets from points reflecting the stationary background in the point cloud data. For example, when a person approaches the radar, the reflection point has a positive frequency shift, while the reflection point of a stationary object such as a wall has a frequency shift close to zero.

[0046] Among them, dynamic point cloud data refers to a set of points with significant frequency shift characteristics extracted from point cloud data, which is used to represent targets in motion in physical space, such as walking human bodies, moving equipment, etc., which are point cloud reflection points formed in radar detection.

[0047] Among them, static point cloud data refers to a set of points extracted from point cloud data with a frequency shift close to zero or within a set threshold range. It is used to represent the background or fixed objects in a stationary state in the physical space, such as walls, floors, tables and chairs, and other non-moving objects, which are formed by point cloud reflection points in radar detection.

[0048] For example, based on the frequency shift characteristics corresponding to each point data in the point cloud data, the dynamics of the entire point cloud set is judged to distinguish reflection points in different motion conditions in the physical space; that is, the frequency shift characteristics reflect the speed information carried in the radar echo, and different frequency shift values ​​correspond to the relative motion state of the reflection point relative to the radar device; when an object has a velocity component relative to the radar, its reflected wave will produce a frequency shift, so the frequency shift value of each data point in the point cloud data can be analyzed to determine whether the spatial object represented by each data point has relative motion behavior. Specifically, it is necessary to read the parameters related to the frequency shift characteristics of each data point and set a certain frequency shift threshold as the boundary between static and dynamic judgments; data points with an absolute frequency shift value higher than the frequency shift threshold are marked as dynamic point cloud data, indicating that they come from physical objects with motion characteristics; and data points with an absolute frequency shift value lower than the frequency shift threshold are regarded as static point cloud data, indicating that they come from the background environment or a stationary object.

[0049] In practical applications, this frequency shift division method can effectively identify the differences between moving targets such as human bodies and vehicles and static backgrounds such as the ground and walls. This processing is completely based on the frequency shift characteristics of the current frame point cloud data itself and does not rely on historical frame comparison. Ultimately, the entire point cloud data is divided into two parts, one is dynamic point cloud data reflecting the moving entity, and the other is static point cloud data reflecting the static scene.

[0050] In step S302, the dynamic point cloud data and the static point cloud data are mapped into the graph space respectively to obtain a first mapping graph corresponding to the dynamic point cloud data and a second mapping graph corresponding to the static point cloud data, and the first mapping graph and the second mapping graph are used as mapping graphs corresponding to the point cloud data.

[0051] Exemplarily, the above-mentioned divided dynamic point cloud data and static point cloud data are respectively mapped into a preset graph space to construct two independent graph structure representations, thereby obtaining a first mapping map and a second mapping map. Specifically, first, for the dynamic point cloud data, according to the set graph space mapping relationship, the point cloud coordinates of each dynamic data point are converted to obtain its corresponding pixel coordinate position in the graph space, and the position is marked or assigned in the graph space to form a first mapping map representing the spatial distribution of the dynamic target. Each graphic element of the first mapping map represents the position mapping of a point in motion in the physical space. Subsequently, the static point cloud data is processed using the same mapping logic, and its point cloud coordinates are mapped into the graph space to obtain a second mapping map that describes the distribution of static background or stationary objects. The first mapping map and the second mapping map are consistent in structure, but their contents correspond to the distribution of point cloud data under different motion conditions.

[0052] After the two mapping images are constructed, they can be combined to generate the final point cloud mapping image. This combination method can be based on the fusion of the primitive level, that is, retaining two independent channels in the graph space, corresponding to the dynamic and static layers respectively, to support the independent processing of motion characteristics in subsequent feature extraction and clustering analysis. In this way, the spatial distribution difference between the dynamic information and static information of the point cloud data can be explicitly retained in the graph structure, so that the point cloud representation not only has the position structure dimension of the physical object, but also has the motion status dimension of the physical object.

[0053] In this embodiment, first, the point cloud data is dynamically divided according to the frequency shift characteristics to obtain dynamic point cloud data and static point cloud data, thereby achieving effective distinction between moving targets and static backgrounds; secondly, the dynamic point cloud data and the static point cloud data are mapped separately to construct a first mapping graph and a second mapping graph that can distinguish the target motion status. Based on this, a point cloud structure representation with the ability to distinguish motion status can be formed in the graph space, thereby improving the processing accuracy and flexibility of dynamic target recognition and background elimination in the subsequent perception process.

[0054] In an exemplary embodiment, feature accumulation and updating processing is performed on the map according to different time levels to obtain feature layers of the map at different feature dimensions, including steps S401 to S403.

[0055] In step S401 , according to the reflection energy of each point data of the point cloud data, Gaussian blur processing is performed on each pixel point in the mapping image according to a preset Gaussian template to obtain a Gaussian feature map corresponding to the corresponding pixel neighborhood of each pixel point.

[0056] Among them, the Gaussian template represents a fixed-shape weight matrix constructed with a two-dimensional Gaussian function as the weight distribution, which is used to perform weighted blur processing on pixels in the image and their surrounding areas. For example, in a 3×3 Gaussian template, the center value is the largest and the edge value is the smallest, which is used to highlight the central pixel while smoothing the neighborhood changes.

[0057] The pixel neighborhood refers to the set of adjacent pixels covered by a central pixel as a reference point. For example, taking a 3×3 pixel neighborhood as an example, the area within the range of one pixel above, below, left and right around the central pixel constitutes its pixel neighborhood.

[0058] Among them, the Gaussian feature map represents a structured image obtained by performing weighted fuzzy calculation on a pixel neighborhood in the mapping image through a Gaussian template, which is used to express the continuous transition change of the energy distribution of the pixel neighborhood in the local space.

[0059] For example, the point cloud reflection energy corresponding to each pixel in the mapping image is used as the feature input, and the entire image structure is Gaussian blurred to enhance the feature continuity and regional transition within the pixel neighborhood in the image. Specifically, first, for each pixel, the reflection energy value of the pixel neighborhood corresponding to the pixel in the original point cloud data needs to be extracted as the initial feature response of the corresponding primitive in the mapping image; then, with the pixel as the center, combined with a fixed-size Gaussian template, the reflection energy values ​​of all pixels in its pixel neighborhood are weighted summed; the role of the Gaussian template is to assign different weights to pixels at different positions in the pixel neighborhood. The closer to the center point, the higher the weight, and the farther away from the center, the lower the weight. This weight distribution ensures that the feature value of the center point is preserved during the blurring process, while the feature influence of the surrounding area is also appropriately considered. After the processing is completed, each pixel no longer represents the reflection intensity of a single point, but integrates the intensity change characteristics of the pixel neighborhood in the surrounding space to generate a Gaussian feature map of the pixel in the current spatial structure.

