CN115170626A - Unsupervised method for robust point cloud registration based on depth features - Google Patents
Unsupervised method for robust point cloud registration based on depth features Download PDFInfo
- Publication number
- CN115170626A CN115170626A CN202210847039.1A CN202210847039A CN115170626A CN 115170626 A CN115170626 A CN 115170626A CN 202210847039 A CN202210847039 A CN 202210847039A CN 115170626 A CN115170626 A CN 115170626A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- point
- tensor
- transformation
- multiplied
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
- 238000000034 method Methods 0.000 title claims abstract description 24
- 238000004364 calculation method Methods 0.000 claims abstract description 9
- 238000012549 training Methods 0.000 claims abstract description 9
- 239000000284 extract Substances 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 42
- 230000009466 transformation Effects 0.000 claims description 41
- 230000006870 function Effects 0.000 claims description 21
- 238000013135 deep learning Methods 0.000 claims description 16
- 230000000694 effects Effects 0.000 claims description 12
- 238000013519 translation Methods 0.000 claims description 10
- 238000013507 mapping Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 230000002776 aggregation Effects 0.000 claims description 5
- 238000004220 aggregation Methods 0.000 claims description 5
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 4
- 241000543540 Guillardia theta Species 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 238000003475 lamination Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000006116 polymerization reaction Methods 0.000 claims description 3
- 238000003786 synthesis reaction Methods 0.000 claims description 3
- 230000001131 transforming effect Effects 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 6
- 241000384062 Armadillo Species 0.000 description 1
- 102000016904 Armadillo Domain Proteins Human genes 0.000 description 1
- 108010014223 Armadillo Domain Proteins Proteins 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an unsupervised method for carrying out robust point cloud registration based on depth characteristics, which comprises the following steps: 1) Acquiring point cloud data; 2) Converting; 3) Extracting characteristics; 4) Point cloud registration; 5) And (5) training. The method is an unsupervised network, the registration framework is trained in an unsupervised mode by combining global and local high-level features to learn and extract the deep features, expensive calculation is not needed to be carried out on point correspondence, and great advantages are achieved in the aspects of precision, initialization robustness and calculation efficiency.
Description
Technical Field
The invention relates to object three-dimensional reconstruction and positioning, in particular to an unsupervised method for robust point cloud registration based on depth features.
Background
Point cloud registration is a problem of estimating a rigid transformation of the alignment of two point clouds. The method has many applications in various fields such as automatic driving, motion and attitude estimation, three-dimensional reconstruction, simultaneous localization and mapping (SLAM), augmented reality and the like.
Recently, some methods based on Deep Learning (DL) have been proposed to handle large rotation angles. (Y.Wang and J.M.Solomon, "Deep closed point: left prediction for point closed prediction," in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019, pp.3523-3532.) (Y.aoki, H.Goforth, R.A. Srivasan, and S.lucey, "Point: robust & impact point closed prediction Point," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp.7163-7172) (V.Sarode, X.Li, H.Goforth, Y.aoki, R.Srivan, S.lucency, H.Chongsis, "for point closed prediction" P.12, P.78. Echo, and "PcJ.12. Participant, P.78. Echo, P.11. Echo, P.12, P.S.communication, P.12, P.S.
Roughly speaking, they can be divided into two categories, supervised and unsupervised methods that rely on ground truth correspondences or class labels. The depth nearest Point (DCP) calculates rigid transformation by Singular Value Decomposition (SVD), wherein a correspondence is constructed by learning soft matching mapping. PointNet uses classical alignment techniques, such as the Lucas-Kanade (LK) algorithm, to align PointNet features, yielding good generalization capability for shapes not seen in training. However, they rely on a large amount of registration tag data, which makes the algorithm impractical because 3D registration tags are very labor intensive. In contrast, achieving registration from unlabeled point cloud data without ground truth correspondence is a significant challenge. PCRnet alleviates the attitude bias shown in PointNelk by replacing the Lucas-Kanade module with a multi-layer perceptron. The PCRNet recovers the transformation parameters directly from the concatenated global descriptors of the source point cloud and the target point cloud. FMR-Net implements an unsupervised framework using encoder-decoder tasks while achieving registration by minimizing feature metric projection errors. While these methods show the more prominent advantages of unsupervised learning, they rely primarily on globally represented depth features, and ignore locally represented depth features. Therefore, the depth features of the point cloud are not fully utilized for registration, and the registration effect is not perfect.
