Dangerous chemical substance warehouse similar stacking point cloud registration method

Abstract

The invention relates to the technical field of hazardous chemical substance point cloud registration, and discloses a hazardous chemical substance warehouse similar stacking point cloud registration method comprising the following steps: obtaining hazardous chemical substance warehouse stacking source point cloud data and target point cloud data, and inputting the data to a feature extraction module, a point cloud feature interaction module and a double-branch feature interaction module; and obtaining a source point cloud rotation final feature, a target point cloud rotation final feature, a source point cloud translation final feature and a target point cloud translation final feature, and performing point cloud registration by using the final features. According to the method, the point cloud features of the similar structure stacks are effectively extracted from three levels of rotation translation decoupling optimization in point clouds, between point clouds and between double branches, and the complex features of the similar structure point clouds are effectively captured. And a position self-attention mechanism containing relative position embedding is provided in feature interaction between point clouds, so that the registration effect under the condition of high similarity of point cloud feature structures in a hazardous chemical substance warehouse scene is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hazardous chemical point cloud registration, and in particular to a method for registering similar stacking point clouds in a hazardous chemical warehouse. Background Art

[0002] Hazardous chemicals are corrosive, harmful, toxic, flammable and explosive. Their storage requires high professionalism and is highly dangerous, and there are great safety risks during the storage process. Therefore, it is of great significance to use 3D reconstruction technology to monitor the safety status of hazardous chemical warehouses. Point cloud registration, as a key step in 3D reconstruction, faces the problem of excessive similarity of point cloud features of similar structure stacks in hazardous chemical warehouse scenarios, which leads to mismatching. At present, point cloud registration methods can be roughly divided into traditional methods, feature descriptor-based methods and end-to-end methods. Although traditional methods are widely used, due to their inherent limitations, they are prone to fall into local optimal solutions in complex scenarios, resulting in unsatisfactory registration effects. The feature descriptor-based method extracts the geometric features of the object for registration, but when faced with highly similar geometric structures, the feature matching accuracy will drop significantly, and it is easy to produce incorrect correspondences, which in turn affects the registration accuracy and reliability.

[0003] The actual application scenarios of hazardous chemical warehouses are quite different from these standard data sets. In hazardous chemical warehouse scenarios, since a large number of items with similar geometric structures (such as stacks and oil tanks) are stored in the warehouse, the point cloud data collected by the camera often contains a large number of similar structures. This similarity leads to a large number of outliers (i.e., mismatched points) in the feature matching between the two point clouds to be registered, which causes incorrect matching of the point cloud correspondence, seriously affecting the accuracy and effect of point cloud registration. Therefore, a method for similar stack point cloud registration in hazardous chemical warehouses is urgently needed. Summary of the Invention

[0004] The purpose of the present invention is to overcome one or more of the above-mentioned existing technical problems and provide a method for aligning similar stack point clouds in a hazardous chemicals warehouse.

[0005] To achieve the above objectives, the present invention provides a method for registering similar stack point clouds in a hazardous chemicals warehouse, comprising:

[0006] Obtain source point cloud data of hazardous chemicals warehouse stacks;

[0007] Preprocess the source point cloud data of hazardous chemicals warehouse stacks to obtain the target point cloud data of hazardous chemicals warehouse stacks;

[0008] The source point cloud data of the hazardous chemicals warehouse stack and the target point cloud data of the hazardous chemicals warehouse stack are input into the feature extraction module, which includes a rotation extraction network branch and a translation extraction network branch to obtain the source point cloud rotation feature, the target point cloud rotation feature, the source point cloud translation feature and the target point cloud translation feature;

[0009] The source point cloud rotation features, target point cloud rotation features, source point cloud translation features, and target point cloud translation features are input into the point cloud feature interaction module. The soft matching algorithm is used to extract the points with high similarity between the two point clouds as key points, and the points with the most similar features are used as reference points.

[0010] After screening the source point cloud rotation features, target point cloud rotation features, source point cloud translation features and target point cloud translation features based on key points and reference points, the position matrix is introduced and feature aggregation is performed to obtain the source point cloud rotation aggregation features, target point cloud rotation aggregation features, source point cloud translation aggregation features and target point cloud translation aggregation features;

[0011] The source point cloud rotation aggregation feature, the target point cloud rotation aggregation feature, the source point cloud translation aggregation feature and the target point cloud translation aggregation feature are input into the dual-branch feature interaction module, and the source point cloud rotation final feature, the target point cloud rotation final feature, the source point cloud translation final feature and the target point cloud translation final feature are obtained through the cascade processing of feature difference and residual calculation;

[0012] Point cloud registration is performed based on the final rotation features of the source point cloud, the final rotation features of the target point cloud, the final translation features of the source point cloud, and the final translation features of the target point cloud.

[0013] According to one aspect of the present invention, the rotation extraction network branch and the translation extraction network branch are set in parallel and have the same network structure, which includes a six-layer graph convolutional network and a three-layer self-attention mechanism embedded between the graph convolutional networks;

[0014] The source point cloud data of the hazardous chemicals warehouse stacking and the target point cloud data of the hazardous chemicals warehouse stacking are respectively input into the rotation extraction network branch to obtain the source point cloud rotation feature and the target point cloud rotation feature. The source point cloud data of the hazardous chemicals warehouse stacking and the target point cloud data of the hazardous chemicals warehouse stacking are respectively input into the translation extraction network branch to obtain the source point cloud translation feature and the target point cloud translation feature. The formula is,

[0015] F1=Self-attention(GCN(U));

[0016] F2=GCN(GCN(Self-attention(GCN(F1))));

[0017] F3=Self-attention(GCN(F2));

[0018] U1=GCN(conact(F1,F2,F3));

[0019] Among them, U represents the input of the network;

[0020] GCN stands for Graph Convolutional Network;

[0021] Self-attention means self-attention mechanism;

[0022] F1 represents the first output feature;

[0023] F2 represents the second output feature;

[0024] F3 represents the third output feature;

[0025] conact means feature fusion;

[0026] U1 represents the output of the network.

[0027] According to one aspect of the present invention, a rotation affinity matrix is obtained based on the transpose of the source point cloud rotation feature and the target point cloud rotation feature, where the formula is:

[0028] B r =(G x,r ) T W(G y,r );

[0029] Among them, B r represents the rotation affinity matrix;

[0030] (G x,r ) T Represents the transpose of the rotation feature of the source point cloud;

[0031] (G y,r ) represents the target point cloud rotation feature;

[0032] W represents the learnable parameters of the affinity layer;

[0033] Based on the rotation affinity matrix, the rotation soft correspondence matrix is obtained by alternating row and column normalization. The rotation soft correspondence matrix is used to transform the target point cloud rotation feature to obtain the rotation transformation feature, where the formula is:

[0034] F tran,r =C r *G y,r ;

[0035] Among them, F tran,r Represents the rotation transformation feature;

[0036] C r represents the rotation soft correspondence matrix;

[0037] The feature difference is calculated based on the rotation feature and rotation transformation feature of the target point cloud, and the L2 norm of the difference is taken. The k=M smallest points are taken as the rotation key points, and the point pair with the smallest difference is taken as the common rotation reference point of the point cloud to be registered. The formula is:

[0038]

[0039] Among them, E1 represents the set of rotation key points;

[0040] x k Indicates the point cloud coordinates corresponding to the key point index in the source point cloud;

[0041] y k Indicates the point cloud coordinates corresponding to the key point index in the target point cloud;

[0042] n represents the index of the source point cloud;

[0043] j represents the index of the target point cloud;

[0044] Represents the first M feature index pairs that minimize the feature distance;

[0045] The source point cloud rotation features and the target point cloud rotation features are filtered based on the key points to obtain the source point cloud rotation filtering features and the target point cloud rotation filtering features.

