CN112381863A - Ground laser point cloud method for forest scene - Google Patents

Abstract

The ground laser point cloud method for the forest scene comprises the following steps: s1: applying the MCS to correlate adjacent scans and predict likely overlapping regions; s2: matching adjacent scanning pairs; s3: all scans are merged to one anchor point. The invention can provide an efficient, automatic and end-to-end solution to register the multi-view and disordered forest point cloud.

Description

Translated fromChinese

森林场景的地面激光点云方法Ground-based laser point cloud method for forest scenes

技术领域technical field

本发明涉及激光点云方法领域，尤其涉及森林场景的地面激光点云方法。The invention relates to the field of laser point cloud methods, in particular to a ground laser point cloud method for forest scenes.

背景技术Background technique

森林场景的点云配准(PCR)技术是一个重要的研究课题，它的研究对象是反映树干及林下信息特征的地面点云。与城市场景不同，因为频繁的遮挡以及没有明显的特征记号，森林场景的配准更加复杂。The point cloud registration (PCR) technology of forest scene is an important research topic. Unlike urban scenes, the registration of forest scenes is more complicated because of frequent occlusions and no obvious feature markers.

在森林调查中，点云配准通常是通过使用人工安置在扫描场中的靶标来完成的。虽然基于靶标的配准可以提供准确的匹配结果，但它非常耗时，需要大量的人力物力。此外，重建完整的场景通常需要数百次的激光扫描，这些扫描可能由不同的团队在不同的时间完成。因此，研究多视图无序森林点云的配准问题迫在眉睫。In forest surveys, point cloud registration is usually done by using targets manually placed in the scan field. Although target-based registration can provide accurate matching results, it is time-consuming and requires a lot of human and material resources. Furthermore, reconstructing a complete scene often requires hundreds of laser scans, which may be done by different teams at different times. Therefore, it is urgent to study the registration problem of multi-view disordered forest point clouds.

现有的森林点云配准方法还没有很好地解决多视图无序点云的配准问题，其中主要存在以下问题：Existing forest point cloud registration methods have not solved the registration problem of multi-view disordered point clouds well. The main problems are as follows:

各扫描之间需要一定的扫描重叠区域以及尽可能少的遮挡；A certain scan overlap area and as little occlusion as possible are required between scans;

点云中树枝与树叶对树干特征提取存在干扰；处理无序点云需要对所有的扫描进行两两匹配，但其时间复杂度太高；考虑到多视图配准的误差累积，在处理时需要扫描网络具有很好的可靠性和一致性。The branches and leaves in the point cloud interfere with the feature extraction of the trunk; processing the disordered point cloud requires pairwise matching of all scans, but the time complexity is too high; considering the accumulation of errors in multi-view registration, it is necessary to process Scanning the network has very good reliability and consistency.

综上，现有方法(全连接网络)的自动化程度低、全局匹配成功率低，时间复杂度高。To sum up, the existing method (fully connected network) has a low degree of automation, a low global matching success rate, and a high time complexity.

发明内容SUMMARY OF THE INVENTION

(一)发明目的(1) Purpose of the invention

为解决背景技术中存在的技术问题，本发明提出森林场景的地面激光点云方法，能够提供高效的、自动的、端到端的解决方案来配准多视图、无序的森林点云。In order to solve the technical problems existing in the background art, the present invention proposes a ground laser point cloud method for forest scenes, which can provide an efficient, automatic, and end-to-end solution to register multi-view and disordered forest point clouds.

(二)技术方案(2) Technical solutions

为解决上述问题，本发明提出了森林场景的地面激光点云方法，包括以下步骤：In order to solve the above problems, the present invention proposes a ground laser point cloud method for forest scenes, including the following steps:

S1：应用MCS关联相邻扫描并预测可能的重叠区域；S1: Apply MCS to correlate adjacent scans and predict possible overlapping areas;

S2：匹配相邻扫描对；S2: Match adjacent scan pairs;

S3：将所有扫描合并到一个锚点。S3: Merge all scans to one anchor.

