CN101350101A - Automatic Registration Method of Multiple Depth Images - Google Patents

Abstract

A method for automatic registration of multiple depth images. (1) Use SIFT features to register any two depth images and judge the correctness of the results. First calculate the SIFT features of the two depth images, bidirectionally cross match the corresponding points, then use the RANSAC algorithm to find the epipolar constraints, filter the wrong match, and then use the ICP algorithm to accurately register and judge the correctness of the results (2) Search the circle space of the model map , to calculate a globally consistent registration result. Firstly, the derived circle base of the model graph is obtained and the adjacency graph of the derived circle base is constructed. Then, a set of bases of the consistent circle space is obtained, and the consistent registration result is obtained. This method can effectively increase the search speed of the circle space, and ideally, the exponential time complexity can be improved to the linear time complexity. (3) Consistency judgment of circles, using a relative error judgment method to judge the consistency of registration results in circles. The invention can reliably and automatically register multiple depth images, and removes wrong registration by searching the coincidence circle to obtain consistent registration results.

Description

Automatic Registration Method of Multiple Depth Images

technical field

The invention belongs to the technical field of computer virtual reality, and in particular relates to automatic registration of multiple depth images to obtain a complete model, and the method can be used for geometric modeling of a three-dimensional model.

Background technique

In recent years, the rapid development of 3D scanning technology makes automatic registration technology more and more important. A depth image is an image with depth information obtained after scanning by a 3D scanning device. Registration refers to the process of transforming depth images scanned from different positions and angles into the same coordinate system through rigid transformation. Through registration, multiple depth images are stitched together to obtain a complete 3D model. The registration calculation process can be mainly divided into registration of two depth images and registration of multiple depth images. The registration of two depth images is to calculate the relative positional relationship between the two images for registration; the registration of multiple depth images is also called global registration, and it is based on the registration of two depth images to remove the wrong pairwise registration. Alignment results, calculate the globally consistent registration results, and then optimize the overall error of the global registration.

The registration of two depth images mainly includes two steps of rough pairwise registration and precise pairwise registration. Accurate pairwise registration mainly uses the ICP (Iterative Closest Point) algorithm (see P.J.Besl, N.D.McKay.A Method for Registration of 3DShapes.IEEE Trans.Pattern Anal[J].and Machine Intell.Vol.14,pp.239- 256, 1992 and Y. Chen, G. Medioni. Object Modeling by Registration of Multiple Range Images [J]. IEEE Conference on Robotics and Automation, pp.2724-2729, 1991. and [Rusinkiewicz 2001] Rusinkiewicz, S., Levoy, M.Efficient Variants of the ICP Algorithm[J].In International Conference on 3D Digital Imaging and Modeling, 2001.), has achieved very good results, but because the ICP algorithm is locally convergent, before using the ICP algorithm, First, two-by-two rough registration is required to obtain a better initial position and avoid falling into local optimum. Pairwise rough registration mainly uses feature descriptors, such as rotating images, integrating spheres, etc., to match at least three pairs or more corresponding points, and then obtain rigid transformations.

Global matching is to judge the correctness of pairwise registration on the basis of pairwise registration, remove erroneous results, and obtain an overall consistent registration result. Huber (see Huber D. Automatic Three-dimensional Modeling from Reality. Carnegie Mellon University, 2002) proposed the first global matching algorithm. He uses the visibility standard to judge whether there is occlusion between the depth images, and then judges whether the registration is correct, and uses the Bayesian probability to obtain the reliability coefficient of each registration, and then uses the greedy method to generate a registration. The minimum spanning tree for the model. However, the overall error must be optimized before each occlusion judgment, and the occlusion calculation complexity is relatively high, so the calculation efficiency is relatively low. Moreover, at the beginning of spanning tree calculation, since only a few depth images are aggregated together, the reliability of occlusion judgment is also low. Hunag (see HuangQ X, Flory S, Gelfand N, Hofer M, Pottmann H. Reassembling fractured objects by geometric matching. ACM Transactions on Graphics (SIGGRAPH), 2006: 569-578) improved Huber's method, and he noticed two more There may be multiple pairwise registration relationships between the sub-models merged by depth images, which is more reliable than a single relationship, so he preferentially merges sub-images with more compatible registration relationships, rather than relying only on a single registration result Finding the minimum spanning tree, but this method still needs to optimize the total error multiple times, the calculation complexity is high, and the reliability is not high in the early stage of model merging. Note that multiple compatible registration relationships between sub-models constitute a loop in essence, and the pairwise registration results must satisfy the consistency constraint in the loop.