[0046] According to one aspect of the present invention, a translation affinity matrix is obtained based on the transpose of the source point cloud translation feature and the target point cloud translation feature, where the formula is:

[0047] B t =(G x,t ) T W(G y,t );

[0048] Among them, B t represents the translation affinity matrix;

[0049] (G x,t ) T Represents the transpose of the translation feature of the source point cloud;

[0050] (G y,t ) represents the translation feature of the target point cloud;

[0051] Based on the translation affinity matrix, the soft translation correspondence matrix is obtained by alternating row and column normalization. The soft translation correspondence matrix is used to transform the translation features of the target point cloud to obtain the translation transformation features, where the formula is:

[0052] F tran,t =C t *Gy,t ;

[0053] Among them, F tran,t Represents translation transformation features;

[0054] C t represents the translation soft correspondence matrix;

[0055] The feature difference is calculated based on the translation feature and translation transformation feature of the target point cloud, and the L2 norm of the difference is taken. The k=M smallest points are taken as the translation key points, and the point pair with the smallest difference is taken as the common translation reference point of the point cloud to be registered. The formula is:

[0056]

[0057] Among them, E2 represents the translation key point set;

[0058] The source point cloud translation features and the target point cloud translation features are filtered based on the key points to obtain the source point cloud translation filtering features and the target point cloud translation filtering features.

[0059] According to one aspect of the present invention, the coordinates of each point of the source point cloud and the target point cloud after the rotation branch screening are calculated relative to the rotation reference point to obtain the first relative distance of the source point cloud and the first relative distance of the target point cloud. The source point cloud rotation distance embedding information and the target point cloud rotation distance embedding information are calculated using a sine function based on the first relative distance of the source point cloud and the first relative distance of the target point cloud, wherein the formula is,

[0060]

[0061] Among them, I r1 Represents the source point cloud rotation distance embedding information;

[0062] I r2 Represents the target point cloud rotation distance embedding information;

[0063] sin represents the sine function;

[0064] d x1 Indicates the first relative distance of the source point cloud;

[0065] d x2 Indicates the first relative distance of the target point cloud;

[0066] d represents the dimension index of the embedding vector;

[0067] D represents the total dimension of the embedding vector;

[0068] For the source point cloud rotation filtering feature and the target point cloud rotation filtering feature, the position matrix is added on the basis of the query matrix, key matrix and value matrix to obtain the source point cloud rotation position feature and the target point cloud rotation position feature, where the formula is,

[0069]

[0070] Q1=G′ x,r W Q ;

[0071] K1=G′ x,r W k ;

[0072] V1=G′ x,r W V ;

[0073] P1=I r1 W P ;

[0074]

[0075] Q2=G′ y,r W Q ;

[0076] K2=G′ y,r W k ;

[0077] V2=G′ y,r W V ;

[0078] P2=I r2 W P ;

[0079] Among them, softmax represents the normalized exponential function;

[0080] F′ x,r Represents the rotation position feature of the source point cloud;

[0081] Q1 represents the first query matrix;

[0082] K1 represents the first bond matrix;

[0083] V1 represents the first value matrix;

[0084] P1 represents the first position matrix;

[0085] T stands for transpose;

[0086] W Q represents the query weight matrix;

[0087] W k represents the bond weight matrix;

[0088] W V represents the value weight matrix;

[0089] W P represents the position weight matrix;

[0090] d k represents the feature dimension;

[0091] G′ x,r Represents the source point cloud rotation screening feature;

[0092] F′ y,r Represents the rotation position characteristics of the target point cloud;

[0093] G′ y,r Indicates the target point cloud rotation screening feature;

[0094] Q2 represents the second query matrix;

[0095] K2 represents the second bond matrix;

[0096] V2 represents the second value matrix;

[0097] P2 represents the second position matrix.

[0098] According to one aspect of the present invention, based on the source point cloud rotation position feature and the target point cloud rotation position feature, a feature cross attention mechanism is used to calculate the attention weight between points, and the feature information from another point cloud is aggregated to obtain the source point cloud rotation aggregation feature and the target point cloud rotation aggregation feature, where the formula is,

[0099]

[0100] Q3=F′ x,r W Q ;

[0101] K3=F′ y,r W k ;

[0102] V3=F′ y,r W V ;

[0103]

[0104] Q4=F′ y,r W Q ;

[0105] K4=F′ x,r W k ;

[0106] V4=F′ x,r WV ;

[0107] Among them, F x,r Represents the rotation aggregation feature of the source point cloud;

[0108] Q3 represents the third query matrix;

[0109] K3 represents the third bond matrix;

[0110] V3 represents the third value matrix;

[0111] F y,r Represents the rotation aggregation feature of the target point cloud;

[0112] Q4 represents the fourth query matrix;

[0113] K4 represents the fourth bond matrix;

[0114] V4 represents the fourth value matrix.

[0115] According to one aspect of the present invention, the coordinates of each point of the source point cloud and the target point cloud after the translation branch screening are calculated relative to the translation reference point to obtain the second relative distance of the source point cloud and the second relative distance of the target point cloud. The source point cloud translation distance embedding information and the target point cloud translation distance embedding information are calculated using a sine function based on the second relative distance of the source point cloud and the second relative distance of the target point cloud, wherein the formula is,

[0116]

[0117] Among them, I t1 Represents the embedded information of the source point cloud translation distance;

[0118] I t2 Represents the target point cloud translation distance embedding information;

[0119] d x3 Indicates the second relative distance of the source point cloud;

[0120] d x4 Indicates the second relative distance of the target point cloud;

[0121] For the source point cloud translation filtering feature and the target point cloud translation filtering feature, the position matrix is added on the basis of the query matrix, key matrix and value matrix to obtain the source point cloud translation position feature and the target point cloud translation position feature, where the formula is,

[0122]

[0123] Q5=G′ x,t W Q ;

[0124] K5=G′ x,t Wk ;

[0125] V5=G′ x,t W V ;

[0126] P5=I t1 W P ;

[0127]

[0128] Q6=G′ y,t W Q ;

[0129] K6=G′ y,t W k ;

[0130] V6=G′ y,t W V ;

[0131] P6=I t2 W P ;

[0132] Among them, F′ x,t Represents the translation position feature of the source point cloud;

[0133] Q5 represents the fifth query matrix;

[0134] K5 represents the fifth bond matrix;

[0135] V5 represents the fifth value matrix;

[0136] P5 represents the fifth position matrix;

[0137] G′ x,t Indicates the source point cloud translation filtering feature;

[0138] F′ y,t Represents the translation position characteristics of the target point cloud;

[0139] G′ y,t Indicates the target point cloud translation filtering feature;

[0140] Q6 represents the sixth query matrix;

[0141] K6 represents the sixth bond matrix;

[0142] V6 represents the sixth value matrix;

[0143] P6 represents the sixth position matrix.

[0144] According to one aspect of the present invention, based on the source point cloud translation position feature and the target point cloud translation position feature, a feature cross attention mechanism is used to calculate the attention weight between points, and the feature information from another point cloud is aggregated to obtain the source point cloud translation aggregation feature and the target point cloud translation aggregation feature, where the formula is,

[0145]

[0146] Q7=F′ x,t W Q ;

[0147] K7=F′ y,t W k ;

[0148] V7=F′ y,t W V ;

[0149]

[0150] Q8=F′ y,t W Q ;

[0151] K8=F′ x,t W k ;

[0152] V8=F′ x,t W V ;

[0153] Among them, F x,t Represents the translation aggregation feature of the source point cloud;

[0154] Q7 represents the seventh query matrix;

[0155] K7 represents the seventh bond matrix;

[0156] V7 represents the seventh value matrix;

[0157] F y,t Represents the translation aggregation features of the target point cloud;

[0158] Q8 represents the eighth query matrix;

[0159] K8 represents the eighth key matrix;

[0160] V8 represents the eighth value matrix.

[0161] According to one aspect of the present invention, the first source point cloud difference feature is obtained by using the i-th point feature of the source point cloud rotation aggregation feature and the i-th point feature of the source point cloud translation aggregation feature to make a difference, where the formula is,

[0162]

[0163] in, Represents the difference feature of the first source point cloud;

[0164] Represents the i-th point feature of the source point cloud rotation aggregation feature;

[0165] Represents the i-th point feature of the source point cloud translation aggregation feature;

[0166] Based on the source point cloud rotation aggregation feature and the first source point cloud difference feature, the source point cloud rotation update feature is obtained, where the formula is:

[0167]

[0168] Among them, γ represents the fusion ratio learnable parameter;

[0169] Indicates the source point cloud rotation update feature;

[0170] Based on the source point cloud rotation update feature and the transposed source point cloud rotation update feature, the first adjacency matrix is obtained, and based on the normalized first adjacency matrix and the source point cloud rotation update feature, the source point cloud rotation final feature is obtained, where the formula is,

[0171]

[0172] Among them, F” x,r Represents the final feature of the source point cloud rotation;

[0173] A'1 represents the normalized first adjacency matrix;

[0174] f a Represents the fully connected layer function;

[0175] f u represents the activation function;

[0176] Use the i-th point feature of the source point cloud translation aggregation feature and the i-th point feature of the source point cloud rotation aggregation feature to make a difference and get the second source point cloud difference feature, where the formula is,

[0177]

[0178] in, Represents the difference feature of the second source point cloud;

[0179] Based on the source point cloud translation aggregation feature and the second source point cloud difference feature, the source point cloud translation update feature is obtained, where the formula is:

[0180]

[0181] in, Represents the source point cloud translation update feature;

[0182] Based on the source point cloud translation update feature and the transposed source point cloud translation update feature, the second adjacency matrix is obtained, and based on the normalized second adjacency matrix and the source point cloud translation update feature, the source point cloud translation final feature is obtained, where the formula is,

[0183]

[0184] Among them, F” x,t Represents the final feature of the source point cloud translation;

[0185] A'2 represents the normalized second adjacency matrix.

