Single tree segmentation method based on UAV lidar point cloud based on collaborative information guidance

Abstract

The present invention discloses a method for segmenting individual trees from UAV LiDAR point clouds based on collaborative information guidance. The method comprises obtaining initial tree crown clusters; constructing geometrically significant features based on the unique distribution of UAV LiDAR point clouds in forest areas to achieve accurate extraction of trunk candidate points and object-oriented detection of tree trunks; classifying the initial tree crown clusters into under-segmentation, correct segmentation, and over-segmentation; outputting the correct segmented tree crowns as the individual tree segmentation results, and further refining the under-segmented clusters; and optimizing the under-segmented tree crowns using mutual reference information between the crowns and trunks. The present invention enhances the ability to refine tree boundary segmentation, effectively improving the accuracy and reliability of individual tree extraction from UAV LiDAR point clouds.

Description

Technical Field

[0001] The present invention relates to a method for segmenting single trees in a UAV laser radar point cloud, specifically a method for extracting single trees by comprehensively considering crown features, trunk features, and the coordinated mutual reference features of crowns and trunks. The method belongs to the technical field of laser radar point cloud data processing and forest resource survey. Background Art

[0002] Forests are the largest terrestrial ecosystems and play a vital role in maintaining global ecological balance. Efficient and accurate acquisition of forest resource information is a prerequisite and foundation for forest resource protection and management. Unmanned aerial vehicle (UAV) LiDAR (LiDAR) technology can rapidly acquire three-dimensional spatial information about forest structure. Its advantages include efficiency, flexibility, high temporal and spatial resolution, cloud-free operation, and low cost, making it a crucial tool for forest resource surveys. Extracting individual trees from UAV LiDAR point clouds is a fundamental step in obtaining structural parameters for individual trees. Accurate structural parameters provide essential data for forest resource surveys and forest management decisions, and also support forestry applications such as forest growth simulation, forest carbon stock assessment, and forest ecosystem function evaluation. However, issues such as the large volume of UAV LiDAR data, the diverse forest site conditions, the variety of tree structures and morphologies, and mutual occlusion have hindered the accuracy of individual tree extraction from UAV LiDAR point clouds, severely restricting the widespread application of UAV LiDAR technology in forest resource surveys, forest protection, and management.

[0003] Currently, the existing single wood splitting methods in the prior art may have the following problems:

[0004] 1) Top-down tree segmentation methods. Due to the bird's-eye view of observations from satellite / manned aircraft platforms, the forest area LiDAR point cloud or image data obtained can better characterize and cover the forest canopy structure. Single tree extraction methods based on satellite and manned aircraft point clouds or images start from the crown characteristics and follow a top-down single tree extraction pattern. This type of method first identifies and locates the crown vertices, and then gradually distributes the crown to individual treetops using a series of different segmentation methods. Common segmentation methods include watershed segmentation, region growing, mean shift, morphological reconstruction, etc. However, the performance of this type of method is completely dependent on the ability to characterize and describe the crown characteristics and is easily affected by heterogeneous canopy structures.

[0005] 2) Bottom-up tree segmentation methods. Terrestrial Laser Scanning (TLS) technology can obtain detailed understory vegetation information. TLS-based tree extraction methods fully utilize trunk features and follow a bottom-up approach. These methods first locate the base or trunk of the tree and then use segmentation methods such as the k-nearest neighbor algorithm, morphological methods, and graph theory to gradually assign crown points to their respective trunks. However, these methods rely on the effectiveness of trunk detection. When trunk identification fails, these methods will be hindered or even terminate prematurely.

[0006] 3) Existing methods for extracting individual trees from drone point clouds typically treat drone point clouds similar to manned aircraft point clouds and employ the same top-down tree segmentation strategy as manned aircraft point clouds. However, drone point clouds can provide more comprehensive and denser tree trunk structural information than manned aircraft point clouds. Furthermore, compared with both manned and ground-based point clouds, drone point clouds can simultaneously capture more complete and detailed information about both crown and trunk structure. These unique advantages of drone laser point cloud data have yet to be fully explored and applied.

