Point cloud quality enhancement method and device based on kernel density analysis, equipment and medium

Abstract

This disclosure relates to the field of autonomous driving technology, and discloses a method, apparatus, device, and medium for point cloud quality enhancement based on kernel density analysis. The method involves generating a perforated raster mask of the same size as the planar raster image corresponding to the dense point cloud to be processed; obtaining a target planar raster image using a preset pixel-level image inpainting technique and the perforated raster mask; analyzing the occupancy probability of each raster in the target planar raster image by the foreground using a preset kernel density analysis method; and performing point cloud selection processing on the dense point cloud to be processed based on the occupancy probability of each raster in the target planar raster image, combined with a preset probability threshold, to obtain a target sparse point cloud corresponding to the dense point cloud to be processed. By combining pixel-level image inpainting technology with kernel density analysis, effective point cloud regions are effectively identified while ensuring data integrity, thus generating a high-quality target sparse point cloud.

Description

Technical Field

[0001] This disclosure relates to the field of autonomous driving technology, specifically to a point cloud quality enhancement method, apparatus, device, and medium based on kernel density analysis. Background Technology

[0002] In the current autonomous driving technology system, the construction of high-precision maps relies heavily on 3D point cloud data collected by LiDAR. The extraction of ground traffic markings (such as lane lines, directional arrows, stop lines, etc.) is mainly based on the reflection intensity of the point cloud. However, the quality of the actual collected point cloud is often affected by multiple factors, such as the inherent noise of radar equipment, changes in vehicle motion attitude, aging of road marking materials or differences in reflectivity, and adverse weather conditions such as rain, fog, and strong light. These factors can all lead to weak reflection intensity in the effective area (foreground) and high reflection intensity in the ineffective area (background) of the point cloud, resulting in lower point cloud quality and further affecting the accuracy of the high-precision map and the reliability of the onboard perception system. In existing technologies, threshold bisection is often used to divide point clouds or their rasterized images into foreground and background to improve point cloud quality. However, the point clouds processed by this method are prone to problems such as loss of edge details and preservation of background noise. For example, low-intensity but valid markers are easily removed and high-intensity non-marked areas are misjudged as valid markers. Therefore, existing point cloud quality enhancement methods limit the integrity and accuracy of the enhanced point clouds and still have the problem of low quality, which cannot provide a high-quality foundation for the construction of high-precision maps. Summary of the Invention

[0003] This disclosure addresses the problems existing in the prior art by providing a point cloud quality enhancement method, apparatus, device, and medium based on kernel density analysis.

[0004] To achieve the above objectives, the technical solution adopted in this disclosure is as follows: A first aspect of this disclosure provides a point cloud quality enhancement method based on kernel density analysis, comprising: generating a holed raster mask of the same size as the planar raster image corresponding to the dense point cloud to be processed; obtaining a target planar raster image using a preset pixel-level image inpainting technique and the holed raster mask; analyzing the occupancy probability of each raster in the target planar raster image being occupied by the foreground using a preset kernel density analysis method, wherein the foreground refers to a region that needs to be retained with a pixel value higher than a preset pixel value; and performing point cloud scavenging processing on the dense point cloud to be processed based on the occupancy probability of each raster in the target planar raster image and a preset probability threshold to obtain a target sparse point cloud corresponding to the dense point cloud to be processed.

[0005] In some embodiments of this disclosure, the step of performing point cloud scavenging on the dense point cloud to be processed based on the occupancy probability of each grid in the target planar grid image and in combination with a preset probability threshold to obtain a target sparse point cloud corresponding to the dense point cloud to be processed includes: taking the grids in the target planar grid image with occupancy probabilities higher than the preset occupancy probability as foreground grids occupied by the foreground; taking the grids with occupancy probabilities lower than the preset occupancy probability as background grids occupied by the background; setting the reflection intensity of multiple points in the dense point cloud to be processed corresponding to each background grid to zero; and for each foreground grid, randomly selecting a reflection intensity from the reflection intensity of multiple points in the dense point cloud to be processed corresponding to the foreground grid as the reflection intensity of the multiple points corresponding to the foreground grid, thereby obtaining the target sparse point cloud.

[0006] In some embodiments of this disclosure, obtaining a target planar raster image using a preset pixel-level image inpainting technique and the hole raster mask includes: performing hole inpainting processing on the planar raster image using the preset pixel-level image inpainting technique combined with the hole raster mask to obtain an inpainted raster image; obtaining a binarized gradient map corresponding to the inpainted raster image; and performing image union processing on the binarized gradient map and the planar raster image to obtain the target planar raster image.

[0007] In some embodiments of this disclosure, obtaining the binarized gradient map corresponding to the repaired raster image includes: performing gradient enhancement processing on the repaired raster image using a preset convolution kernel to obtain a gradient map; performing filtering processing on the gradient map using a first cutoff frequency to obtain a low-noise gradient map; performing secondary repair processing on the low-noise gradient map using a preset fence filter combined with the kernel size of the preset convolution kernel to obtain a repaired gradient map; and performing binarization processing on the repaired gradient map using the Otsu method image binarization to obtain a binarized gradient map.

[0008] In some embodiments of this disclosure, the step of performing image union processing on the binarized gradient map and the planar raster map to obtain a target planar raster map includes: performing secondary filtering processing on the planar raster map using a second cutoff frequency to obtain a filtered planar raster map, wherein the second cutoff frequency is greater than the first cutoff frequency; using the binarized gradient map as a mask map, and combining it with the filtered planar raster map to perform image union processing to obtain the target planar raster map.

[0009] In some embodiments of this disclosure, the step of analyzing the occupancy probability of each grid in the target planar grid image being occupied by the foreground using a preset kernel density analysis method includes: taking grids in the target planar grid image whose reflection intensity grayscale value is not 0 as valid grids; for each valid grid, taking the valid grid as a heat map source, calculating the coverage probability of each grid in the target planar grid image being covered by the heat map source; and for each grid in the target planar grid image, calculating the occupancy probability of the grid being occupied by the foreground based on the multiple coverage probabilities corresponding to the grid.