[0186] According to one aspect of the present invention, the first target point cloud difference feature is obtained by using the i-th point feature of the target point cloud rotation aggregation feature and the i-th point feature of the target point cloud translation aggregation feature to make a difference, where the formula is,

[0187]

[0188] in, Represents the first target point cloud difference feature;

[0189] Represents the i-th point feature of the target point cloud rotation aggregation feature;

[0190] Represents the i-th point feature of the target point cloud translation aggregation feature;

[0191] Based on the target point cloud rotation aggregation feature and the first target point cloud difference feature, the target point cloud rotation update feature is obtained, where the formula is:

[0192]

[0193] in, Indicates the target point cloud rotation update feature;

[0194] Based on the target point cloud rotation update feature and the transposed target point cloud rotation update feature, the third adjacency matrix is obtained, and based on the normalized third adjacency matrix and the target point cloud rotation update feature, the target point cloud rotation final feature is obtained, where the formula is,

[0195]

[0196] Among them, F” y,r Represents the final feature of the target point cloud rotation;

[0197] A'3 represents the normalized third adjacency matrix;

[0198] Use the i-th point feature of the target point cloud translation aggregation feature and the i-th point feature of the target point cloud rotation aggregation feature to make a difference and get the second target point cloud difference feature, where the formula is,

[0199]

[0200] in, Represents the second target point cloud difference feature;

[0201] Based on the target point cloud translation aggregation feature and the second target point cloud difference feature, the target point cloud translation update feature is obtained, where the formula is:

[0202]

[0203] in, Represents the target point cloud translation update feature;

[0204] Based on the target point cloud translation update feature and the transposed target point cloud translation update feature, the fourth adjacency matrix is obtained, and based on the normalized fourth adjacency matrix and the target point cloud translation update feature, the target point cloud translation final feature is obtained, where the formula is,

[0205]

[0206] Among them, F” y,t Represents the final feature of the target point cloud translation;

[0207] A'4 represents the normalized second adjacency matrix.

[0208] Based on this, the beneficial effects of the present invention are as follows: the present application builds a network framework with hierarchical progressive feature interaction, which enables the network to learn spatial information in different receptive fields in the point cloud and reduce the interference of redundant information to ensure accurate point cloud matching;

[0209] A dual-branch feature interaction module is designed for feature interaction between the two branches. Based on the fact that the rotation and translation matrices belong to different solution spaces, the two branches of rotation and translation are decoupled through the two-difference method, so that each path can focus on its own features independently, thereby weakening the feature interference between the two branches. BRIEF DESCRIPTION OF THE DRAWINGS

[0210] Figure 1 This is a flow chart of a method for registering similar stack point clouds in a hazardous chemicals warehouse according to an exemplary embodiment;

[0211] Figure 2 This is an overall network structure diagram of a method for registering similar stack point clouds in a hazardous chemicals warehouse according to an exemplary embodiment;

[0212] Figure 3 This is a structural diagram of a feature extraction module of a method for registering similar stack point clouds in a hazardous chemicals warehouse according to an exemplary embodiment;

[0213] Figure 4 This is a structural diagram of a point cloud feature interaction module of a point cloud registration method for similar stacks in a hazardous chemicals warehouse according to an exemplary embodiment;

[0214] Figure 5 is a schematic diagram showing an added position matrix in a method for registering similar stacks of hazardous chemicals warehouse point clouds according to an exemplary embodiment;

[0215] Figure 6 This is a structural diagram of a dual-branch feature interaction module of a method for similar stack point cloud registration in a hazardous chemicals warehouse according to an exemplary embodiment. DETAILED DESCRIPTION

[0216] The present invention will now be discussed with reference to exemplary embodiments. It should be understood that the embodiments discussed are only for enabling those skilled in the art to better understand and thereby implement the present invention, rather than implying any limitation on the scope of the present invention.

[0217] As used herein, the term “including” and variations thereof are to be interpreted as open-ended terms meaning “including, but not limited to.” The term “based on” is to be interpreted as “based, at least in part, on,” and the terms “one embodiment” and “an embodiment” are to be interpreted as “at least one embodiment.”

[0218] According to one embodiment of the present invention, Figure 1 This is a flow chart of a method for registering similar stacking point clouds in a hazardous chemicals warehouse according to an exemplary embodiment. Figure 1 As shown, to achieve the above purpose, the present invention provides a method for registering similar stack point clouds in a hazardous chemicals warehouse, comprising:

[0219] Obtain source point cloud data of hazardous chemicals warehouse stacks;

[0220] The source point cloud data of the hazardous chemicals warehouse stacking is preprocessed to obtain the target point cloud data of the hazardous chemicals warehouse stacking. Both data are downsampled to 1024 points. A rigid body transformation is generated by randomly generating three Euler angles in the range of [0, 45] degrees and a random translation in [-0.5, 0.5]. The transformed point cloud is used as the target point cloud data of the hazardous chemicals warehouse stacking.

[0221] Figure 2 FIG. 1 is an overall network structure diagram of a method for registering similar stacking point clouds in a hazardous chemicals warehouse according to an exemplary embodiment. Figure 2As shown in the figure, the feature extraction module uses a fusion of graph convolutional networks and self-attention mechanisms. The graph convolutional network is used to extract the local structural features of the point cloud and capture the geometric relationship between neighboring points in the point cloud, while the self-attention mechanism is used to capture the global contextual information of the point cloud and model the dependency between long-distance points in the point cloud. The combination of the two can take into account both local details and global structures, providing a more comprehensive feature representation for point cloud registration. The feature interaction module between point clouds first removes noise through denoising and key point extraction, and extracts local features of overlapping areas based on soft correspondences, and selects significant point pairs as reference points. Subsequently, the relative position information between the points and the reference points is used to extract global position features to provide spatial constraints for registration. The dual-branch feature interaction module uses the sequential two-difference method to let the rotation and translation branches focus on the features they need respectively, eliminating the interference of other branch features on themselves.

[0222] The source point cloud data of the hazardous chemicals warehouse stack and the target point cloud data of the hazardous chemicals warehouse stack are input into the feature extraction module, which includes a rotation extraction network branch and a translation extraction network branch to obtain the source point cloud rotation feature, the target point cloud rotation feature, the source point cloud translation feature and the target point cloud translation feature;

[0223] The source point cloud rotation features, target point cloud rotation features, source point cloud translation features, and target point cloud translation features are input into the point cloud feature interaction module. The soft matching algorithm is used to extract the points with high similarity between the two point clouds as key points, and the points with the most similar features are used as reference points.

[0224] After screening the source point cloud rotation features, target point cloud rotation features, source point cloud translation features and target point cloud translation features based on key points and reference points, the position matrix is introduced and feature aggregation is performed to obtain the source point cloud rotation aggregation features, target point cloud rotation aggregation features, source point cloud translation aggregation features and target point cloud translation aggregation features;

[0225] The source point cloud rotation aggregation feature, the target point cloud rotation aggregation feature, the source point cloud translation aggregation feature and the target point cloud translation aggregation feature are input into the dual-branch feature interaction module, and the source point cloud rotation final feature, the target point cloud rotation final feature, the source point cloud translation final feature and the target point cloud translation final feature are obtained through the cascade processing of feature difference and residual calculation;

[0226] Point cloud registration is performed based on the final rotation features of the source point cloud, the final rotation features of the target point cloud, the final translation features of the source point cloud, and the final translation features of the target point cloud.