Disclosure of Invention
The invention aims to provide an unsupervised method for robust point cloud registration based on depth features, aiming at the defects in the prior art. The method is an unsupervised network, the registration framework is trained in an unsupervised mode by combining global and local high-level features to learn and extract the deep features, expensive calculation is not needed to be carried out on point correspondence, and great advantages are achieved in the aspects of precision, initialization robustness and calculation efficiency.
The technical scheme for realizing the purpose of the invention is as follows:
an unsupervised method for robust point cloud registration based on depth features comprises the following steps:
1) Acquiring point cloud data: acquiring point cloud data to be registered, and uniformly sampling 1024 point data on the surface of the point cloud;
2) Conversion: converting the obtained sampling data types of the two point clouds into tensors, wherein the size of the tensor is 1024 multiplied by 3, and then inputting the tensors into a deep learning framework;
3) Feature extraction: an encoder module in the deep learning framework extracts the point cloud depth features of the input tensor, and finally outputs a one-dimensional tensor representing the point cloud depth features, and the specific process is as follows:
3-1) for the input tensor 1024 × 3, by the EdgeConv module, each point is taken as a central point to represent the edge characteristics of the point and each adjacent point, and the characteristics are aggregated to obtain a new characteristic of the point, namely the local characteristics of the point cloud are obtained by constructing the field of each vertex, and the specific steps are as follows:
h θ (x i ,x j )＝h θ (x i ,x j -x i ) (1)，
wherein x i In vertex set X = { X = ×) 1 ,...x n }∈R F And F represents the feature space dimension information of a point output by a certain layer of the neural network, and then the feature space dimension information is sent to a perceptron to obtain edge features:
e' ijm ＝RELU(θ m ·(x j -x i )+φ m ·x i ) (2)，
wherein:
Θ＝(θ 1 ，...，θ M ，φ 1 ...，φ M ) (3)，
Θ is a learnable parameter, followed by aggregation of the features of each neighboring edge:
wherein
Denotes the polymerization operation, omega denotes the point x i The central point pair set forming the adjacent edge specifically aggregates the operation as follows:
finally, the size of the output tensor of the edgeConv module is 6 multiplied by 1024 multiplied by 20, and the next volume of lamination is input;
3-2) performing five convolutions on the tensor with the local features, wherein the output tensor size of the convolution layer after the first convolution is subjected to 64 convolution kernels with two-dimensional size of 1 is 64 multiplied by 1024 multiplied by 20;
3-3) inspired by the characteristic that an attention mechanism obviously improves the deep learning effect, the output result of the first layer is accessed into a CBAM attention module to weight the importance of each channel and space, and similarly, the output result sequentially passes through convolutional layers with the sizes of 64, 128 and 256;
3-4) splicing the output tensors of the first four layers, modifying the tensor size to be 512 multiplied by 1024 multiplied by 1, then entering 512 convolution kernels with the two-dimensional size of 1 to perform convolution and modify the dimensionality to obtain the tensor size to be 512 multiplied by 1024, flattening the tensor by using a flat function to output the tensor size to be 512, and thus obtaining the depth feature containing the local descriptor and the global descriptor;
4) Point cloud registration: registering two point clouds, i.e. finding a rigid transformation matrix G E SE (3) divided into a rotation matrix R E SO (3) and a translation matrix t E R 3 Then, the minimization objective function F (G (R, t)) is solved:
wherein psi: R 3×N →R K Extracting a function for the characteristic learned by the encoder, wherein k is a characteristic dimension;