[0007] 4) Although the deep learning-based single tree extraction method can obtain better single tree extraction results for the data of a given scene, it requires the collection and labeling of a large amount of sample data for the training of the deep learning model, which is costly to implement and has poor transplantability and robustness to different types of forest structures under different site conditions. Summary of the Invention

[0008] In response to the problems raised in the above background technology, the purpose of the present invention is to fully utilize and explore the characteristics of UAV lidar point cloud, and propose a UAV lidar point cloud single tree segmentation method based on collaborative information guidance. This single tree segmentation method is based on the collaborative information guidance of tree trunks and crowns to improve the accuracy and robustness of single tree segmentation, thereby solving the problems existing in the existing technology.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a method for segmenting single trees from UAV laser radar point clouds based on collaborative information guidance, the method comprising the following steps:

[0010] S1: The UAV LiDAR scans the forest area to obtain a tree crown point cloud. The mean-shift algorithm with a 3D asymmetric kernel is used to perform a preliminary segmentation of the tree crown point cloud to obtain a rough tree crown segmentation result, i.e., the initial tree crown clusters.

[0011] S2: Based on the point density distribution characteristics of tree crowns and trunks in the UAV LiDAR point cloud in the forest area, a height histogram is used to perform a preliminary separation of trunk points and crown points to obtain trunk candidate points. Based on the unique distribution of the UAV LiDAR point cloud in the forest area, geometric saliency features are constructed and the trunk candidate points are further subdivided to achieve accurate extraction of trunk candidate points and object-oriented detection of tree trunks.

[0012] S3: Construct crown and trunk collaborative features and mean shift modulus variance features, use random forest to classify the initial crown clusters into under-segmentation, correct segmentation and over-segmentation, and retain the correct segmentation as the single tree segmentation result;

[0013] S4: Using the mutual coordination and reference information of the crown and trunk, the over-segmented crown is refined. At the same time, vertical region growth is performed on the trunk, and all crown clusters are sorted from high to low, and matched with the trunk respectively. The correctly segmented crown is output as the single tree segmentation result, and the under-segmented clusters are further refined.

[0014] S5: For under-segmented crown clusters after matching, a graph cut segmentation method based on tree center vector deviation angle optimization is constructed, and the mutual reference information between the crown and the trunk is used to optimize the under-segmented crown.

[0015] As a further technical solution of the present invention, S1 specifically includes:

[0016] S11: Normalize the tree crown point cloud obtained by scanning the forest area with the UAV lidar, and eliminate the low-lying vegetation points in the forest area according to the height threshold;

[0017] S12: Using the mean shift algorithm of 3D asymmetric kernel, the tree crown point cloud is preliminarily segmented. The three-dimensional space of the UAV lidar point cloud is divided into two parts: the horizontal domain and the vertical domain. The horizontal domain uses a symmetric Gaussian kernel function to find the local maximum of density, and the vertical domain uses an asymmetric kernel function to find the local maximum of height. For any point X in the UAV lidar point cloud, A , and its mean shift vector is expressed as:

[0018]

[0019] Among them, the vertical kernel function g r It is an asymmetric function. The specific expression is as follows:

[0020]

[0021]

[0022] Where: superscripts s and r represent the horizontal domain and vertical domain respectively; gs and g r Represent the horizontal kernel function and the vertical kernel function respectively; h s and h r Represent the horizontal kernel bandwidth and vertical kernel bandwidth respectively; and Represents X A and X i The horizontal component of the coordinate; and Represents X A and X i The vertical component of the coordinate; each point X i To X A The weight of the mean shift vector is related to the point and X A The distance is related to the kernel function; Used to filter X within a fixed bandwidth i ; Calculate X A and X i The horizontal coordinate distance.