[0010] In some embodiments of this disclosure, the step of using a preset pixel-level image inpainting technique combined with the hole grid mask to perform hole repair processing on the planar grid image to obtain a repaired grid image includes: determining the grid to be repaired in the planar grid image based on the hole marker in the hole grid mask; and for each grid to be repaired, performing hole repair processing on the grid to be repaired using the preset pixel-level image inpainting technique combined with the grayscale value of the reflection intensity of the grids surrounding the grid to be repaired to obtain a repaired grid image.

[0011] In some embodiments of this disclosure, before generating a hole grid mask of the same size as the planar grid map corresponding to the dense point cloud to be processed, the method further includes: acquiring initial single-frame point clouds collected by multiple radars at each time point within a preset time period and pose data of each radar; for each time point within the preset time period, combining the initial single-frame point cloud collected by each radar at that time point and the pose data of the radar to form a single-frame point cloud; and preprocessing the multiple single-frame point clouds within the preset time period using a dual-buffering parallel acceleration strategy to obtain the dense point cloud to be processed.

[0012] A second aspect of this disclosure discloses a point cloud quality enhancement device based on kernel density analysis, comprising: a generation unit for generating a holed raster mask of the same size as the planar raster image corresponding to a dense point cloud to be processed; an acquisition unit for acquiring a target planar raster image using a preset pixel-level image inpainting technique and the holed raster mask; an analysis unit for analyzing the occupancy probability of each raster in the target planar raster image being occupied by the foreground using a preset kernel density analysis method, wherein the foreground refers to a region that needs to be retained with a pixel value higher than a preset pixel value; and a processing unit for performing point cloud rejection processing on the dense point cloud to be processed based on the occupancy probability of each raster in the target planar raster image and a preset probability threshold, to obtain a target sparse point cloud corresponding to the dense point cloud to be processed.

[0013] This disclosure also provides an electronic device, comprising: a memory for storing at least one instruction; and a processor for invoking the instruction stored in the memory to execute the point cloud quality enhancement method based on kernel density analysis in the first aspect and any embodiment of the first aspect.

[0014] This disclosure also provides a computer-readable storage medium and a computer program product, characterized in that the computer-readable storage medium stores at least one executable instruction, which is loaded and executed by a processor to implement the point cloud quality enhancement method based on kernel density analysis in the first aspect and any embodiment of the first aspect; the computer program product includes computer program code, which, when executed by a computer, causes the computer to execute the point cloud quality enhancement method based on kernel density analysis in the first aspect and any embodiment of the first aspect.

[0015] Compared with the prior art, this disclosure has the following beneficial effects: First, the planar raster image is repaired using a generated hole raster mask and pixel-level image inpainting techniques to ensure data integrity and continuity, reduce information loss, and lay a solid foundation for subsequent point cloud quality analysis. Second, kernel density analysis is used to calculate the probability of each raster in the target planar raster image being occupied by the foreground, in order to analyze more accurate foreground information. This accurately identifies the effective point cloud regions that need to be retained, effectively distinguishing between effective point clouds and redundant or noisy point clouds, thus improving the accuracy of foreground identification. Finally, based on the occupancy probability of each raster and a preset probability threshold, targeted selection is performed on the dense point cloud to be processed, eliminating redundant data and noisy point clouds to generate a target sparse point cloud with higher accuracy and reliability. This approach not only optimizes the accuracy and reliability of point cloud data, greatly improving point cloud quality, but also improves subsequent processing efficiency by reducing redundant information, providing strong support for high-precision map production. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a point cloud quality enhancement method based on kernel density analysis provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of a dual-buffered parallel acceleration processing procedure provided according to an embodiment of the present disclosure; Figure 3 This is a block diagram of a point cloud quality enhancement device based on kernel density analysis provided according to an embodiment of the present disclosure. Detailed Implementation

[0017] The present disclosure will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present disclosure and should not be construed as limiting the scope of protection of the present disclosure. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application.

[0018] The acquisition, transmission, storage, use, and processing of data in this disclosed technical solution comply with relevant national laws and regulations. In the embodiments of this disclosure, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this disclosure, and do not imply that the applicant has already used or necessarily used such solutions.

[0019] All terms used in this disclosure have the same meaning as understood by one of ordinary skill in the art to which this disclosure pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein.

[0020] Example 1; The point cloud quality enhancement method based on kernel density analysis in this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. Its specific process can be as follows: Figure 1 As shown, it includes: Step 110: Generate a hole grid mask with the same size as the planar grid map corresponding to the dense point cloud to be processed.

[0021] Specifically, generating a holed raster mask of the same size as the planar raster image corresponding to the dense point cloud to be processed includes: obtaining the planar raster image corresponding to the dense point cloud to be processed, and performing a hole dilation operation on the planar raster image to generate a holed raster mask of the same size as the planar raster image.

[0022] The term "dense point cloud to be processed" refers to a dense point cloud that requires quality enhancement. This dense point cloud is generated from point clouds collected by multiple radars over a preset time period, along with the radar's pose data. The planar raster image refers to projecting the dense point cloud to be processed onto a two-dimensional plane and converting it into a raster image through voxelization. Each grid cell in this planar raster image corresponds to multiple points in the dense point cloud with the same horizontal and vertical coordinates. The hole dilation operation expands the hole regions in the planar raster image to increase the size of the holes that need to be filled, avoiding missed holes and reducing information loss. The hole raster mask is a raster image obtained by using a binary method to label the grids in the planar raster image that need repair and those that do not.