4-1) G (R, t) is regarded as a special lie group and is represented by an exponential mapping as follows:
wherein, T is a generator of exponential mapping G, and epsilon is a lie algebra and is mapped to a transformation matrix G through an exponent;
4-2) the point cloud registration problem is described as Phi (PT) = phi (G.PT), the solving matrix G is the derivative of G obtained by deriving the lie algebra epsilon, so that the lie algebra epsilon is directly adjusted to indirectly optimize the optimal transformation matrix G and minimize the feature space projection error of the two-point cloud to carry out the point cloud registration task, and according to the inspiration of the (IC) formula, firstly carrying out inverse transformation on the Phi (PT) = phi (G.PT) and linearizing the right side of the transformation matrix:
4-3) transformation estimation an iterative run calculates the delta ε, estimating ε at each step by running the inverse synthesis (IC) algorithm:
ε＝(J T J) -1 (J T δ) (9)，
where δ = φ (P) S )-φ(P T ) Is a two-point cloud feature space projection error,
is a jacobian matrix of the projection error delta relative to the transformation parameter epsilon;
4-4) in order to effectively calculate the Jacobian matrix, different finite gradients are adopted to replace the traditional random gradient method to calculate the Jacobian matrix:
wherein t is i For infinite perturbation of the transformation parameters during calculation, when t i When the value is set to a fixed value, a relatively good effect is obtained, t i Three angular parameters are set for rotation and three perturbation parameters are set for translation, t i Size set to 2*e -2 ；
4-5) optimizing the transformation matrix by an iterative operation formula (9) to shorten the projection distance between the source point cloud and the sample point cloud in the feature space:
ΔG·P S →P S (11)，
wherein
Calculating a transformation parameter epsilon for formula (9) to obtain a transformation matrix, and continuously transforming P through multiple iterations S And P T Registering and finally outputting the optimal transformation matrix G est ：
G est ＝ΔG n ·...·ΔG 1 ΔG 0 (12)，
5) Training: in order to realize the deep learning framework to train in an unsupervised mode, a decoder consisting of four fully-connected layers is introduced; meanwhile, reLU is selected as an activation function of the first three layers, and the tanh function is an activation function of the fourth layer;
5-1) in an unsupervised condition, performing model training by using alignment errors between the transformed source point cloud and the target point cloud instead of ground true value transformation, and adopting a robust loss function:
wherein, P T E.g. theta is from the unit square [0,1] 2 A set of points sampled in (a), x is a point cloud feature,
is the ith component in the convolutional layer parameters and phi is the originally input three-dimensional point.
The method has high precision and calculation efficiency and strong initialization robustness, and does not need to perform expensive calculation on the point correspondence.
Drawings
FIG. 1 is a schematic diagram of a point cloud registration network architecture in an embodiment;
FIG. 2 is a schematic and effect diagram of an embodiment performing three-dimensional point cloud registration of bottles;
FIG. 3 is a schematic and effect diagram of an embodiment of three-dimensional point cloud registration of a toilet;
FIG. 4 is a schematic and effect diagram of the embodiment for three-dimensional point cloud registration of a noisy handset;
FIG. 5 is a schematic and effect diagram of the embodiment for performing three-dimensional point cloud registration of a real indoor scene;
fig. 6 is a schematic and effect diagram of an embodiment for point cloud registration of Armadillo in stanford 3d match.
Detailed Description
The invention will be described in further detail with reference to the following drawings and specific examples, but the invention is not limited thereto.