After global matching, the overall error needs to be optimized. Similar to the ICP algorithm, the error is gradually reduced through iteration, but the calculation is more complicated because multiple depth images are involved. Global registration algorithms are mainly divided into two categories, two-by-two optimization Bergevin and Pulli (see Bergevin R., Soucy M., Gagnon H., et al. Towards a General Multi-View Registration Technique. IEEE Trans. Pattern Anal. and Machine Intell , 1996, 18(5): 540-547 and Pulli K. Multiview Registration for Large Data Sets. International Conference on 3D Digital Imaging and Modeling) and Simultaneous Optimization of Krishnan and Neugebauer Algorithms (see Krishnan S., Lee P.Y., Moore J., etal.Global Registration of Multiple 3D Point Sets via Optimization-on-a-Manifold.Proc.ofSymposium on Geometry Processing, 2005: 187-196 and Neugebauer P.J.Geometrical Cloning of3D Objects via Simultaneous Registration of Proceeding-Crange[] ings of the 1997International Conference on Shape Modeling and Applications (SMA'97), 1997, 130). Pairwise optimization repeatedly uses the ICP algorithm to optimize two images, so the convergence speed is slow and may not be able to converge. At the same time, the optimization method directly optimizes the total error, and the convergence speed is fast, but the stability of complex optimization calculations is poor.

Contents of the invention

The technical problem of the present invention is to overcome the deficiencies of the prior art, and provide a method for automatic registration of multiple depth images, which has a fast registration speed and accurate matching for multiple depth images.

The technical solution of the present invention: a method for automatic registration of multiple depth images, including registration of two depth images and a global matching process of multiple depth images, wherein:

The registration of the two depth images is as follows:

(1) Calculate the SIFT feature on the two images separately, and obtain a 128-dimensional feature vector corresponding to each feature point;

(2) according to the feature vector described in step (1), match corresponding point;

(3) After determining the matching corresponding points, use the RANSAC algorithm to solve the basic matrix, and then use the basic matrix to eliminate the corresponding points of matching errors;

(4) sort the corresponding points after removing the matching errors by SIFT feature distance, and remove the 20% corresponding points with the largest distance between the feature vectors of the two images;

(5) According to the corresponding points of the two images, the transformation matrix is obtained by using the quaternion method, so as to obtain a rough registration in pairs;

(6) After the two-by-two rough registration, use the ICP algorithm to accurately register the two depth images;

The global matching process of multiple depth images is as follows:

(7) According to the two depth images obtained in step (6) that are accurately registered in pairs, a model diagram of multiple images is established;

(8) Obtain the derivation circle base of the model graph, set up the adjacency graph Γ of derivation circle base, described derivation circle base refers to the base that does not have the derivation circle formation of chord;

(9) Search the circle space, the circle space is the set of all loops in the model graph, when searching only the circles generated by the connected vertices in the adjacency relationship graph Г derived from the circle base, and when a consistent circle is obtained, then Delete a circle base until the number of vertices remaining in Г is less than the number of vertices to be searched, and the global matching of multiple images is obtained.

In the step (2), the two-way crossover method is used to match the corresponding points, and the steps are: when matching the corresponding points, for the feature point p on the first image, it has a feature vector, and all the features in the second depth image Find the feature point p′ closest to the feature vector of feature point p in the vector, and conversely for p′, find the feature point p″ closest to the feature vector of feature point p′ in all feature vectors of the first depth image, If p=p", then p and p' are matching corresponding points.

After the step (6), it is necessary to judge whether the pairwise registration results are correct. The judgment steps are as follows: first, when matching the corresponding points according to the SIFT feature, if the number of corresponding points matched successfully is less than the set threshold, then it is determined that the registration is not correct. Correct; secondly, when using the ICP algorithm for accurate registration, if the error is too large and greater than the set threshold, it is judged that the registration is incorrect; if the above two conditions cannot determine that the registration result is wrong, the result is considered correct.

