CN111783838A - A point cloud feature space representation method for laser SLAM - Google Patents

Abstract

The invention discloses a point cloud feature space characterization method for laser SLAM, which is used for establishing bidirectional mapping between a point cloud space and a point cloud feature space based on a deep learning network, wherein the feature space realizes key problems of closed-loop detection, relocation and the like in the laser SLAM and map compression storage and transmission. The method comprises the steps of realizing global description feature extraction and compression reconstruction of large scene point clouds through a neural network of a self-encoder structure, designing a characteristic space of the point clouds formed by extracting global description features through the encoder network, giving similarity measurement of scenes according to the distance in the characteristic space, judging whether two or more scene structures are similar, and realizing closed-loop detection and relocation of laser SLAM; reconstructing the point cloud through a designed decoder network, reconstructing an original point cloud from global description features extracted from an encoder network, and realizing compressed storage and low-bandwidth transmission of a point cloud map; the constructed encoder network does not need to be trained in advance according to the point cloud map, and has strong generalization capability.

Description

A point cloud feature space representation method for laser SLAM

technical field

The invention relates to the technical field of laser point cloud mapping and autonomous navigation, in particular to a point cloud feature space representation method for laser SLAM.

Background technique

How to make mobile robots better understand and perceive the surrounding environment and achieve flexible and reliable high-level autonomous navigation, and artificial intelligence technology promoted by deep learning has brought rapid development for mobile robots to understand the environment and autonomous navigation, especially lidar The application of this method enables the robot to directly perceive the 3D environment, but it also brings challenges to the point cloud data processing. Simultaneous Localization And Mapping (SLAM) is one of the basic and key technologies for mobile robots to achieve autonomous navigation and positioning. It aims to establish a local map and determine the robot's position in the map when the robot enters an unknown environment . SLAM mainly includes four parts: front-end odometer, back-end pose graph optimization, closed-loop detection and map construction, and closed-loop detection is one of the key components, which solves the problem of determining whether the mobile robot has returned to the previously visited environment. , Accurately detecting the closed-loop scene is crucial for the SLAM system to correct the long-term cumulative error and build a globally consistent pose estimation and map, and because the drift of the state estimation of the global pose information is unavoidable, in the robot Reliable closed-loop detection is very important, especially for large outdoor scenes. Large-scale position drift and environmental recognition are a major challenge for SLAM systems.

Among them, how to make the robot recognize the environment is an important problem. The closed-loop detection is also a scene recognition problem in nature. Through sensor data, such as cameras, lidars, etc., the environment is recognized and relocated. Scene recognition has always been a hot and difficult issue in the field of pattern recognition, and cameras need to overcome the adverse effects of lighting, perspective and other factors on scene recognition and extract useful information. Compared with cameras, lidar can overcome the influence of light, season, weather, etc., but there are many kinds of objects in the laser point cloud scene data, including a large number of dynamic scenes, even the same scene will be affected by various external factors at different times. There is a big difference, which makes the problem of scene recognition very complicated, and the key of scene recognition is how to find a global description feature that can fully represent the structural information and semantic information of the same type of scene, especially for large-scale laser point clouds. feature characterization method.

Traditional point cloud-based environment recognition algorithms such as bag-of-words (BoW), FV (Fisher Vector), and VLAD usually rely on global, offline, and high-resolution maps, and it is necessary to train a codebook based on the map in advance to achieve high-precision positioning. Segmatch based on deep learning proposes a scene recognition method based on 3D segmentation and matching, which performs local matching through segmented scene blocks, and ensures the reliability of local matching and scene recognition through geometric consistency check. The fast static assumption cannot be satisfied. PointnetVLAD proposes a new aggregation method, which combines PointNet and NetVLAD. The former is used for feature learning and the latter is used for feature aggregation, which can learn to obtain global description features. The PCAN network based on point cloud deep learning is also considered on the basis of PointNetVLAD. The contribution degree of different local features is introduced, and the self-attention mechanism is introduced to learn the contribution weights of different local features during feature aggregation through the network, but the above methods also do not consider local geometric features and semantic context features, point cloud neighborhood relationship and feature space. distribution, and the effect is not good in the encoding of large scene point cloud scene recognition features.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a point cloud feature space representation method for laser SLAM, which can extract the point cloud global description features representing the scene structure information and semantic information from the complex large scene laser point cloud, and extract the point cloud global description features from the complex large scene laser point cloud. The feature restores and reconstructs the original point cloud for closed-loop detection, relocation, and compressed storage and transmission of laser SLAM.

