An intelligent classification method and system based on multi-dimensional feature extraction of point cloud data

Abstract

The application discloses an intelligent classification method and system based on multi-dimensional feature extraction of point cloud data, relates to the technical field of deep learning, and comprises the following steps: collecting point cloud data, and preprocessing the point cloud data; based on the preprocessed point cloud data, features are extracted, and a fused feature vector is generated by using the relative change rate of the features for adaptive fusion; a hybrid activation multilayer perception network model is constructed by using the fused feature vector; the original point cloud data is subjected to multi-scale processing, features are extracted, and a multi-scale fused high-dimensional feature vector is generated by splicing and then delivered to the hybrid activation multilayer perception network model; an attention mechanism is introduced into the hybrid activation multilayer perception network model, and a classification decision is made; the hybrid activation multilayer perception network model can quickly converge and improve classification accuracy in the training process; better classification performance is exhibited when high-dimensional fused feature vectors are processed, and the overall classification effect and robustness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to an intelligent classification method and system based on multi-dimensional feature extraction of point cloud data. BACKGROUND

[0002] With the development of three-dimensional sensing technology, point cloud data has been widely used in automatic driving, robot navigation, 3D reconstruction and other fields. As a discrete representation of three-dimensional space, point cloud data can accurately describe the geometric shape and spatial distribution of objects, and therefore plays an important role in intelligent classification and target recognition. Traditional point cloud data processing methods usually rely on feature extraction and classification models based on geometric features. Although these methods can provide certain accuracy, due to the sparsity, irregularity and high dimensionality of point cloud data, traditional methods have high computational complexity and insufficient feature extraction when processing large-scale point cloud data. Therefore, point cloud classification methods combined with deep learning have gradually attracted attention, and models such as multilayer perceptron and convolutional neural network have been introduced to enhance the automatic feature extraction and classification ability of point cloud data.

[0003] However, the existing technology still has many deficiencies in dealing with multi-scale point cloud feature fusion and classification. Firstly, the existing methods lack an adaptive fusion mechanism for feature changes at different scales, resulting in insufficient feature expression ability. Secondly, most traditional classification models use a single activation function, which cannot effectively combine the nonlinear relationships of different types of features, limiting the classification performance of the model. Thirdly, the existing models do not fully introduce attention mechanisms when processing high-dimensional features, making it difficult to focus on important feature dimensions or spatial regions, resulting in a decrease in classification accuracy. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an intelligent classification method based on multi-dimensional feature extraction of point cloud data to solve the problem of low classification accuracy of point cloud data.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an intelligent classification method based on multi-dimensional feature extraction of point cloud data, which includes collecting point cloud data and preprocessing the point cloud data. Based on the preprocessed point cloud data, features are extracted and the relative change rate of the features is used for adaptive fusion to generate a fused feature vector. The fused feature vector is used to construct a hybrid activation multilayer perception network model. The original point cloud data is processed at multiple scales to extract features, and a multi-scale fused high-dimensional feature vector is generated by splicing and delivered to the hybrid activation multilayer perception network model. An attention mechanism is introduced for the hybrid activation multilayer perception network model to make a classification decision.

[0008] As a preferred scheme of the intelligent classification method based on multi-dimensional feature extraction of point cloud data provided by the application, the point cloud data is collected and preprocessed, and the specific steps are,

[0009] Collecting three-dimensional point cloud data;

[0010] Using statistical filtering for denoising, and downsampling through a voxel grid filter.

[0011] As a preferred scheme of the intelligent classification method based on multi-dimensional feature extraction of point cloud data provided by the application, the feature is extracted, and the relative change rate of the feature is used for adaptive fusion to generate a fusion feature vector, and the specific steps are,

[0012] Using principal component analysis (PCA) to extract geometric features in the three-dimensional point cloud data;

[0013] Based on the reflection intensity in the three-dimensional point cloud data, physical features are obtained;

[0014] Based on the adaptivity of the feature relative change rate, the extracted geometric features and physical features are fused to obtain a fused feature vector F f , the expression is:

[0015]

[0016] Where Z is a normalization factor, Δ geo is the relative change rate of the geometric feature, Δ phy is the relative change rate of the physical feature, f geo (i) is the i-th geometric feature, f phy (j) is the j-th physical feature, i is the index of the geometric feature, j is the index of the physical feature, m is the total number of physical features, and n is the total number of geometric features.

