A general obstacle detection method based on RGBD dense point cloud segmentation

Abstract

This invention discloses a general obstacle detection method based on RGBD dense point cloud segmentation. Building upon traditional point cloud segmentation methods, it utilizes planar and obstacle geometric features to segment RGBD dense point cloud data, accurately extracting ground points and general obstacle points. This invention achieves effective ground removal and general obstacle detection with relatively low computational resources, exhibiting good real-time performance and stability, and has broad application prospects.

Description

Technical Field

[0001] This invention belongs to the field of point cloud processing, and mainly relates to a general obstacle detection method based on RGBD dense point cloud segmentation. Background Technology

[0002] With the development of depth sensor technology, point cloud data has become an important tool for environmental modeling and analysis. Point cloud data provides relatively accurate 3D positional information of objects, and its main characteristic is that it describes the structure of real-world objects or surfaces through a set of three-dimensional coordinate points. Since sensor scanning typically collects a large amount of point cloud data containing the ground, obstacles, and other objects, effectively removing ground point clouds and detecting obstacles has become a significant technical challenge when further analyzing and processing this data.

[0003] In existing technologies, some traditional methods that use height values ​​to filter the ground not only have high requirements for calibration but also cannot handle complex scenarios such as slopes. While advanced algorithms based on machine learning and deep learning are used for point cloud classification and segmentation, these methods typically require large amounts of labeled data for training, have high computational costs, and poor real-time performance. Summary of the Invention

[0004] To address the aforementioned problems, the main technical issue solved by this invention is to provide a general obstacle detection method based on RGBD dense point cloud segmentation. Building upon traditional point cloud segmentation methods, this method utilizes planar and obstacle geometric features to achieve segmentation, enabling effective ground removal and general obstacle detection with relatively low computational resources. It also exhibits good real-time performance and stability, and has broad application prospects.

[0005] To achieve the above objectives, the solution of the present invention is:

[0006] A general obstacle detection method based on RGBD dense point cloud segmentation is proposed, and the specific steps of the method are as follows:

[0007] Step 1: Acquire depth images of the study area in real time using a depth camera and convert them into point cloud data;

[0008] Step 2: Divide the study area into equally spaced two-dimensional grid cells and score each grid cell.

[0009] Step 3: Based on the set scoring threshold, perform an initial judgment on each grid cell: if the score of a grid cell is less than the scoring threshold, the grid cell is judged as a planar grid cell; otherwise, the grid cell is judged as an obstacle grid cell.

[0010] As a further optimization of the present invention, step 1 specifically includes:

[0011] Step 1.1: Perform distortion correction on the depth image;

[0012] Step 1.2: Perform depth correction on the distortion-free depth image using an interpolation algorithm;

[0013] Step 1.3: Based on the depth correction results, acquire point cloud data and filter it;

[0014] Step 1.4: Use the calibrated extrinsic parameters of the depth camera to perform coordinate transformation on the filtered point cloud data.

[0015] As a further optimization of the present invention, step 2 specifically includes:

[0016] Step 2.1: Divide the study area into equally spaced two-dimensional grid cells;

[0017] Step 2.2: For each grid cell, count the number of point clouds M falling within it, the difference between the maximum and minimum Z values ​​∆z, and the color variance Var of the corresponding RBGB image pixels within that grid cell. rgb ;

[0018] Step 2.3: Score each grid cell:

[0019] ,

[0020] Where S is the grid cell score, and w1, w2, and w3 are all weighting coefficients.

[0021] As a further optimization of the present invention, the above method also includes:

[0022] Step 4: For each planar grid cell obtained in the initial judgment, construct a candidate point cloud set based on the point cloud data belonging to the planar grid cell within the 3*3 grid cell centered on it;

[0023] Step 5: For each candidate point cloud in the candidate point cloud set, count the number of point clouds in its neighborhood, and then calculate the erosion expansion radius of each candidate point cloud; where the neighborhood is a sphere with the candidate point cloud as its center.

[0024] Step 6: Based on the erosion dilation radius of each candidate point cloud obtained in Step 5, perform a single-scale morphological erosion dilation operation:

[0025] For each candidate point cloud, the number of points within its erosion expansion radius is queried. If the number is lower than the preset minimum number of points, the corresponding candidate point cloud is removed.