[0023] Further, the aforementioned acquisition of the planar raster image corresponding to the dense point cloud to be processed includes: acquiring the first abscissa and the first ordinate of the bottom left corner point of the bounding box surrounding the dense point cloud to be processed; calculating and processing the coordinates of each point in the dense point cloud to be processed using a preset voxelization formula and a preset resolution in combination with the first abscissa and the first ordinate, to obtain the raster coordinates corresponding to the coordinates of each point in the dense point cloud to be processed, thereby voxelizing the dense point cloud to be processed into an initial planar raster image, the initial planar raster image containing multiple grates, one grate corresponding to multiple points in the dense point cloud to be processed; for each grate in the initial planar raster image, taking the maximum reflection intensity of the multiple points in the dense point cloud to be processed corresponding to the grate as the reflection intensity grayscale value of the grate, thereby obtaining an intermediate planar raster image, wherein one grate coordinate represents one grate, and one point coordinate represents one point; normalizing the reflection intensity grayscale value in the intermediate planar raster image to obtain the planar raster image.

[0024] In this context, a bounding box is the smallest cylinder capable of enclosing a dense point cloud to be processed. The bottom surface of the bounding box is the surface of the bounding box that is parallel to the planar raster image and has a small z-value at its midpoint.

[0025] For example, the coordinates of each point in the dense point cloud to be processed are calculated using a preset voxelization formula, a preset resolution, and the first horizontal and vertical coordinates, to obtain the raster coordinates corresponding to each point in the dense point cloud to be processed. See the following formula: ; in, The first x-coordinate, y is the first ordinate. d is the preset resolution. (u, v) are the raster coordinates of multiple points in the dense point cloud to be processed, with x as the x-coordinate and y as the y-coordinate.

[0026] Further, the aforementioned hole dilation operation on the planar grid image to generate a hole grid mask of the same size as the planar grid image includes: using 1 to identify holes in the planar grid image and using 0 to identify non-holes in the planar grid image, thereby generating a hole binary image; using a preset dilation kernel to expand the hole identifiers in the hole binary image to obtain an expanded hole binary image, wherein the hole identifier is 1; generating a null image of the same size as the planar grid image; and performing a union operation on the null image and the expanded hole binary image to generate a hole grid mask of the same size as the planar grid image.

[0027] Therefore, by using a preset expansion kernel to enlarge the hole markers in the binary hole map, the originally small or large hole areas can be expanded, ensuring more comprehensive coverage of areas where data loss or incomplete acquisition may occur during subsequent repair processing. Furthermore, by clearly marking holes and non-holes to generate a specialized hole grid mask, it is beneficial to accurately locate the hole areas requiring repair during subsequent repair processing, facilitating rapid repair and saving repair time.

[0028] Step 120: Obtain the target planar grid image using a preset pixel-level image inpainting technique and the hole grid mask.

[0029] Among them, the preset pixel-level image inpainting technique can be a fluid dynamics-based image inpainting technique. The target planar raster image refers to the raster image obtained after the planar raster image has undergone hole removal and the preset pixel-level image inpainting technique.

[0030] Specifically, in step 120 above, obtaining the target planar raster image using the preset pixel-level image inpainting technique and the hole raster mask includes: performing hole inpainting processing on the planar raster image using the preset pixel-level image inpainting technique combined with the hole raster mask to obtain the inpainted raster image; obtaining the binarized gradient image corresponding to the inpainted raster image; and performing image union processing on the binarized gradient image and the planar raster image to obtain the target planar raster image.

[0031] In this context, "patching a raster image" refers to obtaining a raster image by repairing the raster cells in a planar raster image that require repair. "Image union" processing refers to a method that performs an "OR" logical check on pixels at the same location in the binarized gradient image and the planar raster image to generate a new target planar raster image. For example, performing the "OR" logical check means comparing the pixel values ​​at the same location in the two images; if one of the two pixels is 1, then the corresponding pixel value at that location in the generated target planar raster image is 1.

[0032] Furthermore, the aforementioned method of using a preset pixel-level image inpainting technique combined with the hole grid mask to perform hole repair processing on the planar grid image to obtain a repaired grid image includes: determining the grid to be repaired in the planar grid image based on the hole marker in the hole grid mask; and for each grid to be repaired, performing hole repair processing on the grid to be repaired using the preset pixel-level image inpainting technique combined with the grayscale value of the reflection intensity of the grids surrounding the grid to be repaired to obtain a repaired grid image.

[0033] The hole marker is used to identify the raster that needs hole repair. The preset pixel-level image inpainting technique is based on fluid dynamics strategies. It adaptively performs pixel-level image inpainting on the raster to be repaired, ensuring consistent brightness lines, by combining the reflection intensity grayscale values ​​of the surrounding rasteres. This results in a smooth change in the reflection intensity grayscale values ​​between the repaired raster and its surrounding rasteres, making the repaired raster image more consistent with the physical characteristics of the real scene and significantly improving the quality and reliability of the repaired raster image.

[0034] Further, obtaining the binarized gradient map corresponding to the repaired raster image includes: performing gradient enhancement processing on the repaired raster image using a preset convolution kernel to obtain a gradient map; performing filtering processing on the gradient map using a first cutoff frequency to obtain a low-noise gradient map; performing secondary repair processing on the low-noise gradient map using a preset fence filter combined with the kernel size of the preset convolution kernel to obtain a repaired gradient map; and performing binarization processing on the repaired gradient map using the Otsu method image binarization to obtain a binarized gradient map.

[0035] The preset convolution kernel can be set according to actual needs, such as a 3×3 Sobel operator or a 5×5 Sobel operator. The Sobel operator is a classic image edge detection operator widely used in computer vision and image processing. It highlights areas of rapid gray-level change, i.e., edges, in an image by calculating approximate gradient values ​​in the horizontal and vertical directions. The first cutoff frequency is used for high-pass filtering of the gradient map to remove background noise. The foreground refers to the target object of the task; for example, in a road scene, the foreground could be ground traffic signs, curbs, speed bumps, etc., and the foreground has a high reflectance gray-level value. The background refers to non-target objects; for example, in a road scene, the background could be ordinary asphalt, concrete pavement, etc., and the background has a low reflectance gray-level value. The low-noise gradient map is the gradient map after removing most of the background noise. Preset fence filtering is used to fill or repair the reflection intensity grayscale values ​​of the center grid cells within the bounded region of a low-noise gradient map defined by a preset convolution kernel. This compensates for defects in the low-noise gradient map or further smooths it, thereby improving its quality and resulting in a higher-quality repaired gradient map. Otsu's image binarization method is a method that automatically determines whether each grid cell in the repaired gradient map belongs to the background or foreground, performs binarization processing on the repaired gradient map, and uses binarization to identify whether a grid cell belongs to the background or foreground.