[0023] As a further technical solution of the present invention, S2 specifically includes:

[0024] S21: Divide the clusters obtained by mean-shift clustering of each asymmetric kernel into twelve intervals according to elevation, count the frequency of points appearing in each interval, analyze the continuous height intervals with low frequency in the height histogram, perform a preliminary separation of crown points and trunk points, and obtain candidate trunk points;

[0025] S22: Calculate linear difference significance features, plane difference significance features, and scattered difference significance features for each of the trunk candidate points, and use random forest to further subdivide the trunk candidate points into trunk points and non-trunk points. In the calculation of the plane difference significance features SoDP and the scattered difference significance features SoDS, P, L, and S are calculated based on the eigenvalues ​​of the covariance matrix of the point cloud segmentation part, representing the plane features, linear features, and scattered point features, respectively, and the formulas are as follows:

[0026] SoDP=P+(1-P)×[P-max(L,S)]

[0027] SoDS=S+(1-S)×[S-max(L,P)]

[0028] S23: Clustering the identified tree trunk points using a 2D mean shift algorithm to obtain tree trunk clusters.

[0029] As a further technical solution of the present invention, S3 specifically includes:

[0030] S31: Based on the initial crown clusters obtained in S1, the crown and trunk collaborative features and the mean shift modulus point variance features are constructed, and the incorrect segmentation in the initial crown clusters is identified based on the crown and trunk collaborative features and the mean shift modulus point variance features;

[0031] The crown and trunk synergistic features include the number of trunks below the initial crown cluster, the crown width predicted by the allometric model based on tree height, and the difference between the crown width predicted based on crown height and the actual crown width of the crown cluster.

[0032] S32: The crown and trunk collaborative features and the mean shift modulus variance features are used in random forests to classify the initial crown clusters into under-segmentation, over-segmentation and correct segmentation.

[0033] As a further technical solution of the present invention, the S4 specifically includes:

[0034] S41: Using the mutual synergy and reference information between crowns and trunks, i.e. the allometric relationship between the trunk and crown sizes, the crown clusters identified as over-segmented are merged with the nearest under-segmented clusters;

[0035] S42: performing vertical region growth on the tree trunk detected in S3 to obtain a complete tree trunk;

[0036] S43: Sort the under-segmented and correctly segmented crowns from high to low, and use the allometric growth relationship between the trunk and crown to match them with the vertically grown trunks. The correctly segmented crowns are output as the single tree segmentation results, while the under-segmented results require further refinement.

[0037] As a further technical solution of the present invention, the S5 specifically includes:

[0038] S51: Using the mutual reference information between the crown and the trunk, i.e. the two-dimensional vector angle consistency between the mean shift modulus of the crown and the center of mass of the trunk, as well as the allometric relationship between the trunk and the crown, we perform graph cut optimization on the under-segmented clusters after matching:

[0039] First, the crown width was estimated based on the crown height of each cluster using an allometric growth model.

[0040] Secondly, the trunk point to be segmented is used as the foreground seed point of the graph cut, the remaining trunks are used as the background seed points of the graph cut, and the crown width estimated according to the allometric growth model is used as the graph cut radius;

[0041] The radius calculation formula is as follows:

[0042] radius=0.446×H 0.854

[0043] Where: H is the normalized height of the trunk;

[0044] S52: Calculate the two-dimensional vector D from each point to the modulus point through the mean shift algorithm n , D n As the prediction vector for each point pointing to the center of the tree;

[0045] S53: Calculate the centroid of the foreground point, use the centroid as the tree center, and calculate the vector D of each point pointing to the tree center (n,K) , D (n,K) as the true vector for each point pointing to the center of the tree;

[0046] S54: Based on the prediction vector D n With the real vector D (n,K) , calculate the tree center vector angle weight w θ :

[0047]

[0048] S55: Deviate the tree center from the angle weight w θ and the original background penalty w in the graph cut back_penalty Add them together to get the mixed weight sink_weight as the new background penalty input into the graph cut algorithm. The new background penalty formula is as follows:

[0049]

[0050] sink_weight=0.5*w back_penalty +0.5*w θ

[0051] Where: dist 2d Indicates the two-dimensional plane distance from the foreground point to the input point cloud;

[0052] S56: Remove the segmented point cloud from the input point cloud, and repeat the above operation until all under-segmentation is performed.