[0060] In step S402, different update ratios are determined according to different time levels, and feature accumulation and update processing is performed on the corresponding pixel neighborhoods of each pixel point in the mapping image at the same time level according to the same update ratio, so as to obtain feature accumulation and update maps corresponding to the corresponding pixel neighborhoods of each pixel point at different time levels.

[0061] Among them, the update ratio represents the weight value used to adjust the update degree or retention degree of existing feature values ​​during the feature accumulation and update process, and is used to simulate the feature response speed at different time levels. For example, when the update ratio is 1, it means that the existing feature values ​​are completely retained at a slower feature response speed; when the update ratio is 0, it means that the existing feature values ​​are completely eliminated at a faster feature response speed.

[0062] Among them, the feature accumulation update map represents a structured image obtained after the feature accumulation update processing of the existing feature values ​​of a pixel neighborhood in the mapping map at a specific update ratio. It is used to represent the degree of feature response retention at different time levels. For example, for a layer with a larger degree of retention at a slower feature response speed, the time accumulation feature is more obvious; for a layer with a smaller degree of retention at a faster feature response speed, the time accumulation feature is less obvious.

[0063] For example, based on the settings of different time levels, feature accumulation and update processing is performed on each pixel neighborhood in the feature map to simulate and express the feature response speed. Specifically, a numerical parameter is set to adjust the trade-off between the degree of feature retention and the degree of update, namely the update ratio. The role of the update ratio is to define the relative weight of the existing eigenvalues ​​of the pixel neighborhood in the process of fusion with the new eigenvalues, thereby simulating the speed and retention of image feature responses at different time levels. Specifically, the larger the update ratio, the slower the time level, and the pixel neighborhood is more inclined to retain the existing eigenvalues, thereby showing a stronger feature retention ability; the smaller the update ratio, the faster the time level, and the pixel neighborhood pays more attention to the new eigenvalues ​​brought by the new input, thereby showing a stronger feature update ability.

[0064] During the implementation process, for each pixel point, the reflected energy value of the pixel neighborhood corresponding to the pixel point in the original point cloud data needs to be extracted as the initial feature response of the corresponding primitive in the mapping diagram; then, according to the update ratio corresponding to different time levels, the reflected energy values ​​of each pixel neighborhood are weighted according to the same update ratio at the same time level, so that pixels at different positions can show the corresponding feature response effect according to the set time level. Finally, the feature accumulation update operation of all pixels is completed at each time level, and the feature accumulation update map corresponding to the corresponding pixel neighborhood of each pixel point at each time level is obtained. That is, at each time level, the corresponding pixel neighborhood of each pixel point corresponds to a feature accumulation update map.

[0065] In step S403, the Gaussian feature maps corresponding to the corresponding pixel neighborhoods of each pixel point are fused with the feature accumulation update maps corresponding to different time levels to obtain feature layers of the mapping maps at different feature dimensions.

[0066] Exemplarily, in the corresponding pixel neighborhood of a pixel point, the Gaussian feature map is fused with the feature accumulation update map corresponding to each time level to obtain the feature fusion map of the corresponding pixel neighborhood of the pixel point in different feature dimensions; then the feature fusion maps corresponding to the corresponding pixel neighborhoods of each pixel point in the same feature dimension are combined according to the corresponding spatial position relationship to obtain the feature layer of the mapping map in the feature dimension. Based on this, according to different feature dimensions, a set of feature layers is generated for the final multi-feature dimension modeling. Specifically, for each pixel point, the corresponding feature map is first located in its pixel neighborhood, and then the feature accumulation update map at the corresponding time level is searched at the same position, and the two are fused to obtain the feature fusion map of the feature dimension corresponding to the corresponding pixel neighborhood of the pixel point at the corresponding time level; the fusion process can adopt weighted averaging, product fusion or other linear transformation-based methods to ensure that each feature fusion map retains the local continuity brought by Gaussian processing and incorporates the response characteristics reflected by time accumulation. Finally, the feature fusion maps corresponding to all pixel neighborhoods on the same feature dimension are combined to form a feature layer on the feature dimension, and all feature layers constitute a multi-dimensional expression of the frame point cloud data under multiple feature perspectives; among them, according to the spatial position relationship of each pixel neighborhood in the image space, the feature fusion maps corresponding to all pixel neighborhoods on the same feature dimension are spliced ​​and superimposed pixel by pixel to obtain a feature layer on the feature dimension.

[0067] Optionally, at each pixel point, the fusion processing of the Gaussian feature map and the feature accumulation update map can refer to formula (3): (3) in, represents the original reflection intensity within a 3×3 pixel neighborhood, Indicates the update ratio corresponding to the specified time level, It represents the feature accumulation update map obtained after the original reflection intensity of the corresponding pixel neighborhood is subjected to feature accumulation and update processing. represents a Gaussian template of size 3×3, Represents the reflection intensity of point cloud data, Represents the Gaussian feature map obtained by Gaussian blurring the reflection intensity of the corresponding pixel neighborhood according to the Gaussian template. It represents the feature fusion map corresponding to the pixel neighborhood obtained by fusing the feature accumulation update map of the corresponding pixel neighborhood with the Gaussian feature map.

[0068] Optionally, the feature layers on different feature dimensions include: a real-time dynamic point cloud mapping layer, which focuses on the properties of moving targets and avoids the phenomenon of ghosting caused by motion by quickly updating the feature capture that satisfies the moving targets; a long-term accumulation layer of dynamic point clouds, which focuses on dynamic targets that do not produce displacement, such as fans, green plants swaying in the wind, and curtains swinging, so as to be effectively applied to the filtering process of non-human targets; a cumulative layer of micro-motion point clouds, which focuses on targets that are weak, have low motion characteristics, or even stationary, so as to be effectively applied to the detection process of stationary targets; a breathing-level Doppler feature layer, which further strengthens the cumulative layer of the above-mentioned micro-motion point clouds to perceive human respiratory signs through long-term Doppler time-frequency analysis, to ensure that the human body can still be effectively detected in a sleeping state or other stationary state.