[0017] As a preferred scheme of the intelligent classification method based on multi-dimensional feature extraction of point cloud data provided by the application, the mixed activation multi-layer perception network model is constructed, and the specific steps are,

[0018] The fused feature vector is input to the multi-layer hidden layer, and linear transformation is performed to introduce non-linear capability, and the mixed activation function is processed to obtain the activation result f(x), and the expression is:

[0019] f(x)=α×ReLU(x)+(1-α)×Tanh(x);

[0020] Where α is an adaptive parameter, ReLU(x) is an activation function, Tanh(x) is a hyperbolic tangent activation function, and x is input data.

[0021] Based on the activation result, the activation result is mapped into the class space of the classification task through a linear transformation;

[0022] A classification threshold is set, and the class whose class score reaches the set threshold in the class space is selected as the prediction result of the hybrid activation multi-layer perception network model;

[0023] The difference between the prediction result of the hybrid activation multi-layer perception network model and the real class label is measured using a cross-entropy loss function;

[0024] The parameters of the hybrid activation multi-layer perception network model are optimized by an Adam optimization algorithm.

[0025] As a preferred scheme of the intelligent classification method based on multi-dimensional feature extraction of point cloud data, wherein: the original point cloud data is subjected to multi-scale processing, features are extracted, and a multi-scale fusion high-dimensional feature vector is generated by splicing and delivered to the hybrid activation multi-layer perception network model, and the specific steps are,

[0026] The original point cloud data is subjected to multi-scale processing, and geometric feature extraction and physical feature extraction are performed on the point cloud data of each scale to generate a feature vector of each scale;

[0027] The feature vectors of different scales are spliced to generate a multi-scale fusion high-dimensional feature vector;

[0028] The multi-scale fusion high-dimensional feature vector is input into the hybrid activation multi-layer perception network model for multi-dimensional feature processing, the multi-scale fusion high-dimensional feature vector is compressed, and a classification result is output.

[0029] As a preferred scheme of the intelligent classification method based on multi-dimensional feature extraction of point cloud data, wherein: the attention mechanism is introduced, and a classification decision is made, and the specific steps are,

[0030] The channel attention mechanism and the spatial attention mechanism are introduced into the hybrid activation multi-layer perception network model;

[0031] The different channels of the multi-scale fusion high-dimensional feature vector are analyzed through the channel attention mechanism, and the most important spatial region of the multi-scale fusion high-dimensional feature vector is highlighted through the spatial attention mechanism to obtain a feature map;

[0032] The feature map is input into the hybrid activation multi-layer perception network model, the multi-scale fusion high-dimensional feature vector is compressed into the same dimension as the number of classes in the class space, and a class probability is generated;

[0033] A probability threshold is set, and the class whose class probability reaches the set threshold is selected as the final prediction result of the hybrid activation multi-layer perception network model.

[0034] As a preferred scheme of the intelligent classification method based on multi-dimensional feature extraction of point cloud data according to the application, wherein: the channel attention mechanism and the spatial attention mechanism are introduced, and the specific steps are,

[0035] The input multi-scale fusion high-dimensional feature vector is globally averaged and pooled to obtain the average pooling value of each channel.

[0036] The global maximum pooling is performed on each channel of the multi-scale fusion high-dimensional feature vector to obtain the maximum pooling value of each channel.

[0037] The average pooling value and the maximum pooling value are spliced to obtain a fusion pooling feature vector.

[0038] The fusion pooling feature vector is input into a fully connected layer to generate a channel attention weight.

[0039] The multi-scale fusion high-dimensional feature vector is weighted using the channel attention weight to generate a weighted multi-scale fusion high-dimensional feature map.

[0040] The weighted multi-scale fusion high-dimensional feature map is input and the spatial dimension is performed.

[0041] The spatial dimension of the weighted multi-scale fusion high-dimensional feature map is globally averaged and pooled to generate a global average pooling spatial map.

[0042] The spatial dimension of the weighted multi-scale fusion high-dimensional feature map is globally maximum pooled to generate a global maximum pooling spatial map.

[0043] The global average pooling spatial map and the global maximum pooling spatial map are spliced to generate a fusion feature map.

[0044] The fusion feature map is subjected to a convolution operation to obtain a spatial attention weight map.