[0036] For example, gradient enhancement processing is performed on the repaired raster image using a preset convolution kernel to obtain a gradient image, as shown in the following formula: ; in, This is the convolution kernel in the x-axis direction, which is the kernel in the horizontal direction of the predefined convolution kernel. This is the convolution kernel in the y-axis direction (the kernel along the preset convolution kernel axis). I represents the patched raster image. GI represents the gradient image.

[0037] Therefore, the gradient enhancement method using a preset convolution kernel can accurately detect and enhance edge information in an image. Based on the repaired raster image containing relatively complete information, the quality of the repaired raster image is further enhanced, resulting in a better gradient image quality. This provides a solid foundation for enhancing point cloud quality and plays an important role in high-precision map production and autonomous driving perception systems.

[0038] Furthermore, the gradient map is filtered in conjunction with the first cutoff frequency to obtain a low-noise gradient map, including: converting the gradient map into a frequency gradient map in the frequency domain using Fourier transform; filtering low-frequency noise in the frequency gradient map based on a preset high-pass Butterworth filter to obtain a low-noise frequency gradient map; and converting the low-noise frequency gradient map into a low-noise gradient map in the spatial domain using inverse Fourier transform.

[0039] The cutoff frequency of the Qualcomm Butterworth filter is the first cutoff frequency.

[0040] For example, low-frequency noise in the frequency gradient map is filtered using a preset high-pass Butterworth filter to obtain a low-noise frequency gradient map, as shown in the following formula: ; in, Represents the coordinates in the frequency gradient plot; This is the cutoff frequency of the high-pass Butterworth filter; for example, the cutoff frequency can be 10. Let be the order of the high-pass Butterworth filter; for example, the order can be 2. This represents the width of the frequency gradient map. H represents the height of the frequency gradient plot. H(u, v) represents the filter response corresponding to the point (u, v) in the frequency gradient plot.

[0041] Therefore, by filtering the gradient map, most of the background noise in the gradient map is removed, which greatly improves the quality of the low-noise gradient map.

[0042] For example, a secondary repair process is performed on the low-noise gradient map using a preset fence filter combined with the preset kernel size to obtain a repaired gradient map. This includes: repeatedly defining the low-noise gradient map using the preset kernel size; for each defined region, calculating the proportion of grayscale values ​​of reflection intensity in the defined region that are higher than those of the center grid of the defined region; if the proportion is higher than a preset proportion, calculating the average grayscale value of reflection intensity in the defined region and replacing the grayscale value of reflection intensity in the center grid of the defined region with the average grayscale value of reflection intensity; if the proportion is lower than a preset proportion, maintaining the grayscale value of reflection intensity in the center grid of the defined region until all grids in the low-noise gradient map are treated as center grids and the preset fence filter process is performed before stopping.

[0043] Therefore, the pre-defined fence filtering simulates the cognitive logic of "filling inward from the boundary," significantly improving the foreground integrity in the low-noise gradient map, especially effectively enhancing the continuity within the foreground. This allows the generated patched gradient map to more accurately reflect the geometric structure of the foreground, providing high-quality input for subsequent Otsu's binarization process and effectively avoiding missegmentation caused by internal breaks or uneven intensity. Furthermore, after initially patching the planar raster map to obtain a patched raster map containing relatively complete information, and then performing gradient enhancement and filtering on the patched raster map to obtain a low-noise gradient map, a second patching using the pre-defined fence filtering is applied. This effectively achieves intelligent filling of holes or weak response regions within the low-noise gradient map, further enhancing its quality. This dual patching approach effectively ensures the quality of the patched gradient map.

[0044] Furthermore, the Otsu method image binarization is used to binarize the repaired gradient map to obtain a binarized gradient map. This includes: for each preset binarization threshold within a preset binarization threshold range, using the preset binarization threshold as a segmentation value to distinguish the background and foreground in the repaired gradient map; designating grids in the repaired gradient map with reflection intensity grayscale values ​​greater than or equal to the preset binarization threshold as foreground grids and counting the first number of foreground grids; designating grids in the repaired gradient map with reflection intensity less than the preset binarization threshold as background grids and counting the second number of background grids; calculating the foreground proportion, average foreground grayscale value, background proportion, and average background grayscale value in the repaired gradient map based on the first grid data and the second grid count; calculating the inter-class variance corresponding to the preset binarization threshold using the foreground proportion, average foreground grayscale value, background proportion, and average background grayscale value; using the preset binarization threshold corresponding to the largest inter-class variance as the target binarization threshold, and binarizing the repaired gradient map based on the target binarization threshold to obtain a binarized gradient map.

[0045] For example, the inter-class variance corresponding to the preset binarization threshold can be calculated using the foreground proportion, the average gray value of the foreground, the background proportion, and the average gray value of the background, as shown in the following formula: ; in, For inter-class variance; For the proportion of prospects; Background proportion; The average gray value of the foreground; This represents the average grayscale value of the background.

[0046] For example, the preset binarization threshold corresponding to the largest inter-class variance can be used as the target binarization threshold, as shown in the following formula: ; in, The target is the binarization threshold.

[0047] Therefore, the Otsu method for image binarization automatically determines the optimal threshold by maximizing the inter-class variance, efficiently binarizing low-noise gradient maps into images with distinct foreground and background. This preserves the complete structure of the foreground identifier while effectively removing residual noise from the background, providing high-quality input for subsequent feature fusion and point cloud selection, and significantly improving the overall performance of point cloud quality enhancement.