[0069] Current technologies for detecting stationary people with human perception radars are extremely limited. They typically achieve this by memorizing the position of a moving point cloud after it disappears. This method often results in target loss or deadlock during multi-target detection, leading to missed and false detections. The micro-motion point cloud accumulation layer and the breathing-level Doppler feature layer involved in this embodiment effectively overcome these issues, achieving absolute detection of stationary targets based on their micro-motion and breathing characteristics.

[0070] Furthermore, in current related technologies, human perception radars usually use a single Doppler to perform feature extraction and classification for filtering non-human targets, which makes the classification accuracy relatively limited; while the feature layers on multiple feature dimensions involved in this embodiment can make full use of the convolutional neural network under the deep learning framework for target classification, thereby improving the classification accuracy.

[0071] In this embodiment, first, Gaussian blur processing is performed on each pixel point based on the reflected energy to obtain a Gaussian feature map corresponding to each pixel neighborhood, thereby enhancing the spatial continuity and local feature expression ability of the pixel neighborhood in the image space; secondly, feature accumulation processing is performed on the pixel neighborhood according to the update ratio set at different time levels to obtain a feature accumulation update map corresponding to each pixel neighborhood at different time levels, thereby realizing feature retention and evolution modeling under different feature response speeds; thirdly, the Gaussian feature map is fused with the feature accumulation update map at different time levels in each pixel neighborhood, thereby constructing a multi-dimensional feature layer with both spatial diffusion characteristics and temporal response characteristics, providing refined, multi-angle image basic support for point cloud perception and target differentiation in complex scenes.

[0072] In an exemplary embodiment, different update ratios are determined according to different time levels, and feature accumulation and update processing is performed on the corresponding pixel neighborhoods of each pixel point in the mapping image at the same time level according to the same update ratio to obtain feature accumulation update maps corresponding to the corresponding pixel neighborhoods of each pixel point at different time levels, including steps S501 to S504; the Gaussian feature maps corresponding to the corresponding pixel neighborhoods of each pixel point are fused with the feature accumulation update maps corresponding to different time levels to obtain feature layers of the mapping image at different feature dimensions, including steps S505 to S506.

[0073] Step S501, determine the first update ratio corresponding to the long time level and the second update ratio corresponding to the short time level.

[0074] Among them, the long time level represents the time scale corresponding to the slower feature response speed and higher degree of historical information retention in the feature accumulation and update process. It is used to simulate the stable evolution state of the image under long-term observation. For example, when the update ratio is 0.9, it means that the current feature value only slightly adjusts the original state, which is suitable for expressing slowly moving or continuously existing background targets.

[0075] Among them, the short-time level represents the time scale corresponding to the faster feature response speed and the higher degree of dominance of new information in the feature accumulation and update process. It is used to simulate the sensitive change state of the image under short-time observation. For example, when the update ratio is 0.2, it means that the current feature value has a significant updating effect on the original state, which is suitable for expressing fast-moving or suddenly appearing targets.

[0076] For example, it is necessary to clearly distinguish the impact of different time levels on the feature response speed, so two representative update ratios are set to represent the feature evolution rate of the long time level and the short time level respectively. Specifically, the first update ratio corresponding to the long time level is usually set to a large value close to 1. Its role is to emphasize the continuity of historical features, so that in the process of image evolution, the existing feature values ​​in the original image can be more fully retained, thereby realizing the accumulation of long-term stable features; the second update ratio corresponding to the short time level is usually set to a smaller value. Its role is to emphasize the responsiveness of the newly added features, so that in the process of image evolution, the existing feature values ​​in the original image can be more fully updated, thereby realizing the accumulation of short-term mutation features. Based on this, different update ratios are set according to different time levels to simulate the feature evolution behavior of time levels of different scales through differences in update speeds.

[0077] Step S502, on a long-term level, performs feature accumulation and update processing on the reflected energy of each pixel point in the mapping image in the corresponding pixel neighborhood according to a first update ratio, and obtains a first feature accumulation update map corresponding to the corresponding pixel neighborhood of each pixel point on a long-term level.

[0078] Step S503, at the short-time level, the reflected energy of each pixel point in the mapping image in the corresponding pixel neighborhood is subjected to feature accumulation and update processing according to the second update ratio, and the second feature accumulation and update map corresponding to the corresponding pixel neighborhood of each pixel point at the short-time level is obtained.

[0079] For example, based on the long-term level, the reflected energy of all pixels in the image space in the corresponding pixel neighborhood is subjected to feature accumulation and update processing, and the update method adopts the first update ratio for weighted fusion. In this process, with each pixel as the center, the reflected energy within the corresponding pixel neighborhood is weightedly fused according to the first update ratio to generate new reflected energy corresponding to the corresponding pixel neighborhood. This is used as the first feature accumulation update map corresponding to the corresponding pixel neighborhood at the long-term level. This first feature accumulation update map reflects the long-term feature accumulation trend and has the ability to fully retain the historical feature state.

[0080] For example, based on the short-term level, the reflected energy of all pixels in the image space in the corresponding pixel neighborhood is subjected to feature accumulation and update processing, and the update method adopts the second update ratio for weighted fusion. In this process, with each pixel as the center, the reflected energy within the corresponding pixel neighborhood is weighted and fused according to the second update ratio to generate new reflected energy corresponding to the corresponding pixel neighborhood. This is used as the second feature accumulation update map corresponding to the corresponding pixel neighborhood at the short-term level. This second feature accumulation update map reflects the short-term feature accumulation trend and has a sufficient update frequency for the historical feature state.

[0081] Step S504: The first feature accumulation update map and the second feature accumulation update map are used as feature accumulation update maps corresponding to the corresponding pixel neighborhood of each pixel point at different time levels.

[0082] Exemplarily, the status of each pixel neighborhood in the first feature accumulation update map and the second feature accumulation update map is recorded separately, and the two are regarded as layer representations at different time levels, forming a set of feature maps with a dual-time response structure; this hierarchical structure allows subsequent processing stages to analyze and judge long-term features and short-term features separately, thereby improving the flexibility and accuracy of feature recognition.

[0083] In step S505, the Gaussian feature maps corresponding to the corresponding pixel neighborhoods of each pixel point are respectively fused with the first feature accumulation update map corresponding to the long-time layer to obtain the feature layer of the mapping map in the long-time accumulation feature dimension.

[0084] Among them, the long-term accumulation feature dimension represents the layer dimension constructed based on the feature accumulation update map at the long-term level during the fusion process. It is used to express the stable structural features formed under a slower feature response speed, such as for identifying background target areas that are continuously stationary or slowly moving. Its feature value more reflects the cumulative impact of historical status.

