A 3D point cloud segmentation method for tree-like crops based on multi-view and voxelization

Abstract

The present invention discloses a method for segmenting a three-dimensional point cloud of tree-like crops based on multi-view and voxelization, and relates to the technical field of laser point cloud data processing. The method comprises: firstly, projecting a three-dimensional canopy point cloud onto the ground plane to obtain a canopy height map, identifying the treetop position in the canopy height map, and correcting it with a point cloud density map, and then obtaining a rough segmentation of adjacent tree clusters through a large-scale watershed algorithm. Based on a side view, a measurement method is used through an angle, and the treetop position obtained in the previous step is used to further segment the tree cluster into individual plants. The tree cluster is voxelized, and then the overlapping parts of the previous segmentation are further finely segmented through the Ncut method to obtain the final single-plant crop point cloud. The segmented single-plant crop point cloud is more accurate than the general method, and provides a basis for the measurement of subsequent observation quantities such as crop height and crown diameter.

Description

Technical Field

[0001] The invention relates to the technical field of laser point cloud data processing, in particular to a three-dimensional point cloud segmentation method for tree-like crops based on multi-viewing angles and voxelization. Background Art

[0002] LiDAR has been increasingly used in object detection. It can emit laser pulses and determine the distance from the object to itself by calculating the time interval from emission to return, thereby obtaining high-density 3D point cloud data. The three-dimensional point cloud data obtained by LiDAR can better describe the contour and height of the crop canopy, and is almost unaffected by environmental factors such as solar radiation, air temperature and humidity, background temperature and light changes, so it has a good application prospect in crop canopy height detection.

[0003] Since agricultural observations also require observations of tree-like crops such as fruit trees, for this type of crop, the canopy represented by the three-dimensional point cloud clearly shows the separation between trees, so it becomes inappropriate to refer to the canopy height in the entire observation area to obtain the representation of the canopy height of a single tree. Summary of the invention

[0004] In view of the above problems and technical requirements, the inventors have proposed a method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization. The technical solution of the present invention is as follows:

[0005] A tree crop 3D point cloud segmentation method based on multi-view and voxelization includes the following steps:

[0006] Obtain the three-dimensional point cloud data of tree crops in the observation area and project it onto the ground plane to obtain a canopy height map from a bird's-eye view;

[0007] The tree top positions are identified in the canopy height map, and the canopy height map is roughly segmented based on the watershed algorithm to obtain multiple canopy boundaries;

[0008] Extract the point cloud data corresponding to each tree crown boundary and generate a laser distance map from the side view. Then traverse each two adjacent measurement points in the laser distance map to confirm whether they are from the same tree, so as to segment the laser distance map into single trees.

[0009] The point cloud data of adjacent overlapping areas between different trees are extracted from the segmented laser distance map and / or canopy height map to construct a three-dimensional voxel structure. The three-dimensional voxel structure is finely segmented based on the Ncut method and merged with the segmentation results of the top view and side view to obtain the final single tree point cloud segmentation result.

[0010] The beneficial technical effects of the present invention are:

[0011] This method uses a combination of top view and side view to first preliminarily locate and segment the tree clusters in the observation area, that is, locate and roughly segment the obviously separated trees, and then perform voxel-based fine segmentation on the preliminary segmentation results, that is, fine segmentation of the overlapping parts in the rough segmentation results of the top view and / or side view based on the weighted updated Ncut method, and merge these three segmentation results to obtain the final single crop point cloud. The segmented single crop point cloud is more accurate than the general method, and provides a basis for the subsequent measurement of crop height, crown diameter and other observations. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flow chart of the method for segmenting tree-like crops three-dimensional point cloud based on multi-view and voxelization provided in this application.

[0013] Figure 2 (a) is the original canopy height image provided by this application, Figure 2 (b) is the canopy height map after filtering.

[0014] Figure 3 (a) is a schematic diagram of searching for tree top positions in the canopy height map provided by this application, Figure 3 (b) is a schematic diagram of searching for the optimal location of the tree top in the point cloud density map.

[0015] Figure 4 (a) is a schematic diagram of the tree top position marked based on the watershed algorithm provided by the present application, Figure 4 (b) is a schematic diagram of multiple tree crown boundaries obtained based on the watershed algorithm.