[0053] The beneficial effects of the present invention are:

[0054] 1) This invention fully utilizes and mines UAV lidar point clouds to obtain the uniqueness of high-density crown and trunk point clouds. Through the clues and guidance of the mutual reference between the trunk and crown, it ensures the high accuracy and robustness of single tree extraction. The invention is applicable to different forest conditions and forest types, including coniferous forests, broad-leaved forests, and mixed coniferous and broad-leaved forests.

[0055] 2) Based on the spatial distribution characteristics of UAV LiDAR point clouds, this paper designs a tree trunk detection method based on the planar difference significance and scattering difference significance features, thereby improving the accuracy of tree trunk detection;

[0056] 3) The present invention utilizes the two-dimensional vector angle consistency between the mean shift modulus of the crown and the trunk centroid, as well as the allometric growth relationship between the trunk and the crown, to enhance the fault tolerance and refinement processing capabilities of over-segmentation and under-segmentation of the crown, improve the accuracy and robustness of single tree extraction in different forest types, diverse tree structures and forms, and mutual occlusion, enhance the accuracy and reliability of UAV lidar technology in forest resource surveys, and promote the promotion and application of UAV lidar technology in forestry surveys. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the overall technical process of the present invention;

[0058] Figure 2a This is a schematic diagram of a tree crown before rough segmentation in an embodiment of the present invention;

[0059] Figure 2b This is a schematic diagram of a roughly segmented tree crown in an embodiment of the present invention;

[0060] Figure 3 A tree crown cluster height histogram constructed in an embodiment of the present invention;

[0061] Figure 4a This is a schematic diagram of a tree trunk before extraction in an embodiment of the present invention;

[0062] Figure 4b This is a schematic diagram of a tree trunk after extraction in an embodiment of the present invention;

[0063] Figure 5 Schematic diagram of a segmentation method based on tree center vector deviation angle optimization in an embodiment of the present invention;

[0064] Figure 6a Schematic diagram of the segmentation result before tree center vector deviation angle optimization in an embodiment of the present invention;

[0065] Figure 6b Schematic diagram of the segmentation result after optimizing the tree center vector deviation angle in an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] Example 1, as Figure 1As shown, this embodiment provides a method for segmenting single trees from UAV lidar point clouds based on collaborative information guidance. The method includes the following steps:

[0068] S1: A UAV lidar scan of the forest area obtains a crown point cloud (crown distribution characteristics). The mean-shift algorithm with a 3D asymmetric kernel is used to perform a preliminary segmentation of the crown point cloud to obtain a rough crown segmentation result (initial crown clusters).

[0069] S1 specifically includes:

[0070] S11: Normalize the tree crown point cloud obtained by UAV lidar scanning and eliminate low-lying vegetation points in the forest area according to the height threshold;

[0071] S12: Use the mean shift algorithm with 3D asymmetric kernel to perform preliminary segmentation of the tree crown point cloud;

[0072] The three-dimensional space of the UAV lidar point cloud is divided into two parts: the horizontal domain and the vertical domain. A symmetric Gaussian kernel function is used in the horizontal domain to find the local maximum of density, while an asymmetric kernel function is used in the vertical domain to find the local maximum of height.