[0085] Exemplarily, the Gaussian feature map is used as the spatial structure basis, and is pixel-level fused with the first feature accumulation update map corresponding to the long-term level. The fusion operation is performed in each pixel neighborhood, that is, the Gaussian response value of the pixel neighborhood of the current pixel point in the Gaussian feature map and its feature value in the first feature accumulation update map are numerically synthesized, and then the layer content is integrated through weighted averaging, interpolation accumulation or regional diffusion to obtain a feature layer of the mapping map in the long-term accumulation feature dimension; this feature layer retains the local smoothness characteristics of the Gaussian template in the spatial structure, and reflects the continuous and stable change characteristics of long-term feature retention in the temporal structure.

[0086] In step S506, the Gaussian feature maps corresponding to the corresponding pixel neighborhoods of each pixel point are fused with the second feature accumulation update map corresponding to the short-time layer to obtain the feature layer of the mapping map in the short-time accumulation feature dimension.

[0087] Among them, the short-term accumulation feature dimension represents the layer dimension constructed based on the feature accumulation update graph at the short-term level during the fusion process. It is used to express sensitive change features formed under a faster feature response speed. For example, it is used to capture fast-moving or suddenly appearing physical objects. Its feature value highlights the dynamic changes of the current state.

[0088] Exemplarily, the Gaussian feature map is used as the spatial structure basis, and is pixel-level fused with the second feature accumulation update map corresponding to the short-time level. The fusion operation is performed in each pixel neighborhood, that is, the Gaussian response value of the pixel neighborhood of the current pixel point in the Gaussian feature map and its feature value in the second feature accumulation update map are numerically synthesized, and then the layer content is integrated through weighted averaging, interpolation accumulation or regional diffusion to obtain a feature layer of the mapping map in the short-time accumulation feature dimension; this feature layer retains the local smoothness characteristics of the Gaussian template in the spatial structure, and reflects the frequent sudden change characteristics of short-time feature retention in the temporal structure.

[0089] Optionally, for the dynamic point cloud data and static point cloud data in the point cloud data, the first mapping map corresponding to the dynamic point cloud data can be updated with feature accumulation at larger and smaller update ratios, and finally feature layers of the first mapping map in the long-term accumulation feature dimension and the short-term accumulation feature dimension are obtained; the second mapping map corresponding to the static point cloud data can be updated with feature accumulation at larger and smaller update ratios, and finally feature layers of the second mapping map in the long-term accumulation feature dimension and the short-term accumulation feature dimension are obtained. Based on this, point cloud clustering processing is performed according to the four feature layers obtained to obtain clustering results corresponding to the point cloud data; on the basis of the above four feature layers, feature layers corresponding to each feature information can be generated according to feature information such as density distribution, curvature, color, and depth of the point cloud data, and point cloud clustering processing is performed comprehensively in combination with the above four feature layers.

[0090] In this embodiment, first, the time level is divided into a long-term level and a short-term level according to the feature response speed, thereby determining a first update ratio corresponding to the long-term level and a second update ratio corresponding to the short-term level, thereby realizing a control basis for feature update behavior at different time scales; second, feature accumulation and update processing is performed on the pixel neighborhood reflection energy according to the first update ratio to obtain a first feature accumulation update map corresponding to the long-term level, thereby enhancing the image's long-term retention ability to stable changes; third, feature accumulation and update processing is performed on the pixel neighborhood reflection energy according to the second update ratio to obtain a second feature accumulation update map corresponding to the short-term level, thereby enhancing the image's sensitive recognition ability to sudden changes; fourth, by fusing the Gaussian feature map with the first feature accumulation update map, a feature layer on the long-term accumulation feature dimension with stability and spatial continuity is constructed; fourth, by fusing the Gaussian feature map with the second feature accumulation update map, a feature layer on the short-term accumulation feature dimension with sensitive response and spatial continuity is generated; based on this, a spatial feature layer structure for multiple time scales can be formed, providing a more hierarchical and adaptive image foundation for the separation perception and joint modeling of static background and dynamic targets.

[0091] In an exemplary embodiment, performing point cloud clustering processing on multiple feature layers to obtain clustering results corresponding to point cloud data includes steps S601 to S605.

[0092] In step S601, data points at the same position in multiple feature layers are used as detection points, and a sliding detection window including a protection area and a reference area is constructed with each detection point as the center.

[0093] Among them, the sliding detection window represents the local spatial analysis area constructed with each detection point as the center, which is used to separate and locally analyze the feature information corresponding to each detection point; the protection area represents the central area closest to the detection point in the sliding detection window, which is used to eliminate the interference of the detection point and its neighboring points on the background noise statistics to ensure the accuracy of the background noise estimation; the reference area represents the external area outside the protection area in the sliding detection window, which is used to extract statistical feature information related to the background noise to calculate the average background intensity and noise level.

[0094] For example, in order to accurately extract point cloud targets from multiple feature layers, it is first necessary to construct a unified detection point structure at the corresponding pixel position of each feature layer based on the constant false alarm rate (CFAR) detection algorithm. Specifically, the feature values ​​of multiple feature layers at the same spatial position are used as a multi-dimensional input vector of a detection point, and a sliding detection window is constructed with the detection point as the center for local area analysis. The sliding detection window includes a protection area and a reference area. The protection area directly surrounds the detection point and is used to shield the influence of the intensity of the detection point and the corresponding neighboring points on the background estimation. The reference area is located outside the protection area and is used to provide the local background information required for threshold calculation. By traversing the entire graph space, the detection points at all positions can obtain adapted local background information support in relatively independent sliding detection windows. This structure provides a spatial basis for the discrimination mechanism of constant false alarm rate detection.

[0095] Step S602 , in the process of traversing each detection point through the sliding detection window, interference elimination processing is performed on the data points in the protection area and background noise statistical processing is performed on the data points in the reference area to obtain a reference point set corresponding to each detection point.

[0096] The reference point set represents the set of all pixel points used for background statistics in the reference area of ​​the sliding detection window, which is used to perform numerical modeling and background estimation on the environmental characteristics around the detection point.