Example (b):
referring to fig. 1, an unsupervised method for robust point cloud registration based on depth features includes the following steps:
1) Acquiring point cloud data: acquiring point cloud data to be registered, and uniformly sampling 1024 point data on the surface of the point cloud;
2) Conversion: converting the obtained sampling data types of the two point clouds into tensors with the size of 1024 multiplied by 3, and then inputting the tensors into a deep learning framework;
3) Feature extraction: an encoder module in the deep learning framework extracts the point cloud depth features of the input tensor, and finally outputs a one-dimensional tensor representing the point cloud depth features, and the specific process is as follows:
3-1) different from other unsupervised methods relying on global descriptors, the method of the present embodiment continuously focuses on local features on the basis of an unsupervised training mode, and for an input tensor 1024 × 3, the edge features of each point and each adjacent point are represented by taking each point as a central point through an EdgeConv module, and then the features are aggregated to obtain a new feature of the point, that is, the local features of the point cloud are obtained by constructing the field of each vertex, and the specific steps are as follows:
h θ (x i ,x j )＝h θ (x i ,x j -x i ) (1)，
wherein x is i In vertex set X = { X = ×) 1 ,...x n }∈R F And F represents the feature space dimension information of a point output by a certain layer of the neural network, and then the feature space dimension information is sent to a perceptron to obtain edge features:
e' ijm ＝RELU(θ m ·(x j -x i )+φ m ·x i ) (2)，
wherein:
Θ＝(θ 1 ，...，θ M ，φ 1 ...，φ M ) (3)，
Θ is a learnable parameter, followed by aggregation of the features of each neighboring edge:
wherein
Denotes the polymerization operation, omega denotes the point x i The specific aggregation operation is as follows:
finally, the size of the output tensor of the edgeConv module is 6 multiplied by 1024 multiplied by 20, and the next volume of lamination is input;
3-2) performing five convolutions on the tensor with the local features, wherein the output tensor size of the convolution layer after the first convolution is subjected to 64 convolution kernels with two-dimensional size of 1 is 64 multiplied by 1024 multiplied by 20;
3-3) inspired by the characteristic that an attention mechanism obviously improves the deep learning effect, the output result of the first layer is accessed into a CBAM attention module to weight the importance of each channel and space, and the weighted result sequentially passes through convolution layers with sizes of 64, 128 and 256 similarly;
3-4) splicing the output tensors of the first four layers, modifying the tensor size to be 512 multiplied by 1024 multiplied by 1, then entering 512 convolution kernels with the two-dimensional size of 1 to perform convolution and modifying the dimensionality to obtain the tensor size to be 512 multiplied by 1024, and flattening the tensor by using a flatten function to obtain the depth feature including the local descriptor and the global descriptor, wherein the output tensor size is 512;
4) Point cloud registration: registering two point clouds, i.e. finding a rigid transformation matrix G E SE (3) divided into a rotation matrix R E SO (3) and a translation matrix t E R 3 Then, the minimization objective function F (G (R, t)) is solved:
wherein psi: R 3×N →R K Extracting a function for the characteristic learned by the encoder, wherein k is a characteristic dimension;
4-1) G (R, t) is regarded as a special lie group and is represented by an exponential mapping as follows:
wherein, T is a generator of exponential mapping G, and epsilon is a lie algebra and is mapped to a transformation matrix G through an exponent;
4-2) the point cloud registration problem is described as Phi (PT) = phi (G.PT), the solving matrix G is the derivative of G obtained by deriving the lie algebra epsilon, so that the lie algebra epsilon is directly adjusted to indirectly optimize the optimal transformation matrix G and minimize the feature space projection error of the two-point cloud to carry out the point cloud registration task, and according to the inspiration of the (IC) formula, firstly carrying out inverse transformation on the Phi (PT) = phi (G.PT) and linearizing the right side of the transformation matrix:
4-3) transformation estimation an iterative run calculates the delta ε, estimating ε at each step by running an inverse synthesis (IC) algorithm:
ε＝(J T J) -1 (J T δ) (9)，
where δ = φ (P) S )-φ(P T ) For two-point cloud feature space projection errors,
is the projection error delta relative to the transform
A Jacobian matrix of parameters ε;
4-4) in order to effectively calculate the Jacobian matrix, different finite gradients are adopted to replace the traditional random gradient method to calculate the Jacobian matrix:
wherein t is i For infinite perturbation of the transformation parameters during calculation, when t i When the value is set to a fixed value, a relatively good effect is obtained, t i Three angular parameters are set for rotation and three perturbation parameters are set for translation, t i Size set to 2*e -2 ；
4-5) optimizing the transformation matrix by an iterative operation formula (9) to shorten the projection distance between the source point cloud and the sample point cloud in the feature space:
ΔG·P S →P S (11)，
wherein
Calculating transformation parameter epsilon for formula (9) to obtain transformation matrix, and continuously transforming P through multiple iterations S And P T Registering, and finally outputting the optimal transformation matrix G est ：
G est ＝ΔG n ·...·ΔG 1 ΔG 0 (12)，
5) Training: in order to realize the deep learning framework to train in an unsupervised mode, a decoder consisting of four fully connected layers is introduced; meanwhile, reLU is selected as an activation function of the first three layers, the tanh function is an activation function of the fourth layer, and the number of layers is shown in FIG. 1;
5-1) under the unsupervised condition, using the alignment error between the transformed source point cloud and the target point cloud instead of the ground true value transformation to train the model, and adopting a robust loss function:
wherein, P T E.g. theta is from the unit square [0,1] 2 A set of points sampled in (b), x is a point cloud feature,
is the ith component in the convolutional layer parameters and phi is the originally input three-dimensional point.
In order to illustrate the effectiveness of the deep learning framework, the most widely used ModelNet40 is adopted as a pre-training set, the ModelNet40 is a data set which has 12311 CAD models and contains 40 object classes, the deep learning framework is trained in the front 20 classes of the ModelNet40,and tested in the last 20 categories. In the training and testing process, a ground truth value transformation matrix G is obtained by randomly generating a rigid transformation matrix G gt And meanwhile, enabling the source point cloud to pass through G to generate the target point cloud. The axis is arbitrarily selected, and the rotation component in G is set to be [0,45 ]]Within a range of degrees, the translation component is [ -0.5,0.5]In-range initialization with true rotational-translation matrix G gt And predicting a rotational-translation matrix G est Mean Square Error (MSE), root Mean Square Error (RMSE) and Absolute average Error (MAE) of the rotation component and the translation component are used as evaluation indexes, and the smaller the value of the MSE, the RMSE and the MAE, the higher the registration accuracy. Meanwhile, in the experiment, the measurement unit of the angle is degree, and the unit of the running registration time is second. The results are shown in the following table:
specific examples are shown in fig. 2-6.
Claims (1)
1. An unsupervised method for robust point cloud registration based on depth features is characterized by comprising the following steps:
1) Acquiring point cloud data: acquiring point cloud data to be registered, and uniformly sampling 1024 point data on the surface of the point cloud;
2) Conversion: converting the types of the obtained sampling data of the two point clouds into tensors with the size of 1024 multiplied by 3, and inputting the tensors into a deep learning frame;
3) Characteristic extraction: an encoder module in the deep learning framework extracts the point cloud depth features of the input tensor, and finally outputs a one-dimensional tensor representing the point cloud depth features, and the specific process is as follows:
3-1) for the input tensor 1024 × 3, by the EdgeConv module, each point is taken as a central point to represent the edge characteristics of the point and each adjacent point, and the characteristics are aggregated to obtain a new characteristic of the point, namely the local characteristics of the point cloud are obtained by constructing the field of each vertex, and the specific steps are as follows:
h θ (x i ,x j )＝h θ (x i ,x j -x i ) (1)，