[0073] For any point X in the UAV lidar point cloud A , its mean shift vector can be expressed as:

[0074]

[0075] Among them, the vertical kernel function g r It is a specially designed asymmetric function, the purpose of which is to give higher points greater weights. The specific expression is as follows:

[0076]

[0077] Where: superscripts s and r represent the horizontal domain and vertical domain respectively; g s and g r Represent the horizontal kernel function and the vertical kernel function respectively; h s and h r Represent the horizontal kernel bandwidth and vertical kernel bandwidth respectively; and Represents X A and X i The horizontal component of the coordinate; and Represents X A and X i The vertical component of the coordinate; each point X i To X A The weight of the mean shift vector is related to the point and X AThe distance is related to the kernel function; Used to filter X within a fixed bandwidth i ; Calculate X A and X i The horizontal coordinate distance.

[0078] S2: Based on the point density distribution characteristics of tree crowns and trunks in the UAV LiDAR point cloud in the forest area, the height histogram is used to perform a preliminary separation of trunk points and crown points to obtain trunk candidate points. Based on the unique distribution of UAV LiDAR point clouds in the forest area, geometric significance features such as planar difference significance and discrete difference significance are constructed to further subdivide the trunk candidate points, achieving accurate extraction of trunk points and object-oriented detection of tree trunks.

[0079] S2 specifically includes:

[0080] S21: Based on the point density distribution characteristics of tree crowns and trunks in the UAV LiDAR point cloud in the forest area, the clusters obtained by mean-shift clustering of each asymmetric kernel are divided into twelve intervals according to elevation. The frequency of points appearing in each interval is counted, and the continuous height intervals with low frequency in the height histogram are analyzed to perform a preliminary separation of tree crown points and tree trunk points to obtain candidate trunk points.

[0081] S22: Calculate the linear difference significance feature, plane difference significance feature and scattering difference significance feature for each trunk candidate point, and use random forest to further subdivide the trunk candidate points into trunk points and non-trunk points;

[0082] Among them, in the calculation of the plane difference significance feature SoDP and the scattered difference significance feature SoDS, P, L, and S are calculated based on the eigenvalues ​​of the covariance matrix of the point cloud segmentation part, representing the plane feature, linear feature, and scattered point feature, respectively. The formula is as follows:

[0083] SoDP=P+(1-P)×[P-max(L,S)]

[0084] SoDS=S+(1-S)×[S-max(L,P)]

[0085] S23: Clustering the identified tree trunk points using the 2D mean shift algorithm to obtain tree trunk clusters, thereby achieving object-oriented detection of tree trunks.

[0086] S3: Construct crown and trunk collaborative features and mean shift modulus variance features, use random forest to divide the crown preliminary segmentation clusters into under-segmentation, correct segmentation and over-segmentation, and retain the correct segmentation as the single tree segmentation result;

[0087] S3 specifically includes:

[0088] S31: For the initial crown clusters obtained in S1, construct crown and trunk collaborative features and mean shift modulus variance features, and identify incorrect segmentation in the clusters based on these features;

[0089] The crown and trunk synergistic features include the number of trunks, the crown width predicted by the allometric model based on tree height, and the difference between the crown width predicted by crown height and the actual crown width of the crown cluster.

[0090] S32: Based on the crown and trunk collaborative features and mean shift modulus variance features in S31, random forest is used to classify crown clusters into under-segmentation, over-segmentation and correct segmentation.

[0091] S4: Utilize the mutual coordination and reference information of the crown and trunk to refine the over-segmented crown; at the same time, perform vertical region growth on the trunk, sort all crown clusters from high to low, and match them with the trunk respectively. The correctly segmented crown is output as the single tree segmentation result, and the under-segmented clusters are further refined.

[0092] S4 specifically includes:

[0093] S41: Using the mutual synergy and reference information between crowns and trunks, i.e. the allometric relationship between the trunk and crown sizes, the crown clusters identified as over-segmented are merged with the nearest under-segmented clusters;

[0094] S42: performing vertical region growth on the tree trunk detected in S3 to obtain a tree trunk that is as complete as possible;

[0095] S43: Sort the under-segmented and correctly segmented crowns from high to low, and use the allometric growth relationship between the trunk and crown to match them with the vertically grown trunks. The correctly segmented crowns are output as the single tree segmentation results, while the under-segmented results require further refinement.