[0097] For example, under the action of the sliding detection window, local feature analysis is performed on each detection point, and the key to the processing is to effectively separate the background information from the interference data. Specifically, for the data points inside the protection area, intensity anomaly detection and elimination should be prioritized, in order to prevent the subsequent statistics from being misled by occasional noise or spike energy near the detection point; and for the data points inside the reference area, statistical feature extraction, including mean, variance, etc., is required to evaluate the overall energy level of the reference area. On this basis, these statistical feature results are not used to directly determine whether the current detection point is the target point, but are used to construct a reference point set corresponding to the detection point. The reference point set is the numerical model of its local background, which provides a basis for subsequent threshold adaptive calculations.

[0098] Step S603 , performing threshold discrimination processing on the reflection characteristics of each detection point according to the reflection intensity statistical characteristics of the reference point set corresponding to each detection point, and obtaining an adaptive threshold corresponding to each detection point.

[0099] The reflection intensity statistical characteristics represent the mean, extreme value, variance or other distribution description information related to the reflection intensity values ​​of each point in the reference point set, and are used to characterize the background noise distribution state.

[0100] The adaptive threshold represents a discrimination threshold dynamically set according to the statistical characteristics of the reflection intensity corresponding to the detection point, and is used to achieve a discrimination benchmark for the reflection intensity of the detection point under a constant false alarm rate.

[0101] For example, a dynamic setting of the reflection intensity threshold is achieved based on the statistical characteristics of the reflection intensity within the reference point set corresponding to each detection point. This threshold is no longer a globally fixed threshold, but rather an adaptive threshold generated based on the background statistics of each detection point within the reference area. The threshold is typically calculated by adding a multiple of the standard deviation to the statistical mean of the corresponding reference area, or some other linear combination, to ensure that the false alarm probability for a given detection point remains consistent across varying noise levels in the local environment. Consequently, each detection point receives an adaptive threshold corresponding to its local environment. This threshold value dynamically adjusts to environmental statistics, avoiding the misjudgment problem associated with fixed thresholds in non-uniform noise fields.

[0102] In step S604, the reflection intensity of each detection point is compared with the corresponding adaptive threshold, and the detection points whose reflection intensity exceeds the corresponding adaptive threshold are screened out as the screening results. The screening results are spatially clustered according to the spatial distribution characteristics of the screening results to obtain a set of category center points.

[0103] Among them, the spatial distribution characteristics of the screening results represent the position distribution of the target points screened by the adaptive threshold in the graph space, which is used to analyze local aggregation or spatial coherence.

[0104] Among them, the category center point set represents the set of category center points of each type of point cluster after the screened target points are clustered according to spatial proximity, which is used as a structured representation of the potential target category. For example, the geometric centroid of each cluster is used as the representative category center point of the corresponding target category.

[0105] For example, the reflection intensity of the detection point is compared with its corresponding adaptive threshold. Only when the reflection intensity is higher than the adaptive threshold, the detection point is identified as the corresponding target point. In this process, the preliminary screening of significant reflection areas in the point cloud data is completed, and the spatial distribution of the screening results is the set of areas where physical objects may exist in the physical space. Furthermore, the target points in the screening results are spatially clustered, that is, target points that are close to each other and have similar features are classified into one category based on spatial proximity and intensity coherence, thereby forming multiple structured point clusters. Each point cluster represents a spatially coherent and feature-significant reflection area, and its center position is determined by calculating the center of mass or intensity center of gravity, thereby forming a set of category center points. In this process, the reflection intensity judgment result of the detection point is converted into a structural expression result of the category center point, providing a data basis for target positioning and recognition.

[0106] Step S605 , predicting and correcting the data state of the category center point set according to the temporal characteristics of the data state of the category center point set, and obtaining the detected objects corresponding to each category center point in the category center point set as the clustering result corresponding to the point cloud data.

[0107] Among them, the data state of the category center point set represents the attribute value set of each category center point in the current image space, such as the coordinate position, height, radial velocity, average reflection intensity and number of points of the corresponding cluster of the category center point; the time series feature represents the trend of the data state of a category center point in the current frame image evolving over time, which is used for target tracking or behavior recognition.

[0108] Among them, the detected object represents the valid target point position determined by state prediction and correction of the category center point, which is used as the final detection result output by point cloud data perception. For example, the center point of a category confirmed by prediction is marked as a real moving target for subsequent tracking or identification.

[0109] For example, a dynamic state analysis is performed on the set of category center points to enhance the stability and traceability of the clustering results in the temporal dimension. Specifically, the position, intensity, area, and other data states of each category center point in the current frame are extracted and correlated with its historical data state in the previous frame. Furthermore, a simple prediction model, such as a linear prediction or a velocity estimation-based model, is constructed to predict the position and data state trend of the category center point in the next frame. Furthermore, the prediction results are matched with the actual observations in the current frame and deviation correction is performed to eliminate drift points or occasional misidentification points, thereby obtaining a more stable set of category center points. Based on the more stable set of category center points, the corresponding detected objects are determined from each category center point. This means that the predicted and corrected stable target instances contain complete spatial position, reflection properties, and temporal continuity, which can be used in subsequent applications such as target recognition, behavior modeling, or trajectory tracking. Based on this, temporal consistency processing is introduced on the basis of the constant false alarm rate detection algorithm, improving the practical application reliability of the detection structure in dynamic environments, making it extremely robust in high-noise scenes and avoiding target false detection.

[0110] In this embodiment, first, a sliding detection window including a protection area and a reference area is constructed, thereby achieving effective separation of interference information and background information in the local environment of the detection point; secondly, interference elimination processing is performed on the data points in the protection area, and background noise statistical processing is performed on the data points in the reference area to obtain a reference point set corresponding to each detection point, thereby ensuring that an adaptive threshold can be set for each detection point based on the local noise level, thereby achieving dynamic target discrimination under a constant false alarm rate; thirdly, the high-reflection points screened out by the adaptive threshold are spatially clustered to form a structured category center point set, and the category center point set is predicted and corrected to obtain the final clustering result of the point cloud data, thereby improving the temporal continuity and state stability of target recognition; based on this, robust target extraction and dynamic tracking of point cloud data based on multiple feature layers can be achieved, thereby improving the accuracy and robustness of detection in multi-target scenarios.

[0111] In an exemplary embodiment, the data state of the category center point set is predicted and corrected based on the temporal characteristics of the data state of the category center point set, and the detected objects corresponding to each category center point in the category center point set are obtained as the clustering results corresponding to the point cloud data, including steps S701 to S704.

[0112] Step S701: Based on the temporal characteristics of the data state of the category center point set, the category center point set is time-series sorted and state initialized to obtain the initial state set corresponding to each category center point. The initial state set includes a state vector and a state covariance determined based on position and velocity.