[0016] Figure 5 (a) is a three-dimensional side view of a point cloud corresponding to a tree crown boundary provided by this application. Figure 5 (b) is the corresponding laser distance diagram from the side view.

[0017] Figure 6 This is a schematic diagram of segmenting the laser distance map provided by the present application.

[0018] Figure 7 It is a schematic diagram of the angle measurement calculation established by two laser beams provided in this application.

[0019] Figure 8 It is a schematic diagram of the internal point cloud distribution of the octree structure provided by this application.

[0020] Fig. 9 This is a schematic diagram of point cloud voxel refinement segmentation provided by this application. DETAILED DESCRIPTION

[0021] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0022] Please refer to Figure 1 As shown, this embodiment provides a tree-like crop 3D point cloud segmentation method based on multi-view and voxelization, comprising the following steps:

[0023] Step 1: Obtain the three-dimensional point cloud data of tree crops in the observation area and project it onto the ground plane to obtain a canopy height map from a bird's-eye view.

[0024] Among them, the canopy height map is a two-dimensional projection of the three-dimensional point cloud, which eliminates the influence of terrain fluctuations on the canopy height, thus reflecting the fluctuation of the canopy height, such as Figure 2 (a) shown.

[0025] Step 2: Identify the tree tops in the canopy height map, including the following steps:

[0026] Step 2.1: Filter the canopy height map.

[0027] Since the generated canopy height map contains noise that may affect the determination of the local maximum, this embodiment preferably uses a circular Gaussian filter with a fixed radius to smooth the canopy height map to reduce the noise, thereby generating a Gaussian smoothed canopy height map, such as Figure 2 Optionally, other filters with better filtering effects may be selected for smoothing filtering, and this application does not impose any specific restrictions on this.

[0028] Step 2.2: In the filtered canopy height map, search for the local maximum through a sliding window as the position information of the top of a single tree, and use the point cloud density map for correction.

[0029] The point cloud of crops sometimes exhibits the following characteristics: (1) The height of the tree top is not significantly higher than other parts of the crown; (2) The tree is crooked, causing the tree's high point to deviate from the center of the crown; (3) The spacing between trees is too small, causing parts of the crown to be connected. Because the selection of the local maximum has a very important impact on the segmentation result, in order to avoid the above situation, this embodiment first selects a relatively large range (i.e., at least larger than a single tree) of local maxima in the Gaussian smoothed canopy height map to perform the first step of coarse segmentation to obtain the aggregation area of ​​multiple crops. Then, narrow the sliding window to obtain the information of the top of a single tree, such as Figure 3 The three points in (a) represent the top of a tree.

[0030] The information of individual tree tops obtained only from the canopy height map is not accurate, and the vertex position of a single tree crown is usually in the grid with the highest point cloud density, such as Figure 3Therefore, in this embodiment, the grid position with the largest density of individual trees is obtained by searching the local maximum value through a sliding window in the filtered point cloud density map, and the grid position is used as the preferred position of the individual tree top, and then the position information of the individual tree top searched in the canopy height map is corrected according to the preferred position of the tree top.

[0031] Step 3: Roughly segment the canopy height map based on the watershed algorithm to obtain multiple tree crown boundaries.

[0032] In this embodiment, the corrected position information of the top of a single tree obtained in the previous step is marked as the lowest point in the watershed algorithm. Then, the watershed line is obtained according to the existing implementation method of the watershed algorithm, thereby generating a group of roughly segmented tree-like crop parts on the canopy height map. The segmentation result is as follows: Figure 4 As shown, in Figure 4 In (a), each treetop is marked with a dot. Figure 4 In (b), the crown boundary is highlighted by a polygon.

[0033] Step 4: Extract the point cloud data corresponding to each tree crown boundary and generate a laser distance map from the side view. Traverse each two adjacent measurement points in the laser distance map in turn to confirm whether they are from the same tree, so as to segment the laser distance map into single trees.