[0096] S5: For under-segmented crown clusters after matching, a graph cut segmentation method based on tree center vector deviation angle optimization is constructed, and the mutual reference information between the crown and the trunk is used to optimize the under-segmented crown.

[0097] S5 specifically includes:

[0098] S51: Utilize the mutual reference information between crown and trunk, i.e., the two-dimensional vector angle consistency between the mean shift modulus of the crown and the center of mass of the trunk, as well as the allometric relationship between the trunk and crown sizes, to perform graph cut optimization on the under-segmented clusters after matching.

[0099] First, the crown width is estimated based on the crown height of each cluster using an allometric growth model.

[0100] Secondly, the trunk point to be segmented is used as the foreground seed point of the graph cut, the remaining trunks are used as the background seed points of the graph cut, and the crown width estimated by the allometric growth model is used as the graph cut radius;

[0101] The radius calculation formula is as follows:

[0102] radius=0.446×H 0.854

[0103] Where: H is the normalized height of the trunk;

[0104] S52: Calculate the two-dimensional vector D from each point to the modulus point through the mean shift algorithm n , D n As the prediction vector for each point pointing to the center of the tree;

[0105] S53: Calculate the centroid of the foreground point, use the centroid as the tree center, and calculate the vector D of each point pointing to the tree center (n,k) , D (n,k) as the true vector for each point pointing to the center of the tree;

[0106] S54: Based on the prediction vector D n With the real vector D (n,k) , calculate the tree center vector angle weight w θ :

[0107]

[0108] S55: Deviate the tree center from the angle weight w θ and the original background penalty w in the graph cut back_penalty Add them together to get the mixed weight sink_weight as the new background penalty input into the graph cut algorithm. The new background penalty formula is as follows:

[0109]

[0110] sink_weight=0.5*w back_penalty +0.5*w θ

[0111] Where: dist 2d Indicates the two-dimensional plane distance from the foreground point to the input point cloud;

[0112] S56: Remove the segmented point cloud from the input point cloud, and repeat the above operation until all under-segmentation is performed.

[0113] Example 2: This embodiment provides a method for segmenting individual trees from UAV laser radar point clouds based on collaborative information guidance. The method comprises the following steps:

[0114] First, the UAV LiDAR scans the forest area to obtain the forest area point cloud (crown point cloud), and preprocesses the forest area point cloud and roughly segments the crown, including: filtering the ground points and normalizing the height of non-ground points; using the mean shift algorithm with a 3D asymmetric kernel to achieve rough single tree segmentation and obtain the initial crown clusters, such as Figure 2a and Figure 2b shown.

[0115] Second: Extract the trunk points of the initial single tree clusters, including: calculating the height histogram of the cluster to obtain potential trunk points, such as Figure 3 ; Calculate the plane difference significance, scattering difference significance and other features, use random forest to further screen potential trunk points (trunk candidate points); use 2D mean shift algorithm to cluster the trunk points, the trunk point extraction results are as follows Figure 4a and Figure 4b shown.

[0116] Third: Segment the initial crown clusters in the first step, including: constructing the collaborative features of the crown and trunk and the mean shift modulus variance features; using random forest to divide the clusters into three segmentation results: over-segmentation, under-segmentation and correct segmentation.

[0117] Fourth: refine the over-segmented and under-segmented clusters, including: merging the over-segmented clusters into the under-segmented clusters; growing the vertical regions of the trunks extracted in the second step, and then matching the trunks and crowns according to the relationship of the allometric growth model.

[0118] Fifth: Optimize the crown segmentation results, including: foreground and background segmentation; calculate the modulus vector of the mean shift algorithm; calculate the true vector of each point pointing to the center of the tree, the vector diagram is as follows Figure 5 As shown in the figure, the tree center vector deflection weight is calculated according to the model point vector and the true vector, and the tree center vector deflection weight and the background penalty weight are mixed to obtain the mixed weight. The above process is repeated, and the graph cut method is iterated until all under-segmented clusters are segmented. The results of using the tree center vector deflection optimization and not using the tree center vector deflection optimization are shown in the figure. Figure 6a and Figure 6b shown.