The model graph G=<V, E> described in the step (7) is an undirected graph, V is a set of vertices, E is a set of edges, and each vertex v _i ∈ V represents a depth image, For each edge e _ij =(v _i , v _j )∈E, if and only if the pairwise registration result T _ij of depth images v _i and v _j exists, and T _ij is an attribute of edge e _ij , where v _i and v _j are the two vertices of the model graph.

The judging method for obtaining a consistent circle in the step (8): randomly pick a depth image from the circle, obtain its diameter d, transform the depth image along the pairwise transformation in the circle for a week, and obtain For a new depth image, calculate the error e between the new depth image and the corresponding point of the initial depth image, and use e/d as the error of the circle. If the error of the circle is less than the set threshold, the circle is consistent, otherwise, it is inconsistent of.

The advantage of the present invention compared with prior art is:

(1) The SIFT-based feature registration calculation speed of the present invention is fast, and the reliability of the cross-matching corresponding points is high, and the corresponding points are filtered by the RANSAC algorithm, which improves the reliability of the corresponding points to the greatest extent.

(2) The present invention sets up the model figure of multiple images, obtains the derivation cycle base of the model figure, establishes the adjacency relation graph Г of the derivation cycle base, and only searches for the connectivity in the adjacency relation graph Г of the derivation cycle base when searching the circle space When the circle generated by the vertices, and a consistent circle is obtained, a circle base is deleted until the number of vertices remaining in Г is less than the number of vertices to be searched, and the global matching of multiple images is obtained, which greatly improves the search speed. In this way, globally consistent registration results of multiple images can be obtained quickly and reliably.

(3) Moreover, the present invention further improves the reliability of multiple images by judging the consistency of the circles.

Description of drawings

Fig. 1 is the flowchart of the inventive method;

Fig. 2 is the schematic diagram of setting up model figure among the present invention;

Fig. 3 adopts the error distribution figure of 64 circles among the present invention;

Figure 4 is a partially enlarged view of Figure 3;

Fig. 5 is a disguised cave map of Maitreya's Lower Life Sutra for automatic registration using the method of the present invention.

Detailed ways

As shown in FIG. 1 , the present invention is divided into two depth image registration and multiple depth image global matching processes.

1. Registration of two depth images

Rough registration of two depth images is a process of coarse registration of partially overlapping two depth images at any position. The goal is to provide a good initial position for precise registration to ensure its convergence. Since the depth image contains color information, the present invention adopts an image corresponding point matching method for registration, and the steps are as follows:

(1) First, calculate SIFT features on two images respectively, each feature point corresponds to a 128-dimensional feature vector, and use Euclidean distance to measure the distance between two SIFT features; in this step, Lowe is used to calculate SIFT features, D. Distinctive image features from scale-invariant keypoints. Int. J. of Computer Vision 60, 2, 91-110.2004. The method is just introduced in the article.

(2) Use two-way cross-matching corresponding points, that is, when matching corresponding points, for the feature point p on the first image, it has a feature vector, and find the distance from the feature point p in all feature vectors of the second depth image The feature point p′ closest to the feature vector, conversely for p’, find the feature point p″ closest to the feature vector of the feature point p′ among all the feature vectors of the first depth image, if p=p″, then p And p' is the matching corresponding point.

(3) After matching, use the RANSAC algorithm to solve the basic matrix to eliminate error points. Use the RANSAC algorithm to solve the basic matrix. See Hartley, R.I., AND ZISSERMAN, A.2004.Multiple View Geometry in Computer Vision.Cambridge University Press, Cambridge, UK .

(4) Then, the corresponding points are sorted according to the SIFT feature distance, and the 20% corresponding points with the largest distance are removed.

(5) Finally, use the quaternion method to obtain the transformation matrix according to the corresponding points. The process of using the quaternion method to solve the transformation matrix can be found in Horn, B.Closed Form Solutions of Absolute Orientation Using Unit Quaternions.Journal of the Optical Society , Vol.4, 629-642, 1987.

