A radar point cloud single-frame denoising method based on point density adaptive voxel division and in-voxel increment bounded sampling

Abstract

The application provides a radar point cloud single frame denoising method based on point density adaptive voxel division and voxel-increment bounded sampling, namely: obtaining a single frame radar point cloud; dividing a radar scanning distance range into multiple distance intervals, and respectively determining target voxel sizes corresponding to each distance interval; determining a target distance interval to which a to-be-processed point belongs based on a spatial distance, and quantizing a coordinate of the to-be-processed point into a voxel index by using the target voxel size; determining a measurement confidence weight according to a point cloud attribute, and updating the to-be-processed point to a sampling storage set with a sample number upper limit in a corresponding voxel based on the weight, and synchronously updating a cumulative fallen point number of the voxel; determining effectiveness based on the cumulative fallen point number of each voxel, and outputting representative points reserved in the sampling storage set in the effective voxel to obtain a denoised point cloud. The application overcomes the statistical distribution failure problem caused by near dense and far sparse of the point cloud, and realizes low memory consumption and high reliable real-time denoising through a streaming processing and a bounded update mechanism.

Description

Technical Field

[0001] This invention relates to the field of radar data processing technology, and more particularly to a single-frame noise reduction method for radar point clouds based on adaptive voxel partitioning of point density and bounded sampling of intra-voxel increments. Background Technology

[0002] In solid-state lidar-based environmental perception systems, a single-frame radar point cloud is the result of a single continuous scan of the surface of an object within the current field of view. The quality of the point cloud data determines the reliability of downstream machine vision algorithms such as target detection, scene segmentation, and 3D spatial topology construction. Because radar equipment faces a complex physical environment during scanning, the original point cloud typically contains interwoven echo signals from the object surface and a large number of discrete interference signals. Therefore, performing high-frequency, real-time, and efficient noise reduction on the acquired single-frame massive point cloud is a fundamental technical step to ensure the accuracy of the 3D perception link data and the overall operational stability of the system.

[0003] Existing point cloud denoising methods typically employ fixed-size statistical voxels or fixed-radius neighborhood search algorithms. In practice, the system divides the entire 3D space into a static grid of absolutely uniform size, or establishes a fixed-radius spherical neighborhood around each candidate point. By counting the number of neighboring point clouds within this fixed spatial range and comparing it with a set threshold, outlier noise points with lower density are identified. However, due to the angular resolution divergence inherent in radar scanning mechanisms, point cloud data exhibits a non-uniform physical characteristic in spatial distribution, with denser points in the vicinity and sparser points in the distance. When using absolutely fixed-size local neighborhoods for statistical measurement, surface points in sparsely distributed areas at a distance are easily misclassified as noise due to insufficient local point counts, while high-density areas at close range can lead to computational redundancy and a decrease in statistical discriminative power. Furthermore, traditional nearest neighbor search mechanisms, when processing massive single-frame point clouds, sometimes require building a global topology tree and performing multiple sorting operations, making it difficult to meet high-frequency throughput requirements on computationally limited real-time edge computing platforms. Additionally, truncating the output can easily lead to spatial order bias.

[0004] In summary, existing 3D point cloud denoising techniques struggle to maintain statistical consistency across distance gradients when dealing with massive, non-uniformly distributed spatial ranging data. Furthermore, current solutions suffer from high computational complexity, resulting in insufficient efficiency and robustness for large-scale streaming processing under complex conditions. Therefore, there is an urgent need to research a method that can improve data processing consistency and real-time streaming computing efficiency in non-uniform spatial perception scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a single-frame noise reduction method for radar point clouds based on adaptive voxel partitioning of point density and bounded sampling of incremental voxel, in order to solve the above-mentioned problems in the prior art.

[0006] Technical solution: A single-frame noise reduction method for radar point clouds based on adaptive voxel partitioning of point density and bounded sampling of intra-voxel increments, comprising: Acquire a single frame of radar point cloud; The pre-configured radar scanning range is divided into multiple range intervals, and the target voxel size corresponding to each range interval is determined. For the point to be processed in a single-frame radar point cloud, the target distance interval to which the point to be processed belongs is determined based on the spatial distance, and the coordinates of the point to be processed are quantized into voxel indexes using the target voxel size corresponding to the target distance interval. The measurement confidence weight is determined based on the point cloud attributes of the point to be processed. Based on the measurement confidence weight, the point to be processed is updated to the sampling storage set of the voxel corresponding to the voxel index. The sampling storage set has an upper limit on the number of samples, and the cumulative number of points falling into the corresponding voxel is updated synchronously during the process of updating the point to be processed. The validity of a voxel is determined based on the cumulative number of points it falls into, and the representative points retained in the sampling storage set of the valid voxels are output to obtain the denoised point cloud.

[0007] Furthermore, before generating the fixed-length hash key, the following steps are also included: Determine whether the calculated triaxial quantized coordinates for each axis exceed the preset representable range; If any of the three-axis quantized coordinates exceeds the representable range, the point to be processed is determined to be an outlier with out-of-bounds coordinates, and subsequent updates and retention operations for the point to be processed are terminated.

[0008] Furthermore, the operations of determining the target distance range, quantizing the voxel index, and updating the sampling storage set for the points to be processed in a single frame of radar point cloud are performed point by point in a streaming manner according to the point cloud scanning order of the radar equipment. In processing a single point to be processed, the cumulative number of points falling into the corresponding voxel is updated and the state is changed independently, without the need to establish a global spatial topological neighborhood of a single-frame radar point cloud.

[0009] Beneficial effects: This invention overcomes the statistical distribution failure problem caused by near-dense and far-sparse point clouds, and achieves low memory consumption and highly reliable real-time noise reduction through streaming processing and bounded update mechanism. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the steps of the radar point cloud single-frame noise reduction method based on point density adaptive voxel partitioning and intra-voxel incremental bounded sampling in the embodiments of this application.

[0013] Figure 2 This is a flowchart illustrating the steps for determining the target voxel size corresponding to each distance interval in the embodiments of this application.

[0014] Figure 3 This is a flowchart illustrating the process steps for obtaining the reference local point spacing corresponding to each distance interval in this embodiment of the application.

[0015] Figure 4 This is a flowchart illustrating the steps of quantizing the coordinates of the point to be processed into a voxel index in an embodiment of this application.