[0119] Working principle and process: The crown point cloud is obtained by scanning the forest area with UAV lidar, and the crown point cloud is preliminarily segmented into single trees; based on the preliminary segmentation results, on the basis of the preliminary separation of the crown and trunk, the difference significance features are designed to realize trunk detection; the collaborative features of the crown and trunk are constructed, the preliminary single tree segmentation results are classified, and the incorrect segmentation is identified; using the mutual collaboration and reference information of the crown and trunk, a segmentation method based on the optimization of the tree center vector deflection angle is constructed to optimize the incorrect segmentation and finally achieve single tree segmentation.

[0120] The present invention targets the characteristics of drone laser point clouds and a single tree extraction method for different forest types. It strengthens the ability to identify incorrect crown segmentation, reduces interference from the understory, improves the ability to refine tree boundary segmentation, and effectively improves the accuracy and reliability of single tree extraction from drone laser point clouds.

[0121] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0122] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for single tree segmentation in UAV laser radar point cloud based on collaborative information guidance, characterized in that: The single wood segmentation method comprises the following steps: S1: The UAV LiDAR scans the forest area to obtain a crown point cloud. The mean-shift algorithm with a 3D asymmetric kernel is used to perform a preliminary segmentation on the crown point cloud to obtain a rough crown segmentation result, i.e., the initial crown clusters. S2: Based on the point density distribution characteristics of tree crowns and trunks in the UAV LiDAR point cloud in the forest area, a height histogram is used to perform a preliminary separation of trunk points and crown points to obtain trunk candidate points. Based on the unique distribution of the UAV LiDAR point cloud in the forest area, geometric saliency features are constructed and the trunk candidate points are further subdivided to achieve accurate extraction of trunk candidate points and object-oriented detection of tree trunks. S3: Constructing crown and trunk collaborative features and mean shift modulus variance features, using random forest to classify the initial crown clusters into under-segmentation, correct segmentation, and over-segmentation, and retaining the correct segmentation as the single tree segmentation result; S4: Refine the over-segmented crown using the mutual coordination and reference information between the crown and trunk; The S4 specifically includes: S41: Using the allometric relationship between crown and trunk sizes, identify the over-segmented crown clusters and merge them with the nearest under-segmented clusters. S42: performing vertical region growth on the tree trunk detected in S3 to obtain a complete tree trunk; S43: Sort the under-segmented and correctly segmented crowns from high to low. Use the allometric relationship between the trunk and crown sizes to match them with the vertically grown trunks. Output the correctly segmented crowns as the single tree segmentation results. The under-segmented results require further refinement. S5: For under-segmented crown clusters after matching, a graph cut segmentation method based on tree center vector deviation angle optimization is constructed, and the mutual reference information between the crown and the trunk is used to optimize the under-segmented crown; The S5 specifically includes: S51: Using the two-dimensional vector angle consistency between the mean shift modulus of the crown and the trunk centroid, as well as the allometric relationship between the trunk and crown sizes, we perform graph cut optimization on the under-segmented clusters after matching: First, the crown width was estimated based on the crown height of each cluster using an allometric growth model. Secondly, the trunk point to be segmented is used as the foreground seed point of the graph cut, the remaining trunks are used as the background seed points of the graph cut, and the crown width estimated by the allometric growth model is used as the graph cut radius; The radius calculation formula is as follows: radius=0.446×H 0.854 Where: H is the normalized height of the trunk; S52: Calculate the two-dimensional vector D from each point to the modulus point through the mean shift algorithm n , D n As the prediction vector for each point pointing to the center of the tree; S53: Calculate the centroid of the foreground point, use the centroid as the tree center, and calculate the vector D of each point pointing to the tree center (n,k) , D (n,k) as the true vector for each point pointing to the center of the tree; S54: Based on the prediction vector D n With the real vector D (n,k) , calculate the tree center vector angle weight w θ : S55: Deviate the tree center from the angle weight w θ and the original background penalty w in the graph cut back_penalty Add them together to get the mixed weight sink_weight as the new background penalty input into the graph cut algorithm. The new background penalty formula is as follows: sink_weight=0.5*w back_penalty +0.5*w θ Where: dist 2d Indicates the two-dimensional plane distance from the foreground point to the input point cloud; S56: Remove the segmented point cloud from the input point cloud, and repeat the above operation until all under-segmentation is performed.