[0227] According to one embodiment of the present invention, the rotation extraction network branch and the translation extraction network branch are arranged in parallel and have the same network structure, which includes a six-layer graph convolutional network and a three-layer self-attention mechanism embedded between the graph convolutional networks;

[0228] Figure 3 FIG. 1 is a structural diagram of a feature extraction module of a method for registering similar stacking point clouds in a hazardous chemicals warehouse according to an exemplary embodiment. Figure 3 As shown in the figure, the feature extraction module designs an independent local-global feature extraction network for rotation and translation branches. By constructing a dual-branch architecture combining the local feature extraction module with the global semantic aggregation module, differentiated feature learning is performed for the rotation and translation transformation characteristics of the point cloud data respectively. At the same time, a multi-layer network structure is used to gradually deepen the feature expression, thereby simultaneously improving the effectiveness of point cloud processing and feature discrimination accuracy at the two levels of spatial geometry understanding and global semantic representation. The module expands the receptive field by introducing a six-layer graph convolutional network and embeds a self-attention mechanism between graph convolutions to achieve efficient collaborative capture and interaction of local and global features. Specifically, the self-attention mechanism of the first two layers can dynamically screen and weighted aggregate graph convolutions. The important neighbor node information in the network enhances the robustness of local feature extraction, and the self-attention mechanism of the last layer further integrates the multi-level feature information extracted by deep graph convolution to strengthen the global context modeling capability. This design enables the module to simultaneously capture the local geometric details and global structural relationships of the point cloud, and provide richer feature representations for point cloud registration. The source point cloud data of the hazardous chemicals warehouse stacking and the target point cloud data of the hazardous chemicals warehouse stacking are respectively input into the rotation extraction network branch to obtain the source point cloud rotation feature and the target point cloud rotation feature. The source point cloud data of the hazardous chemicals warehouse stacking and the target point cloud data of the hazardous chemicals warehouse stacking are respectively input into the translation extraction network branch to obtain the source point cloud translation feature and the target point cloud translation feature. The formula is,

[0229] F1=Self-attention(GCN(U));

[0230] F2=GCN(GCN(Self-attention(GCN(F1))));

[0231] F3=Self-attention(GCN(F2));

[0232] U1=GCN(conact(F1,F2,F3));

[0233] Among them, U represents the input of the network;

[0234] GCN stands for Graph Convolutional Network;

[0235] Self-attention means self-attention mechanism;

[0236] F1 represents the first output feature;

[0237] F2 represents the second output feature;

[0238] F3 represents the third output feature;

[0239] conact means feature fusion;

[0240] U1 represents the output of the network.

[0241] According to one embodiment of the present invention, Figure 4 FIG. 1 is a structural diagram of a feature interaction module between point clouds of a hazardous chemicals warehouse similar stacking point cloud registration method according to an exemplary embodiment. Figure 4 As shown in the figure, the feature interaction module between the source point cloud and the target point cloud consists of two parts. The first part involves the extraction of reference points and key points. A soft matching algorithm (Affinity Instance Normalization Sinkhorn, AIS) is used to extract points with high similarity between the two point clouds as key points, and the points with the most significant feature similarity are used as reference points. The second part calculates the relative position information of each point with respect to the reference point and introduces it as prior knowledge into the Transformer architecture, effectively improving the robustness and accuracy of point cloud registration in the presence of similar structures.

[0242] In point cloud registration, since the target point cloud and the source point cloud cannot completely correspond in real scenes, the feature interaction module between point clouds first removes noise points through reference point and key point extraction algorithms. Then, based on the soft correspondence between point clouds, it extracts local feature information in the overlapping area and uses the point pairs with the most significant overlapping features as reference points.

[0243] The rotation affinity matrix is obtained based on the transpose of the source point cloud rotation feature and the target point cloud rotation feature to measure the similarity between nodes. The formula is:

[0244] B r =(G x,r ) T W(G y,r );

[0245] Among them, B r represents the rotation affinity matrix;

[0246] (G x,r ) T Represents the transpose of the rotation feature of the source point cloud;

[0247] (G y,r ) represents the target point cloud rotation feature;

[0248] W represents the learnable parameters of the affinity layer;

[0249] Based on the rotation affinity matrix, the rotation soft correspondence matrix is obtained by alternating row and column normalization. The rotation soft correspondence matrix is used to transform the target point cloud rotation feature to obtain the rotation transformation feature, where the formula is:

[0250] F tran,r =C r *G y,r ;

[0251] Among them, F tran,r Represents the rotation transformation feature;

[0252] C r represents the rotation soft correspondence matrix;

[0253] The feature difference is calculated based on the rotation feature and rotation transformation feature of the target point cloud, and the L2 norm of the difference is taken. The k=M smallest points are taken as the rotation key points, and the point pair with the smallest difference is taken as the common rotation reference point of the point cloud to be registered. The formula is:

[0254]

[0255] Among them, E1 represents the set of rotation key points;

[0256] x k Indicates the point cloud coordinates corresponding to the key point index in the source point cloud;

[0257] y k Indicates the point cloud coordinates corresponding to the key point index in the target point cloud;

[0258] n represents the index of the source point cloud;

[0259] j represents the index of the target point cloud;

[0260] Represents the first M feature index pairs that minimize the feature distance;

[0261] After the key points and reference points are selected, the relative position information of each key point and reference point is calculated, and a self-attention mechanism is introduced to reduce the mismatch of similar structures during subsequent feature interaction. Secondly, through the cross-attention mechanism, similar structures between the source point cloud and the target point cloud can be more effectively aligned and coordinated, ensuring their synchronization in the feature space and performing feature interaction;

[0262] The source point cloud rotation features and the target point cloud rotation features are filtered based on the key points to obtain the source point cloud rotation filtering features and the target point cloud rotation filtering features.

[0263] According to one embodiment of the present invention, a translation affinity matrix is obtained based on the transpose of the source point cloud translation feature and the target point cloud translation feature, where the formula is:

[0264] B t =(G x,t ) T W(G y,t );

[0265] Among them, B t represents the translation affinity matrix;

[0266] (G x,t ) T Represents the transpose of the translation feature of the source point cloud;

[0267] (G y,t ) represents the translation feature of the target point cloud;

[0268] Based on the translation affinity matrix, the soft translation correspondence matrix is obtained by alternating row and column normalization. The soft translation correspondence matrix is used to transform the translation features of the target point cloud to obtain the translation transformation features, where the formula is:

[0269] F tran,t =C t *G y,t ;

[0270] Among them, F tran,t Represents translation transformation features;

[0271] C t represents the translation soft correspondence matrix;

[0272] The feature difference is calculated based on the translation feature and translation transformation feature of the target point cloud, and the L2 norm of the difference is taken. The k=M smallest points are taken as the translation key points, and the point pair with the smallest difference is taken as the common translation reference point of the point cloud to be registered. The formula is:

[0273]

[0274] Among them, E2 represents the translation key point set;

[0275] The source point cloud translation features and the target point cloud translation features are filtered based on the key points to obtain the source point cloud translation filtering features and the target point cloud translation filtering features.

[0276] According to one embodiment of the present invention, the coordinates of each point of the source point cloud and the target point cloud after the rotation branch screening are calculated relative to the rotation reference point to obtain the first relative distance of the source point cloud and the first relative distance of the target point cloud. The source point cloud rotation distance embedding information and the target point cloud rotation distance embedding information are calculated using a sine function based on the first relative distance of the source point cloud and the first relative distance of the target point cloud, wherein the formula is,

[0277]

[0278] Among them, I r1 Represents the source point cloud rotation distance embedding information;

[0279] I r2 Represents the target point cloud rotation distance embedding information;

[0280] sin represents the sine function;

[0281] d x1 Indicates the first relative distance of the source point cloud;

[0282] d x2 Indicates the first relative distance of the target point cloud;

[0283] d represents the dimension index of the embedding vector;

[0284] D represents the total dimension of the embedding vector;

[0285] Figure 5 FIG. 1 is a schematic diagram showing an example of adding a position matrix to a method for registering similar stacking point clouds in a hazardous chemicals warehouse according to an exemplary embodiment. Figure 5 As shown, for the source point cloud rotation screening feature and the target point cloud rotation screening feature, the position matrix is added on the basis of the query matrix, key matrix and value matrix to obtain the source point cloud rotation position feature and the target point cloud rotation position feature, where the formula is,

[0286]

[0287] Q1=G′ x,r W Q ;

[0288] K1=G′ x,r W k ;

[0289] V1=G′ x,r W V ;

[0290] P1=I r1 W P ;

[0291]

[0292] Q2=G′ y,r W Q ;

[0293] K2=G′ y,r W k ;

[0294] V2=G′ y,r W V ;

[0295] P2=Ir2 W P ;

[0296] Among them, softmax represents the normalized exponential function;

[0297] F′ x,r Represents the rotation position feature of the source point cloud;

[0298] Q1 represents the first query matrix;

[0299] K1 represents the first bond matrix;

[0300] V1 represents the first value matrix;

[0301] P1 represents the first position matrix;

[0302] T stands for transpose;

[0303] W Q represents the query weight matrix;

[0304] W k represents the bond weight matrix;

[0305] W V represents the value weight matrix;

[0306] W P represents the position weight matrix;

[0307] d k represents the feature dimension;

[0308] G′ x,r Represents the source point cloud rotation screening feature;

[0309] F′ y,r Represents the rotation position characteristics of the target point cloud;

[0310] G′ y,r Indicates the target point cloud rotation screening feature;

[0311] Q2 represents the second query matrix;

[0312] K2 represents the second bond matrix;

[0313] V2 represents the second value matrix;

[0314] P2 represents the second position matrix.