[0016] Figure 5 This is a flowchart illustrating the steps for determining the measurement confidence weight based on the point cloud attributes of the points to be processed in this embodiment of the application.

[0017] Figure 6 This is a flowchart illustrating the steps involved in updating the points to be processed to the sampling storage set in this embodiment of the application.

[0018] Figure 7 This is a flowchart illustrating the steps of reserving representative points within the sampling storage set in the output valid voxels in this embodiment. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] Example 1: This embodiment provides a single-frame noise reduction method for radar point clouds based on adaptive voxel partitioning of point density and bounded sampling of intra-voxel increments, such as... Figure 1 As shown, the method includes the following steps: Step 101: Obtain a single-frame radar point cloud.

[0021] Specifically, single-frame radar point clouds can be acquired in real-time through onboard or airborne solid-state LiDAR systems mounted on intelligent vehicles or drone platforms. Single-frame radar point clouds exhibit a near-dense, far-sparse distribution in physical space and are often accompanied by outliers and environmental clutter due to the influence of environmental particles or surface reflection characteristics. Acquiring single-frame radar point clouds provides a purified basic data source for downstream environmental perception systems, enabling them to accurately perform subsequent processing tasks such as target detection, scene segmentation, or obstacle avoidance decisions without significant outlier noise interference.

[0022] Furthermore, the single-frame radar point cloud is acquired in real time by vehicle-mounted or airborne solid-state LiDAR; the denoised point cloud is used as input to the downstream environmental perception system as the basic data source for performing target detection, scene segmentation or obstacle avoidance decisions.

[0023] Step 102: Divide the pre-configured radar scanning range into multiple range intervals and determine the target voxel size corresponding to each range interval.

[0024] After acquiring the raw point cloud data, since fixed-size voxels cannot adapt to changes in the density of a single-frame point cloud, the radar scanning range covered by a single-frame radar point cloud is pre-divided into multiple independent and continuously bounded range intervals through a hierarchical physical spatial segmentation.

[0025] Furthermore, a unique target voxel size is independently set for each range interval, so that the target voxel size assigned to the range interval farther from the radar ranging center is numerically larger. This segmentation and size mapping mechanism aims to normalize the number of surface expectation points inside the voxel at each spatial range gradient, ensuring that the counting-based noise reduction statistics maintain consistent resolution across the entire frame.

[0026] Step 103: For the point to be processed in a single-frame radar point cloud, calculate the spatial distance from the point to be processed to the origin of the radar coordinate system, determine the target distance interval to which the point to be processed belongs based on the spatial distance, and quantize the coordinates of the point to be processed into a voxel index using the target voxel size corresponding to the target distance interval.

[0027] Based on this, the target distance range is determined for the points to be processed in a single frame of radar point cloud, and subsequent coordinate quantization and sampling update operations are performed point by point in a streaming manner according to the point cloud scanning order of the radar equipment.

[0028] Specifically, when the point to be processed arrives in the form of a data stream, the system independently extracts its three-dimensional spatial coordinates and calculates the straight-line spatial distance from the point to be processed to the origin of the radar coordinate system. Based on this spatial distance, the system locates the uniquely matching target distance range.

[0029] Furthermore, the three-axis physical coordinates of the point to be processed are mapped using the target voxel size specific to this target range interval through division and rounding, thereby obtaining a voxel index that can uniquely identify this spatial block. In this process, the entire quantization mapping process can be completed independently relying only on the coordinate data of the point to be processed itself, without the need to pre-build a global spatial topology neighborhood network of a single-frame radar point cloud, thus meeting the requirements of high-frequency real-time throughput.

[0030] Step 104: Determine the measurement confidence weight based on the point cloud attributes of the point to be processed, and update the point to be processed to the sampling storage set of the voxel corresponding to the voxel index based on the measurement confidence weight. The sampling storage set has a preset upper limit on the number of samples, and the cumulative number of points falling into the corresponding voxel is updated synchronously during the process of updating the point to be processed.

[0031] After obtaining a unique voxel index, the system extracts the point cloud attributes carried by the point to be processed for multi-dimensional reliability assessment, and fuses them to generate a measurement confidence weight reflecting the measurement accuracy of the point. The voxel corresponding to the voxel index independently maintains a sampling storage set and an accumulator for global counting in the running memory. The sampling storage set refers to a data buffer structure with a fixed capacity limit that is independently allocated to each voxel. The samples retained in it are dynamically replaced through a continuous competitive screening mechanism, and the final resident sample is the representative point output by that voxel.

[0032] Furthermore, the sample storage set, with its preset upper limit on the number of samples, acts as a bounded local sample retention area. As point cloud data is continuously injected and updated, the system continuously competes and replaces candidate points entering the voxel based on measurement confidence weights, ensuring that points with higher measurement reliability have a greater probability of being ultimately retained in the sample storage set. Simultaneously, regardless of whether the point to be processed is ultimately retained in the storage set, the system continuously increments the cumulative number of points falling through the voxel in the accumulator. This number of falling points represents the true point density level captured in this local spatial region during the scan.

[0033] Based on this, the operations of determining the target distance range, quantizing the voxel index, and updating the sampling storage set for the points to be processed in a single frame of radar point cloud are performed point by point in a streaming manner according to the point cloud scanning order of the radar equipment. In the process of processing a single point to be processed, the cumulative number of points falling into the corresponding voxel and the replacement of the retention state are completed independently, without the need to establish a global spatial topological neighborhood of the single frame radar point cloud.

[0034] Step 105: Determine the validity of each voxel based on the cumulative number of points it falls into, and output the representative points retained in the sampling storage set of the valid voxels to obtain the denoised point cloud.

[0035] After all points in a single-frame radar point cloud have been processed, the system iterates through all activated voxels in the current hash table. By comparing the cumulative number of points with a preset minimum threshold for valid voxels, isolated noise voxels with scarce points and lacking neighborhood support are removed, classified as environmental clutter or flypoints, and discarded entirely. For valid voxels that successfully pass the validity check, the system extracts representative points that have competed multiple times and ultimately reside in the sampling storage set. High-confidence representative points across the entire frame are aggregated and merged to form a spatially uniform and unbiased denoised point cloud, completing the streaming purification output of the single-frame data.

[0036] Example 2: In one possible implementation, in determining the target voxel size corresponding to each distance interval, such as Figure 2 As shown, it includes the following steps: Step 201: Obtain the reference local point spacing corresponding to each distance interval.