2. The method for segmenting single trees from UAV laser radar point cloud based on collaborative information guidance according to claim 1 is characterized in that: Said S1 specifically includes: S11: Normalize the tree crown point cloud obtained by scanning the forest area with the UAV lidar, and eliminate the low-lying vegetation points in the forest area according to the height threshold; S12: Using the mean shift algorithm of 3D asymmetric kernel, the tree crown point cloud is preliminarily segmented. The three-dimensional space of the UAV lidar point cloud is divided into two parts: the horizontal domain and the vertical domain. The horizontal domain uses a symmetric Gaussian kernel function to find the local maximum of density, and the vertical domain uses an asymmetric kernel function to find the local maximum of height. For any point X in the UAV lidar point cloud, A , and its mean shift vector is expressed as: Among them, the vertical kernel function g r It is an asymmetric function. The specific expression is as follows: Where: superscripts s and r represent the horizontal domain and vertical domain respectively; g s and g r Represent the horizontal kernel function and the vertical kernel function respectively; h s and h r Represent the horizontal kernel bandwidth and vertical kernel bandwidth respectively; and Represents X A and X i The horizontal component of the coordinate; and Represents X A and X i The vertical component of the coordinate; each point X i To X A The weight of the mean shift vector is related to the point and X A The distance is related to the kernel function; Used to filter X within a fixed bandwidth i ; Calculate X A and X i The horizontal coordinate distance.

3. The method for segmenting single trees from UAV laser radar point cloud based on collaborative information guidance according to claim 1 is characterized in that: The S2 specifically includes: S21: Divide the clusters obtained by mean-shift clustering of each asymmetric kernel into twelve intervals according to elevation, count the frequency of points appearing in each interval, analyze the continuous height intervals with low frequency in the height histogram, perform a preliminary separation of crown points and trunk points, and obtain candidate trunk points; S22: Calculate linear difference significance features, plane difference significance features, and scattered difference significance features for each of the trunk candidate points, and use random forest to further subdivide the trunk candidate points into trunk points and non-trunk points. In the calculation of the plane difference significance features SoDP and the scattered difference significance features SoDS, P, L, and S are calculated based on the eigenvalues ​​of the covariance matrix of the point cloud segmentation part, representing the plane features, linear features, and scattered point features, respectively, and the formulas are as follows: SoDP=P+(1-P)×[P-max(L,S)] SoDS=S+(1-S)×[S-max(L,P)] S23: Clustering the identified tree trunk points using a 2D mean shift algorithm to obtain tree trunk clusters.

4. The method for segmenting single trees from UAV laser radar point cloud based on collaborative information guidance according to claim 1 is characterized in that: The S3 specifically includes: S31: Based on the initial crown clusters obtained in S1, construct crown and trunk collaborative features and mean shift modulus point variance features, and identify incorrect segmentation in the initial crown clusters based on the crown and trunk collaborative features and the mean shift modulus point variance features; The crown and trunk synergistic features include the number of trunks below the initial crown cluster, the crown width predicted by the allometric model based on tree height, and the difference between the crown width predicted based on crown height and the actual crown width of the crown cluster. S32: The crown and trunk collaborative features and the mean shift modulus variance features are used by random forest to classify the initial crown clusters into under-segmentation, over-segmentation and correct segmentation.

Images