In order to solve the above technical problems, the present invention provides a point cloud feature space characterization method for laser SLAM, which includes the following steps:

s1. Perform filtering and downsampling preprocessing on the point cloud samples in the point cloud data set, and construct a training point cloud sample pair according to the similarity of the point cloud scene, including a positive sample p _pos and a negative sample p _neg ;

s2. Build an encoder network, input the training point cloud sample pair in step s1, and obtain the global description feature f of the input point cloud P through the point cloud adaptive neighborhood feature extraction module, the dynamic graph network aggregation module, and the global description feature generation module. P);

s3. Build a decoder network, input the global description features generated in step s2, and use a hierarchical generation strategy to generate key point contours to generate refined and dense reconstructed point clouds.

s4. Take the sample pair of the training data set in s1, train the encoder network of s2, and obtain the encoder network model and parameters that can measure the similarity of the point cloud scene;

s5. Take the sample pairs of the training data set in s1 and the corresponding global description features generated by the s2 encoder network, train the decoder network of s3, and obtain a decoder network model that can restore and reconstruct the original point cloud from the global description features and parameters;

s6. In the actual SLAM operation, take the encoder network model and parameters in s4, extract the global description feature for the point cloud key frame, and perform NN matching with the global description feature of the historical key frame according to the Euclidean distance, and the matching condition is taken as Closed-loop candidate frame or relocation candidate position;

s7. In the actual SLAM operation, store or transmit the global description features of the point cloud keyframes extracted by s6. The global description features can be restored and reconstructed to the original SLAM point cloud map through the decoder network model and parameters in s5;

s8. In the actual SLAM operation, take the global description features extracted in s6, and perform NN matching with the global description features of the historical key frames of multiple collaborative mapping robots according to the Euclidean distance, through the point cloud feature space representation. The method performs multi-robot laser SLAM collaborative mapping task in point cloud feature space.

Preferably, in step s1, the similarity of the constructed training sample pair is determined by the auxiliary judgment of the position coordinates of the samples in the map. The position interval within 10m is regarded as a positive sample with similar structure, and the distance between 10m and 50m is regarded as a negative sample Select the range, randomly select negative samples, and randomly select samples beyond 50 meters as extremely dissimilar negative samples.

利用邻域点云的协方差矩阵Σ的特征值λ₁≥λ₂≥λ₃≥0计算点云邻域的线性

平面性

和散射性

特征，并通过香农信息熵度量方程E_k自适应的确定最优邻域大小

通过多层感知器特征提取网络提取邻域特征作为点的特征；Preferably, in step s2, the point cloud adaptive neighborhood feature extraction module of the encoder network constructs the kNN neighborhood of each point p _i in the point cloud with the range [k _min ,k _max ]

Using the eigenvalue λ ₁ ≥λ ₂ ≥λ ₃ ≥0 of the covariance matrix Σ of the neighborhood point cloud to calculate the linearity of the point cloud neighborhood

flatness

and scattering

feature, and adaptively determine the optimal neighborhood size through the Shannon information entropy metric equation E _k

Extract neighborhood features as point features through a multi-layer perceptron feature extraction network;

in:

V为特征向量矩阵，E_k＝-L_λlnL_λ-P_λlnP_λ-S_λlnS_λ。

V is an eigenvector matrix, E _k =-L _λ lnL _λ -P _λ lnP _λ -S _λ lnS _λ .

Preferably, in step s2, the dynamic graph network aggregation module is composed of two parts:

(a) Dynamic graph network aggregation in feature space: In step (3), a kNN with fixed k is used to construct a dynamic graph, and feature aggregation is performed through the graph network to update the features of the points; the feature aggregation process of the graph network is: f _i '=ρ(φ _Θ (ε(fi , _{f j )} )=ρ(φ _Θ (fi ‖(f _j -f _{i )} )), φ _Θ is the multilayer perceptron MLP, and ρ is max pooling;

(b) Network aggregation of dynamic graphs in physical space: construct a dynamic graph with kNN of fixed k at all points in the physical space of the point cloud, and use the features of the updated points in the above (a) to update the features through the graph network, and update the features of the points; The dynamic graph G=(V, E) constructed by the center point pi is composed of node V and edge E. The feature of point _pi is f _i , the node V is the feature _{f j of} the neighborhood point, and the edge E is the center point and the neighbor point. The relationship between domain point features ε(f _i , f _j ), where: ε(fi , f _j )=f _i ‖(f _j -f _{i )} , ‖ is the feature splicing by channel.

