Intelligent Classification and Judgment Method for Scrap Steel Piles Based on Mobile LiDAR Point Cloud Data

Abstract

This invention belongs to the field of scrap steel classification technology, specifically relating to an intelligent classification and determination method for scrap steel piles based on mobile lidar point cloud data. The method includes the following steps: Step 1, acquiring point cloud data from the scrap steel unloading site; Step 2, preprocessing the acquired point cloud data; Step 3, classifying the scrap steel according to the preprocessed point cloud data using a pre-defined classification model and a scrap steel classification label library. The classification model incorporates an EdgeConv layer based on the PointNet point cloud classification model to extract neighboring point cloud features; the classification label library stores predefined scrap steel type level labels. This method can accurately obtain the specific classification results for scrap steel.

Description

Technical Field

[0001] This invention belongs to the field of scrap steel classification technology, and in particular relates to an intelligent classification and determination method for scrap steel piles based on mobile lidar point cloud data. Background Technology

[0002] Scrap steel recycling primarily involves the recovery and reuse of metal resources. This not only reduces the consumption of natural resources and environmental pollution but also creates employment opportunities and improves socio-economic benefits. The scrap steel recycling industry is a vital pillar of the national economy, playing a crucial role in resource and environmental protection.

[0003] Traditional scrap steel sorting relies primarily on manual labor, which is labor-intensive and inefficient, failing to meet the demands of fast and efficient production. Furthermore, manual sorting is susceptible to subjectivity and fatigue, potentially affecting accuracy and leading to a decline in scrap steel quality. With the development of deep learning technology, its integration with various industries is becoming increasingly close. How to utilize deep neural networks to improve scrap steel sorting efficiency has become a crucial issue for the scrap steel industry.

[0004] Patent document CN 114782773 A discloses a method for grading unbalanced scrap steel samples based on a tree-like hierarchical network. For example... Figure 1 As shown, this method constructs a correspondence table between scrap steel types and classification grades, and builds a conventional, deep learning-based scrap steel classification model. This addresses the severe long-tail effect present in existing scrap steel classification datasets, thereby avoiding severely biased predictions and improving the overall unbiasedness of the model's predictions. By fusing various representative features stored in the dynamic memory module, the feature representation of the current target is enhanced, further improving the overall prediction accuracy of the model. A tree-based classification network is used to reveal significant and subtle differences between different categories, improving the model's overall generalization ability to tail predicates.

[0005] Patent document CN 114821256 A discloses a scrap steel classification method based on small-objective data augmentation and multi-perspective collaborative reasoning. For example... Figure 2 As shown, this method establishes a multi-view scrap steel image dataset, designs a cropping strategy for high-resolution images, generates feature maps, extracts consistency and complementary feature information between scrap steel images from different perspectives, and achieves classification, detection, and segmentation of small scrap steel, as well as prediction of input images, obtaining classification, detection, and segmentation results. The cropping strategy designed for high-resolution images enhances the image by cropping the input images, increasing the area proportion of small scrap steel in each image. This helps extract sufficient features for recognition, improving the detection rate and classification accuracy of small scrap steel.

[0006] These two patent documents respectively utilize the imbalance of tree-like hierarchical networks and small-data augmentation and multi-view collaboration to improve the accuracy of scrap steel classification. While they improve the accuracy of scrap steel classification to some extent, they still have limitations. First, the above methods can only handle small datasets, which is not practical in actual scrap steel recycling. Furthermore, image data cannot fully reflect all the information of the scrap steel, ultimately affecting the classification results. Second, neither document provides a final grading and classification criterion for the scrap steel; the specific grading and classification cannot be determined based on the extracted scrap steel feature information.

[0007] In summary, how to accurately obtain the specific grading and classification results of scrap steel has become an urgent problem to be solved. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention provides an intelligent classification and determination method for scrap steel piles based on mobile lidar point cloud data, which can accurately obtain the specific grading and classification results of scrap steel.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] The intelligent classification and determination method for scrap steel piles based on mobile LiDAR point cloud data includes the following steps:

[0011] Step 1: Obtain point cloud data of the scrap steel unloading site;

[0012] Step 2: Preprocess the acquired point cloud data;

[0013] Step 3: Using a pre-defined classification model and a scrap steel classification label library, scrap steel is classified according to the pre-processed point cloud data. The classification model introduces an EdgeConv layer on the basis of the PointNet point cloud classification model to extract neighbor point cloud features. The classification label library stores pre-defined scrap steel type level labels.