[0037] As an optional implementation, the reference local point spacing reflects the average physical straight-line distance between adjacent scan points when adjacent laser beams emitted by the radar are projected onto the same physical object surface, due to the inherent divergence characteristics of angular resolution. Under constant angular resolution, the spatial sampling density of the lidar exhibits a non-uniform spatial distribution characteristic of being dense at near distances and sparse at far distances, and this average physical straight-line distance changes positively with increasing spatial distance. Before performing discretization and noise reduction processing, the system can establish a basic physical basis for the subsequent adaptive dynamic mapping of the corresponding spatial quantization grid by obtaining this reference local point spacing.

[0038] Furthermore, regarding the process of obtaining the reference local point spacing corresponding to each distance interval, as described above, such as... Figure 3 As shown, the specific steps include the following: Step 2011: Obtain the pre-configured radar horizontal and vertical angular resolutions.

[0039] Accordingly, during the factory manufacturing or initialization phase, the radar equipment internally stores the hardware calibration parameters of its optical transmission and scanning control components. The system reads the pre-configured horizontal and vertical angular resolutions from the radar's firmware metadata storage interface. These two resolutions together determine the size of the discrete dispersion angle of the spatial beam. Simultaneously, the extraction of these parameters avoids reliance on external manual configuration intervention, providing the fundamental angular input for constructing the theoretical ranging model.

[0040] Step 2012: Divide the distance intervals according to the geometric sequence and determine the center value of the distance in each interval.

[0041] To control the abrupt changes in point density distribution between adjacent range intervals, the system uses a geometric progression to divide the radar's maximum effective detection depth range. Specifically, the system calculates a dimensionless interval scaling factor based on a preset maximum detection range and minimum effective range, and uses this as the base to divide the range into multiple consecutive range segments, while simultaneously calculating the spatial geometric center position of each range segment. The calculation process for dividing the range intervals conforms to the following expression: r m =r0×α m ; Where, r m Let r0 be the boundary distance value of the m-th distance interval, α be the interval scaling factor, and m be the distance interval index.

[0042] The calculation of the interval scaling factor satisfies the following expression: α=(r max / r0) (1 / M) ; Where, r max The maximum effective detection range is pre-configured, r0 is the minimum effective ranging, and M is the number of pre-configured distance intervals. The number of distance intervals M can be adaptively configured by those skilled in the art based on the ranging depth span and computing resources of the actual scene. In this embodiment, M is set to 4.

[0043] In one optional implementation, considering the presence of suspended particles or scattered false targets at extremely close range caused by crosstalk within the sensor in actual operating scenarios, where the measured distance value may be less than the minimum effective distance, the system employs a boundary convergence rule. Specifically, for extremely close points with measured distance values ​​less than the minimum effective distance, the system suspends the conventional logarithmic mapping interval search and assigns the corresponding distance interval index to the innermost starting distance interval. Using this partitioning method, the point density ratio within each distance interval is limited to a constant coefficient window, suppressing the distortion caused by long-distance spans when using linear equal division.

[0044] Step 2013: Calculate the reference local point spacing for each distance interval using the horizontal angular resolution, vertical angular resolution, and the distance center value of each distance interval.

[0045] After performing a proportional division of the spatial detection depth, the system extracts the center distance value corresponding to each distance interval, which represents the average ranging depth of that interval in three-dimensional coordinate space. The system then multiplies this average ranging depth with pre-acquired beam divergence angle parameters to derive the theoretical physical interval between adjacent scanning points at that distance. This theoretical interval is calculated using the following expression: λ m = m×sqrt(△ θ_h ×△ θ_v ); Where, λ m This represents the reference local point spacing corresponding to the m-th distance interval. m Let △ be the center value of the distance in the m-th distance interval. θ_h For horizontal angular resolution, △ θ_v This is the vertical angular resolution. Correspondingly, △ θ_h and △ θ_v The value should be in radians.

[0046] Using the aforementioned mathematical expression, the system transforms the inherent angular measurement characteristics of the sensor into a length measurement value projected onto the object's surface. This computational mechanism eliminates the need for the system to traverse every 3D point in the full-frame point cloud to individually search for the distance between nearest neighbors. Instead, it obtains the theoretical point distance benchmark for the corresponding preset interval by looking up a table, keeping the underlying computational time overhead at a constant time complexity level.

[0047] Step 202: Based on the reference local point spacing and the preset expected number of voxel boundary points, determine the target voxel size corresponding to each distance interval.

[0048] After determining the reference local point spacing reflecting one-dimensional properties, the system expands this spacing into a target voxel size that determines the granularity of the three-dimensional spatial partition by introducing a preset expected number of voxel boundary points. The preset expected number of voxel boundary points is a dimensionless control parameter used to adjust the expected abundance of physical point clouds within the denoising spatial unit. The specific calculation process, through a linear mapping of this control parameter, is as follows: ρ m =C×λ m ; Where, ρ m Let C be the target voxel size corresponding to the m-th distance interval, C be the preset expected number of voxel boundary points, and λ be the target voxel size. m The distance between local points is for reference.

[0049] After implementing this step, the system establishes a density adaptive adjustment mechanism in the three-dimensional grid space, ensuring that the number of radar points contained within each voxel corresponding to each distance interval approaches a uniform numerical distribution at a statistical level. Closer-range intervals are assigned smaller target voxel sizes to suppress redundant calculations of high-density point clouds; longer-range intervals are assigned larger target voxel sizes to prevent low-density echo clusters on the target surface from being misjudged as environmental noise by the subsequent statistical decision module.

[0050] Example 3: In operational scenarios where radar internal calibration parameters are missing or multi-source heterogeneous data are mixed and processed, this embodiment provides an alternative technique for estimation based on data distribution characteristics. The remaining processing steps are the same as in Embodiment 2 and can be referred to the foregoing description, and will not be repeated here.

[0051] Optionally, the reference local point spacing corresponding to each distance interval can be obtained through any of the following methods: In the first execution path, a single frame of radar point cloud is randomly sampled, and the median distance between the sampled point and its nearest neighbor is calculated. The median distance is used as the reference local point spacing.