[0026] For each rejected candidate point cloud, query the number of point clouds within its erosion expansion radius. If the number is greater than the preset point count threshold, then restore the corresponding rejected candidate point cloud.

[0027] Step 7: Based on the morphological erosion and dilation operation results from Step 6, for each planar mesh cell obtained in the initial determination, calculate the slope between it and each planar mesh cell in its 3*3 neighborhood:

[0028] ,

[0029] in, This represents the slope between a planar mesh cell c and the l-th planar mesh cell in its 3x3 neighborhood. This represents the Z-axis median of the point cloud within a planar mesh cell c. This represents the maximum absolute value of the Z-axis coordinate of the point cloud within the l-th planar grid cell in the 3x3 neighborhood of planar grid cell c. ) represents the geometric center coordinates of the planar mesh element c, ( ) represents the geometric center coordinates of the l-th planar mesh cell within the 3x3 neighborhood of planar mesh cell c;

[0030] Step 8: For each planar grid cell obtained in the initial determination, select the maximum value of the slope between it and each planar grid cell in its 3*3 neighborhood as the terrain slope between it and its 3*3 neighborhood.

[0031] Step 9: Based on the set slope threshold, further determine each planar grid cell obtained in the initial determination: if the terrain slope of a certain planar grid cell and its 3*3 neighborhood is greater than the set slope threshold, then the planar grid cell is determined to be an obstacle grid cell; otherwise, it is still a planar grid cell.

[0032] As a further optimization of the present invention, step 5 specifically includes:

[0033] Step 5.1: For each candidate point cloud in the candidate point cloud set, count the number of point clouds in its neighborhood;

[0034] Step 5.2, calculate the point cloud density within the neighborhood of each candidate point cloud:

[0035] ρ(p) = N(p) / V(R0),

[0036] Where ρ(p) represents the point cloud density in the p-neighborhood of the candidate point cloud, N(p) represents the number of points in the p-neighborhood of the candidate point cloud, R0 represents the neighborhood radius, and V(R0) represents the neighborhood volume;

[0037] Step 5.3, calculate the erosion expansion radius of each candidate point cloud:

[0038] R(p) = α*( 1 / ρ(p)) β ,

[0039] Where α is the scaling factor and β is the exponential control parameter.

[0040] To achieve the above objectives, another solution of the present invention is:

[0041] A general obstacle detection method based on RGBD dense point cloud segmentation is proposed, and the specific steps of the method are as follows:

[0042] Step 1: Acquire depth images of the study area in real time using a depth camera and convert them into point cloud data;

[0043] Step 2: Divide the study area into equally spaced two-dimensional grid cells and score each grid cell.

[0044] Step 3: Based on the set scoring threshold, perform an initial judgment on each grid cell: if the score of a grid cell is less than the scoring threshold, the grid cell is judged as a planar grid cell; otherwise, the grid cell is judged as an obstacle grid cell.

[0045] Step 4: For each planar grid cell obtained in the initial judgment, construct a candidate point cloud set based on the point cloud data belonging to the planar grid cell within the 3*3 grid cell centered on it;

[0046] Step 5: Set the neighborhood radius sequence R = {R1, R2, ..., Rn}, R1 > R2 > ... > Rn. For each candidate point cloud in the candidate point cloud set, count the number of point clouds in the neighborhood of its different neighborhood radii, and then calculate the N erosion expansion radii of each candidate point cloud; where the neighborhood is a sphere with the candidate point cloud as its center, N is the number of elements in the neighborhood radius sequence, and Rn represents the nth neighborhood radius, n = 1, 2, ..., N;

[0047] Step 6: Based on the erosion dilation radius of each candidate point cloud obtained in Step 5, perform a scale-progressive erosion dilation operation:

[0048] For each candidate point cloud, query the number of points within its erosion expansion radius Rn. If the number is lower than the preset minimum number of points, the corresponding candidate point cloud is removed.

[0049] For each candidate point cloud that has been removed, query the number of point clouds within its erosion expansion radius Rn. If the number is greater than the preset point count threshold, then restore the corresponding candidate point cloud.