[0048] Further, the aforementioned image union processing of the binarized gradient map and the planar raster map to obtain the target planar raster map includes: performing a secondary filtering process on the planar raster map using a second cutoff frequency to obtain a filtered planar raster map, wherein the second cutoff frequency is greater than the first cutoff frequency; using the binarized gradient map as a mask map, and combining it with the filtered planar raster map to perform image union processing to obtain the target planar raster map.

[0049] The second cutoff frequency can be 50.

[0050] Therefore, by filtering the gradient map using the first cutoff frequency, low-frequency background noise is effectively suppressed, generating a high-quality, low-noise gradient map. Furthermore, a precise binary gradient map is obtained through fence filtering and Otsu's image binarization method, accurately depicting the foreground boundary. Based on this, a higher second cutoff frequency is used to perform a secondary high-pass filter on the planar raster map, further eliminating mid-to-low-frequency interference, such as road textures, minor stains, or slowly varying noise caused by sensor drift, while retaining areas with stronger reflective features, thus obtaining a cleaner filtered planar raster map. Subsequently, using the binary gradient map as a mask, an image union operation is performed with the filtered planar raster map. This retains the high-confidence intensity information in the filtered planar raster map while incorporating the precise edge localization provided by the binary gradient map. The final generated target planar raster map shows significant improvements in structural integrity, edge sharpness, and semantic accuracy. This dual-band collaborative filtering strategy balances edge sensitivity and intensity fidelity, effectively avoiding information loss or redundancy caused by a single threshold or single filtering. It provides a highly robust input for subsequent kernel density analysis and point cloud screening, significantly improving the accuracy and reliability of foreground extraction in high-precision maps.

[0051] Step 130: Analyze the occupancy probability of each grid cell in the target planar grid image being occupied by the foreground using a preset kernel density analysis method, wherein the foreground refers to the area that needs to be retained when the pixel value is higher than the preset pixel value.

[0052] Specifically, in step 130 above, the step of analyzing the occupancy probability of each grid in the target planar grid image being occupied by the foreground using a preset kernel density analysis method includes: taking grids in the target planar grid image whose reflection intensity grayscale value is not 0 as valid grids; for each valid grid, taking the valid grid as a heat map source, calculating the coverage probability of each grid in the target planar grid image being covered by the heat map source; and for each grid in the target planar grid image, calculating the occupancy probability of the grid being occupied by the foreground based on the multiple coverage probabilities corresponding to the grid.

[0053] In this context, the heatmap source refers to each valid raster with non-zero grayscale in the target plane raster image. It serves as a sample point for kernel density estimation, propagating probability weights to its neighborhood through a kernel function to construct a continuous foreground occupancy probability field. The preset kernel density analysis method is density estimation based on the Quartic kernel function.

[0054] Further, calculating the coverage probability of each grid cell in the target planar grid image being covered by the heat map source includes: drawing a circle with the heat map source as the center and a preset radius as the radius to obtain the kernel density calculation region corresponding to the heat map source; setting the coverage probability of the grid cells outside the kernel density calculation region being covered by the heat map source as 0; for the grid cells within the kernel density calculation region, taking the ratio of the Euclidean distance between the grid cell and the heat map source to the preset radius as the distance between the grid cell and the heat map source, and calculating the coverage probability of the grid cell being covered by the heat map source based on this distance.

[0055] For example, using the effective grid as a heatmap source, the coverage probability of each grid in the target planar grid map being covered by the heatmap source can be calculated using the following formula: ; Where x is the distance between the grid and the heat map source, and x is the ratio of the Euclidean distance between the grid and the heat map source to the preset radius. Let x be the coverage probability of a grid cell at a distance x from the heatmap source.

[0056] For example, for each grid cell in the target planar grid image, the occupancy probability of the grid cell being occupied by the foreground is calculated based on the multiple coverage probabilities corresponding to the grid cell, as shown in the following formula: ; in, For grid The occupancy probability of being occupied by the foreground. For grid The probability of being covered by heatmap source k.

[0057] Furthermore, after calculating the occupancy probability of the grid being occupied by the foreground based on the multiple coverage probabilities corresponding to the grid, the method further includes: using a nonlinear mapping function to enhance the occupancy probability to obtain an enhanced occupancy probability, and using the enhanced occupancy probability as the occupancy probability.

[0058] For example, the occupancy probability is enhanced using a nonlinear mapping function to obtain the enhanced occupancy probability, as shown in the following formula: ; in, This is the enhanced occupancy probability of the occupancy probability p.

[0059] Therefore, by using a Sigmoid-type nonlinear mapping function to enhance the occupancy probability obtained from kernel density estimation, the contrast of the intermediate probability interval is effectively stretched, and the ability to distinguish between the foreground and background is improved. This enhances the semantic clarity of the foreground in the point cloud and the robustness of subsequent processing, providing a more reliable basis for high-quality map construction.

[0060] Step 140: Based on the occupancy probability of each grid in the target planar grid map, and combined with a preset probability threshold, perform point cloud selection and rejection processing on the dense point cloud to be processed to obtain the target sparse point cloud corresponding to the dense point cloud to be processed.

[0061] Specifically, in step 140 above, the step of performing point cloud scavenging on the dense point cloud to be processed based on the occupancy probability of each grid in the target planar grid image and in combination with a preset probability threshold to obtain a target sparse point cloud corresponding to the dense point cloud to be processed includes: taking the grids in the target planar grid image with occupancy probabilities higher than the preset occupancy probability as foreground grids occupied by the foreground; taking the grids with occupancy probabilities lower than the preset occupancy probability as background grids occupied by the background; setting the reflection intensity of multiple points in the dense point cloud to be processed corresponding to each background grid to zero; and for each foreground grid, randomly selecting a reflection intensity from the reflection intensity of multiple points in the dense point cloud to be processed corresponding to the foreground grid as the reflection intensity of the multiple points corresponding to the foreground grid, thereby obtaining the target sparse point cloud.