优选的，S1包括以下步骤：Preferably, S1 includes the following steps:

S11：获取树干的分布图：对于单株树木，胸径和树干的中心位置记为与透视无关的不变性特征，利用公式A对树干点云进行超圆拟合，获得更加准确的树干属性；并且，通过公式A，拟合出树干模型，并且记录下单次扫描的树干分布图；S11: Obtain the distribution map of the trunk: For a single tree, the diameter at breast height and the center position of the trunk are recorded as invariant features independent of perspective, and formula A is used to perform hypercircle fitting on the trunk point cloud to obtain more accurate trunk attributes; and , through formula A, fit the trunk model, and record the trunk distribution map of a single scan;

公式A：

Formula A:

其中，

in,

每一个树干属性表示为：A＝{DBH，Cpt2，Hdem}；Each trunk attribute is expressed as: A={DBH, Cpt2, Hdem};

其中，Cpt2表示树干中心在二维空间中的位置；Among them, Cpt2 represents the position of the trunk center in two-dimensional space;

Hdem表示树干中心DEM的高度；Hdem represents the height of the DEM in the center of the trunk;

S12：基于上述的树干分布图，搜索每个树干的K个相邻树干，以生成三角形；对于每个三角形，计算出相应的质心；S12: Based on the above-mentioned trunk distribution map, search for K adjacent trunks of each trunk to generate a triangle; for each triangle, calculate the corresponding centroid;

对每个三角形用特征向量

识别，特征向量

为：Eigenvectors for each triangle

identification, feature vector

for:

其中，A_j是树干j的特征向量；Among them, A_j is the feature vector of trunk j;

d_ij是树干i和j的特征向量；d_ij is the feature vector of trunk i and j;

a_j表示树干i对应的角度；a_j represents the angle corresponding to the trunk i;

是第一轮中三角形i的质心；

is the centroid of triangle i in the first round;

比较点云Pⁱ和P^j中的每个三角形检索近似全等三角形对{Tr_i，Tr_i’}；Compare each triangle in point clouds Pⁱ and P^j to retrieve approximately congruent triangle pairs {T_i ,_Tri '};

根据上述三角形对，得到一个近似的公共子图；According to the above triangle pair, an approximate common subgraph is obtained;

S13：每个{Tr_i，Tr_i’}都有对应的{w_r，w’_r}，其中，

是

和

之间的距离；S13: Each {T_i ,_Tri '} has a corresponding {w_r , w'_r }, where,

Yes

and

the distance between;

通过程序验证所有质心，去除伪质心，将保留的质心作为新的树干分布图继续迭代；Verify all the centroids through the program, remove the false centroids, and continue to iterate with the retained centroids as the new trunk distribution map;

当迭代n和n+1中没有伪质心时，仍然有不少于3个质心点存在，则判定两个扫描是相邻的；When there are no pseudo centroids in iterations n and n+1, and there are still no less than 3 centroid points, it is determined that the two scans are adjacent;

其中，

则判定这对质心为伪质心。in,

Then the pair of centroids are determined to be pseudo-centroids.

优选的，S2包括以下步骤：Preferably, S2 includes the following steps:

S21：在两个相邻的扫描中，预测一个初步的重叠区域，即对应的公共子图区域，并将其表示为

和

利用Semantic Keypoint 4-Points Congruent Sets 方法，进行扫描对的匹配；S21: In two adjacent scans, predict a preliminary overlapping region, that is, the corresponding common subgraph region, and denote it as

and

Use the Semantic Keypoint 4-Points Congruent Sets method to match scan pairs;

在

中选择四个非共面的点，即树干中心点，生成一个集合s，确定s中没有三个点共线；exist

Select four non-coplanar points, namely the trunk center point, to generate a set s, and determine that no three points in s are collinear;

s＝{a，b，c，d}；s={a,b,c,d};

其中，每个点利用特征向量a记录其语义；Among them, each point uses the feature vector a to record its semantics;