[0034] Since the point cloud collected by the LiDAR contains not only the top surface information of the canopy but also the side information, the point cloud after the rough segmentation in the previous step is restored to the three-dimensional side view angle, such as Figure 5 As shown in (a), the positioning of individual crops is further corrected by the side view of the point cloud. Specifically, the laser distance map of the three-dimensional side view point cloud is first generated, such as Figure 5 As shown in (b), the key to single tree segmentation is to determine whether the laser beam formed by the measuring point in the distance map and the distance measurement origin O comes from the same tree. Then, starting from the search starting point of the laser distance map, each measuring point is traversed in the specified order. When an unmarked measuring point is encountered, a breadth-first search is performed on the measuring point. The goal is to mark all measuring points belonging to the same tree as the measuring point and assign the same label until all measuring points in the laser distance map are marked. The result of single tree segmentation of the laser distance map is output according to the label, as shown in Figure 6 As shown in the figure, we can see that there are some areas where crops are adjacent to each other. In the next step, we will perform more detailed segmentation on the 3D point clouds of these adjacent overlapping areas.

[0035] The breadth-first search for the measuring point includes:

[0036] <1> Establish a search queue and push unmarked measurement points into the search queue;

[0037] <2> If the search queue is not empty, the elements in the queue are traversed in sequence; otherwise, the next measurement point of the laser distance map is traversed in sequence until a new unmarked measurement point is reached;

[0038] <3> For each measurement point in the neighborhood of the current element, determine whether the current element and the neighboring measurement point are from the same tree based on the angle measurement form, including:

[0039] Take a point farther from the origin of the distance measurement between the current element and the neighboring measurement points and connect it with the origin of the distance measurement, and calculate the angle between the connecting line and the connecting line between the current element and the neighboring measurement points β , if it is greater than the set threshold θ , then the current element and the neighboring measurement point are considered to be from the same tree; if it is not greater than the set threshold θ , it is considered that the current element and the neighboring measurement points come from different trees, and the traversal continues to the next element in the queue.

[0040] like Figure 7 As shown, assuming A , B The points are any measuring points of the current element and its neighborhood, respectively. A , B The coordinates of the distance measurement origin O The y-axis is the center of the two laser beams. OA , OB In the longer coordinate system, the angle required is β for OA , AB The angle between them. According to the nature of laser ranging, the distance || OA || corresponds to the first laser beam, and the measured value is recorded as d 1 , distance|| OB || corresponds to the second laser beam, and the measured value is recorded as d 2 , then use the following formula to calculate β :

[0041] (1)

[0042] in, φ It is the angle formed by the lines connecting the current element and the neighboring measurement point with the distance measurement origin.

[0043] Normal angle β For most objects, the depth difference between adjacent measurement points in a given distance map is relatively large, and only when the depth difference between adjacent measurement points is significantly greater than the depth difference caused by the angle increment of the laser beam. β and θ The size between determines whether to divide two points in the distance graph into two clusters or merge them into one cluster.β No more than θ , it is considered that the depth change is too large and the two points point to different trees.

[0044] Optionally, the search starting point set in this embodiment is to start from the upper left corner measuring point of the distance map and pass through each measuring point in order from top to bottom and from left to right. It is stipulated that the neighborhood of the current element consists of the four measuring points of the top, bottom, left and right of the current element.

[0045] <4> Push the neighboring measurement points from the same tree into the search queue and assign them the same label as the current element as marked, and delete the current element from the queue.

[0046] Step 5: Extract point cloud data of adjacent overlapping areas between different trees from the segmented laser distance map and / or canopy height map and construct a three-dimensional voxel structure.

[0047] Compared with airborne laser radar, the laser point cloud data of ground agricultural gas stations has a higher point density, so the tree structure details can be described in more detail. Therefore, establishing a spatial index can achieve better point cloud processing. This embodiment preferably constructs an octree structure. In order to reduce the bloated nodes in the existing octree structure, this application proposes an internal point cloud distribution rule to determine whether further splitting is required. Specifically, it includes:

[0048] Step 5.1: Based on the segmentation information of the top view and the side view, extract the point cloud data of the adjacent overlapping areas between different trees in the laser distance map and / or the canopy height map to construct the root node. The following is a detailed description using the segmented laser distance map output in step 4 as an example:

[0049] Firstly, according to the point cloud data of the adjacent overlapping areas between different trees in the segmented laser distance map, the limit values ​​of the point cloud distribution in the XYZ directions (MaxX, MinX, MaxY, MinY, MaxZ, MinZ) are obtained.