(6) After the two-by-two rough registration, use the ICP (Iterative Closest Point) algorithm to accurately register the two depth images. ICP algorithm see P.J.Besl, N.D.McKay.A Method for Registration of 3D Shapes.IEEE Trans.Pattern Anal[J].and Machine Intell.Vol.14, pp.239-256, 1992. Each iteration is dynamically modified according to the error The distance threshold, thus obtaining the result of accurate registration of the two depth images.

(7) Perform pairwise registration consistency judgment on the precise registration results of the two depth images. Mismatching of feature points or partial overlap of two depth images may cause errors in pairwise registration results. Therefore, pairwise registration is to judge whether the result is correct or not. The judgment steps are as follows: first, when matching corresponding points according to SIFT features, if the number of corresponding points that match successfully is less than the set threshold, it is judged that the registration is incorrect; secondly, use When the ICP algorithm is accurately registered, if the error is too large and greater than the set threshold, it is judged that the registration is incorrect. If neither of the above two conditions can determine that the registration result is wrong, the result is considered correct, and the pairwise registration result is added to the model map of the subsequent global matching of multiple images.

2. Global matching of multiple depth images

(1) According to the obtained two depth images that are accurately registered in pairs, a model map of multiple images is established;

After two-two registration, the model graph G is obtained, but the result obtained only by the consistency judgment of two-two registration is unreliable. For example, due to the self-symmetry of the model, registration errors are caused, but it is difficult to Judge it. Therefore, after pairwise registration, a global consistency judgment is required. Since any loop of the model graph G should satisfy the consistency constraint, otherwise it is incorrect, so the registration result in a loop can be self-checked, and the present invention proposes a loop-based global matching method.

The model graph G=<V, E> is an undirected graph, each vertex v _i ∈ V represents a depth image, and the edge e _ij = (v _i , v _j )∈E if and only if the depth image v _i and The pairwise registration result T _ij of v _j exists, and T _ij is an attribute of edge e _ij . It is worth noting that any loop must satisfy the consistency constraint, that is, the total transformation should be a unit transformation. As shown in Figure 2, in the loop 16521, T ₁₂ ·T ₂₅ ·T ₅₆ ·T ₆₁ =I should be satisfied. A cycle is called consistent if it satisfies the consistency constraints, otherwise it is called inconsistent.

(2) Obtain the derived circle base of the model graph, and establish the adjacency graph Г of the derived circle base;

First a brief introduction to the concepts of graphs, circles and circle spaces. A graph G=<V,E>, where V=v(G) represents the set of vertices and E=ε(G) represents the set of edges. A path in which each vertex appears no more than once is called a basic path (Path). A basic path whose starting point and end point are the same vertex is called a cycle. An edge connecting two vertices in a circle, but it is not an edge in the circle, is called a chord. A cycle without strings is called an Induced Cycle. If graphs _G1 and _G2 have no common edges, then ${\mathrm{<figure-callout\; id="1"\; label="G"\; filenames="A20081022209500132.png,A20081022209500133.png"\; state="\{\{state\}\}">G</figure-callout>}}_1\&cup;{\mathrm{<figure-callout\; id="2"\; label="G"\; filenames="A20081022209500133.png"\; state="\{\{state\}\}">G</figure-callout>}}_2$ The edges called graphs _G1 and _G2 are not merged. A cycle means that the edges of the circle and the circle do not overlap. Obviously a circle is a cycle, and a circle is connected, but a cycle is not necessarily connected. The symmetric difference (Symmetric Difference) operation of graphs G ₁ and G ₂ is defined as ${\mathrm{<figure-callout\; id="1"\; label="G"\; filenames="A20081022209500132.png,A20081022209500133.png"\; state="\{\{state\}\}">G</figure-callout>}}_1\&CirclePlus;{\mathrm{<figure-callout\; id="2"\; label="G"\; filenames="A20081022209500133.png"\; state="\{\{state\}\}">G</figure-callout>}}_2=({\mathrm{<figure-callout\; id="1"\; label="G"\; filenames="A20081022209500132.png,A20081022209500133.png"\; state="\{\{state\}\}">G</figure-callout>}}_1\&cup;G_2)-\left({\mathrm{<figure-callout\; id="1"\; label="G"\; filenames="A20081022209500132.png,A20081022209500133.png"\; state="\{\{state\}\}">G</figure-callout>}}_1{\&cap;G}_2\right).$ The symmetric difference operation satisfies commutative and associative laws.