[0052] Specifically, without relying on radar angular resolution hardware parameters, the system extracts a preset number of candidate sample points within a single frame of data space according to the scanning sequence of the point cloud recording, using either equal-interval step sizes or by calling a standard uniform pseudo-random number generator. After extraction, the system constructs a topology query tree for the extracted candidate sample points using a spatial indexing algorithm. For each candidate sample point, the system traverses the topology query tree to extract its nearest neighbor with the smallest Euclidean distance in 3D space and records the distance between the two points. After obtaining the nearest neighbor distance set for all candidate sample points, the system sorts the distance set, extracts the value located at the quantile midpoint, and assigns it as the reference local point spacing. By outputting the median distance, the system eliminates the bias interference caused by locally clustered noise points and discrete edge jump points on the mean calculation, ensuring that the estimation result converges to the statistical expectation of the point density on the physical surface. This execution path limits the time cost of extracting local point spacing characteristics to a local range independent of the total number of points in the entire frame.

[0053] In the second execution path, the Euclidean distance between adjacent scan points within the same scan line in a single-frame radar point cloud is calculated, and the median distance is used as the reference local point spacing.

[0054] Specifically, for solid-state LiDAR data sources with row and column scanning temporal structure characteristics, the system parses the underlying timestamps of a single-frame radar point cloud or the row and column index codes of the scanning array to separate point cloud data sequences belonging to the same continuous scan line. For two scan points with consecutively adjacent column numbers within the same scan line, the system extracts their three-axis coordinate parameters and calculates the three-dimensional Euclidean distance. After traversing one or more scan line sequences, the system performs median filtering and extraction operations on the collected set of adjacent point distances, assigning the output median distance as the reference local point distance. This execution path reduces the spatial nearest neighbor search task in the three-dimensional spatial domain to a difference calculation operation of a one-dimensional temporal sequence, reducing the system computing power and memory consumption caused by constructing a topology structure tree.

[0055] Example 4: In one possible implementation, the coordinates of the point to be processed are quantized into voxel indices, such as... Figure 4 As shown, it includes the following steps: Step 401: Calculate the triaxial quantized coordinates of the point to be processed based on the target voxel size corresponding to the target distance interval and the pre-configured coordinate offset.

[0056] In the original physical coordinate system output by the radar equipment, the coordinate values ​​of the point to be processed are typically distributed symmetrically with the radar itself as the origin. To avoid wraparound errors caused by negative modulo operations in subsequent spatial indexing and bitwise operations, the system introduces a pre-configured coordinate offset to perform a global translation mapping of the spatial coordinate space. The value of this pre-configured coordinate offset is greater than the maximum detection bounding box size of the radar in each axis. The system extracts the single-axis coordinates of the point to be processed in the three-dimensional physical space, superimposes the pre-configured coordinate offset to convert it into a non-negative floating-point value, and performs a division operation using the target voxel size unique to the target range interval, rounding down to obtain the discretized grid index. The calculation expression for this process is: n j =floor((p j -o j +offset) / ρ m ); Where, n j p is a three-axis quantized coordinate system for a specified axis. j The physical coordinates of the point to be processed along this specified axis, o j Here, ρ represents the coordinates of the voxel mesh origin along this specified axis, offset is the pre-configured coordinate offset, and ρ is the coordinate value of the voxel mesh origin along this axis. m is the target voxel size corresponding to the target distance interval, and floor is the floor function.

[0057] By applying a translation operation to the above arithmetic expression, the system ensures that the produced three-axis quantized coordinates are all positive integers, thus eliminating sign bit interference for subsequent efficient bit assembly operations of the underlying hardware registers. In some optional implementations, the pre-configured coordinate offset can be adaptively configured according to the detection radius of the platform. For example, for autonomous driving environments with small detection radii, it can be preset to 1000 meters to provide sufficient space redundancy.

[0058] According to one aspect of this application, before generating the fixed-length hash key, the method further includes: Optionally, it can be determined whether the calculated three-axis quantized coordinates for each axis exceed the preset representable range.

[0059] Accordingly, after obtaining the discrete coordinate values ​​after translation and rounding, the system needs to perform spatial validity verification. Since the bit encoding space used to accommodate the coordinate values ​​has a bit width limit, unrestricted coordinate input may cause high-bit overflow and overwrite other control flags. The system extracts the maximum bit width reserved for each coordinate axis in the bit packing architecture, and uses this to calculate the maximum unsigned integer threshold as the preset representable range. Then, it performs a Boolean logic comparison of the quantized coordinates of each axis with this range.

[0060] Optionally, if any of the three-axis quantized coordinates in any axis exceeds the representable range, the point to be processed is determined to be an outlier with out-of-bounds coordinates, and the subsequent update and retention operations for the point to be processed are terminated.

[0061] In this embodiment, an interception mechanism for extreme abnormal data input is provided. During the system's comparison process, if a discrete coordinate value along a certain axis is found to be greater than the maximum bit width that can be represented, it indicates that the physical spatial location of the point to be processed has exceeded the maximum coverage bounding box designed by the algorithm. Since the target voxel size allocated to different target distance intervals increases with distance, its single-axis coverage range is much greater than the radar's nominal ranging capability. Therefore, points that trigger this out-of-bounds state are all characterized as spatial flying points caused by sensor hardware failure or multipath reflection. At this time, the system discards the data packet structure, releases the temporary memory resources it occupies, and returns to the next candidate point in the processing data stream. This mechanism cuts off the channel for abnormal points to invade the hash table, avoiding the error of silent overlap and merging of two distant spatial voxels due to integer overflow truncation.

[0062] Step 402: The interval identifiers generated when dividing the target distance intervals are bit-packed with the three-axis quantized coordinates to generate a fixed-length hash key as a voxel index. The interval identifiers corresponding to different target distance intervals do not overlap in the high-bit region of the fixed-length hash key.

[0063] After passing the out-of-bounds validity check, the system enters the final addressing encoding generation stage. To achieve constant-time voxel localization addressing in a single-pass streaming process and avoid the memory waste caused by using dense 3D arrays, the system employs a spatial hash architecture to map 3D spatial indices to 1D storage addresses. The system pushes the interval identifier of the target distance interval into the highest few bits of a blank 64-bit unsigned integer, divides the remaining bits into three segments, and fills them with the quantized coordinates in the X, Y, and Z axes using bitwise OR operations, assembling them to form a unique fixed-length hash key. The bit-packing encoding process is as follows: ; Where h is the generated fixed-length hash key, and m is the interval identifier of the target distance interval. Assign bits to the preset interval identifier. Assign bits to each preset axial coordinate. The quantized coordinates are in the X-axis direction. The quantized coordinates are in the Y-axis direction. The Z-axis coordinate is the quantized coordinate, and mod is the modulo operator.