[0045] The spatial attention weight map is used to adjust the weighted multi-scale fusion high-dimensional feature map to obtain a final weighted multi-scale fusion high-dimensional feature map.

[0046] In a second aspect, the present application provides an intelligent classification system based on multi-dimensional feature extraction of point cloud data, comprising a collection and processing module, an adaptive fusion module, a construction module, a multi-scale processing module and a decision module; the collection and processing module is used for collecting point cloud data and pre-processing the point cloud data; the adaptive fusion module is used for extracting features based on the pre-processed point cloud data and generating a fusion feature vector by adaptive fusion of the relative change rate of the features; the construction module is used for constructing a hybrid activation multi-layer perception network model by using the fused feature vector; the multi-scale processing module is used for multi-scale processing of the original point cloud data, extracting features, and generating a multi-scale fusion high-dimensional feature vector by splicing and delivering the multi-scale fusion high-dimensional feature vector to the hybrid activation multi-layer perception network model; and the decision module is used for introducing an attention mechanism for the hybrid activation multi-layer perception network model and making a classification decision.

[0047] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any step of the intelligent classification method based on multi-dimensional feature extraction of point cloud data according to the first aspect of the present application is implemented.

[0048] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, any step of the intelligent classification method based on multi-dimensional feature extraction of point cloud data according to the first aspect of the present application is implemented.

[0049] The present application has the following advantages: the present application combines the advantages of ReLU and Tanh activation functions, enhances the nonlinear expression ability of different features, ReLU can provide better gradient propagation when processing sparse data, and Tanh performs well when processing smooth nonlinear relationships; moreover, the adaptive parameter alpha and the network can flexibly adjust the combination ratio of the two activation functions according to different types of features, improving the generalization ability of the hybrid activation multi-layer perception network model; in addition, the combination of cross-entropy loss function and Adam optimization algorithm enables the hybrid activation multi-layer perception network model to quickly converge and improve the classification accuracy during the training process; the hybrid activation mechanism effectively solves the limitations of single activation function and performs better classification performance when processing high-dimensional fusion feature vectors, improving the overall classification effect and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0051] Figure 1 Flow chart of the intelligent classification method based on multi-dimensional feature extraction of point cloud data in embodiment 1.

[0052] Figure 2 Module diagram of the intelligent classification system based on multi-dimensional feature extraction of point cloud data in embodiment 1. DETAILED DESCRIPTION

[0053] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0054] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0055] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is separate or alternative to other embodiments.

[0056] Embodiment 1, reference Figure 1 and Figure 2 , the first embodiment of the present application, the embodiment provides an intelligent classification method based on multi-dimensional feature extraction of point cloud data, comprising the following steps:

[0057] S1, collecting point cloud data, and pre-processing the point cloud data.

[0058] Further, three-dimensional point cloud data is collected using a laser radar; each point contains three-dimensional coordinates (x, y, z) and additional attributes (such as reflection intensity); when collecting, each angle of the target object should be covered to ensure the comprehensiveness and density of the data;

[0059] Statistical filtering is used for denoising, the distance of each point to its neighborhood points is calculated, and those points with a distance greater than a set threshold to the neighborhood points are removed to ensure the smoothness and effectiveness of the data;

[0060] The number of points of the point cloud data is usually very large, and direct processing will increase the computational overhead; down-sampling is performed through a voxel grid filter; the voxel grid filter divides the point cloud into fixed-size cubic grids (such as a voxel size of 0.05 meters), and uses the centroid point in the voxel to represent all points in the voxel, thereby reducing the number of points while retaining the overall structure of the point cloud;

[0061] The points with too low reflection intensity of the point cloud data are removed, which usually do not contain valuable information and can interfere with subsequent feature extraction and classification.

[0062] It should be noted that the point cloud data is comprehensively collected and after noise removal, downsampling and invalid data removal, the effectiveness and quality of the data are significantly improved; the statistical filtering effectively removes the noise points, the voxel grid filter reduces the computational overhead while retaining the overall structure of the point cloud, and the removal of low reflection intensity points ensures the effectiveness of the data, finally making the subsequent feature extraction and classification more efficient and accurate, and improving the overall processing performance and robustness.

[0063] S2, based on the preprocessed point cloud data, extracting features and using the relative change rate of the features to adaptively fuse to generate a fused feature vector.