The set of all loops and empty sets of the connected graph G=<V, E> is C={C ₁ , C ₂ ..., C _n ), on which the symmetric difference operation and multiplication operation 0C _i =Φ are defined, 1·C _i ＝C _i , then it constitutes a number field A |E|-|V|+1-dimensional linear space on , denoted as C(G), is called a cycle space (Cycle Space). Any circle space can be derived from the circle base to generate the entire circle space.

(3) search circle space;

整个圈空间C(G)由

张成，因此C(G)的大小为2^|E|-|V|+1，是指数复杂度，搜索效率太低。模型图中所有一致圈构成了一个线性子空间，称为一致子空间。本发明提出一种快速求得一致子空间的一组基的方法，大大提高了圈空间的搜索效率，理想情况下是线性复时间杂度。Let the derived cycle base of the model graph G be

The entire circle space C(G) consists of

Zhang Cheng, so the size of C(G) is 2 ^|E|-|V|+1 , which is exponential complexity, and the search efficiency is too low. All the consistent cycles in the model diagram form a linear subspace, called the consistent subspace. The invention proposes a method for quickly obtaining a set of bases of the consistent subspace, which greatly improves the search efficiency of the circle space, and ideally has a linear complex time complexity.

边(B_i，B_j)∈L当且仅当ε(B_i)⌒ε(B_j)≠Φ。因此有以下结论：设 $C=B_{k_1}\&CirclePlus;B_{k_2}\&CirclePlus;\&CenterDot;\&CenterDot;\&CenterDot;\&CirclePlus;B_{k_n},$ 如果C是圈，那么在中

是连通的。注意由Г中连通的顶点所对应的圈基生成的环路也可能不是圈，因此还要检验其是不是圈。这样在搜索圈空间时，只要搜索由任意多个连通的基生成的所有圈，判断它们的一致性，这样可以大大减少搜索次数。Note that if the symmetric difference of two cycles is still a cycle, then the two cycles must have a common edge. In order to speed up the search, first define the adjacency graph of the circle base

Edge (B _i , B _j )∈L if and only if ε(B _i )⌒ε(B _j )≠Φ. Therefore, the following conclusions are drawn: $C={\mathrm{<figure-callout\; id="1"\; label="B\; k"\; filenames="A20081022209500132.png,A20081022209500133.png"\; state="\{\{state\}\}">B</figure-callout>}}_{k_{}}\&CirclePlus;{\mathrm{<figure-callout\; id="2"\; label="B\; k"\; filenames="A20081022209500133.png"\; state="\{\{state\}\}">B</figure-callout>}}_{k_{}}\&CirclePlus;\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;\&CirclePlus;B_{k_{\mathrm{no}}},$ If C is a circle, then in middle

is connected. Note that the cycle generated by the cycle base corresponding to the connected vertices in Γ may not be a cycle, so it is necessary to check whether it is a cycle. In this way, when searching the circle space, we only need to search all the circles generated by any number of connected bases and judge their consistency, which can greatly reduce the number of searches.

并且生成的邻接关系图采用导出圈基是因为它的长度较小，因而一致的可能性较大，有助于提高搜索效率。然后求S(G)的一组基，首先，判断每个圈C＝B_i是否一致，如果一致，则将C加入到S(G)的基中，并从Г中删除B_i；然后判断由任何两个连通的圈基生成的圈 $C=B_i\&CirclePlus;B_j$ 是否一致，如果一致，将C加入到S(G)的基中，并从Г中删除B_i；然后判断由任何三个连通的圈基生成的圈，这样下去，直到Г中剩余的顶点小于要搜索的连通顶点个数为止。然后，将所有一致圈的边合并得到一致模型图G′，详细算法在算法1中给出。该算法在最好情况下是线性复杂度，与指数复杂度相比，大大提高了计算效率。如果G′只有一个连通子图，则全局配准计算完成。如果G′有多个连通子图，则连通子图之间不再存在一致回路，本发明采用基于可见性的判断方法——自由空问冲突(Free Space Violation)继续全局配准[Huber 2002]，生成连通子图之间的最小生成树，以合并更多的子图。因为一般情况下，每个连通子图都含有多幅深度图像，所以，相对于单幅图像而言，可见性判断具有更高的可靠性。When searching for the coincidence circle, firstly, the derived circle base is obtained according to the model graph G of pairwise registration