[0064] By employing the aforementioned bit combination strategy, since the interval identifier representing the spatial scale exclusively occupies the high-order control domain of the fixed-length hash key, grids with different voxel sizes in physical space exhibit numerical discontinuities when mapped to the one-dimensional key-value space. This structural design allows various voxels in a multi-scale system to coexist within the same hash table without collisions, simplifying the maintenance complexity of the underlying data structure. The system uses this fixed-length hash key as a retrieval index, retrieving the memory pointer of the corresponding voxel from the hash table in a query time expected to be constant, providing an efficient access point for subsequent attribute accumulation and sample update operations.

[0065] Example 5: In one possible implementation, in determining the measurement confidence weights based on the point cloud attributes of the points to be processed, such as Figure 5 As shown, it includes the following steps: Step 501: Obtain the reflection intensity, distance measurement value, and echo identifier of the point to be processed.

[0066] After receiving the underlying data packets from the front-end digital-to-analog converter, the system parses and extracts the physical echo parameters from the radar hardware. Reflection intensity characterizes the backscattering capability of an object's surface to the laser beam; the ranging value reflects the spatial depth determined by the signal's round-trip time; and echo markers are used to distinguish multiple echo sequences generated when the laser beam encounters semi-transparent or edge-obstructed objects. Obtaining these multi-dimensional physical properties provides a multi-dimensional data foundation for subsequent assessment of the uncertainty in the physical spatial positioning of this three-dimensional coordinate point.

[0067] In other alternative implementations, point cloud attributes may include, but are not limited to, physical measurement parameters output by radar equipment such as signal-to-noise ratio, pulse width, multi-echo superposition intensity difference, and scanning time deviation. Those skilled in the art can select one or more combinations of these parameters to construct measurement confidence weights based on the data fields actually provided by the radar equipment used.

[0068] Step 502: Determine the intensity component based on the preset reference reflection intensity and reflection intensity, determine the ranging variance component based on the preset reference ranging and ranging value, and determine the echo component based on the echo identifier.

[0069] As an optional implementation, to address the dimensional differences of physical properties across different dimensions, the system employs an independent mapping mechanism to normalize them into independent factors representing reliability. Specifically, the system extracts a factory-calibrated preset reference reflection intensity as a benchmark, calculates the ratio of the actual reflection intensity carried by the point to be processed to this benchmark, and truncates and restricts this ratio to a range with a maximum value of 1, thereby obtaining an intensity component characterizing the signal-to-noise ratio, resulting in low-reflectivity target points receiving lower numerical scores.

[0070] For distance factors, the system introduces a reference distance measurement whose sensor ranging variance equals the nominal accuracy. It calculates the square of the ratio of the distance measured at the point to be processed to the reference distance, adds 1 to this square, and takes the reciprocal of the whole to obtain the ranging variance component. This component characterizes the physical attenuation law of the ranging error variance expanding with increasing distance. This mapping relationship characterizes the physical attenuation trend of ranging error increasing with distance. For echo factors, the system assigns preset constant weight coefficients to different echo identifiers by looking up a table. For example, it assigns the highest coefficient to single echoes or the first echo, and assigns progressively decreasing coefficients to intermediate echoes or the last echo, which have a higher probability of being affected by spurious interference. Based on this, the echo component is generated.

[0071] The calculation formulas for each of the above items are as follows: w i int =min(I i / I ref ,1); Among them, w i int For the strength component, I i I represents the reflection intensity at the point to be processed. ref is the preset reference reflection intensity, and min is the function to take the minimum value.

[0072] w i range =1 / (1+(r i / r ref ) 2 ); Among them, w i range For the distance measurement variance component, r i r is the distance measurement value of the point to be processed. ref This is the preset reference distance measurement.

[0073] Step 503: Integrate the intensity component, the ranging variance component, and the echo component to obtain the measurement confidence weight.

[0074] After obtaining the three independent evaluation components that have undergone dimensionless processing, the system performs a compound superposition in the probability space using a multiplication operation. The calculation formula for this process is as follows: wi =w i_int ×w i_range ×w i_echo ; Among them, w i w represents the measurement confidence weight for the points to be processed. i_int For the intensity component, w i_range For the distance measurement variance component, w i_echo This is the echo component.

[0075] This operation aggregates multi-dimensional hardware measurement uncertainties into a comprehensive scalar value. The calculated measurement confidence weights are located in the range of real numbers greater than zero and less than or equal to one; the larger the value, the closer the three-dimensional coordinate measurement result of the candidate point is to the object's reflective surface.

[0076] According to one aspect of this application, when updating the point to be processed to the sampling storage set of the voxel corresponding to the voxel index based on the measurement confidence weight, the specific steps include: Optionally, random evaluation key values ​​that are positively correlated with the measurement confidence weights are generated for the points to be processed.

[0077] To transform deterministic physical evaluation metrics into dynamic retention probabilities in a streaming data environment, the system establishes a weighted mapping mechanism by introducing random variables. The system calls a standard uniform pseudo-random number generator to extract uniformly distributed random floating-point numbers in the 0-1 interval. Using the reciprocal of the measurement confidence weight as the exponent, the system exponentiation is applied to these uniformly distributed random floating-point numbers to obtain the random evaluation key value. The expression for this operation is: key i =u i ^(1 / w i ); Among them, key i The random evaluation key value generated for the point to be processed, u i w is a random number uniformly distributed within the interval 0 to 1. i ^ represents the measurement confidence weight of the point to be processed, and ^ is the superscript symbol.

[0078] Based on this mathematical property, when the measurement confidence weight of a candidate point approaches 1, its corresponding reciprocal exponent approaches its minimum value. After the underlying uniform random numbers are exponentially processed by this minimum exponent, the resulting random evaluation key values ​​tend to cluster in the high-value range in statistical distribution. This mechanism, while maintaining the characteristics of global random exploration, assigns higher weight to highly reliable observation data.

[0079] Optionally, if the sampling storage set does not reach the upper limit of the number of samples, the points to be processed can be added to the sampling storage set.