[0062] Therefore, the point cloud selection strategy refines the selection based on the occupancy probability of each grid in the target planar raster image and a preset probability threshold, significantly improving the quality and usability of the output point cloud. By classifying high-occupancy-probability grids as foreground and low-probability grids as background, the system can effectively distinguish between key semantic regions such as the foreground and irrelevant noise such as the background. Furthermore, the reflection intensity of the point cloud corresponding to the background grid is set to zero, achieving precise removal of redundant information and significantly reducing data redundancy and interference. Simultaneously, for each foreground grid, the reflection intensity of one point is randomly retained from its multiple corresponding original points, maintaining the geometric integrity of the foreground structure while achieving reasonable sparsity of the point cloud, avoiding the computational burden caused by excessive density. The final generated target sparse point cloud retains high semantic value information while possessing lower data volume, higher signal-to-noise ratio, and stronger structural clarity. This not only improves the production efficiency of high-precision maps but also provides lightweight, high-quality input data for vehicle-mounted real-time perception systems, balancing accuracy, efficiency, and robustness.

[0063] In some examples, before generating a hole grid mask of the same size as the planar grid map corresponding to the dense point cloud to be processed in step 110, the method further includes: acquiring initial single-frame point clouds collected by multiple radars at each time point within a preset time period, as well as the pose data of each radar; for each time point within the preset time period, combining the initial single-frame point cloud collected by each radar at that time point with the pose data of the radar to form a single-frame point cloud; and preprocessing the multiple single-frame point clouds within the preset time period using a dual-buffering parallel acceleration strategy to obtain the dense point cloud to be processed.

[0064] Here, pose data refers to the radar's pose data. The dual-buffering parallel acceleration strategy refers to processing a single frame of point cloud twice to obtain the dense point cloud to be processed.

[0065] Specifically, a dual-buffered parallel acceleration strategy is used to preprocess multiple single-frame point clouds within the preset time period to obtain the dense point cloud to be processed. This includes: grouping multiple single-frame point clouds into a continuous parallel basic unit, and using four single-frame point clouds as one processing core; during the first preprocessing, the first three single-frame point clouds in the processing core at the beginning of the parallel basic unit are used as the first processing unit, and the fourth single-frame point cloud is used as the first buffer unit; the first single-frame point cloud in the processing core at the end of the parallel basic unit is used as the first buffer unit, and the second to fourth single-frame point clouds are used as the first processing units; the second and third single-frame point clouds in the processing cores that are neither at the beginning nor the end of the parallel basic unit are used as the first processing units, and the first and fourth single-frame point clouds are used as the first buffer units; during the first preprocessing, targeting... For each processing core, each of the first processing units in the processing core is processed in conjunction with each of the first cache units in the processing core to obtain a first enhanced point cloud corresponding to each of the first processing units in the processing core; in the second preprocessing, each of the first cache units is used as a second processing unit in the second preprocessing, and each of the first processing units is used as a second cache unit in the second preprocessing; in the second preprocessing, for each processing core, each of the second processing units in the processing core is processed in conjunction with each of the second cache units in the processing core to obtain a second enhanced point cloud corresponding to each of the second processing units in the processing core; in chronological order, each of the first enhanced point clouds and the second enhanced point clouds are combined to obtain a dense point cloud to be processed.

[0066] For example, such as Figure 2As shown, white squares are considered as processing units and gray squares as buffer units. In the second preprocessing, the first buffer unit in the first preprocessing is used as the second processing unit for processing, so that all single-frame point clouds are processed after the two preprocessing steps.

[0067] Therefore, this dual-caching parallel acceleration strategy divides multiple single-frame point clouds into processing kernels of four frames each, and dynamically allocates "processing units" and "caching units" in two preprocessing stages, achieving efficient context fusion and parallel enhancement of temporal point cloud data. This mechanism not only ensures that each frame of point cloud can fully utilize the spatiotemporal context of the preceding and following frames, effectively repairing data loss caused by occlusion, sparsity, or noise, but also avoids data dependency conflicts through structured parallel design, significantly improving processing throughput. Furthermore, in the second processing, the first cache unit from the first iteration is used as the second processing unit, ensuring that all single-frame point clouds are processed in both iterations. More importantly, this strategy pre-fuses the original sparse and discontinuous single-frame point clouds into structurally complete and semantically coherent dense point clouds to be processed, significantly reducing the computational burden of subsequent processing of these dense point clouds. Therefore, this method significantly improves the processing speed and overall system efficiency of large-scale point cloud data while maintaining the quality of the dense point clouds to be processed, providing strong support for real-time high-precision map construction and low-latency operation of vehicle-mounted perception systems.

[0068] In summary, planar raster images are repaired using hole dilation and pixel-level image inpainting techniques to ensure data integrity and continuity while minimizing information loss. Next, gradient enhancement is performed using specific convolutional kernels, and noise is reduced through filtering, enhancing foreground features while eliminating background noise. Subsequently, binarization and image union operations are used to obtain a target planar raster image containing accurate foreground information. Finally, kernel density analysis is employed to calculate the probability of each cell in the target planar raster image being occupied by the foreground, further validating the foreground information in the target planar raster image to obtain more accurate foreground information. This further filters the dense point cloud to be processed, thereby generating a more accurate target sparse point cloud. This method not only optimizes the accuracy and reliability of point cloud data, greatly improving point cloud quality, but also enhances subsequent processing efficiency by reducing redundant information, providing strong support for high-precision map production.

[0069] Example 2: Another embodiment of this application relates to a point cloud quality enhancement device based on kernel density analysis. The implementation details of this embodiment's point cloud quality enhancement device are described below. The following details are for ease of understanding and are not essential for implementing this solution. A schematic diagram of the point cloud quality enhancement device 30 based on kernel density analysis in this embodiment can be seen as follows: Figure 3As shown, it includes a generation unit 300, an acquisition unit 310, an analysis unit 320, and a processing unit 330.

[0070] The generation unit 300 is used to generate a hole grid mask with the same size as the planar grid map corresponding to the dense point cloud to be processed.

[0071] The acquisition unit 310 is used to acquire a target planar raster image using a preset pixel-level image repair technique and the hole raster mask.