特征向量a为：a＝{x，y，z，Hdem，DBH}；The feature vector a is: a={x, y, z, Hdem, DBH};

S22：利用每个匹配的候选基准，计算出一个刚性变换T，并将

中所有的树干中心转换到

如下：S22: Using each matched candidate datum, a rigid transformation T is calculated, and the

Transform all trunk centers in to

as follows:

将

中的树干中心

利用刚性变换T得到：

Will

trunk center in

Using rigid transformation T to get:

在

中搜索距离

最近的点

在利用公式B、公式C和公式D进行判断，其中：公式B为：exist

medium search distance

nearest point

Using formula B, formula C and formula D for judgment, where: formula B is:

如果

if

公式C为：

Formula C is:

公式D为：Formula D is:

如果

if

R₂，表示不正确匹配的树木的百分比；R₂ , the percentage of incorrectly matched trees;

R₁，表示重叠区域之外树木所占百分比；R₁ , representing the percentage of trees outside the overlapping area;

其中，如果迭代没有终止，则记录得分最高的T来转换

进行粗匹配；Among them, if the iteration does not terminate, record the T with the highest score to convert

make a rough match;

S23：引入法向量的方向一致性来避免对齐树状点云产生局部最小问题；S23: Introduce the direction consistency of the normal vector to avoid the local minimum problem caused by aligning the tree point cloud;

应用树干模型来约束变换，以解决基于点的配准方法产生的破坏树干模型的问题，具体的约束条件如公式E：The trunk model is applied to constrain the transformation to solve the problem of destroying the trunk model generated by the point-based registration method. The specific constraints are as follows:

公式E为：Formula E is:

ROT和t分别表示旋转矩阵和平移矢量中的6个自由度；ROT and t represent the 6 degrees of freedom in the rotation matrix and translation vector, respectively;

(q’_i，q_i)表示相应的点；(q'_i , q_i ) represents the corresponding point;

(n’_i，n_i)表示其法向量(n'_i , n_i ) represents its normal vector

是二维空间中树干j的中心；

is the center of trunk j in two-dimensional space;

是两次扫描中的点对；

is the point pair in two scans;

p_j是树干j上，在|Hdem_pj-Hdem_j|≤ε₄条件下和

相关的点；p_j is on the trunk j, under the condition of |Hdem_pj -Hdem_j |≤ε₄ and

relevant points;

当ΔROT＜10^-6且Δt＜10^-4时，迭代终止，得到所求的旋转矩阵和平移矢量。When ΔROT<^10-6 and Δt<^10-4 , the iteration is terminated, and the required rotation matrix and translation vector are obtained.

优选的，S3包括以下步骤：Preferably, S3 includes the following steps:

S31：应用最小环路拓展，合并扫描网络中的环路；当扫描网络中不再存在最小环路时，采用并行合并策略来合并剩余的扫描；S31: Apply the minimum loop extension to merge the loops in the scanning network; when the minimum loop no longer exists in the scanning network, adopt a parallel merging strategy to merge the remaining scans;

S32：并行合并策略基于最小生成树MST进行；MST边的权重由相邻扫描之间的重叠率来定义；S32: The parallel merging strategy is performed based on the minimum spanning tree MST; the weight of the MST edge is defined by the overlap ratio between adjacent scans;

S33：对于单中心扫描和侧向扫描的组合，并行合并策略将同时把所有侧向扫描配准到中心扫描上。S33: For the combination of single center scan and lateral scan, the parallel merge strategy will simultaneously register all lateral scans to the center scan.

本发明，能够提供高效的、自动的、端到端的解决方案来配准多视图、无序的森林点云。The present invention can provide an efficient, automatic, end-to-end solution to register multi-view, unordered forest point clouds.

具体实施方式Detailed ways

为使本发明的目的、技术方案和优点更加清楚明了，下面结合具体实施方式对本发明进一步详细说明。应该理解，这些描述只是示例性的，而并非要限制本发明的范围。此外，在以下说明中，省略了对公知结构和技术的描述，以避免不必要地混淆本发明的概念。In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments. It should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. Also, in the following description, descriptions of well-known structures and techniques are omitted to avoid unnecessarily obscuring the concepts of the present invention.