[0113] Among them, the initial state set represents a dynamic state structure set constructed according to the historical position and change trend of the center point of each category, which is used to provide the initial point state information in the Kalman algorithm. For example, the initial point state information of the center point of a certain category includes its current position, movement speed and state uncertainty.

[0114] Among them, the state vector represents an ordered set of parameters used to describe the point state of the category center point, which usually includes variables such as position and velocity, and is used for state prediction and correction calculations; the state covariance represents the matrix structure composed of the variance and covariance of the estimation errors of each variable in the state vector, which is used to quantify the degree of uncertainty in the point state prediction or observation of the category center point.

[0115] For example, in order to achieve dynamic tracking and state evolution of the category center point set in the time dimension, it is necessary to perform time sequence sorting and initialization processing on the point state of each category center point based on the basic structure of the Kalman algorithm. Specifically, the spatial position information of each category center point in the current frame is combined with the motion trend in the historical frame, and a complete initial state set is formed by constructing a state vector and a state covariance matrix. Among them, the state vector usually contains two elements, position and velocity, and its function is to simultaneously characterize the static coordinates and dynamic change trends of the target, while the state covariance matrix is ​​used to express the uncertainty of the current point state estimation and provide an error control basis for the subsequent prediction and correction stages. In addition, in order to ensure the adaptability of the prediction model, the initial point states of all category center points must be uniformly formatted so that they have mathematical structural characteristics that can be connected to the Kalman algorithm and are continuous in space and time.

[0116] Step S702 : Based on the initial state set and in combination with a preset state transition model, forward prediction processing of the category center point set with a preset time step is performed to obtain a predicted state set of each category center point in the current frame.

[0117] Among them, the state transition model represents a mathematical model used to recursively update the state vector according to a preset time step, which is used to realize the prediction of the current point state to the point state at the next moment. For example, the linear model under the assumption of uniform linear motion directly superimposes the current velocity on the position to estimate the position of the next frame.

[0118] Among them, the predicted state set represents the predicted state of the center point of each category in the current frame obtained after the forward evolution of the initial state set under the action of the state transfer model, which is used as the basis for observation correction. For example, the predicted point state includes the estimated position and estimated speed, which are used for comparison and correction with the currently observed point state.

[0119] Exemplarily, based on the initial state set, a state prediction process is performed on each category center point to infer the spatial position and dynamic properties of each category center point in the future frame at a preset time step. Specifically, this process uses a preset state transition model as the calculation basis, that is, the state vector at the previous moment is substituted into the linear state transition model, and recursive calculation is performed based on the time step to obtain the predicted values ​​of parameters such as the current position and velocity. At the same time, the corresponding covariance matrix can be updated according to a preset state error propagation model to reflect the growth of uncertainty in the predicted point state; the state error propagation model represents a mathematical model used to describe how the uncertainty of the state estimate evolves with the state transition process over time.

[0120] During the prediction process, a mathematical prior estimate is provided for the state of points that have not been observed in the current frame, so that a stable reference basis is available when observations arrive, so that the state evolution has a stable trend continuity in consecutive frames and can maintain trajectory coherence when there is no observation or the observation is abnormal.

[0121] Step S703 , performing observation correction processing based on the fusion prediction value and the observation value on the position of each category center point observed in the current frame according to the prediction state set, and obtaining a detection object set corresponding to each category center point in the current frame.

[0122] Among them, the detection object set represents the final target state set obtained by fusing the category center point actually observed in the current frame with the predicted state, which is used to represent the spatial and dynamic information of each identified target in the current frame. For example, the category center point position corrected by Kalman filtering is included in the detection object set as a valid output result.

[0123] For example, the predicted point states in the predicted state set are fused with the currently observed point states to achieve dynamic correction of the point states of each category center point. Specifically, the core of this fusion process lies in the calculation and application of the Kalman gain, which determines the weight distribution of the predicted and observed values ​​of the point states during the update process by comparing the proportional relationship between the predicted state covariance and the observation error covariance. Furthermore, for each category center point, the observation point with the highest degree of match with its predicted point state is found in the current frame, the residual between the predicted position and the actual position is calculated, and the residual is weighted and corrected using the Kalman gain to form a new state vector. The state covariance matrix is ​​also updated on this basis to reflect the improvement in state confidence.

[0124] Through this correction mechanism, not only can the prediction error be effectively compensated, but the real-time accuracy of state estimation can also be improved by fusing the current observation values. The fused detection results contain both the prediction trend and the response to the real observation values, forming the most credible point state expression in the current frame, thereby constructing the detection object set of the current frame.

[0125] Step S704 : in the detection object set, the detected objects corresponding to the respective category center points in the category center set are obtained as clustering results corresponding to the point cloud data.

[0126] For example, first, within the detection object set, all fused state vectors are verified for consistency to ensure state continuity is intact and to determine whether the category center remains within the valid tracking range. Furthermore, category center points with stable states and small prediction and observation errors are identified as valid targets in the current frame and marked as detected. Based on this, the detection object set serves as the clustering output for the point cloud data of the current frame, representing not only the spatial category distribution but also the temporal target trajectory and recognition continuity.

[0127] In this embodiment, first, the category center point set is time-series sorted and state initialized, so as to establish a complete initial state set for subsequent dynamic prediction and uncertainty quantification; secondly, the initial state set is forward predicted according to the state transition model to obtain the predicted state set, thereby realizing the continuity inference of the current frame state; thirdly, the detection object set is obtained based on the observation correction processing of the fused prediction value and observation value of the predicted state set, thereby improving the real-time accuracy and robustness of the state estimation; thirdly, the detection object set is screened and confirmed, thereby outputting the detected objects with dynamic continuity and spatial stability; based on this, the state tracking and robust identification of the point cloud target in the time series dimension can be realized, thereby improving the temporal consistency and application reliability of the clustering results.

[0128] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0129] Based on the same inventive concept, embodiments of the present application also provide a millimeter-wave radar indoor sensing-based point cloud clustering device for implementing the aforementioned millimeter-wave radar indoor sensing-based point cloud clustering method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the millimeter-wave radar indoor sensing-based point cloud clustering device provided below can be found in the limitations of the millimeter-wave radar indoor sensing-based point cloud clustering method described above and will not be repeated here.