wherein x i In vertex set X = { X = ×) 1 ,...x n }∈R F And F represents the feature space dimension information of a point output by a certain layer of the neural network, and then the feature space dimension information is sent to a perceptron to obtain edge features:
e' ijm ＝RELU(θ m ·(x j -x i )+φ m ·x i ) (2)，
wherein:
Θ＝(θ 1 ，...，θ M ，φ 1 ...，φ M ) (3)，
Θ is a learnable parameter, followed by aggregation of the features of each neighboring edge:
wherein
Denotes the polymerization operation, omega denotes the point x i Forming a point pair set of adjacent edges for the center, wherein the specific aggregation operation is as follows:
finally, the size of the output tensor of the edgeConv module is 6 multiplied by 1024 multiplied by 20, and the next volume of lamination is input;
3-2) performing five convolutions on the tensor with the local features, wherein the output tensor size of the convolution layer after the first convolution is subjected to 64 convolution kernels with two-dimensional size of 1 is 64 multiplied by 1024 multiplied by 20;
3-3) inspired by the characteristic that an attention mechanism obviously improves the deep learning effect, the output result of the first layer is accessed into a CBAM attention module to weight the importance of each channel and space, and the weighted result sequentially passes through convolution layers with sizes of 64, 128 and 256 similarly;
3-4) splicing the output tensors of the first four layers, modifying the tensor size to be 512 multiplied by 1024 multiplied by 1, then entering 512 convolution kernels with the two-dimensional size of 1 to perform convolution and modify the dimensionality to obtain the tensor size to be 512 multiplied by 1024, flattening the tensor by using a flat function to output the tensor size to be 512, and thus obtaining the depth feature containing the local descriptor and the global descriptor;
4) Point cloud registration: registering two point clouds, i.e. finding a rigid transformation matrix G E SE (3) divided into a rotation matrix R E SO (3) and a translation matrix t E R 3 Then, the minimization objective function F (G (R, t)) is solved:
wherein psi: R 3×N →R K Extracting a function for the characteristic learned by the encoder, wherein k is a characteristic dimension;
4-1) G (R, t) is regarded as a special lie group and is represented by an exponential mapping as follows:
wherein, T is a generator of exponential mapping G, epsilon is a lie algebra and is mapped to a transformation matrix G through an exponent;
4-2) the point cloud registration problem is described as Phi (PT) = phi (G · PT), solving the matrix G, namely obtaining the derivative of G by deriving the lie algebra epsilon, so as to directly adjust the lie algebra epsilon to indirectly optimize the optimal transformation matrix G and minimize the feature space projection error of the two-point cloud, and as inspired by the (IC) formula, firstly performing inverse transformation on Phi (PT) = phi (G · PT) and linearizing the right side thereof:
4-3) transformation estimation an iterative run calculates the delta ε, estimating ε at each step by running an inverse synthesis (IC) algorithm:
ε＝(J T J) -1 (J T δ) (9)，
where δ = φ (P) S )-φ(P T ) Is a two-point cloud feature space projection error,
is a jacobian matrix of the projection error delta relative to the transformation parameter epsilon;
4-4) in order to effectively calculate the Jacobian matrix, different finite gradients are adopted to replace the traditional random gradient method to calculate the Jacobian matrix:
wherein t is i For infinite perturbation of the transformation parameters during calculation, when t i When the value is set to a fixed value, a relatively good effect is obtained, t i Three angular parameters are set for rotation and three perturbation parameters are set for translation, t i Size set to 2*e -2 ；
4-5) optimizing the transformation matrix by an iterative operation formula (9) to shorten the projection distance between the source point cloud and the sample point cloud in the feature space:
ΔG·P S →P S (11)，
wherein
Calculating a transformation parameter epsilon for formula (9) to obtain a transformation matrix, and continuously transforming P through multiple iterations S And P T Registering, and finally outputting the optimal transformation matrix G est ：
G est ＝ΔG n ·...·ΔG 1 ΔG 0 (12)，
5) Training: in order to realize the deep learning framework to train in an unsupervised mode, a decoder consisting of four fully-connected layers is introduced; meanwhile, reLU is selected as an activation function of the first three layers, and the tanh function is an activation function of the fourth layer;