本发明提出的森林场景的地面激光点云方法，包括以下步骤：The ground laser point cloud method for forest scene proposed by the present invention includes the following steps:

S1：应用MCS关联相邻扫描并预测可能的重叠区域；S1: Apply MCS to correlate adjacent scans and predict possible overlapping areas;

S2：匹配相邻扫描对；S2: Match adjacent scan pairs;

S3：将所有扫描合并到一个锚点。S3: Merge all scans to one anchor.

本发明，能够提供高效的、自动的、端到端的解决方案来配准多视图、无序的森林点云。The present invention can provide an efficient, automatic, end-to-end solution to register multi-view, unordered forest point clouds.

在一个可选的实施例中，S1包括以下步骤：In an optional embodiment, S1 includes the following steps:

S11：获取树干的分布图：对于单株树木，胸径和树干的中心位置记为与透视无关的不变性特征，利用公式A对树干点云进行超圆拟合，获得更加准确的树干属性；并且，通过公式A，拟合出树干模型，并且记录下单次扫描的树干分布图；S11: Obtain the distribution map of the trunk: For a single tree, the diameter at breast height and the center position of the trunk are recorded as invariant features independent of perspective, and formula A is used to perform hypercircle fitting on the trunk point cloud to obtain more accurate trunk attributes; and , through formula A, fit the trunk model, and record the trunk distribution map of a single scan;

公式A：

Formula A:

其中，

in,

每一个树干属性表示为：A＝{DBH，Cpt2，Hdem}；Each trunk attribute is expressed as: A={DBH, Cpt2, Hdem};

其中，Cpt2表示树干中心在二维空间中的位置；Among them, Cpt2 represents the position of the trunk center in two-dimensional space;

Hdem表示树干中心DEM的高度；Hdem represents the height of the DEM in the center of the trunk;

S12：基于上述的树干分布图，搜索每个树干的K个相邻树干，以生成三角形；对于每个三角形，计算出相应的质心；S12: Based on the above-mentioned trunk distribution map, search for K adjacent trunks of each trunk to generate a triangle; for each triangle, calculate the corresponding centroid;

对每个三角形用特征向量

识别，特征向量

为：Eigenvectors for each triangle

identification, feature vector

for:

其中，A_j是树干j的特征向量；Among them, A_j is the feature vector of trunk j;

d_ij是树干i和j的特征向量；d_ij is the feature vector of trunk i and j;

a_j表示树干i对应的角度；a_j represents the angle corresponding to the trunk i;

是第一轮中三角形i的质心；

is the centroid of triangle i in the first round;

比较点云Pⁱ和P^j中的每个三角形检索近似全等三角形对{Tr_i，Tr_i’}；Compare each triangle in point clouds Pⁱ and P^j to retrieve approximately congruent triangle pairs {T_i ,_Tri '};

根据上述三角形对，得到一个近似的公共子图；According to the above triangle pair, an approximate common subgraph is obtained;

S13：每个{Tr_i，Tr_i’}都有对应的{w_r，w’_r}，其中，

是

和

之间的距离；S13: Each {T_i ,_Tri '} has a corresponding {w_r , w'_r }, where,

Yes

and

the distance between;

通过程序验证所有质心，去除伪质心，将保留的质心作为新的树干分布图继续迭代；Verify all the centroids through the program, remove the false centroids, and continue to iterate with the retained centroids as the new trunk distribution map;

当迭代n和n+1中没有伪质心时，仍然有不少于3个质心点存在，则判定两个扫描是相邻的；When there are no pseudo centroids in iterations n and n+1, and there are still no less than 3 centroid points, it is determined that the two scans are adjacent;

其中，

则判定这对质心为伪质心。in,

Then the pair of centroids are determined to be pseudo-centroids.