Preferably, in step s2, the global description feature generation module generates the global description feature through the NetVLAD Layer.

Preferably, in step s3, the hierarchical generation strategy includes two parts:

(a) Keypoint contour generation network: Generate rough keypoint contours of points through a three-layer fully connected network and Reshape;

并按通道拼接所述步骤s2中生成的全局描述特征，通过MLP生成稠密重建点云；(b) Refinement of dense point cloud generation network: a 2D planar grid of η × η points is stitched at each point in the key points generated in (a) above

And splicing the global description features generated in step s2 according to the channel, and generating dense reconstructed point clouds through MLP;

in:

η is the number of points on the edge of the plane grid, and ξ is the size of the plane grid points.

半监督训练使得输入样本与正样本p_pos之间全局描述特征的欧式距离δ_pos更近，与负样本p_neg之间的全局描述特征的欧氏距离δ_neg更远，极负样本

对用于防止负样本对之间的全局描述特征距离错误的拉近，极负样本与负样本的欧式距离为

其中：

α,β为超参数。Preferably, in step s4, use the metric learning method to train the encoder network, and use the quadruple loss function

Semi-supervised training makes the Euclidean distance δ _pos of the global description feature between the input sample and the positive sample p _pos closer, and the Euclidean distance δ _neg of the global description feature between the negative sample p _neg is farther, and the extremely negative sample

To prevent the error of the global description feature distance between pairs of negative samples, the Euclidean distance between extremely negative samples and negative samples is

in:

α and β are hyperparameters.

约束重建点云与真值点云的欧氏距离，其中

应用在步骤s2的(a)生成的关键点轮廓，用于约束点云的物理空间分布，

应用于稠密重建点云；Preferably, in step s5, the decoder network is trained in a supervised manner, using the loss function

Constrains the Euclidean distance between the reconstructed point cloud and the ground truth point cloud, where

The keypoint contours generated in (a) of step s2 are applied to constrain the physical spatial distribution of the point cloud,

applied to densely reconstructed point clouds;

in:

ψ(p)为

中p的对应点，

ψ(p) is

The corresponding point of p in ,

Preferably, in steps s6 and s8, the NN matching of the global description feature of the point cloud key frame is determined by matching the sequence of N consecutive current point cloud frames with N consecutive historical point cloud frames. If the consecutive sequence matches the key frame If the Euclidean distance between the global description features is less than the set threshold and the gradient change compared with other historical keyframe feature distances is greater than the set threshold, it can be regarded as a closed-loop detection candidate sequence keyframe or relocation candidate position; Provide the current continuous point cloud frame and historical key frame, then it is multi-robot collaborative mapping; laser SLAM closed-loop task, relocation task and multi-robot collaborative task are carried out in the feature space of point cloud through the point cloud feature space representation method.

Preferably, in step s7, the encoder network described in s4 is used to extract the global description features of the key frames in the laser point cloud map as compression codes for storage or communication transmission, and then the decoder network described in s5 is used to store or transmit and receive the features. The features are restored, and the original laser point cloud map is reconstructed. The described point cloud feature space representation method can be applied to embedded systems or low bandwidth situations.

The beneficial effects of the present invention are as follows: (1) An encoder network is designed to extract the information of geometric structure features and semantic context features for the laser point cloud data of large scenes, fully excavate the scene information of the point cloud, and generate the global description feature to form the point cloud and the distance in this feature space gives the similarity measure of the scene, which is used to judge whether two or more scene structures are similar, and the closed-loop detection and relocation of laser SLAM can be realized in the point cloud feature space; ( 2) Reconstruct the point cloud through the designed decoder network, reconstruct the original point cloud from the global description features extracted from the encoder network, and realize the compressed storage and low-bandwidth transmission of the point cloud map; (3) The constructed encoder network Road does not need to be trained according to the point cloud map in advance, and has a strong generalization ability.