[0315] According to one embodiment of the present invention, based on the source point cloud rotation position feature and the target point cloud rotation position feature, a feature cross attention mechanism is used to calculate the attention weight between points, and the feature information from another point cloud is aggregated to obtain the source point cloud rotation aggregation feature and the target point cloud rotation aggregation feature, where the formula is,

[0316]

[0317] Q3=F′ x,r W Q ;

[0318] K3=F′ y,r W k ;

[0319] V3=F′ y,r W V ;

[0320]

[0321] Q4=F′ y,r W Q ;

[0322] K4=F′ x,r W k ;

[0323] V4=F′ x,r W V ;

[0324] Among them, F x,r Represents the rotation aggregation feature of the source point cloud;

[0325] Q3 represents the third query matrix;

[0326] K3 represents the third bond matrix;

[0327] V3 represents the third value matrix;

[0328] F y,r Represents the rotation aggregation feature of the target point cloud;

[0329] Q4 represents the fourth query matrix;

[0330] K4 represents the fourth bond matrix;

[0331] V4 represents the fourth value matrix.

[0332] According to one embodiment of the present invention, the coordinates of each point of the source point cloud and the target point cloud after the translation branch screening are calculated relative to the translation reference point to obtain the second relative distance of the source point cloud and the second relative distance of the target point cloud. The source point cloud translation distance embedding information and the target point cloud translation distance embedding information are calculated using a sine function based on the second relative distance of the source point cloud and the second relative distance of the target point cloud, wherein the formula is,

[0333]

[0334] Among them, I t1 Represents the embedded information of the source point cloud translation distance;

[0335] I t2 Represents the target point cloud translation distance embedding information;

[0336] d x3 Indicates the second relative distance of the source point cloud;

[0337] d x4 Indicates the second relative distance of the target point cloud;

[0338] For the source point cloud translation filtering feature and the target point cloud translation filtering feature, the position matrix is added on the basis of the query matrix, key matrix and value matrix to obtain the source point cloud translation position feature and the target point cloud translation position feature, where the formula is,

[0339]

[0340] Q5=G′ x,t W Q ;

[0341] K5=G′ x,t W k ;

[0342] V5=G′ x,t W V ;

[0343] P5=I t1 W P ;

[0344]

[0345] Q6=G′ y,t W Q ;

[0346] K6=G′ y,t W k ;

[0347] V6=G′ y,t W V ;

[0348] P6=I t2 W P ;

[0349] Among them, F′ x,t Represents the translation position feature of the source point cloud;

[0350] Q5 represents the fifth query matrix;

[0351] K5 represents the fifth bond matrix;

[0352] V5 represents the fifth value matrix;

[0353] P5 represents the fifth position matrix;

[0354] G′ x,t Indicates the source point cloud translation filtering feature;

[0355] F′ y,t Represents the translation position characteristics of the target point cloud;

[0356] G′ y,t Indicates the target point cloud translation filtering feature;

[0357] Q6 represents the sixth query matrix;

[0358] K6 represents the sixth bond matrix;

[0359] V6 represents the sixth value matrix;

[0360] P6 represents the sixth position matrix.

[0361] According to one embodiment of the present invention, based on the source point cloud translation position feature and the target point cloud translation position feature, a feature cross attention mechanism is used to calculate the attention weight between points, and the feature information from another point cloud is aggregated to obtain the source point cloud translation aggregation feature and the target point cloud translation aggregation feature, where the formula is,

[0362]

[0363] Q7=F′ x,t W Q ;

[0364] K7=F′ y,t W k ;

[0365] V7=F′ y,t W V ;

[0366]

[0367] Q8=F′ y,t W Q ;

[0368] K8=F′ x,t W k ;

[0369] V8=F′ x,t W V ;

[0370] Among them, F x,t Represents the translation aggregation feature of the source point cloud;

[0371] Q7 represents the seventh query matrix;

[0372] K7 represents the seventh bond matrix;

[0373] V7 represents the seventh value matrix;

[0374] F y,t Represents the translation aggregation features of the target point cloud;

[0375] Q8 represents the eighth query matrix;

[0376] K8 represents the eighth key matrix;

[0377] V8 represents the eighth value matrix.

[0378] According to one embodiment of the present invention, Figure 6 FIG. 1 is a structural diagram of a dual-branch feature interaction module of a method for registration of similar stacking point clouds in a hazardous chemicals warehouse according to an exemplary embodiment. Figure 6 As shown in the figure, the solution space of point cloud registration can be divided into two parts: rotation space and translation space. Due to the significant differences in the geometric characteristics of rotation and translation, this method adopts a dual-branch collaborative architecture and realizes cross-branch feature optimization through the cascade processing of feature difference and residual calculation. Among them, each branch dynamically analyzes the high-level features output by the other branch, extracts the geometric representation differences by feature difference operation, and then strengthens the spatial characteristic separation of rotation and translation components in rigid transformation by residual calculation. This mechanism constructs a feature mutual correction link between the two channels, which can not only suppress feature redundancy through difference comparison, but also explicitly expand the geometric distribution differences between rotation and translation space, thereby improving the feature representation differences of point clouds under different spatial transformations. Each branch uses the feature path information of the other branch to make point-level feature-specific adjustments to the features extracted by itself to suppress cross-branch redundancy. The i-th point feature of the source point cloud rotation aggregation feature and the i-th point feature of the source point cloud translation aggregation feature are used to make a difference to obtain the first source point cloud difference feature, where the formula is,

[0379]

[0380] in, Represents the difference feature of the first source point cloud;

[0381] Represents the i-th point feature of the source point cloud rotation aggregation feature;

[0382] Represents the i-th point feature of the source point cloud translation aggregation feature;

[0383] This application designs a residual mechanism to fuse the original features with the corrected features to ensure that the branch output contains both cross-path optimization information and retains the core characteristics of its own structure. This design reduces feature coupling while maintaining the independent representation capabilities of the rotation and translation branches, thereby enhancing the model's decoupling learning efficiency for complex geometric transformations. Based on the source point cloud rotation aggregation feature and the first source point cloud difference feature, the source point cloud rotation update feature is obtained, where the formula is:

[0384]

[0385] Among them, γ represents the fusion ratio learnable parameter;

[0386] Indicates the source point cloud rotation update feature;

[0387] Based on the source point cloud rotation update feature and the transposed source point cloud rotation update feature, the first adjacency matrix is obtained, and based on the normalized first adjacency matrix and the source point cloud rotation update feature, the source point cloud rotation final feature is obtained, where the formula is,

[0388]

[0389] Among them, F” x,r Represents the final feature of the source point cloud rotation;

[0390] A'1 represents the normalized first adjacency matrix;

[0391] f a Represents the fully connected layer function;

[0392] f u represents the activation function;

[0393] Use the i-th point feature of the source point cloud translation aggregation feature and the i-th point feature of the source point cloud rotation aggregation feature to make a difference and get the second source point cloud difference feature, where the formula is,

[0394]

[0395] in, Represents the difference feature of the second source point cloud;

[0396] Based on the source point cloud translation aggregation feature and the second source point cloud difference feature, the source point cloud translation update feature is obtained, where the formula is:

[0397]

[0398] in, Represents the source point cloud translation update feature;

[0399] Based on the source point cloud translation update feature and the transposed source point cloud translation update feature, the second adjacency matrix is obtained, and based on the normalized second adjacency matrix and the source point cloud translation update feature, the source point cloud translation final feature is obtained, where the formula is,

[0400]

[0401] Among them, F” x,t Represents the final feature of the source point cloud translation;

[0402] A'2 represents the normalized second adjacency matrix.