[0072] Analysis unit 320 is used to analyze the occupancy probability of each grid cell in the target planar grid image being occupied by the foreground using a preset kernel density analysis method, wherein the foreground refers to the area that needs to be retained with a pixel value higher than a preset pixel value.

[0073] The processing unit 330 is used to perform point cloud selection and rejection processing on the dense point cloud to be processed based on the occupancy probability of each grid in the target planar grid map and in combination with a preset probability threshold, so as to obtain a target sparse point cloud corresponding to the dense point cloud to be processed.

[0074] In some examples, when the device performs point cloud scavenging on the dense point cloud to be processed based on the occupancy probability of each grid in the target planar grid image and a preset probability threshold to obtain a target sparse point cloud corresponding to the dense point cloud to be processed, it specifically performs the following: Grids in the target planar grid image with occupancy probabilities higher than the preset occupancy probability are designated as foreground grids occupied by the foreground; grids with occupancy probabilities lower than the preset occupancy probability are designated as background grids occupied by the background; the reflection intensity of multiple points in the dense point cloud to be processed corresponding to each background grid is set to zero; and for each foreground grid, a reflection intensity is randomly selected from the reflection intensities of multiple points in the dense point cloud to be processed corresponding to the foreground grid as the reflection intensity of the multiple points corresponding to the foreground grid, thereby obtaining the target sparse point cloud.

[0075] In some examples, when the device is used to obtain a target planar raster image using a preset pixel-level image inpainting technique and the hole raster mask, it is specifically used to: perform hole inpainting processing on the planar raster image using the preset pixel-level image inpainting technique combined with the hole raster mask to obtain a repaired raster image; obtain a binarized gradient image corresponding to the repaired raster image; and perform image union processing on the binarized gradient image and the planar raster image to obtain the target planar raster image.

[0076] In some examples, when the device is used to obtain the binarized gradient map corresponding to the repaired raster image, it is specifically used to: perform gradient enhancement processing on the repaired raster image using a preset convolution kernel to obtain a gradient map; perform filtering processing on the gradient map in combination with a first cutoff frequency to obtain a low-noise gradient map; perform secondary repair processing on the low-noise gradient map using a preset fence filter combined with the kernel size of the preset convolution kernel to obtain a repaired gradient map; and perform binarization processing on the repaired gradient map using the Otsu method image binarization to obtain a binarized gradient map.

[0077] In some examples, when the device is used to perform image union processing on the binarized gradient map and the planar raster map to obtain a target planar raster map, it specifically performs the following: performs secondary filtering processing on the planar raster map using a second cutoff frequency to obtain a filtered planar raster map, wherein the second cutoff frequency is greater than the first cutoff frequency; and uses the binarized gradient map as a mask map, combined with the filtered planar raster map, to perform image union processing to obtain the target planar raster map.

[0078] In some examples, when the device is used to analyze the occupancy probability of each grid in the target planar grid image being occupied by the foreground using a preset kernel density analysis method, it is specifically used to: identify grids in the target planar grid image whose reflection intensity grayscale value is not 0 as valid grids; for each valid grid, use the valid grid as a heatmap source and calculate the coverage probability of each grid in the target planar grid image being covered by the heatmap source; and for each grid in the target planar grid image, calculate the occupancy probability of the grid being occupied by the foreground based on the multiple coverage probabilities corresponding to the grid.

[0079] In some examples, when the device is used to perform hole repair processing on the planar grid image using a preset pixel-level image inpainting technique combined with the hole grid mask to obtain a repaired grid image, it specifically performs the following steps: determining the grid to be repaired in the planar grid image based on the hole marker in the hole grid mask; and for each grid to be repaired, performing hole repair processing on the grid to be repaired using the preset pixel-level image inpainting technique combined with the grayscale value of the reflection intensity of the grids surrounding the grid to be repaired to obtain a repaired grid image.

[0080] In some examples, before generating a hole-filled grid mask of the same size as the planar grid map corresponding to the dense point cloud to be processed, the device is further configured to: acquire initial single-frame point clouds collected by multiple radars at each time point within a preset time period, as well as the pose data of each radar; for each time point within the preset time period, combine the initial single-frame point cloud collected by each radar at that time point and the pose data of the radar to form a single-frame point cloud; and preprocess the multiple single-frame point clouds within the preset time period using a dual-buffering parallel acceleration strategy to obtain the dense point cloud to be processed.

[0081] It is worth mentioning that all units involved in this embodiment are logical units. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0082] Example 3: This disclosure also provides an electronic device, comprising: a memory for storing at least one instruction; and a processor for invoking the instruction stored in the memory to execute the point cloud quality enhancement method based on kernel density analysis in any of the above embodiments.

[0083] Example 4: This disclosure also provides a computer-readable storage medium and a computer program product, wherein the computer-readable storage medium stores at least one executable instruction, which is loaded and executed by a processor to implement the point cloud quality enhancement method based on kernel density analysis in any of the above embodiments; the computer program product includes computer program code, which, when executed by a computer, causes the computer to perform the point cloud quality enhancement method based on kernel density analysis in any of the above embodiments.

[0084] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0089] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0090] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0091] It should be noted that the terms "first," "second," and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different parts. Terms such as "including" or "contains" mean that the element preceding the word covers the element listed after the word, and do not exclude the possibility of covering other elements as well.

[0092] Although operations are described in a specific order in the accompanying drawings in this disclosure, it should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all of the shown operations to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0093] Finally, it should be noted that the above content is only used to illustrate the technical solution of this disclosure, and is not intended to limit the scope of protection of this disclosure. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of this disclosure do not depart from the substance and scope of the technical solution of this disclosure.

Claims

1. A point cloud quality enhancement method based on kernel density analysis, characterized in that, include: Generate a hole grid mask of the same size as the planar grid map corresponding to the dense point cloud to be processed; The target planar grid image is obtained using a preset pixel-level image inpainting technique and the hole grid mask; The occupancy probability of each grid cell in the target planar grid image being occupied by the foreground is analyzed using a preset kernel density analysis method. The foreground refers to the area that needs to be retained when the pixel value is higher than a preset pixel value. Based on the occupancy probability of each grid in the target planar grid image, and combined with a preset probability threshold, the dense point cloud to be processed is subjected to point cloud selection and rejection processing to obtain the target sparse point cloud corresponding to the dense point cloud to be processed.