Description of drawings

FIG. 1 is a schematic flow chart of the method of the present invention.

FIG. 2 is a schematic structural diagram of an encoder network model of the present invention.

FIG. 3 is a schematic structural diagram of a decoder network model of the present invention.

Detailed ways

As shown in Figure 1, a point cloud feature space representation method for laser SLAM includes the following steps:

Step 1: Perform preprocessing such as filtering and downsampling on the point cloud samples in the point cloud data set, and construct training point cloud sample pairs according to the similarity of the point cloud scene, including the positive sample p _pos and the negative sample p _neg .

The similarity of the constructed training sample pair is determined by the auxiliary judgment of the position coordinates of the samples in the map. The position interval within 10m is regarded as a positive sample with similar structure, and the distance between 10m and 50m is regarded as a negative sample selection range, and a negative sample is randomly selected. , and randomly selected samples at a distance of 50 meters are regarded as extremely dissimilar negative samples.

Specifically: select the Oxford RobotCar dataset, use a pass-through filter to set the point cloud range to [x:±30m, y:±20m, z:0-10m], and use a random voxel filter to downsample the point cloud to 4096 points and normalized to [-1～1].

Step 2: The network structure shown in Figure 2 is designed to construct the encoder network, input the training point cloud sample pair of step 1, and use the point cloud adaptive neighborhood feature extraction module, dynamic graph network aggregation module, and global description feature generation module Obtain the global description feature f(P) of the input point cloud P.

利用邻域点云的协方差矩阵Σ的特征值λ₁≥λ₂≥λ₃≥0计算点云邻域的线性

平面性

和散射性

特征，并通过香农信息熵度量方程E_k自适应的确定最优邻域大小

通过MLP(多层感知器)特征提取网络提取邻域特征作为点的特征。The point cloud adaptive neighborhood feature extraction module of the encoder network builds a kNN neighborhood of each point pi in the point cloud in the range [ _{k min} ,k _max ]

Using the eigenvalue λ ₁ ≥λ ₂ ≥λ ₃ ≥0 of the covariance matrix Σ of the neighborhood point cloud to calculate the linearity of the point cloud neighborhood

flatness

and scattering

feature, and adaptively determine the optimal neighborhood size through the Shannon information entropy metric equation E _k

Neighborhood features are extracted as point features through an MLP (Multilayer Perceptron) feature extraction network.

in:

V为特征向量矩阵

V is the eigenvector matrix

E _k = -L _λ lnL _λ -P _λ lnP _λ -S _λ lnS _λ

Specifically: set the point cloud adaptive neighborhood feature range [20-100], the step size is 10, and calculate to determine the optimal neighborhood size.

The dynamic graph network aggregation module consists of two parts:

(a) Dynamic graph network aggregation in feature space: kNN with fixed k is constructed at all points in the feature space of claim 3 to construct a dynamic graph, and feature aggregation is performed through the graph network to update the features of the points;

(b) Physical space dynamic graph network aggregation: A kNN with a fixed k is used to construct a dynamic graph at all points in the physical space of the point cloud, and the features of the updated points (a) above are used to update the features through the graph network to update the features of the points. The dynamic graph G=(V, E) constructed by the center point pi is composed of node V and edge E. The feature of point _pi is f _i , the node V is the feature _{f j of} the neighborhood point, and the edge E is the center point and the neighbor point. The relation ε(f _i ,f _j ) of the domain point features.

Among them: ε(f _i , f _j )=fi ‖(f _j -f _{i )} , ‖ is the feature splicing by channel;

The feature aggregation process of the graph network is:

f _i '=ρ(φ _Θ (ε(fi , _{f j )} )=ρ(φ _Θ (fi ‖(f _j -f _{i )} )), φ _Θ is MLP (multi-layer perceptron), and ρ is max pooling.

Specifically: kNN with fixed k=20 is used for dynamic graph construction, and φ _Θ is an MLP with two layers (64,64).

Preferably, the global description feature generation module generates the global description feature through the NetVLAD Layer.

Specifically: NetVLAD Layer selects the central feature dimension D=256, and selects K=64 central features in total.

The global description feature generation module generates global description features through the NetVLAD Layer.