[0403] According to one embodiment of the present invention, the first target point cloud difference feature is obtained by performing a difference between the i-th point feature of the target point cloud rotation aggregation feature and the i-th point feature of the target point cloud translation aggregation feature, where the formula is:

[0404]

[0405] in, Represents the first target point cloud difference feature;

[0406] Represents the i-th point feature of the target point cloud rotation aggregation feature;

[0407] Represents the i-th point feature of the target point cloud translation aggregation feature;

[0408] Based on the target point cloud rotation aggregation feature and the first target point cloud difference feature, the target point cloud rotation update feature is obtained, where the formula is:

[0409]

[0410] in, Indicates the target point cloud rotation update feature;

[0411] Based on the target point cloud rotation update feature and the transposed target point cloud rotation update feature, the third adjacency matrix is obtained, and based on the normalized third adjacency matrix and the target point cloud rotation update feature, the target point cloud rotation final feature is obtained, where the formula is,

[0412]

[0413] Among them, F” y,r Represents the final feature of the target point cloud rotation;

[0414] A'3 represents the normalized third adjacency matrix;

[0415] Use the i-th point feature of the target point cloud translation aggregation feature and the i-th point feature of the target point cloud rotation aggregation feature to make a difference and get the second target point cloud difference feature, where the formula is,

[0416]

[0417] in, Represents the second target point cloud difference feature;

[0418] Based on the target point cloud translation aggregation feature and the second target point cloud difference feature, the target point cloud translation update feature is obtained, where the formula is:

[0419]

[0420] in, Represents the target point cloud translation update feature;

[0421] Based on the target point cloud translation update feature and the transposed target point cloud translation update feature, the fourth adjacency matrix is obtained, and based on the normalized fourth adjacency matrix and the target point cloud translation update feature, the target point cloud translation final feature is obtained, where the formula is,

[0422]

[0423] Among them, F” y,t Represents the final feature of the target point cloud translation;

[0424] A'4 represents the normalized second adjacency matrix.

[0425] According to one embodiment of the present invention, this application uses two datasets for experiments: a self-built dataset of hazardous chemicals and a public dataset of ModelNet40. ModelNet40 is a classic dataset in the field of point cloud registration, with 12,311 CAD models and a total of 40 different categories of objects. The simulated hazardous chemicals warehouse model is 3.5 meters long, 1.7 meters wide, and 1.7 meters high, with facilities such as explosion-proof lights and yellow lines for stacking distance warnings. There are also models of hazardous chemicals such as cartons, wooden boxes, and oil tanks. The dataset covers 36 different scenes, including neatly arranged scenes with similar geometric structures, such as oil tanks, wooden boxes, and stacks, as well as some messy scenes. Each category of scenes is collected from different perspectives to ensure sufficient coverage of point cloud scenes at different angles.

[0426] This application conducts comprehensive comparative experiments with learning-based point cloud registration algorithms DCP, PRNet, DeepGMR, and FINet on self-built and public datasets, focusing on verifying their performance under unseen scenes and partially overlapping point clouds (partial-to-partial) conditions. This application uses the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) as the evaluation index: When RMSE and MAE approach 0 and R 2 When it approaches 1, it indicates that the rotation matrix and translation vector estimated by the rigid transformation are more accurate and the registration effect is better.

[0427] According to one embodiment of the present invention, in an experiment under an unseen category, the present application trains a model on the first 20 categories and evaluates it on a validation set of another 20 categories to test the robustness of the trained model in registration on unknown data. The predicted rotation and translation matrices are compared with the true values of the rotation and translation matrices using evaluation indicators, as shown in Table 1.

[0428] Table 1 Registration results of public datasets without seeing species

[0429] Model RMSE(R) MAE(R) R2(R) RMSE(t) MAE(t) R2(t) DCP 6.541682 4.409162 0.743471 0.028090 0.021503 0.990493 prnet 3.580264 1.804074 0.924199 0.010592 0.008063 0.998632 DeepGMR 7.928523 4.207549 0.622921 0.003873 0.002570 0.999818 FINet 4.123788 2.558949 0.897849 0.009244 0.006495 0.998972 This application 0.023906 0.013683 0.999997 0.001126 0.000607 0.999985

[0430] On the self-built dataset, unlike the public dataset, the experiment is trained on the first 80% of the categories and evaluated on the validation set of the last 20% of the categories. The predicted rotation and translation matrices are compared with the true values of the rotation and translation matrices using evaluation indicators, as shown in Table 2.

[0431] Table 2 Registration effect of self-built dataset when no species are seen

[0432] Model RMSE(R) MAE(R) R2(R) RMSE(t) MAE(t) R2(t) DCP 16.069927 8.666111 -0.646582 0.227168 0.136299 0.362123 prnet 7.066190 4.162383 0.682730 0.185269 0.105002 0.586234 DeepGMR 3.319152 2.066561 0.929596 0.085727 0.045318 0.912517 FINet 11.033032 9.137153 0.232469 0.143757 0.115104 0.745365 This application 0.128996 0.091966 0.999895 0.004033 0.002883 0.999799

[0433] It can be seen that this application shows significant advantages in unseen category scenarios and achieves the best registration effect in two datasets compared with other algorithms. The improvement of rotation matrix accuracy is particularly prominent. The rotation mean absolute error (MAE) of the public dataset is greatly improved from 2.5589 to 1.37×10 -2This breakthrough is mainly due to the synergy of the rotation and translation decoupling design and the relative position information embedding mechanism - by separating the feature learning paths of rotation and translation, the extraction of rotation-related point cloud features is targeted and strengthened, and at the same time, relative position information is embedded to constrain the spatial invariance of rotation features, thereby significantly improving the estimation accuracy of the rotation matrix. In contrast, other algorithms rely on local geometric structure matching. In scenes with a large number of similar stacks, such as hazardous chemical warehouses, it is difficult to distinguish local similar structures. This application uses global relative position information modeling to capture subtle geometric differences to enhance generalization capabilities and maintain high accuracy in public datasets. This application achieves stable registration performance across scenes and categories through the dual optimization of decoupled learning and position invariance constraints, showing stronger robustness in complex geometric environments.

[0434] According to one embodiment of the present invention, under the condition of partially overlapping point clouds, in the modelnet40 experiment, a plane is randomly sampled and moved to retain 80% of the point cloud, that is, 819 points out of 1024 points. This simulates the situation in real scenes where the occlusion of two-angle point clouds is different to test the robustness of the algorithm under the condition of partial overlap of point clouds. The registration effect of the public dataset under the condition of partially overlapping point clouds is shown in Table 3.

[0435] Table 3 Registration results of public datasets under partially overlapping point clouds

[0436] Model RMSE(R) MAE(R) R2(R) RMSE(t) MAE(t) R2(t) DCP 4.1820 2.7480 0.8961 0.02066 0.01492 0.99482 prnet 2.7203 1.5232 0.9560 0.01689 0.01142 0.99654 DeepGMR 7.9381 4.4034 0.6264 0.02857 0.02000 0.99009 FINet 4.0227 2.5616 0.9038 0.01481 0.01058 0.99734 This application 0.599557 0.361392 0.997872 0.03320 0.02183 0.98660

[0437] On the self-built dataset, this application is trained on the first 80% of the categories and evaluated on the validation set of the last 20% of the categories. This part of the setting is the same as the case of unseen categories. On this basis, this application also randomly samples a plane and moves it to retain 80% of the point cloud, that is, retaining 819 points out of 1024 points. This simulates the situation in which the occlusion parts of two point clouds at different angles are different in real scenes. Under the condition of partially overlapping point clouds, as shown in Table 4,

[0438] Table 4 Registration effect of self-built dataset under the condition of partially overlapping point clouds

[0439] Model RMSE(R) MAE(R) R2(R) RMSE(t) MAE(t) R2(t) DCP 10.979244 8.219366 0.235005 0.171758 0.125639 0.638984 prnet 9.261734 5.483089 0.463417 0.206039 0.133191 0.474095 DeepGMR 9.727736 3.552378 0.400510 0.197213 0.077747 0.516180 FINet 12.195807 10.242543 0.067012 0.192143 0.159759 0.543893 This application 1.755124 1.087534 0.980435 0.062645 0.039664 0.950564

[0440] This application achieves optimal rotation estimation accuracy on both self-built and public datasets. This is due to the embedding mechanism of global relative position information. By modeling the stable spatial relationship between point clouds, the geometric invariance of the rotation feature is effectively enhanced. In the more challenging hazardous chemical warehouse scenario, the registration accuracy of other algorithms drops significantly; however, this method can maintain robust registration performance even in complex storage environments by integrating global relative position constraints with rotation-translation decoupling design. Experiments have shown that the introduction of relative position information not only strengthens the stability of rotation estimation, but also improves the model's ability to discriminate local similar structures through cross-level feature interaction, thereby achieving a double breakthrough in accuracy and generalization capabilities in a variety of scenarios.