[0014] Preferably, in step 2, the preprocessing includes mesh voxelization, nearest neighbor statistical analysis, and point cloud registration in sequence.

[0015] Preferably, in step 2, the nearest neighbor statistical analysis process includes:

[0016] Suppose any point x in the point cloud set Q i The three-dimensional coordinates are (x i y i , z i ); where i = 1, 2, ..., n; n is the number of points in the point cloud; use a kd-tree to search for each point x in the point cloud set. i The k-neighborhood;

[0017] Assume sampling point x i A point x in the k-neighborhood of . ij Where j = 1, 2, ..., k; k is the sampling point x i The number of points in the k-neighborhood of x; ij The three-dimensional coordinates are (x ij y ij , z ij ); Calculate the distance from each point in the k-neighborhood to the sampling point x. i Average distance

[0018]

[0019] Calculate the set of average distances again Follows a Gaussian distribution. In the formula, μ D for The mean, The corresponding variance;

[0020]

[0021] If the calculation yields Not in T∈(μ D -ξσ D μ D +ξσ D If the range is within a certain range, the corresponding point cloud data is identified as a noise point and removed; where ξσ D This indicates a deviation from the threshold, where the value of ξ is determined based on experience and experimentation.

[0022] Preferably, in step 2, the iterative nearest-point algorithm is used for point cloud registration; during point cloud registration, assuming the source point cloud is P and the target point cloud is Q, the point cloud registration process includes:

[0023] 1) In two point clouds P and Q, let point p in point cloud P be... i Choose the point q that is closest to it. i ;

[0024] 2) When the error function is minimized, solve for the optimal values ​​of the rotation matrix R and the translation vector t, and denote the current minimum error function value as ε1; the error function is:

[0025]

[0026] In the formula, N p The value represents the number of point clouds, and T represents the translation matrix.

[0027] 3) Apply the optimal rotation matrix R and translation vector t to the point cloud P to obtain the point cloud P':

[0028] P` = RP + t ；

[0029] 4) Replace point cloud P with point cloud P' and repeat steps 1)-3); the absolute value of the difference between the two error values ​​is |ε k -ε k-1 If the number of iterations is less than the threshold or greater than the preset maximum number of iterations, stop the iteration and use the final point cloud as the registered point cloud; where k≥2.

[0030] Preferably, in step 3, the working process of the hierarchical classification model includes:

[0031] S1. Spatial transformation processing of point cloud data is performed through the spatial transformation module;

[0032] S2. Extract the features of neighboring point clouds from the spatially transformed point cloud data through the EdgeConv layer to obtain the corresponding feature vectors; the features of neighboring point clouds include the three-dimensional coordinates, color, layer, and surface normals of the point clouds.

[0033] S3. Pool the feature vector obtained in S2 to obtain the feature representation of the center point; obtain the global features of the point cloud through a max pooling layer, and map the global features to the hierarchical classification label library through a fully connected layer to obtain the corresponding scrap steel type level label, which serves as the hierarchical classification result of scrap steel.

[0034] Preferably, S2 includes:

[0035] S2.1, Represent the point cloud data as an undirected graph using the EdgeConv convolutional layer; Define the local undirected point cloud graph as... in, Represents a set of point clouds. n is the number of points in the point cloud; ε represents the edge set, e ij ∈ε, j∈(1,2,...,k), k represents point v i The number of edges in the k-neighborhood;

[0036] S2.2. Within a local area, feature extraction is performed on the point cloud data using a multilayer perceptron; for each point v i Let the edges e in its k-neighborhood ij The features are extracted using a shared neural network to obtain the corresponding feature vectors.

[0037] Preferably, the specific process of S2.2 is as follows:

[0038] For point v iDefine point cloud x i and x j edge features e ij =h Θ (x i x j ), where h Θ :R F ×R F ;R F h is a nonlinear function with a set of learnable parameters Θ; h is calculated using an asymmetric edge function. Θ :

[0039] h Θ (x i x j ) = h Θ (x i x j -x i );

[0040] Among m different filters, there are edge features:

[0041] e ijm =ReLU(θ) m .(x i -x i )+φ m ·x i );

[0042] Where, θ m and φ m Represents the learning parameters under different filters, Θ=(θ1, θ2, ..., θ M φ1, φ2, ..., φ m );

[0043] ReLU(·) represents the activation function;

[0044] Then, the output of the i-th point cloud is obtained by aggregating the edge features of the i-th point cloud:

[0045]

[0046] Preferably, in step 3, during the training process of the hierarchical classification model, the loss calculation uses the cross-entropy loss function:

[0047]

[0048] Among them, y i For real labels, For predicted labels.