[0130] In an exemplary embodiment, Figure 2 As shown, a point cloud clustering device based on millimeter wave radar indoor perception is provided, including: a mapping module 201, an updating module 202 and a clustering module 203, wherein: A mapping module 201 is configured to obtain point cloud data detected by a millimeter-wave radar in a physical space and map the point cloud data to a preset graph space to obtain a mapping graph corresponding to the point cloud data. An updating module 202 is configured to perform feature accumulation and updating processing on the map according to different time layers to obtain feature layers of the map at different feature dimensions; The clustering module 203 is used to perform point cloud clustering processing on multiple feature layers to obtain clustering results corresponding to the point cloud data. The clustering results are used to characterize the distribution of the point cloud data in various feature dimensions to detect and track physical objects in the physical space.

[0131] In an exemplary embodiment, the mapping module 201 is further used to: determine the reference point positions in the image space and the scaling ratio of the image space corresponding to the point cloud data based on the geometric characteristics of the image space; convert the corresponding point cloud coordinates of each point data in the point cloud data in the preset radar coordinate system into pixel coordinates in the image space in combination with the reference point positions and the scaling ratio; and obtain a mapping map corresponding to the point cloud data based on the pixel coordinates of each point data in the point cloud data in the image space.

[0132] In an exemplary embodiment, the mapping module 201 is also used to: determine the dynamic point cloud data and static point cloud data in the point cloud data based on different frequency shift characteristics of the point cloud data, and the different frequency shift characteristics reflect the different motion conditions of the physical objects in the physical space relative to the millimeter wave radar; map the dynamic point cloud data and the static point cloud data into the graph space respectively, and obtain a first mapping graph corresponding to the dynamic point cloud data and a second mapping graph corresponding to the static point cloud data, and use the first mapping graph and the second mapping graph as the mapping graphs corresponding to the point cloud data.

[0133] In an exemplary embodiment, the update module 202 is also used to: perform Gaussian blur processing on each pixel point in the mapping image according to the reflection energy of each point data of the point cloud data according to a preset Gaussian template, and obtain a Gaussian feature map corresponding to the corresponding pixel neighborhood of each pixel point; determine different update ratios according to different time levels, and perform feature accumulation update processing on the corresponding pixel neighborhood of each pixel point in the mapping image according to the same update ratio at the same time level, and obtain feature accumulation update maps corresponding to the corresponding pixel neighborhood of each pixel point at different time levels; fuse the Gaussian feature map corresponding to the corresponding pixel neighborhood of each pixel point with the feature accumulation update maps corresponding to different time levels, and obtain feature layers of the mapping image at different feature dimensions.

[0134] In an exemplary embodiment, the update module 202 is also used to: determine a first update ratio corresponding to the long time level and a second update ratio corresponding to the short time level; perform feature accumulation and update processing on the reflected energy of each pixel point in the mapping image in the corresponding pixel neighborhood according to the first update ratio on the long time level, and obtain a first feature accumulation and update map corresponding to the corresponding pixel neighborhood of each pixel point on the long time level; perform feature accumulation and update processing on the reflected energy of each pixel point in the mapping image in the corresponding pixel neighborhood according to the second update ratio on the short time level, and obtain a second feature accumulation and update map corresponding to the corresponding pixel neighborhood of each pixel point on the short time level; use the first feature accumulation and update map and the second feature accumulation and update map as the feature accumulation and update maps corresponding to the corresponding pixel neighborhood of each pixel point at different time levels; fuse the Gaussian feature map corresponding to the corresponding pixel neighborhood of each pixel point with the first feature accumulation and update map corresponding to the long time level, and obtain a feature layer of the mapping image in the long-time accumulation feature dimension; fuse the Gaussian feature map corresponding to the corresponding pixel neighborhood of each pixel point with the second feature accumulation and update map corresponding to the short time level, and obtain a feature layer of the mapping image in the short-time accumulation feature dimension.

[0135] In an exemplary embodiment, the clustering module 203 is further configured to: use data points at the same position in multiple feature layers as detection points, and construct a sliding detection window including a protection area and a reference area with each detection point as the center; in the process of traversing each detection point through the sliding detection window, perform interference elimination processing on the data points in the protection area and perform background noise statistical processing on the data points in the reference area to obtain a reference point set corresponding to each detection point; perform threshold discrimination processing on the reflection characteristics of each detection point based on the reflection intensity statistical characteristics of the reference point set corresponding to each detection point to obtain an adaptive threshold corresponding to each detection point; compare the reflection intensity of each detection point with the corresponding adaptive threshold, filter out detection points with reflection intensity exceeding the corresponding adaptive threshold as screening results, perform spatial clustering processing on the screening results based on the spatial distribution characteristics of the screening results to obtain a class center point set; predict and correct the data state of the class center point set based on the temporal characteristics of the data state of the class center point set, and obtain detected objects corresponding to each class center point in the class center point set as the clustering results corresponding to the point cloud data.

[0136] In an exemplary embodiment, the clustering module 203 is also used to: perform time series sorting and state initialization processing on the category center point set according to the time series characteristics of the data state of the category center point set, and obtain an initial state set corresponding to each category center point, the initial state set including a state vector and a state covariance determined based on the position and velocity; perform forward prediction processing on the category center point set with a preset time step in combination with a preset state transition model according to the initial state set, and obtain a predicted state set of each category center point in the current frame; perform observation correction processing on the position of each category center point observed in the current frame based on the predicted state set based on the fusion prediction value and the observation value, and obtain a detection object set corresponding to each category center point in the current frame; in the detection object set, obtain the detected objects corresponding to each category center point in the category center set as the clustering result corresponding to the point cloud data.

[0137] Each module in the millimeter-wave radar indoor perception-based point cloud clustering device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0138] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in any of the above embodiments when executing the computer program.

[0139] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in any of the above embodiments are implemented.

[0140] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0141] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0142] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A point cloud clustering method based on millimeter wave radar indoor perception, characterized in that: The method comprises: Acquire point cloud data obtained based on millimeter-wave radar detection in a physical space, and map the point cloud data into a preset graph space to obtain a mapping graph corresponding to the point cloud data; Perform feature accumulation and updating processing on the mapping graph according to different time levels to obtain feature layers of the mapping graph at different feature dimensions; Point cloud clustering processing is performed on multiple feature layers to obtain clustering results corresponding to the point cloud data, and the clustering results are used to characterize the distribution status of the point cloud data in various feature dimensions to detect and track physical objects in the physical space.