Claims

1. A method for registering similar stack point clouds in a hazardous chemicals warehouse, characterized by: include: Obtain source point cloud data of hazardous chemicals warehouse stacks; Preprocess the source point cloud data of hazardous chemicals warehouse stacks to obtain the target point cloud data of hazardous chemicals warehouse stacks; The source point cloud data of the hazardous chemicals warehouse stack and the target point cloud data of the hazardous chemicals warehouse stack are input into the feature extraction module, which includes a rotation extraction network branch and a translation extraction network branch to obtain the source point cloud rotation feature, the target point cloud rotation feature, the source point cloud translation feature and the target point cloud translation feature; The source point cloud rotation features, target point cloud rotation features, source point cloud translation features, and target point cloud translation features are input into the point cloud feature interaction module. The soft matching algorithm is used to extract the points with high similarity between the two point clouds as key points, and the points with the most similar features are used as reference points. After screening the source point cloud rotation features, target point cloud rotation features, source point cloud translation features and target point cloud translation features based on key points and reference points, the position matrix is introduced and feature aggregation is performed to obtain the source point cloud rotation aggregation features, target point cloud rotation aggregation features, source point cloud translation aggregation features and target point cloud translation aggregation features; The source point cloud rotation aggregation feature, the target point cloud rotation aggregation feature, the source point cloud translation aggregation feature and the target point cloud translation aggregation feature are input into the dual-branch feature interaction module, and the source point cloud rotation final feature, the target point cloud rotation final feature, the source point cloud translation final feature and the target point cloud translation final feature are obtained through the cascade processing of feature difference and residual calculation; Point cloud registration is performed based on the final rotation features of the source point cloud, the final rotation features of the target point cloud, the final translation features of the source point cloud, and the final translation features of the target point cloud.

2. The method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 1, characterized in that: The rotation extraction network branch and the translation extraction network branch are set in parallel and have the same network structure, which includes a six-layer graph convolutional network and a three-layer self-attention mechanism embedded in the graph convolutional network; The source point cloud data of the hazardous chemicals warehouse stacking and the target point cloud data of the hazardous chemicals warehouse stacking are respectively input into the rotation extraction network branch to obtain the source point cloud rotation feature and the target point cloud rotation feature. The source point cloud data of the hazardous chemicals warehouse stacking and the target point cloud data of the hazardous chemicals warehouse stacking are respectively input into the translation extraction network branch to obtain the source point cloud translation feature and the target point cloud translation feature. The formula is, F1=Self-attention(GCN(U)); F2=GCN(GCN(Self-attention(GCN(F1)))); F3=Self-attention(GCN(F2)); U1=GCN(conact(F1,F2,F3)); Among them, U represents the input of the network; GCN stands for Graph Convolutional Network; Self-attention means self-attention mechanism; F1 represents the first output feature; F2 represents the second output feature; F3 represents the third output feature; conact means feature fusion; U1 represents the output of the network.

3. The method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 2, wherein: The rotation affinity matrix is obtained based on the transpose of the source point cloud rotation feature and the target point cloud rotation feature, where the formula is: B r =(G x,r ) T W(G y,r ); Among them, B r represents the rotation affinity matrix; (G x,r ) T Represents the transpose of the rotation feature of the source point cloud; (G y,r ) represents the target point cloud rotation feature; W represents the learnable parameters of the affinity layer; Based on the rotation affinity matrix, the rotation soft correspondence matrix is obtained by alternating row and column normalization. The rotation soft correspondence matrix is used to transform the target point cloud rotation feature to obtain the rotation transformation feature, where the formula is: F tran,r =C r *G y,r ; Among them, F tran,r Represents the rotation transformation feature; C r represents the rotation soft correspondence matrix; The feature difference is calculated based on the rotation feature and rotation transformation feature of the target point cloud, and the L2 norm of the difference is taken. The k=M smallest points are taken as the rotation key points, and the point pair with the smallest difference is taken as the common rotation reference point of the point cloud to be registered. The formula is: Among them, E1 represents the set of rotation key points; x k Indicates the point cloud coordinates corresponding to the key point index in the source point cloud; y k Indicates the point cloud coordinates corresponding to the key point index in the target point cloud; n represents the index of the source point cloud; j represents the index of the target point cloud; Represents the first M feature index pairs that minimize the feature distance; The source point cloud rotation features and the target point cloud rotation features are filtered based on the key points to obtain the source point cloud rotation filtering features and the target point cloud rotation filtering features.

4. A method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 3, characterized in that: The translation affinity matrix is obtained based on the transpose of the source point cloud translation feature and the target point cloud translation feature, where the formula is: B t =(G x,t ) T W(G y,t ); Among them, B t represents the translation affinity matrix; (G x,t ) T Represents the transpose of the translation feature of the source point cloud; (G y,t ) represents the translation feature of the target point cloud; Based on the translation affinity matrix, the soft translation correspondence matrix is obtained by alternating row and column normalization. The soft translation correspondence matrix is used to transform the translation features of the target point cloud to obtain the translation transformation features, where the formula is: F tran,t =C t *G y,t ; Among them, F tran,t Represents translation transformation features; C t represents the translation soft correspondence matrix; The feature difference is calculated based on the translation feature and translation transformation feature of the target point cloud, and the L2 norm of the difference is taken. The k=M smallest points are taken as the translation key points, and the point pair with the smallest difference is taken as the common translation reference point of the point cloud to be registered. The formula is: Among them, E2 represents the translation key point set; The source point cloud translation features and the target point cloud translation features are filtered based on the key points to obtain the source point cloud translation filtering features and the target point cloud translation filtering features.

5. The method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 4, characterized in that: The coordinates of each point of the source point cloud and the target point cloud after the rotation branch screening are calculated relative to the rotation reference point to obtain the first relative distance of the source point cloud and the first relative distance of the target point cloud. The sine function is used to calculate the source point cloud rotation distance embedding information and the target point cloud rotation distance embedding information based on the first relative distance of the source point cloud and the first relative distance of the target point cloud, where the formula is: Among them, I r1 Represents the source point cloud rotation distance embedding information; I r2 Represents the target point cloud rotation distance embedding information; sin represents the sine function; d x1 Indicates the first relative distance of the source point cloud; d x2 Indicates the first relative distance of the target point cloud; d represents the dimension index of the embedding vector; D represents the total dimension of the embedding vector; For the source point cloud rotation filtering feature and the target point cloud rotation filtering feature, the position matrix is added on the basis of the query matrix, key matrix and value matrix to obtain the source point cloud rotation position feature and the target point cloud rotation position feature, where the formula is, Q1=G′ x,r W Q ; K1=G′ x,r IN k ; V1=G′ x,r W V ; P1=I r1 IN P ; <h2 style=";text-align:left;direction:ltr">Q2=G′<h2 style=";text-align:left;direction:ltr"> y,r <h2 style=";text-align:left;direction:ltr"> W<h2 style=";text-align:left;direction:ltr"> Q <h2 style=";text-align:left;direction:ltr"> ; K2=G′ y,r W k ; <h2 style=";text-align:left;direction:ltr">V2=G′<h2 style=";text-align:left;direction:ltr"> y,r <h2 style=";text-align:left;direction:ltr"> W<h2 style=";text-align:left;direction:ltr"> V <h2 style=";text-align:left;direction:ltr"> ; P2=I r2 W P ; Among them, softmax represents the normalized exponential function; F′ x,r Represents the rotation position feature of the source point cloud; Q1 represents the first query matrix; K1 represents the first bond matrix; V1 represents the first value matrix; P1 represents the first position matrix; T stands for transpose; W Q represents the query weight matrix; W k represents the bond weight matrix; W V represents the value weight matrix; W P represents the position weight matrix; d k represents the feature dimension; G′ x,r Represents the source point cloud rotation screening feature; F′ y,r Represents the rotation position characteristics of the target point cloud; G′ y,r Indicates the target point cloud rotation screening feature; Q2 represents the second query matrix; K2 represents the second bond matrix; V2 represents the second value matrix; P2 represents the second position matrix.