[0049] Preferably, in step 1, point cloud data of the scrap steel unloading site is acquired by using a lidar mounted on a drone.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. This method incorporates an EdgeConv layer into existing deep neural network classification models. This helps the model capture the spatial relationships between different points in point cloud data, better extracting local features and improving the accuracy of scrap steel grading and classification. Furthermore, this method provides the specific working process of the grading and classification model, as well as its synergistic effect with the scrap steel grading and classification label library. Using this method, after acquiring and preprocessing the point cloud data from the scrap steel unloading site, the grading and classification results of the scrap steel can be accurately obtained through the pre-set grading and classification model and the scrap steel grading and classification label library.

[0052] 2. To enable the grading and classification of scrap steel, this method pre-stores specific scrap steel type labels (equivalent to establishing a relatively comprehensive scrap steel grading and classification standard). In practice, after extracting point cloud features based on a deep neural network-based classification model, the specific grading and classification results can be obtained according to the scrap steel type level labels defined in the scrap steel grading and classification label library.

[0053] 3. In this invention, the preprocessing of the acquired point cloud data includes, in sequence, mesh voxelization, nearest neighbor statistical analysis, and point cloud registration. Mesh voxelization reduces the amount of point cloud data to be processed while preserving the object's shape characteristics, thus improving overall processing efficiency. Nearest neighbor statistical analysis in this method removes discrete noise from the scrap steel point cloud data, ensuring the accuracy of subsequent processing. Furthermore, since the acquired point cloud data may come from different devices or different perspectives of the same device, point cloud registration in this method unifies the point cloud data into a specified coordinate system, further ensuring the accuracy of subsequent processing.

[0054] In summary, this method can accurately obtain the specific grading and classification results of scrap steel. Attached Figure Description

[0055] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0056] Figure 1 This is a schematic diagram of the first prior art patent in the background art of this invention;

[0057] Figure 2 This is a schematic diagram of the second prior art patent in the background section of this invention;

[0058] Figure 3 Flowchart for an embodiment;

[0059] Figure 4 This is a schematic diagram of the scrap steel stockpile point cloud data acquisition device in the embodiment;

[0060] Figure 5 This is a schematic diagram of the point cloud data preprocessing process in the embodiment;

[0061] Figure 6 This is a schematic diagram of the workflow of the hierarchical classification model in the embodiment;

[0062] Figure 7 This is a schematic diagram illustrating the scrap steel grading and classification labels in the example. Detailed Implementation

[0063] The following detailed explanation illustrates the specific implementation methods:

[0064] Example:

[0065] like Figure 3 As shown in the figure, this embodiment discloses an intelligent classification and determination method for scrap steel piles based on mobile lidar point cloud data, including the following steps:

[0066] Step 1: Obtain point cloud data of the scrap steel unloading site. For example... Figure 4 In practice, point cloud data of the scrap steel unloading site can be acquired using a lidar mounted on a drone. Those skilled in the art can also set up other data acquisition methods based on the specific conditions of the site, which will not be elaborated upon here.

[0067] Step 2: Preprocess the acquired point cloud data. For example... Figure 5 As shown, in specific implementation, the preprocessing includes mesh voxelization, nearest neighbor statistical analysis, and point cloud registration in sequence.

[0068] The initial point cloud data acquired from the lidar on the drone is large in volume and contains a lot of noise. This noise not only affects the computational efficiency of the scrap steel grading and classification model and increases memory consumption, but also affects the accuracy of the final scrap steel grading and classification. Therefore, the point cloud data of the scrap steel pile needs to be preprocessed before performing grading and classification based on point cloud data.

[0069] Mesh voxelization

[0070] To reduce the traversal time when inputting scrap metal stockpile data into a neural network, the point cloud data first needs to be simplified. Mesh voxelization involves creating a 3D voxel grid within the input target point cloud. Within each voxel of the grid, the centroids of all point cloud data within that voxel are used to replace all points within the voxel, and the centroids are extracted to form a simplified new point cloud dataset. Mesh voxelization reduces the amount of point cloud data to be processed while preserving the object's shape characteristics.

[0071] Nearest neighbor statistical analysis

[0072] Noise is often introduced during point cloud acquisition due to factors such as equipment and working environment. In order to avoid noise data affecting the accuracy of point cloud classification, it is necessary to reduce the noise of the simplified point cloud data.