[0050] Step 7: Based on the morphological erosion and dilation operation results from Step 6, for each planar mesh cell obtained in the initial determination, calculate the slope between it and each planar mesh cell in its 3*3 neighborhood:

[0051] ,

[0052] in, This represents the slope between a planar mesh cell c and the l-th planar mesh cell in its 3x3 neighborhood. This represents the Z-axis median of the point cloud within a planar mesh cell c. This represents the maximum absolute value of the Z-axis coordinate of the point cloud within the l-th planar grid cell in the 3x3 neighborhood of planar grid cell c. ) represents the geometric center coordinates of the planar mesh element c, ( ) represents the geometric center coordinates of the l-th planar mesh cell within the 3x3 neighborhood of planar mesh cell c;

[0053] Step 8: For each planar grid cell obtained in the initial determination, select the maximum value of the slope between it and each planar grid cell in its 3*3 neighborhood as the terrain slope between it and its 3*3 neighborhood.

[0054] Step 9: Based on the set slope threshold, further determine each planar grid cell obtained in the initial determination: if the terrain slope of a certain planar grid cell and its 3*3 neighborhood is greater than the set slope threshold, then the planar grid cell is determined to be an obstacle grid cell; otherwise, it is still a planar grid cell.

[0055] As a further optimization of the present invention, step 5 specifically includes:

[0056] Step 5.1: Set the neighborhood radius sequence R = {R1, R2, ..., Rn}, where R1 > R2 > ... > Rn;

[0057] Step 5.2: For each candidate point cloud in the candidate point cloud set, count the number of point clouds in the neighborhood of different neighborhood radii.

[0058] Step 5.3, calculate the point cloud density within the neighborhood of each candidate point cloud:

[0059] ρ(p) = N(p) / V(Rn),

[0060] Where ρ(p) represents the point cloud density in the p-neighborhood of the candidate point cloud, N(p) represents the number of points in the p-neighborhood of the candidate point cloud, and V(Rn) represents the volume of the neighborhood with a radius of Rn.

[0061] Step 5.3, calculate the erosion expansion radius of each candidate point cloud:

[0062] R(p) = α*( 1 / ρ(p)) β ,

[0063] Where α is the scaling factor and β is the exponential control parameter.

[0064] As a further optimization of the present invention, the above method also includes:

[0065] Step 10: Delete the point cloud corresponding to the obstacle mesh cell obtained in step 9 whose Z value does not exceed the set Z value threshold;

[0066] Step 11: Merge the deletion result from Step 10 with the point cloud within the obstacle mesh cell obtained in Step 3 to obtain the final result of obstacle detection.

[0067] The present invention also proposes a computer-readable storage medium for storing one or more programs, said one or more programs including instructions that, when executed by a computing device, cause the computing device to perform the method described above.

[0068] The present invention also proposes an electronic device comprising one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for performing the methods described above.

[0069] The significant advantages of this invention compared to existing technologies are:

[0070] (1) Low computational overhead: The traditional geometric segmentation algorithm is adopted, which has high computational efficiency and can realize real-time processing on devices with limited computing resources;

[0071] (2) Good stability: This method is applicable to various terrains, especially slopes and other flat surfaces, which can be used as ground filters and general obstacles for extraction. It has strong adaptability to various environments;

[0072] (3) No need for large-scale training data: It does not rely on a large amount of labeled data and deep learning models, thus avoiding the cost of data collection and training. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the present invention.

[0074] Figure 2 These are image data acquired by the depth camera in this embodiment of the invention, where (a) is an RGB image and (b) is a depth image.

[0075] Figure 3 This is a schematic diagram of obstacle detection results according to an embodiment of the present invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0077] This invention proposes a general obstacle detection method based on RGBD dense point cloud segmentation. The aim is to accurately extract ground points and general obstacle points by segmenting RGBD dense point cloud data, thereby achieving ground removal and obstacle detection. This method combines the geometric characteristics of point clouds with traditional segmentation techniques, providing an efficient and accurate point cloud processing solution even with limited computing resources.

[0078] like Figure 1 As shown, this method achieves ground removal and obstacle detection through the following steps.