在一个可选的实施例中，S2包括以下步骤：In an optional embodiment, S2 includes the following steps:

S21：在两个相邻的扫描中，预测一个初步的重叠区域，即对应的公共子图区域，并将其表示为

和

利用Semantic Keypoint 4-Points Congruent Sets 方法，进行扫描对的匹配；S21: In two adjacent scans, predict a preliminary overlapping region, that is, the corresponding common subgraph region, and denote it as

and

Use the Semantic Keypoint 4-Points Congruent Sets method to match scan pairs;

在

中选择四个非共面的点，即树干中心点，生成一个集合s，确定s中没有三个点共线；exist

Select four non-coplanar points, namely the trunk center point, to generate a set s, and determine that no three points in s are collinear;

s＝{a，b，c，d}；s={a,b,c,d};

其中，每个点利用特征向量a记录其语义；Among them, each point uses the feature vector a to record its semantics;

特征向量a为：a＝{x，y，z，Hdem，DBH}；The feature vector a is: a={x, y, z, Hdem, DBH};

S22：利用每个匹配的候选基准，计算出一个刚性变换T，并将

中所有的树干中心转换到

如下：S22: Using each matched candidate datum, a rigid transformation T is calculated, and the

Transform all trunk centers in to

as follows:

将

中的树干中心

利用刚性变换T得到：

Will

trunk center in

Using rigid transformation T to get:

在

中搜索距离

最近的点

在利用公式B、公式C和公式D进行判断，其中：公式B为：exist

medium search distance

nearest point

Using formula B, formula C and formula D for judgment, where: formula B is:

如果

if

公式C为：

Formula C is:

公式D为：Formula D is:

如果

if

R₂，表示不正确匹配的树木的百分比；R₂ , the percentage of incorrectly matched trees;

R₁，表示重叠区域之外树木所占百分比；R₁ , representing the percentage of trees outside the overlapping area;

其中，如果迭代没有终止，则记录得分最高的T来转换

进行粗匹配；Among them, if the iteration does not terminate, record the T with the highest score to convert

make a rough match;

S23：引入法向量的方向一致性来避免对齐树状点云产生局部最小问题；S23: Introduce the direction consistency of the normal vector to avoid the local minimum problem caused by aligning the tree point cloud;

应用树干模型来约束变换，以解决基于点的配准方法产生的破坏树干模型的问题，具体的约束条件如公式E：The trunk model is applied to constrain the transformation to solve the problem of destroying the trunk model generated by the point-based registration method. The specific constraints are as follows:

公式E为：Formula E is:

ROT和t分别表示旋转矩阵和平移矢量中的6个自由度；ROT and t represent the 6 degrees of freedom in the rotation matrix and translation vector, respectively;

(q’_i，q_i)表示相应的点；(q'_i , q_i ) represents the corresponding point;

(n’_i，n_i)表示其法向量(n'_i , n_i ) represents its normal vector

是二维空间中树干j的中心；

is the center of trunk j in two-dimensional space;

是两次扫描中的点对；

is the point pair in two scans;

p_j是树干j上，在

条件下和

相关的点；p_j is on trunk j, at

condition and

relevant points;

当ΔROT＜10^-6且Δt＜10^-4时，迭代终止，得到所求的旋转矩阵和平移矢量。When ΔROT<^10-6 and Δt<^10-4 , the iteration is terminated, and the required rotation matrix and translation vector are obtained.

在一个可选的实施例中，S3包括以下步骤：In an optional embodiment, S3 includes the following steps:

S31：应用最小环路拓展，合并扫描网络中的环路；当扫描网络中不再存在最小环路时，采用并行合并策略来合并剩余的扫描；S31: Apply the minimum loop extension to merge the loops in the scanning network; when the minimum loop no longer exists in the scanning network, adopt a parallel merging strategy to merge the remaining scans;

S32：并行合并策略基于最小生成树MST进行；MST边的权重由相邻扫描之间的重叠率来定义；S32: The parallel merging strategy is performed based on the minimum spanning tree MST; the weight of the MST edge is defined by the overlap ratio between adjacent scans;

S33：对于单中心扫描和侧向扫描的组合，并行合并策略将同时把所有侧向扫描配准到中心扫描上。S33: For the combination of single center scan and lateral scan, the parallel merge strategy will simultaneously register all lateral scans to the center scan.