[0050] Then take (MinX, MinY, MinZ) as the lower left corner position of the root node, and set the side length of the root node to: (2)

[0051] Among them, MaxD=Max[(MaxX- MinX),(MaxY- MinY)], and the specific value of N is obtained according to the inequality. Then the root node established is a cubic space structure, and all point clouds are added to the root node.

[0052] Step 5.2: Build a worker node queue and push the root node into the worker node queue.

[0053] Step 5.3: If the work node queue is not empty, traverse the elements in the queue in turn and go to step 5.4. If the work node queue is empty, output the octree structure, and regard each leaf node of the octree structure as a voxel, thereby obtaining a three-dimensional voxel structure.

[0054] Step 5.4: If the edge length of the current element is not greater than the predetermined minimum node edge length, mark the current element as a leaf node and delete it from the queue, and return to step 5.3 to continue traversing the next element in the queue; otherwise, go to step 5.5.

[0055] Step 5.5: If the edge length of the current element is greater than the predetermined minimum node edge length, and the number of point clouds of the current element is greater than the predetermined minimum number of points, proceed to step 5.7, otherwise proceed to step 5.6.

[0056] Step 5.6: If the side length of the current element is greater than the predetermined minimum node side length, and the number of point clouds of the current element is not greater than the predetermined minimum number of points, determine whether to continue splitting the current element according to the internal point cloud distribution rule of the current element.

[0057] Specifically, assuming that the current element is divided into eight equal-capacity subnodes, the distribution of the point cloud in the current element is determined according to the searched empty subnodes. For example, if there are four coplanar empty subnodes among the searched empty subnodes, it is considered that the point cloud in the current element is distributed on the other side, such as Figure 8 If the point cloud is only distributed on one side or a specific corner of the current element, go to step 5.7; otherwise, the point cloud is considered to be randomly distributed, and go to step 5.4 to mark the current element as a leaf node.

[0058] Step 5.7: Divide the current element into eight equal-capacity child nodes, update the point cloud of each equal-capacity child node, push the non-empty child nodes into the queue, delete the current element from the queue, and return to step 5.3 to continue traversing the next element in the queue.

[0059] Through the above method, fewer boundary type nodes appear in the expanded nodes, and the average position accuracy of the nodes in the boundary area of ​​the object is improved.

[0060] Step 6: Based on the Ncut method, the 3D voxel structure is finely segmented and merged with the segmentation results of the top view and the side view to obtain the final single tree point cloud segmentation result.

[0061] After the point cloud is voxelized, according to the segmentation information of the top view and the side view, this application uses the Ncut segmentation method to segment the voxels in the overlapping area. Specifically, it includes: constructing a weighted graph G={V,E}, where V represents the set of nodes in the weighted graph, and the node is each voxel in the three-dimensional voxel structure; E represents the edge set between the nodes in the weighted graph, and the weight of each edge isw ij Represents the similarity between nodes i and j.

[0062] The purpose of Ncut is to cut the graph G into two disjoint node parts C and D, that is, by maximizing the similarity of the internal nodes of C and D while minimizing the node similarity between C and D. The general expression is:

[0063] (3)

[0064] in, represents the sum of the weights of the edges between C and D, and represents the sum of the weights of all edges with one end in C, Similarly, the segmentation result is obtained by solving the feature value that minimizes formula (3). The solution method is based on the conventional Ncut algorithm and will not be repeated here.

[0065] However, the existing single tree segmentation Ncut method has two disadvantages. First, the integrity of the individual. If the point set contains more than two trees, the segmentation based on the global cost may forcibly split the single tree into two, thus destroying the integrity of the single tree. Second, the node similarity only considers the distance factor, which increases the segmentation error at the top of the branch.