and generate The adjacency graph of The reason for adopting the derived circle base is that its length is small, so the probability of consistency is high, which helps to improve the search efficiency. Then seek a set of bases of S(G), first, judge whether each circle C=B _i is consistent, if consistent, then add C to the base of S(G), and delete B _i from Г; then judge A cycle generated by any two connected cycle bases $C=B_i\&CirclePlus;B_j$ Whether it is consistent, if consistent, add C to the base of S(G), and delete B _i from Г; then judge the circle generated by any three connected circle bases, and so on, until the remaining vertices in Γ are less than The number of connected vertices to be searched. Then, merge the edges of all consistent circles to obtain the consistent model graph G′, and the detailed algorithm is given in Algorithm 1. The algorithm is linear complexity in the best case, which greatly improves computational efficiency compared with exponential complexity. If G′ has only one connected subgraph, the global registration calculation is complete. If G′ has multiple connected subgraphs, there is no consistent loop between the connected subgraphs, and the present invention adopts a judgment method based on visibility - Free Space Violation to continue global registration [Huber 2002] , to generate a minimum spanning tree between connected subgraphs to merge more subgraphs. Because in general, each connected subgraph contains multiple depth images, so compared with a single image, the visibility judgment has higher reliability.

Algorithm 1:

Input: pairwise registered model graph G Output: globally consistent model graph G'Procedure to find the derived cycle base B of G to establish the adjacency relationship graph of B Г<B, L>consistent_cycle←ΦN←1while the number of vertices in Г＞ ＝The circle generated by N connected vertices in Nfor each Г Cif C consistent then inserts C into consistent_cycle and deletes one of the N vertices. Add all the vertices of the consistent_cycle to G' and add all the edges of all cycles in consistent_cycle to G'

(4) Consistency judgment of the circle;

After obtaining a circle, it is necessary to judge whether it is consistent. A circle is consistent, which means that the total pairwise transformation in the circle approximates the unit transformation. Set the circle C=V ₁ V ₂ ... V _n V ₁ , take any depth image V _i , and obtain the total transformation in the loop from V _i to V _i T = T _{i, i+1} ... T _{n, 1} ... T _{i -1, i} , T _{i, j} represent the transformation matrix from V _j to V _i . The total transformation error in the calculation circle is $\mathrm{\&epsiv;}=\frac{1}{\mathrm{no}}\underset{k=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}\left|\right|p_k-{\mathrm{Tp}}_k\left|\right|,$ Where p _k ∈ _{V i} is the point in the depth image V _i , and n represents the number of points in V _i .

Let the diameter of the enclosing sphere of the depth image V _i be d. If ε<<d, then cycle C is consistent, otherwise cycle C is inconsistent. Defining r = ε/d as the error in the circle, r is scale invariant such that it is independent of the scaling transformation of the depth image. If r is less than the set threshold circle is consistent, otherwise, it is inconsistent.

(5) Test results;

When judging the consistency of circles, the error distribution of each loop is shown in Figures 3 and 4. It can be seen that there is an obvious break in the error distribution between 0.03 and 0.3. Therefore, the present invention sets the error threshold for judging the consistency circle to be 0.03 .

As shown in Figure 5, there are 81 depth images in the Dazu Rock Carvings Maitreya Sutra Disguised Phase Cave Model, including 22,477,248 points, with an average of about 277,000 points in each depth image. Using the same parameter settings as above, after a total of 3240 pairwise registrations, the obtained model graph contains 217 edges, so the dimension of the circle space is 137. When searching the circle space, a total of 140 circles were searched and the search was completed, and 124 consistent circles were obtained, that is, the consistent subspace was 124 dimensions. The obtained consistent model graph contained 184 edges and was connected, and the search circle space took 42.5 seconds. The entire registration calculation takes about 35 minutes. It can be seen that most of the time is used for pairwise registration to obtain a complete model.

In a word, the present invention can reliably and automatically register multiple depth images, remove wrong registration by searching the coincidence circle, and obtain consistent registration results.

The contents not described in detail in the description of the present invention belong to the prior art known to those skilled in the art.