2. The method according to claim 1, characterized in that Mapping the point cloud data to a preset graph space to obtain a mapping graph corresponding to the point cloud data includes: Determining, based on geometric features of the image space, reference points in the image space and a scaling ratio of the image space corresponding to the point cloud data; In combination with the reference point position and the scaling ratio, the corresponding point cloud coordinates of each point data in the point cloud data in the preset radar coordinate system are converted into pixel coordinates in the image space; A mapping image corresponding to the point cloud data is obtained according to the pixel coordinates of each point data of the point cloud data in the image space.

3. The method according to claim 1, characterized in that Mapping the point cloud data to a preset graph space to obtain a mapping graph corresponding to the point cloud data includes: determining dynamic point cloud data and static point cloud data in the point cloud data according to different frequency shift characteristics of the point cloud data, wherein the different frequency shift characteristics reflect different motion conditions of the physical object in the physical space relative to the millimeter wave radar; The dynamic point cloud data and the static point cloud data are respectively mapped into the graph space to obtain a first mapping graph corresponding to the dynamic point cloud data and a second mapping graph corresponding to the static point cloud data, and the first mapping graph and the second mapping graph are used as mapping graphs corresponding to the point cloud data.

4. The method according to claim 1, wherein The feature accumulation and updating processing is performed on the map according to different time layers to obtain feature layers of the map at different feature dimensions, including: According to the reflection energy of each point data of the point cloud data, each pixel point in the mapping image is subjected to Gaussian blur processing according to a preset Gaussian template to obtain a Gaussian feature map corresponding to the corresponding pixel neighborhood of each pixel point; Determining different update ratios according to different time levels, performing feature accumulation and update processing on the corresponding pixel neighborhood of each pixel point in the mapping image at the same time level according to the same update ratio, thereby obtaining feature accumulation and update maps corresponding to the corresponding pixel neighborhood of each pixel point at different time levels; The Gaussian feature maps corresponding to the corresponding pixel neighborhoods of each pixel point are fused with the feature accumulation update maps corresponding to different time levels to obtain the feature layers of the mapping maps at different feature dimensions.

5. The method according to claim 4, characterized in that The method of determining different update ratios according to different time levels, performing feature accumulation and update processing on the corresponding pixel neighborhood of each pixel point in the mapping image at the same time level according to the same update ratio, and obtaining feature accumulation and update maps corresponding to the corresponding pixel neighborhood of each pixel point at different time levels, includes: Determine a first update ratio corresponding to the long time level and a second update ratio corresponding to the short time level; Performing feature accumulation and updating processing on the reflected energy of each pixel point in the mapping image in the corresponding pixel neighborhood at a long-term level according to the first update ratio, thereby obtaining a first feature accumulation and updating map corresponding to the corresponding pixel neighborhood of each pixel point at a long-term level; Performing feature accumulation and updating processing on the reflected energy of each pixel point in the mapping image in the corresponding pixel neighborhood according to the second update ratio at a short-time level to obtain a second feature accumulation and updating map corresponding to the corresponding pixel neighborhood of each pixel point at a short-time level; The first feature accumulation update map and the second feature accumulation update map are used as feature accumulation update maps corresponding to the corresponding pixel neighborhood of each pixel point at different time levels; The Gaussian feature map corresponding to the corresponding pixel neighborhood of each pixel point is fused with the feature accumulation update map corresponding to different time levels to obtain the feature layers of the mapping map at different feature dimensions, including: The Gaussian feature maps corresponding to the corresponding pixel neighborhood of each pixel point are respectively fused with the first feature accumulation update map corresponding to the long-term layer to obtain the feature layer of the mapping map in the long-term accumulation feature dimension; The Gaussian feature maps corresponding to the corresponding pixel neighborhood of each pixel point are fused with the second feature accumulation update map corresponding to the short-time level to obtain the feature layer of the mapping map in the short-time accumulation feature dimension.

6. The method according to claim 1, wherein The performing point cloud clustering processing on the plurality of feature layers to obtain a clustering result corresponding to the point cloud data includes: The data points at the same position in multiple feature layers are used as detection points. A sliding detection window containing the protection area and the reference area is constructed with each detection point as the center. In the process of traversing each detection point through the sliding detection window, interference elimination processing is performed on the data points in the protection area and background noise statistical processing is performed on the data points in the reference area to obtain a reference point set corresponding to each detection point; According to the statistical characteristics of the reflection intensity of the reference point set corresponding to each detection point, the reflection characteristics of each detection point are threshold-judged to obtain the adaptive threshold corresponding to each detection point; Comparing the reflection intensity of each detection point with the corresponding adaptive threshold, screening the detection points whose reflection intensity exceeds the corresponding adaptive threshold as the screening results, and performing spatial clustering processing on the screening results according to the spatial distribution characteristics of the screening results to obtain a set of category center points; According to the temporal characteristics of the data state of the category center point set, the data state of the category center point set is predicted and corrected, and the detected objects corresponding to each category center point in the category center point set are obtained as the clustering results corresponding to the point cloud data.

7. The method according to claim 6, characterized in that The method of predicting and correcting the data state of the class center point set based on the temporal characteristics of the data state of the class center point set to obtain the detected objects corresponding to the respective class center points in the class center point set as the clustering results corresponding to the point cloud data includes: According to the time series characteristics of the data state of the category center point set, the category center point set is subjected to time series arrangement and state initialization processing to obtain an initial state set corresponding to each category center point, wherein the initial state set includes a state vector and a state covariance determined based on position and velocity; According to the initial state set, in combination with a preset state transition model, a forward prediction process of a preset time step is performed on the category center point set to obtain a predicted state set of each category center point in the current frame; Performing observation correction processing based on the fusion prediction value and the observation value on the position of each category center point observed in the current frame according to the predicted state set, to obtain a detection object set corresponding to each category center point in the current frame; In the detection object set, the detected objects corresponding to the respective category center points in the category center set are obtained as clustering results corresponding to the point cloud data.

8. A point cloud clustering device based on millimeter wave radar indoor perception, characterized in that: The device comprises: A mapping module is used to obtain point cloud data obtained based on millimeter wave radar detection in a physical space, and map the point cloud data into a preset graph space to obtain a mapping graph corresponding to the point cloud data; An updating module, configured to perform feature accumulation and updating processing on the map according to different time layers, to obtain feature layers of the map at different feature dimensions; A clustering module is used to perform point cloud clustering processing on multiple feature layers to obtain clustering results corresponding to the point cloud data. The clustering results are used to characterize the distribution of the point cloud data in various feature dimensions to detect and track physical objects in the physical space.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images