本发明中，通过对FGI_Forest、JX_Forest_1、JX_Forest_2三个数据集的实验，对于包含较多扫描次数的数据集，RegisMUF相较于完全连通的配准方法减少了70％-80％的两两匹配次数，但配准的准确率没有降低，同时在点云数据存在遮挡的情况下配准网络的完整性也达到了90％以上。In the present invention, through experiments on three data sets of FGI_Forest, JX_Forest_1 and JX_Forest_2, for data sets containing more scan times, RegisMUF reduces the number of pairwise matching by 70%-80% compared to the fully connected registration method , but the accuracy of the registration does not decrease, and the integrity of the registration network also reaches more than 90% when the point cloud data is occluded.

本发明，通过测试验证了RegisMUF(多视图无序点云配准)的旋转和平移误差分别低于0.15°和0.1m。In the present invention, the rotation and translation errors of RegisMUF (multi-view unordered point cloud registration) are verified by tests to be lower than 0.15° and 0.1m, respectively.

本发明，将RegisMUF应用于FGI_Forest、JX_Forest_1和JX_Forest_2 数据集对RegisMUF的性能进行了评估；In the present invention, RegisMUF is applied to FGI_Forest, JX_Forest_1 and JX_Forest_2 data sets to evaluate the performance of RegisMUF;

FGI_Forest数据集包含了5次扫描结果，JX_Forest_1和JX_Forest_2数据集分别包含了15次和23次扫描结果。The FGI_Forest dataset contains 5 scans, and the JX_Forest_1 and JX_Forest_2 datasets contain 15 and 23 scans, respectively.

如果采用全连接的配准方式则三个数据集的扫描分别需要进行10次、105 次和253次两两匹配。通过应用RegisMUF来配准这三个数据集的点云分别进行了10次、25次和73次扫描的两两匹配；If the fully-connected registration method is adopted, the scans of the three datasets require 10, 105, and 253 pairwise matchings, respectively. By applying RegisMUF to register the point clouds of these three datasets, pairwise matching of 10, 25, and 73 scans was performed;

通过表1可以很明显的看出RegisMUF在提高多视图无序点云配准效率的同时保持了很好的配准准确度：From Table 1, it can be clearly seen that RegisMUF maintains a good registration accuracy while improving the registration efficiency of multi-view disordered point clouds:

表1：RegisMUF的最大公共子图引导网络性能Table 1: Maximum Common Subgraph Guided Network Performance of RegisMUF

需要说明的是，为了评估RegisMUF对扫描对间两两配准的性能，将RegisMUF与以下五种方法进行了比较：It should be noted that, in order to evaluate the performance of RegisMUF for pairwise registration between scan pairs, RegisMUF was compared with the following five methods:

(1)快速匹配剪枝分枝定界(FMP+BNB)；(1) Fast matching pruning branch and bound (FMP+BNB);

(2)快速全局配准(FGR)；(2) Fast global registration (FGR);

(3)博弈论方法(GTA)；(3) Game Theory Approach (GTA);

(4)关键点4点一致集(key-point 4PCS)；(4) Key-point 4-point consistent set (key-point 4PCS);

(5)随机一致性采样(RANSAC)。(5) Random Consistent Sampling (RANSAC).

通过比较三个数据集的旋转和平移误差，RegisMUF在这两个指标中都能达到足够准确的结果，其旋转和平移误差分别没有超过0.15°和0.1m，具有很好的稳定性。在配准精度方面本方法优于上述五种方法。By comparing the rotation and translation errors of the three datasets, RegisMUF can achieve sufficiently accurate results in these two metrics, and its rotation and translation errors do not exceed 0.15° and 0.1m, respectively, with good stability. In terms of registration accuracy, this method is superior to the above five methods.