[0073] In this embodiment, the nearest neighbor statistical analysis method is used to remove discrete noise from the scrap steel point cloud data. The specific steps are as follows:

[0074] Suppose any point x in the point cloud set Q i The three-dimensional coordinates are (x i y i , z i ); where i = 1, 2, ..., n; n is the number of points in the point cloud; use a kd-tree to search for each point x in the point cloud set. i The k-neighborhood;

[0075] Assume sampling point x i A point x in the k-neighborhood of . ij Where j = 1, 2, ..., k; k is the sampling point x i The number of points in the k-neighborhood of x; ij The three-dimensional coordinates are (x ij y ij , z ij ); Calculate the distance from each point in the k-neighborhood to the sampling point x. i Average distance

[0076]

[0077] Calculate the set of average distances again Follows a Gaussian distribution. In the formula, μ D for The mean, The corresponding variance;

[0078]

[0079] If the calculation yields Not in T∈(μ D -ξσ D μ D +ξσ D If the range is within a certain range, the corresponding point cloud data is identified as a noise point and removed; where ξσ D This indicates a deviation from the threshold, where the value of ξ is determined based on experience and experimentation.

[0080] Point cloud registration

[0081] Since the collected point cloud data may come from different devices or different viewpoints of the same device, it is necessary to perform rotation, translation, and other transformations on the point cloud to integrate the data into a specified coordinate system. Specifically, this can be achieved using the Iterative Closest Point (ICP) algorithm. The ICP algorithm uses the point found in the origin set that is closest to the target point set, and through multiple iterations, minimizes the mean square distance metric between the corresponding points to determine the transformation matrix, thereby achieving registration of the two point clouds.

[0082] In practice, during point cloud registration, assuming the source point cloud is P and the target point cloud is Q, the point cloud registration process includes:

[0083] 1) In two point clouds P and Q, let point p in point cloud P be... i Choose the point q that is closest to it. i ;

[0084] 2) When the error function is minimized, solve for the optimal values ​​of the rotation matrix R and the translation vector t, and denote the current minimum error function value as ε1; the error function is:

[0085]

[0086] In the formula, N p The value represents the number of point clouds, and T represents the translation matrix.

[0087] 3) Apply the optimal rotation matrix R and translation vector t to the point cloud P to obtain the point cloud P':

[0088] P` = RP + t;

[0089] 4) Replace point cloud P with point cloud P' and repeat steps 1)-3); the absolute value of the difference between the two error values ​​is |ε k -ε k-1 If the number of iterations is less than the threshold or greater than the preset maximum number of iterations, stop the iteration and use the final point cloud as the registered point cloud; where k≥2.

[0090] Step 3: Using a pre-defined classification model and a scrap steel classification label library, scrap steel is classified according to the pre-processed point cloud data. The classification model introduces an EdgeConv layer on the basis of the PointNet point cloud classification model to extract neighbor point cloud features. The classification label library stores pre-defined scrap steel type level labels.

[0091] like Figure 6 As shown, the working process of the hierarchical classification model includes:

[0092] S1. Spatial transformation processing of point cloud data is performed through the spatial transformation module.

[0093] S2. Extract the features of neighboring point clouds from the spatially transformed point cloud data through the EdgeConv layer to obtain the corresponding feature vectors; the features of neighboring point clouds include the three-dimensional coordinates, color, layer, and surface normals of the point clouds.

[0094] In specific implementation, S2 includes:

[0095] S2.1, Represent the point cloud data as an undirected graph using the EdgeConv convolutional layer; Define the local undirected point cloud graph as... in, Represents a set of point clouds. n is the number of points in the point cloud; ε represents the edge set, e ij ∈ε, j∈(1,2,...,k), k represents point v i The number of edges in the k-neighborhood;

[0096] S2.2. Within a local area, feature extraction is performed on the point cloud data using a multilayer perceptron; for each point v i Let the edges e in its k-neighborhood ij The features are extracted using a shared neural network to obtain the corresponding feature vectors.