[0080] Before updating data, the system monitors the memory usage of the sampling storage set allocated within the current target voxel. At the initial stage of the scanning task, the total number of points falling into the voxel grid is relatively small, and there are still available slots for the upper limit of the sample number. The determination of the upper limit of the sample number must meet the following constraints: its value should be less than or equal to the minimum effective voxel point threshold to ensure that the number of representative points retained within the voxel that passes the validity judgment does not exceed the original number of falling points; simultaneously, it should be greater than or equal to the total number of sub-grids to ensure that each sub-grid has a chance to obtain at least one representative sample. The upper limit of the sample number can be adaptively configured by those skilled in the art based on the actual memory budget and output point density requirements.

[0081] At this point, the system ignores the numerical comparison of random key values, writes the three-dimensional coordinates of the point to be processed, the associated point cloud attributes, and the generated random evaluation key values ​​as a whole structure into the buffer array, and updates the internal survival counter of the voxel synchronously.

[0082] Optionally, when the sample storage set reaches the upper limit of the number of samples, and the random evaluation key value of the point to be processed is greater than the minimum key value of the existing samples in the sample storage set, the existing sample corresponding to the minimum key value is replaced by the point to be processed; when the sample storage set reaches the upper limit of the number of samples, and the random evaluation key value of the point to be processed is less than or equal to the minimum key value, the point to be processed is discarded without updating the sample storage set.

[0083] As point cloud data streams continue to be input, the sampling storage set reaches full capacity, and the system transitions to a dynamic update process based on key-value competition. The system iterates through the random evaluation key values ​​of all surviving sample records within the current sampling storage set, retrieves and locks the lower bound of the values ​​(i.e., the minimum key value), and simultaneously locates the memory index of the old sample corresponding to this minimum key value. The system determines whether the random evaluation key value carried by the newly arrived point to be processed is greater than the recorded minimum key value. When the comparison result is true, the system overwrites and erases the old sample data corresponding to the minimum key value from storage, writes the structure data of the point to be processed to that memory address, and updates the lower bound record in the storage set. When the comparison result is false, the system discards the point to be processed. By continuously maintaining a fixed set of maximum key values ​​within a single frame scan cycle, the system ensures that samples in the storage set at any given time conform to a retention expectation based on weighted proportions, without needing to know in advance the final number of points the voxel will fall into. This avoids the temporal bias of the scan order and the deviation of spatial data distribution caused by the quota truncation method.

[0084] Example 6: In one possible implementation, during the process of updating the points to be processed to the sample storage set, such as Figure 6 As shown, it includes the following steps: Step 601: Divide the voxels corresponding to the voxel indexes into multiple subgrids and maintain the placeholder identifier of each subgrid in the sampling storage set.

[0085] To suppress local spatial clustering of retained samples within voxels, the system establishes a micro-diversion structure within the voxels. Specifically, the system divides the three-axis spatial dimensions of the target voxel into two equal parts, uniformly dividing the independent 3D voxel entity into eight sub-mesh of equal volume. For any reached point to be processed, the system extracts its local coordinate offset within the voxel and calculates its bisection identifier along a specified axis. The calculation formula is as follows: s j =floor(2×(p j -o j +offset-n j ×ρ m ) / ρ m ); Among them, s j To specify the bisection identifier for the specified axis, p j The physical coordinates of the point to be processed along this specified axis, o j Here, n represents the coordinates of the grid origin along this specified axis, offset is the pre-configured coordinate offset, and n is the coordinates of the grid origin along this axis. j ρ represents the triaxial quantized coordinates of the point to be processed along this specified axis. m is the target voxel size corresponding to the target distance interval, and floor is the floor function.

[0086] After obtaining the binary identifiers for each of the three axes, the system linearly combines them into a one-dimensional subgrid index ranging from 0 to 7 using a preset carry-weight matrix. The fusion calculation formula is as follows: s=s x ×4+s y ×2+s z ×1; Where s is the index of the fused one-dimensional subgrid, s x The bisection identifier in the X-axis direction, s y The binary identifier in the Y-axis direction, s z This is a binary identifier in the Z-axis direction.

[0087] Based on the calculated one-dimensional subgrid index, the system independently allocates an 8-bit binary variable as a placeholder identifier in the data structure of the sample storage set. When a sample candidate point is included in the sample storage set, the system sets the binary bit at the corresponding index position in this variable to a 1 state to record that the microscopic local space has been covered.

[0088] Step 602: Determine whether the placeholder identifier of the target subgrid to which the point to be processed belongs is empty.

[0089] The system extracts the one-dimensional subgrid index calculated from the newly arrived point to be processed and performs a bitwise AND check operation on the placeholder identifiers maintained within the storage system. Through this state detection mechanism, the system confirms whether the physical subgrid space mapped to the point to be processed has been occupied by other representative samples previously reserved in the storage set, and determines the subsequent competitive retention mechanism branch accordingly.

[0090] In the first execution path, if the placeholder marker of the target subgrid is empty, the sampling storage set has reached its maximum sample size, and multiple existing samples from the same subgrid exist in the sampling storage set, then the existing samples with duplicate placeholders in the sampling storage set will be replaced with the points to be processed first. Specifically, when new input data is detected to originate from a blank subgrid region not yet covered by existing samples, in order to improve the spatial coverage representativeness of the sample distribution within the voxel, the system skips the conventional numerical comparison competition based on measurement confidence weights. Instead, the system traverses the sample source markers within the current sampling storage set to find redundant data blocks that have contributed two or more samples from the same subgrid space.

[0091] Furthermore, the system overwrites the memory address space corresponding to the redundant samples, records the point to be processed and its multidimensional attributes, and simultaneously sets the placeholder flag bit position corresponding to the new target subgrid to a valid overwrite state. After the replacement operation is completed, if the replaced sample is the unique representative of its original subgrid in the sampling storage set, the placeholder flag bit corresponding to the original subgrid is cleared.

[0092] In the second execution path, if the placeholder identifier of the target subgrid is not empty, or if there is no existing sample with duplicate placeholder in the sampling storage set, then the random evaluation key value is compared with the minimum key value of the existing sample in the sampling storage set and the corresponding replacement or discard operation is performed.

[0093] Specifically, the system determines that the current voxel is already in a state of high spatial dispersion coverage, or that the newly injected point cloud cannot provide additional spatial topology gain. At this time, the system cancels the spatial priority constraint mechanism, extracts the random evaluation key value calculated by the point to be processed, and independently performs numerical comparison and condition replacement operations with the sample corresponding to the smallest key value in the sampled storage set.