FMP+BNB、RANSAC和FGR在计算稳定性和配准精度方面虽然略好于其他两种方法，但配准精度仍然不足以支持后续的重建工作。Although FMP+BNB, RANSAC and FGR are slightly better than the other two methods in terms of computational stability and registration accuracy, the registration accuracy is still not enough to support subsequent reconstruction work.

如表2所示，表2详细说明了RegisMUF在所有实验中的配准精度和时间性能。在森林环境中，RegisMUF提供的配准精度平均旋转误差为0.15°，平移误差为0.11m，虽然这些结果略逊于在城市应用中获得的结果，但它们足以进行树木的三维重建。As shown in Table 2, which details the registration accuracy and time performance of RegisMUF in all experiments. In a forest environment, RegisMUF provides registration accuracy with an average rotation error of 0.15° and translation error of 0.11m, although these results are slightly inferior to those obtained in urban applications, they are sufficient for 3D reconstruction of trees.

此外，在我们的测试中，RegisMUF在效率方面表现良好：分别需要21.3min、45.1min和101.7分钟来收敛FGI_Forest、JX_Forest_1和JX_Forest_2三个数据集的扫描网络。这是引导网络在无序情况下所起到的重要作用，它有效地剔除了完全连通情况下不可避免的错误扫描对(即没有重叠区域的扫描对)。Furthermore, in our tests, RegisMUF performs well in terms of efficiency: it takes 21.3min, 45.1min and 101.7min, respectively, to converge the scanning network for the three datasets FGI_Forest, JX_Forest_1 and JX_Forest_2. This is an important role played by the bootstrap network in the disordered case, which effectively eliminates the unavoidable false scan pairs (that is, scan pairs without overlapping regions) in the fully connected case.

表2：配准精度的定量评估Table 2: Quantitative evaluation of registration accuracy

表3比较了不同的收敛合并扫描策略，显示出了并行合并策略的潜力。 RegisMUF中的扫描网络收敛策略结合了MLE(最小环路拓展)和并行合并 (parallel)策略，因此我们在表3中将其表示为“MLE+P”Table 3 compares different convergent merge scan strategies, showing the potential of parallel merge strategies. The scanning network convergence strategy in RegisMUF combines MLE (minimum loop expansion) and parallel merging (parallel) strategy, so we denote it as "MLE+P" in Table 3

表3：端到端配准精度的定量评估Table 3: Quantitative evaluation of end-to-end registration accuracy

综上，通过上述对RegisMUF的性能进行的全面实验评估，结果表明， RegisMUF为森林环境中的多视图无序点云配准处理提供了可靠且准确的解决方法。In summary, through the above comprehensive experimental evaluation of the performance of RegisMUF, the results show that RegisMUF provides a reliable and accurate solution for multi-view unordered point cloud registration processing in forest environments.

应当理解的是，本发明的上述具体实施方式仅仅用于示例性说明或解释本发明的原理，而不构成对本发明的限制。因此，在不偏离本发明的精神和范围的情况下所做的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。此外，本发明所附权利要求旨在涵盖落入所附权利要求范围和边界、或者这种范围和边界的等同形式内的全部变化和修改例。It should be understood that the above-mentioned specific embodiments of the present invention are only used to illustrate or explain the principle of the present invention, but not to limit the present invention. Therefore, any modifications, equivalent replacements, improvements, etc. made without departing from the spirit and scope of the present invention should be included within the protection scope of the present invention. Furthermore, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims, or the equivalents of such scope and boundaries.

Claims (4)

1. The ground laser point cloud method for the forest scene is characterized by comprising the following steps:

s1: applying the MCS to correlate adjacent scans and predict likely overlapping regions;

s2: matching adjacent scanning pairs;

s3: all scans are merged to one anchor point.