[0097] The specific process of S2.2:

[0098] For point v i Define point cloud x i and x j edge features e ij =h Θ (x i x j ), where h Θ :R F ×R F ;R F h is a nonlinear function with a set of learnable parameters Θ; h is calculated using an asymmetric edge function. Θ :

[0099] h Θ (x i x j ) = h Θ (x i x j -x i );

[0100] Among m different filters, there are edge features:

[0101] e ijm=ReLU(θ) m .(x i -x i )+φ m ·x i );

[0102] Where, θ m and φ m Represents the learning parameters under different filters, Θ=(θ1, θ2, ..., θ M φ1, φ2, ..., φ m );

[0103] ReLU(·) represents the activation function;

[0104] Then, the output of the i-th point cloud is obtained by aggregating the edge features of the i-th point cloud:

[0105]

[0106] During the training of the hierarchical classification model, the cross-entropy loss function is used for loss calculation:

[0107]

[0108] Among them, y i For real labels, For predicted labels.

[0109] S3. Pool the feature vector obtained in S2 to obtain the feature representation of the center point; obtain the global features of the point cloud through a max pooling layer, and map the global features to the hierarchical classification label library through a fully connected layer to obtain the corresponding scrap steel type level label, which serves as the hierarchical classification result of scrap steel.

[0110] To facilitate a better understanding of the graded classification label library in this invention, the definition of scrap steel graded classification labels in this embodiment is as follows: Figure 7 As shown.

[0111] In actual scrap steel unloading, scrap steel is classified into three categories: baled blocks, crushed scrap, and ring-shaped scrap. Baled blocks can be further divided into: pure material briquettes, ordinary briquettes, rebar briquettes, and newly formed briquettes based on their raw materials. Crushed scrap can be classified into four categories based on its particle size. Ring-shaped scrap is classified into four categories based on its maximum radius.

[0112] Based on the above, different types of scrap steel are labeled. Based on the scrap steel type, they can be defined as Block, Grain, and Circle.

[0113] Based on the different raw materials of the baled scrap steel, it can be divided into four categories: Block_A (pure material), Block_B (ordinary material), Block_C (reinforcing steel), and Block_D (new material).

[0114] Based on the degree of crushing of scrap steel, it is divided into four categories. Specifically, crushed material with a thickness of less than 10cm and a length of less than 50cm is classified as Category A, labeled as Grain_A; crushed material with a thickness of more than 10cm and less than 30cm and a length of more than 50cm and less than 100cm is classified as Category B, labeled as Grain_B; crushed material with a thickness of more than 30cm and less than 60cm and a length of more than 100cm and less than 150cm is classified as Category C, labeled as Grain_C; and the rest is classified as Category D, labeled as Grain_D.

[0115] Based on the circular shape of the scrap steel, it is divided into four categories according to its maximum radius: Category A circular scrap steel with a maximum radius of less than 30cm is labeled Circle_A; Category B circular scrap steel with a maximum radius of more than 30cm and less than 60cm is labeled Circle_B; Category C circular scrap steel with a maximum radius of more than 60cm and less than 90cm is labeled Circle_C; and the rest are Category D circular scrap steel, labeled Circle_D.