[0064] Further, principal component analysis (PCA) is used to extract geometric features from the three-dimensional point cloud data; for each point, a covariance matrix is calculated based on its neighborhood point set, and the normal vector and curvature of the point are calculated through the distribution of eigenvalues; the curvature reflects the local geometric shape features of the point, such as flat, convex or concave;

[0065] Based on the reflection intensity in the three-dimensional point cloud data, physical features are obtained; it reflects the physical properties of the object surface, such as material, color, smoothness, etc.; the ability of the object surface to reflect the laser beam will affect the reflection intensity, which makes the reflection intensity an important physical feature of the object surface characteristics;

[0066] By adapting based on the relative change rate of the features, the extracted geometric features and physical features are fused to obtain a fused feature vector F f , the expression is:

[0067]

[0068] Where Z is a normalization factor, Z = Δ geo + Δ phy , to ensure the numerical stability of the fused feature vector, Δ geo is the relative change rate of the geometric features, Δ phy is the relative change rate of the physical features, f geo (i) is the i-th geometric feature (normal vector, curvature, etc.), f phy (j) is the j-th physical feature (such as reflection intensity, etc.), i is the index of the geometric feature, j is the index of the physical feature, m is the total number of physical features, i.e. the total number of physical features extracted from the point cloud data, n is the total number of geometric features, i.e. the total number of geometric features extracted from the point cloud data.

[0069] It should be noted that the relative change rate of the geometric feature is assumed to be 0.6, the relative change rate of the physical feature is 0.4, the geometric feature is 3, respectively f geo (1) = 1.2, f geo (2) = 1.5, f geo (3) = 1.3, the physical feature is 2, respectively f phy (1) = 0.8, f phy (2) = 0.9, the normalization factor is 1;

[0070] First, calculate the weighted coefficient of the geometric feature:

[0071] Calculate the weighted coefficient of the physical feature:

[0072] Then calculate the sum of the geometric features:

[0073]

[0074] Calculate the sum of the physical features:

[0075] Substitute the formula to calculate the fused feature vector:

[0076] F f = 1 x (0.6 x 4 + 0.4 x 1.7) = 3.08;

[0077] It should be noted that the geometric features of the point cloud data are extracted by principal component analysis, and the physical features (such as surface material and smoothness) are extracted based on the reflection intensity, then the relative change rates of the geometric features and the physical features are used for adaptive fusion to generate a fused feature vector; this fusion considers both local geometric shape and physical properties, ensuring the accuracy and comprehensiveness of feature expression, and improving the accuracy and robustness of subsequent classification.

[0078] S3, using the fused feature vector, a hybrid activation multilayer perception network model is constructed.

[0079] Further, the input layer is defined as the fused feature vector, and the output layer is defined as the prediction result;

[0080] Define multiple hidden layers; here, three hidden layers are defined;

[0081] The fused feature vector is input into the multiple hidden layers, and linear transformation is performed to introduce non-linear capability, and the hybrid activation function is used for processing to obtain the activation result f(x), the expression is:

[0082] f(x) = α x ReLU(x) + (1-α) x Tanh(x);

[0083] where f(x) is a weighted combination of ReLU and Tanh functions, a is an adaptive parameter with a value range of [0, 1] that controls the weight proportion of ReLU and Tanh functions in the combination, ReLU(x) is an activation function used to process input x, which sets the negative part to 0 and keeps the positive part unchanged, Tanh(x) is a hyperbolic tangent activation function, x is the input data, which is a smooth nonlinear function that maps input x to the interval [0, 1], and 1-a is the weight of the Tanh function, representing the contribution of Tanh in the combination, which changes with a, thereby adjusting the influence of Tanh in the combination;

[0084] The second hidden layer receives the activation results of the first layer, generates a new feature vector, and processes it again through the hybrid activation function to obtain the activation results input to the third hidden layer.

[0085] The third hidden layer receives the activation results of the second layer, generates a new feature vector, and processes it again through the hybrid activation function to obtain the activation results as the output results.

[0086] Based on the activation results, the activation results are mapped to the class space of the classification task through linear transformation.