2. The method of claim 1, wherein S1 comprises the steps of:

s11: obtaining a distribution map of the trunk: for a single tree, marking the breast diameter and the central position of the trunk as invariant features irrelevant to perspective, and performing hypercircular fitting on the trunk point cloud by using a formula A to obtain more accurate trunk attributes; fitting a trunk model through a formula A, and recording a trunk distribution diagram of a single scanning;

formula A:

wherein,

each trunk attribute is represented as: a ═ { DBH, Cpt2, Hdem };

where Cpt2 represents the location of the trunk center in two-dimensional space;

hdem represents the height of the trunk center DEM;

s12: searching K adjacent trunks of each trunk based on the trunk distribution diagram to generate a triangle; for each triangle, calculating a corresponding centroid;

feature vectors for each triangle

Identification, feature vectors

Comprises the following steps:

wherein A is_jIs the feature vector of trunk j;

d_ijis the feature vector of trunks i and j;

a_jrepresenting the corresponding angle of the trunk i;

is the centroid of triangle i in the first round;

comparing the point clouds PⁱAnd P^jEach triangle in (1) retrieves approximately congruent triangle pairs { Tr_i，Tr′_i}；

Obtaining an approximate public subgraph according to the triangle pair;

s13: each { Tr_i，Tr′_iAll have a corresponding { w }_r，w'_rAnd (c) the step of (c) in which,

is that

And

the distance between them;

verifying all centroids through a program, removing pseudo centroids, and continuously iterating the reserved centroids as a new trunk distribution map;

when no pseudo centroid exists in the iterations n and n +1, no less than 3 centroid points still exist, and two scans are judged to be adjacent;

wherein,

the pair of centroids is determined to be a pseudo centroid.

3. The method of claim 1, wherein S2 comprises the steps of:

s21: in two adjacent scans, a preliminary overlap region, i.e. the corresponding common sub-picture region, is predicted and represented as

And

matching scanning pairs by using a method of Semantic Keypoint 4-Points Congrent Sets;

in that

Selecting four non-coplanar points, namely, the central point of the trunk, generating a set s, and determining that no three points in the set s are collinear;

s＝{a，b，c，d}；

wherein, each point records the semanteme by using the feature vector a;

the feature vector a is: a ═ { x, y, z, Hdem, DBH };

s22: using each matched candidate basis, a rigid transformation T is calculated and

in which all trunk centers are switched to

The following were used:

will be provided with

Middle trunk center

Using the rigid transformation T, we obtain:

in that

Search for distance in

Nearest point

Judging by using a formula B, a formula C and a formula D, wherein: the formula B is:

if it is not

The formula C is:

the formula D is:

if it is not

R₂Represents the percentage of incorrectly matched trees;

R₁representing the percentage of trees outside the overlapping area;

wherein if the iteration is not terminated, the T with the highest score is recorded for conversion

Performing coarse matching;

s23: the direction consistency of the normal vector is introduced to avoid the local minimum problem caused by aligning the tree-shaped point cloud;

the stem model is applied to constrain the transformation to solve the problem of destroying the stem model generated by the point-based registration method, and the specific constraint conditions are as the following formula E:

the formula E is:

ROT and t represent 6 degrees of freedom in the rotation matrix and translation vector, respectively;

(q′_i，q_i) Representing the corresponding point;

(n′_i，n_i) Represents its normal vector

Is the center of trunk j in two-dimensional space;

is a point pair in two scans;

p_jis on the trunk j of

Under the conditions of

A point of relevance;

when Delta ROT < 10^-6And Δ t < 10^-4And then, terminating iteration to obtain the rotation matrix and the translation vector.

4. The method of claim 1, wherein S3 comprises the steps of:

s31: merging loops in the scanning network by applying minimum loop expansion; when the minimum loop does not exist in the scanning network, adopting a parallel combination strategy to combine the rest scanning;

s32: the parallel merging strategy is carried out based on a Minimum Spanning Tree (MST); the weight of the MST edge is defined by the overlap ratio between adjacent scans;

s33: for a combination of single center and side scans, a parallel merge strategy would register all side scans onto the center scan simultaneously.