[0116] This method incorporates an EdgeConv layer into existing deep neural network classification models, which helps the model capture the spatial relationships between different points in point cloud data, better extract local features from the point cloud data, and improve the accuracy of scrap steel grading and classification. Furthermore, this method provides the specific working process of the grading and classification model and its synergistic effect with the scrap steel grading and classification label library. Using this method, after acquiring and preprocessing the point cloud data from the scrap steel unloading site, the grading and classification results of the scrap steel can be accurately obtained through the pre-set grading and classification model and the scrap steel grading and classification label library. To enable grading and classification of scrap steel types, this method pre-stores specific scrap steel type labels (equivalent to establishing a relatively complete scrap steel grading and classification criterion). In practice, after the point cloud features are extracted by the deep neural network-based classification model, the specific grading and classification results can be obtained according to the scrap steel type level labels defined in the scrap steel grading and classification label library. In addition, the preprocessing of the acquired point cloud data in this invention includes grid voxelization, nearest neighbor statistical analysis, and point cloud registration in sequence. By using mesh voxelization, the amount of point cloud data to be processed can be reduced while ensuring that the shape characteristics of the object remain unchanged, thus improving the overall processing efficiency. Nearest neighbor statistical analysis in this method can remove discrete noise from the scrap steel point cloud data, ensuring the accuracy of subsequent processing. Furthermore, since the collected point cloud data may come from different devices or different viewpoints of the same device, point cloud registration in this method can unify and integrate the point cloud data into a specified coordinate system, further ensuring the accuracy of subsequent processing.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent classification and determination of scrap material piles based on mobile laser radar point cloud data, characterized in that, The method comprises the following steps: Step 1, obtaining point cloud data of a scrap steel unloading site; Step 2, preprocessing the obtained point cloud data; Step 3, classifying the scrap steel according to the preprocessed point cloud data through a preset hierarchical classification model and a scrap steel hierarchical classification label library; wherein the hierarchical classification model introduces an EdgeConv layer based on a PointNet point cloud classification model, which is used to extract neighbor point cloud features; the hierarchical classification label library stores pre-defined scrap steel type level labels; In step 3, the working process of the hierarchical classification model comprises: S1, performing spatial transformation processing on the point cloud data through a spatial transformation module; S2, extracting neighbor point cloud features through the EdgeConv layer to obtain corresponding feature vectors after the spatial transformation processing of the point cloud data; the neighbor point cloud features include three-dimensional coordinates, color, layers and surface normal of the point cloud; S3, obtaining the feature representation of the center point by pooling the feature vectors obtained in S2; obtaining the global feature of the point cloud through a max-pooling layer, mapping the global feature to the hierarchical classification label library through a fully connected layer, and obtaining the corresponding scrap steel type level label as the hierarchical classification result of the scrap steel; S2 comprises: S2.1, represent the point cloud data as an undirected graph by an EdgeConv convolutional layer; define the local point cloud undirected graph as ; wherein, denotes a point cloud set, n is the number of points in the point cloud; denotes an edge set, k denotes the number of edges in the k-neighborhood of the point . S2.2, in the local area, feature extraction is performed on the point cloud data by a multi-layer perception; for each point the features of the edges in the k-neighborhood of the point are extracted by a shared neural network to obtain the corresponding feature vector; the features of the edges in the k-neighborhood of the point are extracted by a shared neural network to obtain the corresponding feature vector; The specific process of S2.2 is as follows: For a point , define the edge feature of the point cloud where, ; is a nonlinear function with a set of learnable parameters ; compute the edge feature using an asymmetric edge function: Among the m different filters, there are edge features: wherein, and denote learning parameters under different filters, ; denotes an activation function; Then, the output of the i-th point cloud is obtained by aggregating the edge features of the i-th point cloud:

2. The mobile LiDAR point cloud data based scrap yard intelligent classification determination method according to claim 1, characterized in that: In step 2, the preprocessing comprises grid voxelization, neighborhood statistical analysis and point cloud registration in sequence.

3. The mobile LiDAR point cloud data based scrap yard intelligent classification determination method of claim 2, wherein: In step 2, the process of neighborhood statistical analysis comprises: Suppose any point in the point cloud set Q The three-dimensional coordinates are ( );in, n is the number of points in the point cloud; a kd-tree is used to search for each point in the point cloud set. of Neighborhood; Assume the sampling point ; ; ; ; ; ; ; ; ; ; ; ; Recalculate distance average set ; obeys a Gaussian distribution, , where, is the mean of , and is the corresponding variance; If the calculated value is not within the range, the corresponding point cloud data is determined as a noise point, and the noise point is removed; in the formula, not in the range, the corresponding point cloud data is determined as a noise point, and the noise point is removed; in the formula, represents determined according to experience and experiments.

4. The method of claim 3, wherein: In step 2, the point cloud registration is performed using an iterative closest point algorithm; when the point cloud registration is performed, it is assumed that the source point cloud is P and the target point cloud is Q, and the process of the point cloud registration comprises: 1) In two point clouds P and Q, for a point p in point cloud P select the closest point to it ; 2) Solve the optimal values of the rotation matrix R and the translation vector t at the minimum of the error function, and record the current minimum error function value as ; the error function is: In the formula, represents the number of point clouds, represents a translation matrix; 3) apply the optimal rotation matrix R and translation vector t to the point cloud P to obtain the point cloud : 4) Using point clouds Replace point cloud P and repeat steps 1)-3); then take the absolute value of the difference between the current and subsequent error values. When the number of iterations is less than the threshold or greater than the preset maximum number of iterations, the iteration stops, and the final point cloud is used as the registered point cloud; where k≥2.

5. The mobile LiDAR point cloud data based scrap yard intelligent classification decision method of claim 1, wherein: In step 3, in the training process of the hierarchical classification model, the loss calculation adopts a cross-entropy loss function: wherein, is the true label, is the predicted label.

6. The mobile LiDAR point cloud data based scrap yard intelligent classification decision method of claim 1, wherein: In step 1, the point cloud data of the scrap steel unloading site is obtained by a laser radar carried by a UAV.

Images