[0079] 1. Point cloud acquisition and coordinate transformation

[0080] For real-time input depth images from a depth camera, the depth values ​​and camera intrinsic parameters are converted into point clouds. Since the original input point cloud data is large and may contain noise, distortion correction is first performed on the depth map, followed by depth value correction through interpolation to obtain relatively accurate point cloud data. Because the depth camera produces dense point clouds, filtering is also required to improve data quality and subsequent processing efficiency. At this point, the 3D coordinates of the point cloud are in the camera coordinate system. For more intuitive processing, the point cloud coordinates are transformed to the vehicle coordinate system using calibrated extrinsic parameters, where X represents the front of the vehicle, Y represents the left side of the vehicle, and Z represents the top of the vehicle.

[0081] 2. Planar Extraction

[0082] Traditional LiDAR-based point cloud planar extraction methods typically rely solely on the geometric features of the point cloud, primarily using height differences and normal vector consistency to determine planar regions. These methods are prone to misjudgment when dealing with scenes featuring complex textures or minimal geometric changes, and they struggle to fully leverage the auxiliary role of color information in RGBD point cloud data.

[0083] To address this, this invention proposes a planar extraction method based on point cloud meshing and multi-feature fusion. This method combines the color consistency of RGB images with point cloud density and geometric height difference features, effectively improving the accuracy and robustness of planar extraction. The implementation process is as follows:

[0084] 2.1 Regional Gridding Processing

[0085] The preset region of interest is divided into equally spaced two-dimensional grid cells (in the embodiment of the present invention, the size of the grid cell is 15cm*15cm), and all point cloud data falling into each grid are summarized.

[0086] 2.2 Extraction of Multidimensional Statistical Features

[0087] For each grid cell, the following statistical information is extracted as the feature of the grid cell: number of points. (Reflecting local density), the difference between the maximum and minimum values ​​in the Z direction (Reflecting geometric flatness), the color variance of pixels within the corresponding RGB image of this region. (Reflects color consistency).

[0088] 2.3 Design of Weighted Plane Scoring Function

[0089] By introducing a scoring function, the above three features are integrated into one, in the following form:

[0090] ,

[0091] in , , These are weight coefficients obtained through parameter tuning and optimization on a self-collected depth camera dataset, used to balance the importance of each feature.

[0092] 2.4 Planar Mesh Determination Mechanism

[0093] Based on experience, a planar scoring threshold T is set. When S < T, the grid is determined to be a planar grid cell; otherwise, it is determined to be an obstacle grid cell.

[0094] The above-mentioned judgment mechanism can effectively eliminate misjudgment problems caused by color inconsistency or sparse point clouds.

[0095] 2.5 Planar / Non-planar Application Division

[0096] Planar mesh cells are used as the ground to be processed for ground removal; obstacle mesh cells are used as obstacles for obstacle extraction.

[0097] 3. Ground removal

[0098] Using the point cloud data within the planar grid cells identified in the previous stage as input, slope analysis is used to separate, extract, and reconstruct obstacles of different scales, performing secondary screening of the ground and obstacles. Due to the characteristics of RGBD cameras, the generated dense point clouds have certain errors compared to radar, so traditional methods of calculating the slope of adjacent points or normal vectors cannot accurately remove the ground. Therefore, this invention adopts two-dimensional grid slope calculation, as follows:

[0099] 3.1 Morphological Filtering

[0100] To avoid interference from noise and non-planar points that were not fully extracted in the previous stage, a progressive adaptive morphological filter is first applied to the point cloud within the planar grid cell to filter out noise and non-planar points.

[0101] 3.1.1 Adaptive Corrosion / Expansion Radius

[0102] Based on the point cloud data belonging to planar grid cells within a 3x3 grid cell centered on the target grid eye, a candidate point cloud set is constructed. For each candidate point cloud in the candidate point cloud set, the number of point clouds in its neighborhood is counted, and then the erosion expansion radius of each candidate point cloud is calculated.

[0103] Specifically, a fixed sphere radius R0 is set as the neighborhood radius, the number of points N(p) in the neighborhood of a candidate point cloud is calculated, and then the point cloud density in the neighborhood of the candidate point cloud is estimated:

[0104] ρ(p) = N(p) / V (R0) ,

[0105] Where V (R0) Represents the neighborhood volume, i.e. .