Step 3: The network structure design shown in Figure 3 is used to construct the decoder network, input the global description features generated in step 2, and through the hierarchical generation strategy, from generating key point contours to generating refined and dense reconstructed point clouds

The hierarchical generation strategy consists of two parts:

(a) Keypoint contour generation network: Generate rough keypoint contours of points through a three-layer fully connected network and Reshape;

并按通道拼接所述步骤2中生成的全局描述特征，通过MLP(多层感知器)生成稠密重建点云。(b) Refinement of dense point cloud generation network: a 2D planar grid of η × η points is stitched at each point in the key points generated in (a) above

And the global description features generated in step 2 are spliced by channel, and the dense reconstructed point cloud is generated by MLP (multi-layer perceptron).

in:

η为平面网格边上点的数量，ξ为平面网格点的尺寸。

η is the number of points on the edge of the plane grid, and ξ is the size of the plane grid points.

Specifically: the key point contour generation network generates a rough contour containing 4096 key points, and splices a 2×2 point plane network with a size of ξ=0.05.

Step 4: Take the sample pair of the training data set in Step 1, train the encoder network in Step 2, and obtain the encoder network model and parameters that can measure the similarity of the point cloud scene.

半监督训练使得输入样本与正样本p_pos之间全局描述特征的欧式距离δ_pos更近，与负样本p_neg之间的全局描述特征的欧氏距离δ_neg更远，极负样本

对用于防止负样本对之间的全局描述特征距离错误的拉近，极负样本与负样本的欧式距离为

Train an encoder network using metric learning methods, utilizing a quadruple loss function

Semi-supervised training makes the Euclidean distance δ _pos of the global description feature between the input sample and the positive sample p _pos closer, and the Euclidean distance δ _neg of the global description feature between the negative sample p _neg is farther, and the extremely negative sample

To prevent the error of the global description feature distance between pairs of negative samples, the Euclidean distance between extremely negative samples and negative samples is

in:

α,β为超参数。

α and β are hyperparameters.

Specifically: set the hyperparameters α=0.5, β=0.2, select 2 positive samples, 18 negative samples and 1 randomly selected extremely negative sample during training.

Step 5: Take the sample pair of the training data set in step 1 and the corresponding global description feature generated by the encoder network in step 2, train the decoder network in step 3, and obtain the original point cloud that can be recovered and reconstructed from the global description feature. Decoder network model and parameters.

约束重建点云与真值点云的欧氏距离，其中

应用在生成的关键点轮廓，用于约束点云的物理空间分布，

应用于稠密重建点云。In step 5, the decoder network is trained in a supervised manner, using the loss function

Constrains the Euclidean distance between the reconstructed point cloud and the ground truth point cloud, where

Applied to the generated keypoint contours to constrain the physical spatial distribution of the point cloud,

Applied to densely reconstructed point clouds.

in:

ψ(p)为

中p的对应点

ψ(p) is

The corresponding point of p in

Step 6: In the actual SLAM job, take the encoder network model and parameters in step 4, extract the global description feature for the point cloud key frame, and perform NN matching with the global description feature of the historical key frame according to the Euclidean distance, which meets the matching conditions is used as a closed-loop candidate frame or a relocation candidate position.

In step 6, the NN matching of the global description features of the point cloud key frames is determined by the sequence matching of N consecutive current point cloud frames and consecutive N consecutive historical point cloud frames, if the consecutive sequences match the global description features between key frames. The Euclidean distance of is less than the set threshold and the gradient change compared with other historical key frame feature distances is greater than the set threshold, it can be regarded as a closed loop detection candidate sequence key frame or relocation candidate position. The laser SLAM closed-loop task and relocation task are performed in the feature space of the point cloud by the described point cloud feature space characterization method.

Step 7: In the actual SLAM operation, store or transmit the global description features of the point cloud keyframes extracted in Step 6. The global description features can be restored and reconstructed to the original SLAM point cloud map through the decoder network model and parameters in Step 5.

Use the encoder network described in step 4 to extract the global description features from the key frames in the laser point cloud map as compression codes for storage or communication transmission, and then use the decoder network described in step 5 to restore the storage features or transmission and reception features, reconstruct Raw laser point cloud map, the described point cloud feature space characterization method can be applied to embedded systems or low bandwidth situations.