2. The point cloud quality enhancement method based on kernel density analysis according to claim 1, characterized in that, The step of performing point cloud rejection processing on the dense point cloud to be processed based on the occupancy probability of each grid in the target planar grid image and in combination with a preset probability threshold to obtain the target sparse point cloud corresponding to the dense point cloud to be processed includes: The grid cells in the target planar grid that have an occupancy probability higher than the preset occupancy probability are taken as the foreground grid cells occupied by the foreground. The grids with an occupancy probability lower than the preset occupancy probability are designated as background grids occupied by the background. Set the reflection intensity of multiple points in the dense point cloud to be processed, corresponding to each of the background grid cells, to zero. For each of the foreground grids, a reflection intensity is randomly selected from the reflection intensities of multiple points corresponding to the foreground grid in the dense point cloud to be processed, and used as the reflection intensity of the multiple points corresponding to the foreground grid, thereby obtaining the target sparse point cloud.

3. The point cloud quality enhancement method based on kernel density analysis according to claim 1, characterized in that, The method of obtaining the target planar raster image using a preset pixel-level image inpainting technique and the hole raster mask includes: The planar grid image is repaired by using a preset pixel-level image repair technique combined with the hole grid mask to obtain a repaired grid image. Obtain the binarized gradient map corresponding to the repaired raster image; The binarized gradient map and the planar raster map are subjected to image union processing to obtain the target planar raster map.

4. The point cloud quality enhancement method based on kernel density analysis according to claim 3, characterized in that, The step of obtaining the binarized gradient map corresponding to the repaired raster image includes: The repaired raster image is subjected to gradient enhancement processing using a preset convolution kernel to obtain a gradient image; The gradient map is filtered using the first cutoff frequency to obtain a low-noise gradient map. The low-noise gradient map is repaired by using a preset fence filter combined with the kernel size of the preset convolution kernel to obtain a repaired gradient map. The repair gradient map is binarized using the Otsu method for image binarization to obtain a binarized gradient map.

5. The point cloud quality enhancement method based on kernel density analysis according to claim 4, characterized in that, The step of performing image union processing on the binarized gradient map and the planar raster map to obtain the target planar raster map includes: The planar grid image is subjected to secondary filtering using a second cutoff frequency to obtain a filtered planar grid image, wherein the second cutoff frequency is greater than the first cutoff frequency; The binarized gradient map is used as a mask map, and combined with the filtered plane raster map, image union processing is performed to obtain the target plane raster map.

6. The point cloud quality enhancement method based on kernel density analysis according to claim 1, characterized in that, The step of analyzing the occupancy probability of each grid cell in the target planar grid image being occupied by the foreground using a preset kernel density analysis method includes: Grids in the target planar grid image whose reflection intensity grayscale value is not 0 are considered as valid grids; For each valid grid cell, the valid grid cell is used as a heatmap source, and the coverage probability of each grid cell in the target planar grid cell is covered by the heatmap source is calculated. For each grid cell in the target planar grid image, the occupancy probability of the grid cell being occupied by the foreground is calculated based on the multiple coverage probabilities corresponding to the grid cell.

7. The point cloud quality enhancement method based on kernel density analysis according to claim 3, characterized in that, The process of using a preset pixel-level image inpainting technique combined with the hole grid mask to perform hole repair processing on the planar grid image to obtain a repaired grid image includes: Based on the void markers in the void grid mask, the grid cells to be repaired corresponding to the void markers in the planar grid diagram are determined. For each of the grid cells to be repaired, a hole repair process is performed on the grid cell to be repaired by using a preset pixel-level image repair technique combined with the grayscale value of the reflection intensity of the grid cells surrounding the grid cell to be repaired, so as to obtain a repaired grid image.

8. The point cloud quality enhancement method based on kernel density analysis according to claim 1, characterized in that, Before generating a hole-filled raster mask of the same size as the planar raster image corresponding to the dense point cloud to be processed, the method further includes: Acquire the initial single-frame point cloud data collected by multiple radars at each time point within a preset time period, as well as the pose data of each radar. For each moment in the preset time period, the initial single-frame point cloud collected by each radar at that moment and the pose data of the radar are combined to form a single-frame point cloud. A dual-buffering parallel acceleration strategy is used to preprocess multiple single-frame point clouds within the preset time period to obtain the dense point cloud to be processed.

9. A point cloud quality enhancement device based on kernel density analysis, characterized in that, include: The generation unit is used to generate a hole grid mask with the same size as the planar grid map corresponding to the dense point cloud to be processed; The acquisition unit is used to acquire a target planar grid image using a preset pixel-level image inpainting technique and the hole grid mask; The analysis unit is used to analyze the occupancy probability of each grid cell in the target planar grid image being occupied by the foreground using a preset kernel density analysis method, wherein the foreground refers to the area that needs to be retained when the pixel value is higher than a preset pixel value; The processing unit is used to perform point cloud rejection processing on the dense point cloud to be processed based on the occupancy probability of each grid in the target planar grid map and in combination with a preset probability threshold, so as to obtain the target sparse point cloud corresponding to the dense point cloud to be processed.

10. An electronic device, characterized in that, include: Memory, used to store at least one instruction; as well as A processor is configured to invoke instructions stored in the memory to execute the point cloud quality enhancement method based on kernel density analysis as described in any one of claims 1-8.

11. A computer-readable storage medium and a computer program product, characterized in that, The computer-readable storage medium stores at least one executable instruction, which is loaded and executed by a processor to implement the point cloud quality enhancement method based on kernel density analysis as described in any one of claims 1-8; The computer program product includes computer program code, which, when executed by a computer, causes the computer to perform the point cloud quality enhancement method based on kernel density analysis as described in any one of claims 1-8.

Images