The above is only a preferred embodiment of the present invention, it should be pointed out that, for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (6)

1. The method for automatic registration of multiple depth images, the process of registration of two depth images and the global matching process of multiple depth images, is characterized in that the steps are as follows: The registration of the two depth images is as follows: (1) Calculate the SIFT feature on the two images separately, and obtain a 128-dimensional feature vector corresponding to each feature point; (2) according to the feature vector described in step (1), match corresponding point; (3) After determining the matching corresponding points, use the RANSAC algorithm to solve the basic matrix, and then use the basic matrix to eliminate the corresponding points of matching errors; (4) sort the corresponding points after removing the matching errors by SIFT feature distance, and remove the 20% corresponding points with the largest distance between the feature vectors of the two images; (5) According to the corresponding points of the two images, the transformation matrix is obtained by using the quaternion method, so as to obtain a rough registration in pairs; (6) After the two-by-two rough registration, use the ICP algorithm to accurately register the two depth images; The global matching process of multiple depth images is as follows: (7) According to the two depth images obtained in step (6) that are accurately registered in pairs, a model diagram of multiple images is established; (8) Obtain the derivation circle base of the model diagram, establish the adjacency graph Γ of the derivation circle base, and the derivation circle base refers to the base formed by the derivation circle without chord; (9) Search circle space, said circle space is the set of all loops in the model graph, when searching, only search the circle generated by the connected vertices in the adjacency relationship graph Γ of the derived circle base, and when a consistent circle is obtained, then Delete a circle base until the number of remaining vertices in Γ is less than the number of vertices to be searched, and the global matching of multiple images is obtained. 2. The method for automatic registration of multiple depth images according to claim 1, characterized in that: in the step (2), a two-way crossover method is used to match corresponding points, and the steps are: when matching corresponding points, for the first image The feature point p on the , which has a feature vector, finds the feature point p′ closest to the feature vector of the feature point p in all the feature vectors of the second depth image, and conversely for p′, the first depth image Find the feature point p″ closest to the feature vector of the feature point p′ among all the feature vectors of , if p=p″, then p and p′ are matching corresponding points. 3. The method for automatic registration of multiple depth images according to claim 1, characterized in that: after the step (6), it is necessary to judge whether the pairwise registration results are correct, and the judgment step is: first, according to the SIFT feature When matching the corresponding points, if the number of corresponding points that match successfully is less than the set threshold, it is judged that the registration is incorrect; secondly, when using the ICP algorithm for precise registration, if the error is too large and greater than the set threshold, it is judged that the registration is incorrect; if If neither of the above two conditions can determine that the registration result is wrong, the result is considered correct. 4. The method for automatic registration of multiple depth images according to claim 1, characterized in that: the model graph G=<V, E> described in the step (7) is an undirected graph, and V is a vertex set, E is a set of edges, each vertex v _i ∈ V represents a depth image, each edge e _ij = (v _i , v _j ) ∈ E, if and only if two depth images v _i and v _j Two registration results T _ij exist, and T _ij is an attribute of edge e _ij , where v _i and v _j are two vertices of the model graph. 5. The method for automatic registration of multiple depth images according to claim 1, characterized in that: in the step (8), the judging method for obtaining a consistent circle: randomly select a depth image from the circle, and obtain Its diameter d transforms the depth image along the pairwise transformation in the circle to obtain a new depth image, calculates the error e between the new depth image and the corresponding point of the initial depth image, and takes e/d as The error of the circle, if the error of the circle is less than the set threshold, the circle is consistent, otherwise, it is inconsistent. 6. The method for automatic registration of multiple depth images according to claim 1, wherein the method for searching the circle space in the step (9) is as follows: Input: model graph G for pairwise registration Output: globally consistent model graph G' Procedure Find the derived cycle base B of G Build B's adjacency graph Г<B, L> consistent_cycle←Φ N←1 while the number of vertices in Γ>=N The circle C generated by N connected vertices in for each Γ ifC agrees then Insert C into consistent_cycle delete one of these N vertices If the number of vertices in Γ < N then break end if end if end for N←N+1 end while G’←Φ Add the vertices in G to G' Add all edges of all cycles in consistent_cycle to G'.

Images