[0094] Example 7: In one possible implementation, the validity of a voxel is determined based on the cumulative number of points it falls into, and the representative points retained in the sampling storage set of the valid voxels are output, such as... Figure 7 As shown, it includes the following steps: Step 701: Compare the cumulative number of points that each voxel falls into with the preset minimum effective voxel point threshold.

[0095] At the end of a single-frame point cloud data stream, after all point cloud data has undergone spatial quantization and sampling update operations, the system freezes the count update state of all voxels in the current hash table. At this point, each activated voxel maintains a cumulative count of recorded landing points. The system iterates through the hash table, extracts this count value, and compares it with a preset minimum effective voxel count threshold, which serves as the validity evaluation benchmark. The setting of the minimum effective voxel count threshold is mathematically related to the preset expected voxel boundary count in the distance interval division process.

[0096] Specifically, for conventional radar scanning scenarios where point clouds are scattered on the surface of 2D objects, the square of the preset expected number of voxel boundaries is greater than or equal to the minimum effective voxel count threshold. For 3D volume scattering scenarios where point clouds are scattered on structures including tree canopies or multi-layered vegetation, the cube of the preset expected number of voxel boundaries is greater than or equal to the minimum effective voxel count threshold. By maintaining the above inequality constraints, the system increases the fault tolerance of voxels generated on the object surface, preventing them from being misjudged due to statistical fluctuations.

[0097] Step 702: Voxels whose cumulative number of points is less than the minimum effective voxel threshold are identified as noise voxels, and all points within the noise voxels are discarded.

[0098] When the comparison results show that the cumulative number of landing points for a certain voxel has not reached the minimum threshold for effective voxels, the system determines that the 3D spatial region lacks sufficient local physical neighborhood support. In the imaging physics model of solid-state lidar, such sparse point density regions correspond to randomly occurring isolated flying points, environmental rain and snow particle scattering, or jitter measurement results of near-ground low reflectivity targets. Based on this, the system classifies the grid cell as a noise voxel with no engineering value. At the memory processing level, the system releases the storage unit occupied by the voxel, overwrites and erases all remaining candidate sample records in its internal sampling storage set, and cuts off the data link for noise data to enter the downstream processing stage.

[0099] Step 703: Voxels whose cumulative number of points falling into the voxel are greater than or equal to the minimum threshold of effective voxels are determined as effective voxels, and the spatial coordinates and point cloud attributes of each representative point within the effective voxel are output simultaneously.

[0100] When the comparison result meets or exceeds the set numerical threshold, the system confirms that the spatial region enclosed by the voxel has a physical reflective surface and classifies it as a valid voxel. The system then accesses the memory address of the voxel and extracts the representative point data that finally resides in the sampling storage set after the previous weighted competition and subgrid dispersion constraints.

[0101] During output operations, the system extracts the three-axis spatial coordinates of each representative point in the radar coordinate system. Simultaneously, it packages and extracts multi-dimensional point cloud attributes acquired during the hardware analysis phase, such as reflection intensity, ranging values, and echo identifiers. These point cloud attributes are then bound to the spatial coordinates as auxiliary variables in the output. All output representative points converge on the system bus, forming a denoised discrete point cloud set. By employing a synchronous output of multi-dimensional attributes, the system transmits complete physical echo features to the autonomous driving terminal or target detection module, providing underlying data input for downstream algorithms to evaluate target material and recognize environmental semantic information.

[0102] In a typical application scenario, this method was applied to a single frame of approximately 300,000 points in a raw point cloud acquired by a solid-state lidar platform. The processing results show that distant target surface points are effectively preserved, isolated flying points are effectively filtered out, and the spatial distribution uniformity of the output point cloud is superior to that of the fixed voxel denoising method. Furthermore, the single-frame processing latency meets real-time requirements. Those skilled in the art will understand that specific performance indicators may vary depending on the hardware and the application scenario.

[0103] Example 8: When processing cross-source point clouds generated by different radar sorties or different scanning configurations on the same physical surface, this embodiment provides a density normalization method for a single global interval.

[0104] When processing multi-source point cloud data, the spatial point density of the same physical surface varies under different radar hardware configurations. The system assigns a value of 1 to the range interval division parameter, which means that the entire physical space between the farthest effective detection boundary and the nearest effective range of the radar scan is identified as a single global range interval.

[0105] The system acquires the three-dimensional coordinate distribution data of the radar point cloud in the current single frame, calculates the statistical median value of the distance from all points to be processed to the radar space origin, and assigns it as the representative distance within the frame.

[0106] Furthermore, the system uses the beam divergence angle parameter and the representative distance within the frame to derive the density adaptive parameters applicable to the entire point cloud within that frame. The expression for this calculation process is as follows: ε=(C× ×sqrt(△ θ_h ×△ θ_v )) / d; Where ε is the global adaptive factor, and C is the preset expected number of voxel boundary points. To calculate the representative intra-frame distance, △ θ_h For the pre-configured horizontal angular resolution, △ θ_v d is the pre-configured vertical angular resolution, and d is the pre-configured reference voxel size.

[0107] After obtaining the global adaptive factor calculated above, the system multiplies the pre-configured baseline voxel size with this global adaptive factor to obtain a unified target voxel size applicable to the current single-frame point cloud. In the subsequent three-axis coordinate quantization and voxel index generation stages, the system calls this unified target voxel size to perform coordinate discretization operations on all points to be processed in the entire frame. Using this implementation method, the system uniformly maps the scale differences in local point spacing between cross-source data to the scaling ratio of the grid size, so that the upper limit of the downstream sample number and the threshold of the minimum number of effective voxels do not need to be adjusted with the change of radar hardware parameters.

[0108] This scheme improves the problem of statistical failure of gradients across distances caused by non-uniform spatial distribution by dividing the radar scanning space into layers according to distance intervals and matching adaptively enlarged target voxel sizes to different intervals based on beam divergence characteristics. This mechanism achieves fine voxels at near distances to preserve high-frequency geometric details and wide voxels at far distances to encompass discrete echoes. It ensures a constant and uniform expected number of points on the internal surface of voxels across all gradients of physical ranging, thus maintaining statistical consistency of the noise reduction decision threshold across the global space and avoiding the false rejection of distant targets.

[0109] Meanwhile, to address the computational conflict between high-frequency throughput of point clouds and offline global topology search, the physical coordinates are discretized and quantized using adaptive voxel size. The interval identifier representing the spatial scale and the three-axis quantized coordinates are integrated and encapsulated into a high-order mutually exclusive fixed-length hash key, avoiding the static memory waste caused by dense three-dimensional arrays. Spatial positioning is compressed into hash addressing with constant time complexity, enabling massive point cloud data to be streamed in a non-blocking single pass according to the radar hardware scanning order, reducing the computational load on the edge computing platform.