[0106] Then, the erosion / expansion radius of the candidate point cloud is calculated based on the point cloud density:

[0107] Rp = α*( 1 / ρ (p) ) β ,

[0108] Where α is the scaling factor and β is the exponential control parameter, both of which are empirical parameters.

[0109] 3.1.2 Morphological Filtering

[0110] By utilizing morphological erosion and dilation operations on point clouds, for each candidate point cloud, the number of neighboring points within its adaptive erosion / dilation radius is queried. If this number is lower than a preset minimum number of points, the corresponding candidate point cloud is discarded. For each discarded candidate point cloud, the number of points within its erosion / dilation radius is queried. If this number is greater than a preset point count threshold, the corresponding discarded candidate point cloud is restored. The erosion operation effectively removes points with a small number of points in the neighborhood, eliminating noise or discrete non-planar points. The dilation operation, by detecting the neighborhood of the discarded points, determines whether they belong to a continuous structure, and is used for restoring planar boundary points.

[0111] Furthermore, a set of neighborhood radius sequences R can be set. The data are arranged from largest to smallest, and multi-scale erosion and dilation operations are performed. An adaptive radius mechanism is used at each scale to ensure consistent processing across regions of different densities; large radii are used to extract discrete noise points or non-planar points, while small radii are used to extract denser noise points.

[0112] 3.2 Calculation of Mesh Slope

[0113] 3.2.1 Slope Statistics

[0114] For the target mesh, calculate the slope between it and each planar mesh cell in its 3x3 neighborhood.

[0115] Specifically, let's take a planar grid cell (hereinafter referred to as the neighborhood grid) within the target grid's neighborhood as an example:

[0116] First, extract the Z-axis coordinates of all point clouds within the target mesh, and then calculate their Z-axis median ( This median value represents the elevation value of the current grid. The geometric center of the target grid is used as the X and Y coordinates of the target grid. ), constituting the spatial coordinates of the target grid ( ).

[0117] Subsequently, the Z-axis coordinates of all point clouds within the neighborhood grid are extracted, and the Z-axis coordinate with the largest absolute value is selected. This value represents the elevation of the neighboring grid; simultaneously, the geometric center of the neighboring grid is used as its X and Y coordinates. ), constituting the spatial coordinates of the neighborhood grid ( ).

[0118] Based on the spatial coordinate difference between the target mesh and its neighboring meshes, calculate the slope in the corresponding direction:

[0119] .

[0120] 3.2.2 Select the maximum slope value obtained in 3.2.1 as the terrain slope between the target grid and its 3*3 neighborhood. ),like If the slope threshold is set, the target mesh becomes the obstacle mesh, and the point cloud within it becomes the obstacle point cloud. This slope threshold can be empirically set based on the terrain environment and taking into account the actual maximum slope information.

[0121] 4. Process the obstacle points extracted in stage 3. Points with a height Z-value greater than a set Z-value threshold are identified as obstacle points. Merge these points with the obstacle point cloud obtained in stage 2 to output the final obstacle point cloud result.

[0122] 5. To avoid the impact of scattered noise on the final result, the obstacle point cloud in stage 4 can be further filtered to remove scattered noise.

[0123] In one embodiment, the depth camera acquires, for example... Figure 2 The image data shown, when used with the obstacle detection method proposed in this invention, yields the following obstacle detection results: Figure 3 As shown. By Figure 3 It is evident that the detection method of this invention can effectively segment ground points (smaller points) and obstacle points (larger points). Specifically, it can accurately identify the position and outline of large obstacles (such as a distant blackboard or chair), medium obstacles (such as pedestrians), and small obstacles (such as a small cardboard box), providing reliable perceptual information support for subsequent path planning and motion control. Compared to traditional methods, the detection method of this invention exhibits higher accuracy and robustness in perceiving obstacles of various sizes, maintaining good segmentation performance even in complex scenes.

[0124] This application also provides an electronic device, including: a memory and a processor; the memory stores a computer program, and when the computer program is executed by the processor, it implements the above-described general obstacle detection method based on RGBD dense point cloud segmentation.