Step 8: In the actual SLAM job, take the global description features extracted in s6, and perform NN matching with the global description features of the historical key frames of multiple collaborative mapping robots according to the Euclidean distance, through the point cloud feature space described The characterization method performs a multi-robot laser SLAM collaborative mapping task in the point cloud feature space.

In step 8, if multiple robots provide the current continuous point cloud frame and historical key frames, the multi-robot collaborative mapping is performed, and NN matching is performed on the global description features of all point cloud historical key frames, so that multi-robot collaboration can be realized. Build a map. The laser SLAM multi-robot collaborative task is carried out in the feature space of the point cloud through the described point cloud feature space characterization method.

The beneficial effects of the invention are as follows: (1) An encoder network is designed to extract information of geometric structure features and semantic context features for laser point cloud data in large scenes, and the point cloud feature space representation method realizes scene recognition by means of feature retrieval. , in the scene recognition task of the Oxford RobotCar dataset, the average recall rate reaches 93.35%; (2) The point cloud is reconstructed through the designed decoder network, and the original point cloud is reconstructed from the global description features extracted by the encoder network to realize the point cloud. Map compression storage and low-bandwidth transmission; (3) The constructed encoder network does not need to be trained according to the point cloud map in advance, and has a strong generalization ability, which can be directly applied in various data sets such as KITTI without training.

Claims (10)

1. a point cloud feature space characterization method for laser SLAM, is characterized in that, comprises the steps: s1. Perform filtering and downsampling preprocessing on the point cloud samples in the point cloud data set, and construct a training point cloud sample pair according to the similarity of the point cloud scene, including a positive sample p _pos and a negative sample p _neg ; s2. Build an encoder network, input the training point cloud sample pair in step s1, and obtain the global description feature f of the input point cloud P through the point cloud adaptive neighborhood feature extraction module, the dynamic graph network aggregation module, and the global description feature generation module. P); s3. Build a decoder network, input the global description features generated in step s2, and use a hierarchical generation strategy to generate key point contours to generate refined and dense reconstructed point clouds.

s4. Take the sample pair of the training data set in s1, train the encoder network of s2, and obtain the encoder network model and parameters that can measure the similarity of the point cloud scene; s5. Take the sample pairs of the training data set in s1 and the corresponding global description features generated by the s2 encoder network, train the decoder network of s3, and obtain a decoder network model that can restore and reconstruct the original point cloud from the global description features and parameters; s6. In the actual SLAM operation, take the encoder network model and parameters in s4, extract the global description feature for the point cloud key frame, and perform NN matching with the global description feature of the historical key frame according to the Euclidean distance, and the matching condition is taken as Closed-loop candidate frame or relocation candidate position; s7. In the actual SLAM operation, store or transmit the global description features of the point cloud keyframes extracted by s6. The global description features can be restored and reconstructed to the original SLAM point cloud map through the decoder network model and parameters in s5; s8. In the actual SLAM operation, take the global description features extracted in s6, and perform NN matching with the global description features of the historical key frames of multiple collaborative mapping robots according to the Euclidean distance, through the point cloud feature space representation. The method performs multi-robot laser SLAM collaborative mapping task in point cloud feature space. 2. The point cloud feature space representation method for laser SLAM as claimed in claim 1, characterized in that, in step s1, the similarity of constructing the training sample pair is determined by the auxiliary judgment of the position coordinates of the sample in the map, and the position interval Within 10m, it is regarded as a positive sample with similar structure, and the distance from 10m to 50m is regarded as the negative sample selection range, and the negative sample is randomly selected. 3. The point cloud feature space characterization method for laser SLAM as claimed in claim 1, wherein in step s2, the point cloud adaptive neighborhood feature extraction module of the encoder network uses the range [k _min , km _max ] ] Construct the kNN neighborhood N _i of each point p _i in the point cloud, and use the eigenvalue λ ₁ ≥λ ₂ ≥λ ₃ ≥0 of the covariance matrix Σ of the neighborhood point cloud to calculate the linearity of the point cloud neighborhood

flatness

and scattering

feature, and adaptively determine the optimal neighborhood size through the Shannon information entropy metric equation E _k

Extract neighborhood features as point features through a multi-layer perceptron feature extraction network; in:

V is an eigenvector matrix, E _k =-L _λ ln L _λ -P _λ ln P _λ -S _λ lnS _λ . 4. The point cloud feature space characterization method for laser SLAM as claimed in claim 1, wherein in step s2, the dynamic graph network aggregation module is composed of two parts: (a) Dynamic graph network aggregation in feature space: In step (3), a kNN with fixed k is used to construct a dynamic graph, and feature aggregation is performed through the graph network to update the features of the points; the feature aggregation process of the graph network is: f _i '=ρ(φ _Θ (ε(fi , _{f j )} )=ρ(φ _Θ (fi ‖(f _j -f _{i )} )), φ _Θ is the multilayer perceptron MLP, and ρ is max pooling; (b) Network aggregation of dynamic graphs in physical space: construct a dynamic graph with kNN of fixed k at all points in the physical space of the point cloud, and use the features of the updated points in the above (a) to update the features through the graph network, and update the features of the points; The dynamic graph G=(V, E) constructed by the center point pi is composed of node V and edge E. The feature of point _pi is f _i , the node V is the feature _{f j of} the neighborhood point, and the edge E is the center point and the neighbor point. The relationship between domain point features ε(f _i , f _j ), where: ε(fi , f _j )=f _i ‖(f _j -f _{i )} , ‖ is the feature splicing by channel. 5. The point cloud feature space representation method for laser SLAM as claimed in claim 1, wherein in step s2, the global description feature generation module generates global description features through NetVLAD Layer. 6. The point cloud feature space characterization method for laser SLAM as claimed in claim 1, wherein in step s3, the hierarchical generation strategy comprises two parts: (a) Keypoint contour generation network: Generate rough keypoint contours of points through a three-layer fully connected network and Reshape; (b) Refinement of dense point cloud generation network: a 2D planar grid of η × η points is stitched at each point in the key points generated in (a) above

And splicing the global description features generated in step s2 according to the channel, and generating dense reconstructed point clouds through MLP; in:

η is the number of points on the edge of the plane grid, and ξ is the size of the plane grid points. 7. The point cloud feature space representation method for laser SLAM as claimed in claim 1, wherein in step s4, the encoder network is trained using a metric learning method, and a quadruple loss function is used.

Semi-supervised training makes the Euclidean distance δ _pos of the global description feature between the input sample and the positive sample p _pos closer, and the Euclidean distance δ _neg of the global description feature between the negative sample p _neg is farther, and the extremely negative sample

To prevent the error of the global description feature distance between pairs of negative samples, the Euclidean distance between extremely negative samples and negative samples is

in:

α and β are hyperparameters. 8. The point cloud feature space representation method for laser SLAM as claimed in claim 1, characterized in that, in step s5, the decoder network is trained in a supervised manner, using a loss function

Constrains the Euclidean distance between the reconstructed point cloud and the ground truth point cloud, where

The keypoint contours generated in (a) of step s2 are applied to constrain the physical spatial distribution of the point cloud,

applied to densely reconstructed point clouds; in:

ψ(p) is

The corresponding point of p in ,

9. The point cloud feature space characterization method for laser SLAM as claimed in claim 1, wherein in steps s6 and s8, performing NN matching on the global description feature of the point cloud key frame is performed by consecutive N current points It is judged that the cloud frame matches the sequence of N consecutive historical point cloud frames. If the Euclidean distance of the global description feature between the consecutive sequence matching key frames is less than the set threshold and the gradient change compared with the feature distances of other historical key frames is greater than the set threshold , it can be regarded as a closed-loop detection candidate sequence key frame or a repositioning candidate position; if multiple robots provide the current continuous point cloud frame and historical key frame, it is a multi-robot collaborative mapping; through the point cloud feature space representation The method performs laser SLAM closed-loop tasks, relocation tasks and multi-robot collaborative tasks in the feature space of point clouds. 10. The point cloud feature space characterization method for laser SLAM as claimed in claim 1, wherein in step s7, the encoder network described in s4 is used to extract the global description feature for the key frame in the laser point cloud map as Compression coding is used for storage or communication transmission, and then the decoder network described in s5 is used to restore the storage features or transmission and reception features to reconstruct the original laser point cloud map. The point cloud feature space representation method can be applied to embedded systems or low bandwidth situation.

Images