[0110] Furthermore, to address the spatial bias and loss of high-value target points caused by conventional capacity quota truncation, a confidence level is constructed by integrating multi-dimensional physical measurement factors such as reflection intensity and ranging variance. A dynamic random key-value competition elimination mechanism positively correlated with this confidence level is introduced into the capacity-bounded sampling and storage set, and a micro-subgrid occupancy constraint to suppress local clustering is superimposed. This processing mechanism not only ensures the high probability of retaining high-reliability echo points statistically, but also disrupts the spatial clustering pattern of the point cloud, ultimately outputting a denoised point cloud with uniform and bounded density distribution and high confidence in reflecting the physical reflective surface, providing a data source for downstream environmental sensing.

[0111] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0112] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0113] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A single-frame noise reduction method for radar point clouds based on adaptive voxel partitioning of point density and bounded sampling of intra-voxel increments, characterized in that, include: Acquire a single frame of radar point cloud; The pre-configured radar scanning range is divided into multiple range intervals, and the target voxel size corresponding to each range interval is determined. For the point to be processed in a single-frame radar point cloud, the target distance interval to which the point to be processed belongs is determined based on the spatial distance, and the coordinates of the point to be processed are quantized into voxel indexes using the target voxel size corresponding to the target distance interval. The measurement confidence weight is determined based on the point cloud attributes of the point to be processed. Based on the measurement confidence weight, the point to be processed is updated to the sampling storage set of the voxel corresponding to the voxel index. The sampling storage set has an upper limit on the number of samples, and the cumulative number of points falling into the corresponding voxel is updated synchronously during the process of updating the point to be processed. The validity of a voxel is determined based on the cumulative number of points it falls into, and the representative points retained in the sampling storage set of the valid voxels are output to obtain the denoised point cloud.

2. The method according to claim 1, characterized in that, Determining the target voxel size corresponding to each distance interval includes: Obtain the reference local point spacing corresponding to each distance interval; Based on the reference local point spacing and the preset expected number of voxel boundaries, the target voxel size corresponding to each distance interval is determined.

3. The method according to claim 2, characterized in that, In obtaining the reference local point spacing corresponding to each distance interval, the following is included: Obtain the pre-configured radar horizontal and vertical angular resolution; Divide the distance intervals according to the geometric sequence and determine the center value of the distance in each interval; Using the horizontal angular resolution, vertical angular resolution, and the distance center value of each distance interval, the spacing between reference local points corresponding to each distance interval is calculated.

4. The method according to claim 2, characterized in that, The reference local point spacing corresponding to each distance interval can be obtained through any of the following methods: Randomly sample a single frame of radar point cloud, calculate the median distance between the sampled point and its nearest neighbor, and use the median distance as the reference local point spacing. Alternatively, the Euclidean distance between adjacent scan points within the same scan line in a single-frame radar point cloud can be calculated, and the median distance can be used as a reference local point spacing.

5. The method according to claim 1, characterized in that, The coordinates of the point to be processed are quantized into voxel indices, including: Based on the target voxel size corresponding to the target distance interval and the pre-configured coordinate offset, calculate the three-axis quantized coordinates of the point to be processed; The interval identifier of the target distance interval is bit-packed with the three-axis quantized coordinates to generate a fixed-length hash key as a voxel index. The interval identifiers corresponding to different target distance intervals do not overlap in the high-bit region of the fixed-length hash key.

6. The method according to claim 1, characterized in that, In determining the measurement confidence weights based on the point cloud attributes of the points to be processed, the following are included: Acquire the reflection intensity, distance measurement value, and echo identifier of the point to be processed; The intensity component is determined based on the preset reference reflection intensity and reflection intensity, the ranging variance component is determined based on the preset reference ranging and ranging value, and the echo component is determined based on the echo identifier. The measurement confidence weights are obtained by integrating the intensity component, the ranging variance component, and the echo component.

7. The method according to claim 1, characterized in that, When updating the points to be processed to the sampling storage set of the voxels corresponding to the voxel index based on the measurement confidence weight, the specific steps include: Generate random evaluation key values ​​that are positively correlated with the measurement confidence weights for the points to be processed; When the sampling storage set has not reached the upper limit of the number of samples, the points to be processed are added to the sampling storage set; When the sample storage set reaches the upper limit of the number of samples, and the random evaluation key value of the point to be processed is greater than the minimum key value of the existing samples in the sample storage set, the existing sample corresponding to the minimum key value is replaced by the point to be processed.

8. The method according to claim 7, characterized in that, The process of updating the points to be processed to the sampling storage set also includes: The voxels corresponding to the voxel indexes are divided into multiple sub-grids, and the placeholder identifier of each sub-grid in the sampling storage set is maintained. Determine if the placeholder marker of the target subgrid to which the point to be processed belongs is empty; If the placeholder of the target subgrid is empty and the sampling storage set has reached the upper limit of the number of samples, then the existing samples with duplicate placeholders in the sampling storage set will be replaced with the points to be processed first. If the placeholder identifier of the target subgrid is not empty, or if there is no existing sample with duplicate placeholder in the sampling storage set, then the random evaluation key value is compared with the minimum key value of the existing sample in the sampling storage set and the corresponding replacement or discard operation is performed.

9. The method according to claim 1, characterized in that, The validity of a voxel is determined based on the cumulative number of points it falls into, and the representative points retained in the sampling storage set of the valid voxels are output, including: The cumulative number of points that each voxel falls into is compared with the preset minimum effective voxel point threshold. Voxels whose cumulative number of points is less than the minimum effective voxel count threshold are identified as noise voxels, and all points within the noise voxels are discarded. Voxels whose cumulative number of points falling on them is greater than or equal to the minimum threshold of effective voxels are identified as effective voxels, and the spatial coordinates and point cloud attributes of each representative point within the effective voxel are output simultaneously.

10. The method according to any one of claims 1 to 9, characterized in that, Single-frame radar point cloud is acquired in real time by vehicle-mounted or airborne solid-state lidar; The denoised point cloud is used as input to the downstream environmental perception system as the basic data source for performing target detection, scene segmentation or obstacle avoidance decisions.

Images