5-1) under the unsupervised condition, using the alignment error between the transformed source point cloud and the target point cloud instead of the ground true value transformation to train the model, and adopting a robust loss function:
wherein, P T E.g. theta is from the unit square [0,1] 2 A set of points sampled in (a), x is a point cloud feature,
is the ith component in the convolutional layer parameters and phi is the originally input three-dimensional point.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210847039.1A CN115170626A (en) | 2022-07-07 | 2022-07-07 | Unsupervised method for robust point cloud registration based on depth features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210847039.1A CN115170626A (en) | 2022-07-07 | 2022-07-07 | Unsupervised method for robust point cloud registration based on depth features |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115170626A true CN115170626A (en) | 2022-10-11 |
Family
ID=83494865
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210847039.1A Withdrawn CN115170626A (en) | 2022-07-07 | 2022-07-07 | Unsupervised method for robust point cloud registration based on depth features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115170626A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116188543A (en) * | 2022-12-27 | 2023-05-30 | 中国人民解放军61363部队 | Point cloud registration method and system based on deep learning unsupervised |
-
2022
- 2022-07-07 CN CN202210847039.1A patent/CN115170626A/en not_active Withdrawn
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116188543A (en) * | 2022-12-27 | 2023-05-30 | 中国人民解放军61363部队 | Point cloud registration method and system based on deep learning unsupervised |
| CN116188543B (en) * | 2022-12-27 | 2024-03-12 | 中国人民解放军61363部队 | Point cloud registration method and system based on deep learning unsupervised |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112837356B (en) | WGAN-based non-supervision multi-view three-dimensional point cloud joint registration method | |
| CN110533722B (en) | Robot rapid repositioning method and system based on visual dictionary | |
| Ranftl et al. | Deep fundamental matrix estimation | |
| CN107229757B (en) | Video retrieval method based on deep learning and Hash coding | |
| CN107169117B (en) | Hand-drawn human motion retrieval method based on automatic encoder and DTW | |
| CN112819853B (en) | Visual odometer method based on semantic priori | |
| CN115170626A (en) | Unsupervised method for robust point cloud registration based on depth features | |
| CN109919215B (en) | Target detection method for improving characteristic pyramid network based on clustering algorithm | |
| CN110097499B (en) | Single-frame image super-resolution reconstruction method based on spectrum mixing kernel Gaussian process regression | |
| CN111260702A (en) | Laser three-dimensional point cloud and CT three-dimensional point cloud registration method | |
| Chen et al. | An improved iterative closest point algorithm for rigid point registration | |
| Jain et al. | Robustifying the multi-scale representation of neural radiance fields | |
| Li et al. | Sparse-to-local-dense matching for geometry-guided correspondence estimation | |
| Trottier et al. | Sparse dictionary learning for identifying grasp locations | |
| Mei | Point cloud registration with self-supervised feature learning and beam search | |
| Li et al. | Unsupervised partial point set registration via joint shape completion and registration | |
| CN111832399A (en) | Attention mechanism fused cross-domain road navigation mark registration algorithm | |
| CN112396089B (en) | Image matching method based on LFGC network and compression excitation module | |
| CN112784800B (en) | Face key point detection method based on neural network and shape constraint | |
| Li et al. | Few-shot meta-learning on point cloud for semantic segmentation | |
| Wang et al. | Lgvit: Local-global vision transformer for breast cancer histopathological image classification | |
| Zhang et al. | RecFRCN: Few-Shot Object Detection With Recalibrated Faster R-CNN | |
| CN112906432A (en) | Error detection and correction method applied to human face key point positioning task | |
| CN111353509A (en) | Key point extractor generation method of visual SLAM system | |
| Wenzhi et al. | FVLoc-NeRF: Fast Vision-Only Localization within Neural Radiation Field |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20221011 |