[0066] In order to overcome the above shortcomings, this embodiment improves the Ncut segmentation method on the octree result and only performs segmentation in the overlapping area. On the one hand, the information of the single crown obtained in the previous step can ensure the integrity of the single crown, and on the other hand, it can greatly reduce the amount of calculation. The weight of each edge designed by this application is w ij , adding horizontal distance, vertical distance and node connectivity, the expression is:

[0067] (4)

[0068] in, represents the horizontal distance between nodes i and j, represents the vertical distance between nodes i and j, represents the standard deviation of the corresponding numerator, represents the horizontal influence radius of similarity; represents the shortest path between nodes i and j. If the two nodes are connected (for example, the two nodes can be connected through adjacent nodes), then , The shortest path from node i to node j passes through k The length of the edge of each node, mns represents the predetermined minimum node edge length in the three-dimensional voxel structure; if two nodes are not connected (for example, some nodes are in a corner and may not be connected to other nodes), then Set to a predetermined large value.

[0069] The formula (4) provides w ij Instead of the original weights in the Ncut method, the weighted graph is cut into two non-intersecting node parts, thereby achieving a refined segmentation of the point cloud in the adjacent overlapping area. The segmentation results of the overlapping part are combined with the existing segmentation results of the previous top view and side view to obtain the final voxel segmentation result as shown in the figure. Fig. 9 As shown, each point represents the geometric center of the node. After annotating the corresponding point cloud according to the voxel label, the point cloud segmentation result is obtained.

[0070] It should be noted that the refined segmentation based on the Ncut method proposed in the present application is not limited to the three-dimensional voxel structure generated based on the adjacent overlapping areas of the side view, but can also be used to generate the corresponding three-dimensional voxel structure according to the method of step 5 for the point cloud of adjacent overlapping areas of different tree crown boundaries based on the top view output in step 3, and perform refined segmentation on it according to the method of step 6.

[0071] The above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims

1. A 3D point cloud segmentation method for tree-like crops based on multi-view and voxelization, characterized in that: The method comprises: Obtain the three-dimensional point cloud data of tree crops in the observation area and project it onto the ground plane to obtain a canopy height map from a bird's-eye view; Identifying tree top positions in the canopy height map, and roughly segmenting the canopy height map based on a watershed algorithm to obtain a plurality of tree crown boundaries; Extracting the point cloud data corresponding to each of the tree crown boundaries and generating a laser distance map from a side perspective, traversing each two adjacent measurement points in the laser distance map in turn to confirm whether they are from the same tree, so as to perform single tree segmentation on the laser distance map; Extracting point cloud data of adjacent overlapping areas between different trees from the segmented laser distance map and / or canopy height map and constructing a three-dimensional voxel structure, performing fine segmentation on the three-dimensional voxel structure based on the Ncut method, and merging the segmentation results of the top view and the side view to obtain the final single tree point cloud segmentation result; The method of sequentially traversing every two adjacent measurement points in the laser distance map to confirm whether they are from the same tree to perform single tree segmentation on the laser distance map includes: Starting from the search starting point of the laser distance map, traversing each measuring point in a specified order, when encountering an unmarked measuring point, establishing a search queue and pushing the unmarked measuring point into the search queue; If the search queue is not empty, the elements in the queue are traversed in sequence; otherwise, the next measurement point of the laser range map is traversed in sequence until a new unmarked measurement point is reached; For each measurement point in the neighborhood of the current element, determine whether the current element and the neighboring measurement point are from the same tree based on the angle measurement form, push the neighboring measurement points from the same tree into the search queue, assign the same label as the current element as marked, and delete the current element from the queue; When all the measuring points in the laser distance map have been marked, the result of performing single tree segmentation on the laser distance map is output according to the labels.

2. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 1, characterized in that: The method for identifying the tree top position in the canopy height map comprises: Performing filtering on the canopy height map; In the filtered canopy height map, the local maximum is searched through a sliding window as the position information of the top of a single tree, and the point cloud density map is used for correction.

3. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 2, characterized in that: The method for correcting using a point cloud density map comprises: In the filtered point cloud density map, the local maximum is searched through a sliding window to obtain the grid position with the largest density of single trees, which is used as the preferred position of the top of the single tree; the position information of the top of the single tree searched in the canopy height map is corrected according to the preferred position.

4. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 1, characterized in that: The method for determining whether the current element and the neighboring measurement point are from the same tree based on the angle measurement form includes: Take a point farther from the distance measurement origin from the current element and the neighboring measurement points and connect it with the distance measurement origin, calculate the angle between the line and the line between the current element and the neighboring measurement points, if it is greater than the set threshold, it is considered that the current element and the neighboring measurement points are from the same tree, otherwise it is considered that the current element and the neighboring measurement points are from different trees, and continue to traverse the next element in the queue.

5. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 4, characterized in that: The angle is recorded as β , the calculation expression is: ； in, d 1 is the larger value of the distance between the current element and the distance from the origin of the distance measurement and the distance between the neighborhood measurement point and the origin of the distance measurement, d 2 is a smaller value. φ It is the angle formed by the lines connecting the current element and the neighboring measurement point with the distance measurement origin respectively.

6. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 1, characterized in that: The method for constructing a three-dimensional voxel structure comprises: Extracting point cloud data of adjacent overlapping areas between different trees from the segmented laser distance map and / or canopy height map to construct a root node, and adding all point clouds to the root node, wherein the root node is a cubic spatial structure; Building a work node queue and pushing the root node into the work node queue; If the working node queue is not empty, the elements in the queue are traversed in sequence; otherwise, an octree structure is output, and each leaf node of the octree structure is regarded as a voxel, thereby obtaining a three-dimensional voxel structure; If the edge length of the current element is not greater than the predetermined minimum node edge length, the current element is marked as a leaf node and deleted from the queue, and the next element in the queue is traversed; If the side length of the current element is greater than the predetermined minimum node side length, and the number of point clouds of the current element is greater than the predetermined minimum number of points, the current element is divided into eight equal-capacity child nodes, the point cloud of each equal-capacity child node is updated, the non-empty child nodes are pushed into the queue, and the current element is deleted from the queue, and the next element in the queue is traversed; If the side length of the current element is greater than the predetermined minimum node side length, and the number of point clouds of the current element is not greater than the predetermined minimum number of points, it is determined whether to continue splitting the current element according to the internal point cloud distribution rule of the current element.

7. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 6, characterized in that: The methods for constructing the root node include: Obtaining the limit values ​​MaxX, MinX, MaxY, MinY, MaxZ, MinZ of the point cloud distribution in the XYZ directions according to the point cloud data of the adjacent overlapping areas between different trees in the segmented laser distance map and / or the canopy height map; Take (MinX, MinY, MinZ) as the lower left corner of the root node, and set the side length of the root node to: ; Among them, MaxD=Max[(MaxX- MinX),( MaxY- MinY)].

8. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 6, characterized in that: The method for determining whether to continue splitting the current element according to the internal point cloud distribution rule of the current element includes: Assume that the current element is divided into eight equal-capacity sub-nodes, and determine the distribution of the point cloud within the current element based on the searched empty sub-nodes. If the point cloud is only distributed on one side of the current element or in a specific corner, enter the step of dividing the current element into eight equal-capacity sub-nodes; otherwise, the point cloud is regarded as randomly distributed, and enter the step of marking the current element as a leaf node.

9. The method for segmenting tree-like crops 3D point cloud based on multi-view and voxelization according to claim 1, characterized in that: The method for finely segmenting the three-dimensional voxel structure based on the Ncut method includes: Construct a weighted graph G={V,E}, where V represents the set of nodes in the weighted graph, and a node is each voxel in the three-dimensional voxel structure; E represents the set of edges connecting the nodes in the weighted graph, and the weight of each edge is w ij Represents the similarity between nodes i and j, and is designed as: ； in, represents the horizontal distance between nodes i and j, represents the vertical distance between nodes i and j, represents the standard deviation of the corresponding numerator, represents the horizontal influence radius of similarity; represents the shortest path between nodes i and j. If the two nodes are connected, then , The shortest path from node i to node j passes through k The length of the side of a node, mns represents the predetermined minimum node side length in the three-dimensional voxel structure; if two nodes are not connected, Set to a predetermined large value; Use the above w ij Instead of the original weights in the Ncut method, the weighted graph is cut into two non-intersecting node parts, thereby achieving refined segmentation of the point cloud of adjacent overlapping areas.

Images