[0125] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the aforementioned general obstacle detection method based on RGBD dense point cloud segmentation. The computer-readable storage medium may include various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0126] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0127] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A general obstacle detection method based on RGBD dense point cloud segmentation, characterized in that, The specific steps of this method are as follows: Step 1: Acquire depth images of the study area in real time using a depth camera and convert them into point cloud data; Step 2: Divide the study area into equally spaced two-dimensional grid cells and score each grid cell. Step 3: Based on the set scoring threshold, perform an initial judgment on each grid cell: if the score of a grid cell is less than the scoring threshold, the grid cell is judged as a planar grid cell; otherwise, the grid cell is judged as an obstacle grid cell. Step 4: For each planar grid cell obtained in the initial judgment, construct a candidate point cloud set based on the point cloud data belonging to the planar grid cell within the 3*3 grid cell centered on it; Step 5: For each candidate point cloud in the candidate point cloud set, count the number of point clouds in its neighborhood, and then calculate the erosion expansion radius of each candidate point cloud; where the neighborhood is a sphere with the candidate point cloud as its center. Step 6: Based on the erosion dilation radius of each candidate point cloud obtained in Step 5, perform a single-scale morphological erosion dilation operation: For each candidate point cloud, the number of points within its erosion expansion radius is queried. If the number is lower than the preset minimum number of points, the corresponding candidate point cloud is removed. For each rejected candidate point cloud, query the number of point clouds within its erosion expansion radius. If the number is greater than the preset point count threshold, then restore the corresponding rejected candidate point cloud. Step 7: Based on the morphological erosion and dilation operation results from Step 6, for each planar mesh element obtained in the initial determination, calculate the slope between it and each planar mesh element in its 3*3 neighborhood: ， in, This represents the slope between a planar mesh cell c and the l-th planar mesh cell in its 3x3 neighborhood. This represents the Z-axis median of the point cloud within a planar mesh cell c. This represents the maximum absolute value of the Z-axis coordinate of the point cloud within the l-th planar grid cell in the 3x3 neighborhood of planar grid cell c. ) represents the geometric center coordinates of the planar mesh element c, ( ) represents the geometric center coordinates of the l-th planar mesh cell within the 3x3 neighborhood of planar mesh cell c; Step 8: For each planar grid cell obtained in the initial determination, select the maximum value of the slope between it and each planar grid cell in its 3*3 neighborhood as the terrain slope between it and its 3*3 neighborhood. Step 9: Based on the set slope threshold, further determine each planar grid cell obtained in the initial determination: if the terrain slope of a certain planar grid cell and its 3*3 neighborhood is greater than the set slope threshold, then the planar grid cell is determined to be an obstacle grid cell; otherwise, it is still a planar grid cell.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Perform distortion correction on the depth image; Step 1.2: Perform depth correction on the distortion-free depth image using an interpolation algorithm; Step 1.3: Based on the depth correction results, acquire point cloud data and filter it; Step 1.4: Use the calibrated extrinsic parameters of the depth camera to perform coordinate transformation on the filtered point cloud data.

3. The method according to claim 2, characterized in that, Step 2 specifically includes: Step 2.1: Divide the study area into equally spaced two-dimensional grid cells; Step 2.2: For each grid cell, count the number of point clouds M falling within it, the difference between the maximum and minimum Z values ​​∆z, and the color variance Var of the corresponding RGB image pixels within that grid cell. rgb ; Step 2.3: Score each grid cell: ， Where S is the grid cell score, and w1, w2, and w3 are all weighting coefficients.

4. The method according to claim 1, characterized in that, Step 5 specifically involves: Step 5.1: For each candidate point cloud in the candidate point cloud set, count the number of point clouds in its neighborhood; Step 5.2, calculate the point cloud density in the neighborhood of each candidate point cloud: ρ(p) = N(p) / V(R0), Where ρ(p) represents the point cloud density in the p-neighborhood of the candidate point cloud, N(p) represents the number of points in the p-neighborhood of the candidate point cloud, R0 represents the neighborhood radius, and V(R0) represents the neighborhood volume; Step 5.3, calculate the erosion expansion radius of each candidate point cloud: R(p) = α*( 1 / ρ(p)) β , Where α is the scaling factor and β is the exponential control parameter.