6. A method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 5, characterized in that: Based on the source point cloud rotation position feature and the target point cloud rotation position feature, the feature cross attention mechanism is used to calculate the attention weight between points and aggregate the feature information from another point cloud to obtain the source point cloud rotation aggregation feature and the target point cloud rotation aggregation feature. The formula is: <h2 style=";text-align:left;direction:ltr">Q3=F′<h2 style=";text-align:left;direction:ltr"> x,r <h2 style=";text-align:left;direction:ltr"> W<h2 style=";text-align:left;direction:ltr"> Q <h2 style=";text-align:left;direction:ltr"> ; <h2 style=";text-align:left;direction:ltr">K3=F′<h2 style=";text-align:left;direction:ltr"> y,r <h2 style=";text-align:left;direction:ltr"> W<h2 style=";text-align:left;direction:ltr"> k <h2 style=";text-align:left;direction:ltr"> ; V3=F′ y,r W V ; Q4=F′ y,r W Q ; K4=F′ x,r W k ; V4=F′ x,r W V ; Among them, F x,r Represents the rotation aggregation feature of the source point cloud; Q3 represents the third query matrix; K3 represents the third bond matrix; V3 represents the third value matrix; F y,r Represents the rotation aggregation features of the target point cloud; Q4 represents the fourth query matrix; K4 represents the fourth bond matrix; V4 represents the fourth value matrix.

7. A method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 6, characterized in that: The coordinates of each point of the source point cloud and the target point cloud after the translation branch screening are calculated relative to the translation reference point to obtain the second relative distance of the source point cloud and the second relative distance of the target point cloud. The sine function is used to calculate the source point cloud translation distance embedding information and the target point cloud translation distance embedding information based on the second relative distance of the source point cloud and the second relative distance of the target point cloud, where the formula is: Among them, I t1 Represents the embedded information of the source point cloud translation distance; I t2 Represents the target point cloud translation distance embedding information; d x3 Indicates the second relative distance of the source point cloud; d x4 Indicates the second relative distance of the target point cloud; For the source point cloud translation filtering feature and the target point cloud translation filtering feature, the position matrix is added on the basis of the query matrix, key matrix and value matrix to obtain the source point cloud translation position feature and the target point cloud translation position feature, where the formula is, Q5=G′ x,t W Q ; K5=G′ x,t W k ; <h2 style=";text-align:left;direction:ltr">V5=G′<h2 style=";text-align:left;direction:ltr"> x,t <h2 style=";text-align:left;direction:ltr"> W<h2 style=";text-align:left;direction:ltr"> V <h2 style=";text-align:left;direction:ltr"> ; P5=I t1 IN P ; Q6=G′ y,t W Q ; K6=G′ y,t W k ; V6=G′ y,t W V ; P6=I t2 W P ; Among them, F′ x,t Represents the translation position feature of the source point cloud; Q5 represents the fifth query matrix; K5 represents the fifth bond matrix; V5 represents the fifth value matrix; P5 represents the fifth position matrix; G′ x,t Indicates the source point cloud translation filtering feature; F′ y,t Represents the translation position characteristics of the target point cloud; G′ y,t Indicates the target point cloud translation filtering feature; Q6 represents the sixth query matrix; K6 represents the sixth bond matrix; V6 represents the sixth value matrix; P6 represents the sixth position matrix.

8. The method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 7, wherein: Based on the source point cloud translation position feature and the target point cloud translation position feature, the feature cross attention mechanism is used to calculate the attention weight between points and aggregate the feature information from another point cloud to obtain the source point cloud translation aggregation feature and the target point cloud translation aggregation feature. The formula is: Q7=F′ x,t W Q ; K7=F′ y,t IN k ; V7=F′ y,t W V ; <h2 style=";text-align:left;direction:ltr">Q8=F′<h2 style=";text-align:left;direction:ltr"> y,t <h2 style=";text-align:left;direction:ltr"> W<h2 style=";text-align:left;direction:ltr"> Q <h2 style=";text-align:left;direction:ltr"> ; K8=F′ x,t W k ; V8=F′ x,t W V ; Among them, F x,t Represents the translation aggregation feature of the source point cloud; Q7 represents the seventh query matrix; K7 represents the seventh bond matrix; V7 represents the seventh value matrix; F y,t Represents the translation aggregation features of the target point cloud; Q8 represents the eighth query matrix; K8 represents the eighth key matrix; V8 represents the eighth value matrix.

9. A method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 8, characterized in that: Use the i-th point feature of the source point cloud rotation aggregation feature and the i-th point feature of the source point cloud translation aggregation feature to make a difference and get the first source point cloud difference feature, where the formula is, in, Represents the difference feature of the first source point cloud; Represents the i-th point feature of the source point cloud rotation aggregation feature; Represents the i-th point feature of the source point cloud translation aggregation feature; Based on the source point cloud rotation aggregation feature and the first source point cloud difference feature, the source point cloud rotation update feature is obtained, where the formula is: Among them, γ represents the fusion ratio learnable parameter; Indicates the source point cloud rotation update feature; Based on the source point cloud rotation update feature and the transposed source point cloud rotation update feature, the first adjacency matrix is obtained, and based on the normalized first adjacency matrix and the source point cloud rotation update feature, the source point cloud rotation final feature is obtained, where the formula is, Among them, F” x,r Represents the final feature of the source point cloud rotation; A'1 represents the normalized first adjacency matrix; f a Represents the fully connected layer function; f u represents the activation function; Use the i-th point feature of the source point cloud translation aggregation feature and the i-th point feature of the source point cloud rotation aggregation feature to make a difference and get the second source point cloud difference feature, where the formula is, in, Represents the difference feature of the second source point cloud; Based on the source point cloud translation aggregation feature and the second source point cloud difference feature, the source point cloud translation update feature is obtained, where the formula is: in, Represents the source point cloud translation update feature; Based on the source point cloud translation update feature and the transposed source point cloud translation update feature, the second adjacency matrix is obtained, and based on the normalized second adjacency matrix and the source point cloud translation update feature, the source point cloud translation final feature is obtained, where the formula is, Among them, F” x,t Represents the final feature of the source point cloud translation; A'2 represents the normalized second adjacency matrix.

10. A method for registering similar stack point clouds in a hazardous chemicals warehouse according to claim 9, characterized in that: Use the i-th point feature of the target point cloud rotation aggregation feature and the i-th point feature of the target point cloud translation aggregation feature to make a difference and get the first target point cloud difference feature, where the formula is, in, Represents the first target point cloud difference feature; Represents the i-th point feature of the target point cloud rotation aggregation feature; Represents the i-th point feature of the target point cloud translation aggregation feature; Based on the target point cloud rotation aggregation feature and the first target point cloud difference feature, the target point cloud rotation update feature is obtained, where the formula is: in, Indicates the target point cloud rotation update feature; Based on the target point cloud rotation update feature and the transposed target point cloud rotation update feature, the third adjacency matrix is obtained, and based on the normalized third adjacency matrix and the target point cloud rotation update feature, the target point cloud rotation final feature is obtained, where the formula is, Among them, F” y,r Represents the final feature of the target point cloud rotation; A'3 represents the normalized third adjacency matrix; Use the i-th point feature of the target point cloud translation aggregation feature and the i-th point feature of the target point cloud rotation aggregation feature to make a difference and get the second target point cloud difference feature, where the formula is, in, Represents the second target point cloud difference feature; Based on the target point cloud translation aggregation feature and the second target point cloud difference feature, the target point cloud translation update feature is obtained, where the formula is: in, Represents the target point cloud translation update feature; Based on the target point cloud translation update feature and the transposed target point cloud translation update feature, the fourth adjacency matrix is obtained, and based on the normalized fourth adjacency matrix and the target point cloud translation update feature, the target point cloud translation final feature is obtained, where the formula is, Among them, F” y,t Represents the final feature of the target point cloud translation; A'4 represents the normalized second adjacency matrix.

Images