5. A general obstacle detection method based on RGBD dense point cloud segmentation, characterized in that, The specific steps of this method are as follows: Step 1: Acquire depth images of the study area in real time using a depth camera and convert them into point cloud data; Step 2: Divide the study area into equally spaced two-dimensional grid cells and score each grid cell. Step 3: Based on the set scoring threshold, perform an initial judgment on each grid cell: if the score of a grid cell is less than the scoring threshold, the grid cell is judged as a planar grid cell; otherwise, the grid cell is judged as an obstacle grid cell. Step 4: For each planar grid cell obtained in the initial judgment, construct a candidate point cloud set based on the point cloud data belonging to the planar grid cell within the 3*3 grid cell centered on it; Step 5: Set the neighborhood radius sequence R = {R1, R2, ..., Rn}, R1 > R2 > ... > Rn. For each candidate point cloud in the candidate point cloud set, count the number of point clouds in the neighborhood of its different neighborhood radii, and then calculate the N erosion expansion radii of each candidate point cloud; where the neighborhood is a sphere with the candidate point cloud as its center, N is the number of elements in the neighborhood radius sequence, and Rn represents the nth neighborhood radius, n = 1, 2, ..., N; Step 6: Based on the erosion dilation radius of each candidate point cloud obtained in Step 5, perform a scale-progressive erosion dilation operation: For each candidate point cloud, query the number of points within its erosion expansion radius Rn. If the number is lower than the preset minimum number of points, the corresponding candidate point cloud is removed. For each candidate point cloud that has been removed, query the number of point clouds within its erosion expansion radius Rn. If the number is greater than the preset point count threshold, then restore the corresponding candidate point cloud. Step 7: Based on the morphological erosion and dilation operation results from Step 6, for each planar mesh element obtained in the initial determination, calculate the slope between it and each planar mesh element in its 3*3 neighborhood: ， in, This represents the slope between a planar mesh cell c and the l-th planar mesh cell in its 3x3 neighborhood. This represents the Z-axis median of the point cloud within a planar mesh cell c. This represents the maximum absolute value of the Z-axis coordinate of the point cloud within the l-th planar grid cell in the 3x3 neighborhood of planar grid cell c. ) represents the geometric center coordinates of the planar mesh element c, ( ) represents the geometric center coordinates of the l-th planar mesh cell within the 3x3 neighborhood of planar mesh cell c; Step 8: For each planar grid cell obtained in the initial determination, select the maximum value of the slope between it and each planar grid cell in its 3*3 neighborhood as the terrain slope between it and its 3*3 neighborhood. Step 9: Based on the set slope threshold, further determine each planar grid cell obtained in the initial determination: if the terrain slope of a certain planar grid cell and its 3*3 neighborhood is greater than the set slope threshold, then the planar grid cell is determined to be an obstacle grid cell; otherwise, it is still a planar grid cell.

6. The method according to claim 5, characterized in that, Step 5 specifically involves: Step 5.1: Set the neighborhood radius sequence R = {R1, R2, ..., Rn}, where R1 > R2 > ... > Rn; Step 5.2: For each candidate point cloud in the candidate point cloud set, count the number of point clouds in the neighborhood of different neighborhood radii. Step 5.3, calculate the point cloud density within the neighborhood of each candidate point cloud: ρ(p) = N(p) / V(Rn), Where ρ(p) represents the point cloud density in the p-neighborhood of the candidate point cloud, N(p) represents the number of points in the p-neighborhood of the candidate point cloud, and V(Rn) represents the volume of the neighborhood with a radius of Rn. Step 5.4, calculate the erosion expansion radius of each candidate point cloud: R(p) = α*( 1 / ρ(p)) β , Where α is the scaling factor and β is the exponential control parameter.

7. The method according to any one of claims 1 to 6, characterized in that, The method also includes: Step 10: Delete the point cloud corresponding to the obstacle mesh cell obtained in step 9 whose Z value does not exceed the set Z value threshold; Step 11: Merge the deletion result from Step 10 with the point cloud within the obstacle mesh cell obtained in Step 3 to obtain the final result of obstacle detection.

8. A computer-readable storage medium storing one or more programs, said one or more programs comprising instructions, characterized in that, When the instruction is executed by the computing device, it causes the computing device to perform the method as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, It includes one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for performing the method as described in any one of claims 1 to 7.

Images