[0087] A classification threshold is set, and the class whose score reaches the set threshold in the class space is selected as the prediction result of the hybrid activation multi-layer perception network model; the output layer generates a vector, where each element represents the score of a certain class, and through the Softmax function, the class with the highest score (reaching the set threshold) is selected as the final prediction result; threshold setting refers to setting a fixed probability value (such as 0.5) in the classification task, when the model predicts a score of a certain class exceeds the threshold, it is taken as the final prediction result, if no class reaches the threshold, the result is output through selecting the maximum classification or other processing mechanism.

[0088] The cross-entropy loss function is used to measure the difference between the prediction results of the hybrid activation multi-layer perception network model and the true class labels; the cross-entropy loss function compares the model's prediction probability with the true class label and calculates the current error of the model.

[0089] The parameters of the hybrid activation multi-layer perception network model are optimized through the Adam optimization algorithm. In order to minimize the cross-entropy loss, we use the backpropagation algorithm to calculate the parameter gradient of each layer. Then, the Adam optimization algorithm is used to update the parameters of the network according to these gradients. The Adam optimization algorithm is an adaptive gradient descent algorithm that can automatically adjust the learning rate to make the model converge faster.

[0090] It should be noted that the multi-layer perception network inputting the three-layer hidden layer through the fusion feature vector, using the adaptive parameter control combination of ReLU and Tanh activation function, enhances the non-linear expression ability of the network; the multi-layer hidden layer gradually extracts deep features, measures the prediction error through the cross-entropy loss function, and combines the Adam optimization algorithm to adaptively adjust the learning rate to speed up the model convergence; the model can not only accurately capture complex features, but also improve the accuracy and robustness of classification.

[0091] S4, the original point cloud data is processed in multiple scales, features are extracted, and a multi-scale fusion high-dimensional feature vector is generated by splicing and input into a hybrid activation multi-layer perception network model.

[0092] In order to capture the global and local features of the point cloud data, the original point cloud data is processed in multiple scales, and the geometric features and physical features of the point cloud data at each scale are extracted to generate a feature vector at each scale; different scales of point clouds are generated through downsampling of different voxel resolutions (such as 0.02 meters, 0.05 meters, and 0.1 meters), the smaller voxel size retains more detailed information, and the larger voxel size reflects global information, which helps to capture features of different granularities; by extracting geometric features such as normal vectors and curvatures, and physical features such as reflectance intensity, a feature vector at each scale is generated;

[0093] The feature vectors at different scales are spliced to generate a multi-scale fusion high-dimensional feature vector; this spliced multi-scale fusion high-dimensional feature vector contains multi-scale information from micro to macro, and can more comprehensively describe the shape and material of the object;

[0094] The multi-scale fusion high-dimensional feature vector is input into the hybrid activation multi-layer perception network model for multi-dimensional feature processing, compression of the multi-scale fusion high-dimensional feature vector, and output of the classification result using the previously trained classification ability.

[0095] It should be noted that the point cloud data is downsampled at different resolutions through multi-scale processing, which can capture multi-scale features from details to global, and generate multi-scale feature vectors containing geometric features and physical features; these feature vectors are spliced into multi-scale fusion high-dimensional feature vectors to comprehensively describe the shape and material of the object; the spliced multi-scale fusion high-dimensional feature vectors are input into the hybrid activation multi-layer perception network, which effectively compresses the data through multi-dimensional feature processing and outputs the classification result, improving the accuracy of classification and the perception ability of multi-scale information.

[0096] S5, introduce an attention mechanism for the hybrid activation multi-layer perception network model to make a classification decision.

[0097] Furthermore, a channel attention mechanism and a spatial attention mechanism are introduced into the hybrid activation multi-layer perception network model.

[0098] The channel attention mechanism analyzes different channels of the multi-scale fusion high-dimensional feature vector, and automatically assigns higher weights to important channels. The weight of each channel is generated by a global pooling operation (average pooling and maximum pooling), which is used to control the importance of each feature channel. The weighted multi-scale fusion high-dimensional feature map is obtained. The spatial attention mechanism highlights the most important spatial region of the multi-scale fusion high-dimensional feature vector. The spatial attention generates spatial weights through convolution operation, so that the classification model can focus more on the key region of the input data, thereby improving the classification accuracy, and obtaining the final multi-scale fusion high-dimensional feature map.

[0099] The multi-scale fusion high-dimensional feature map is input into the hybrid activation multi-layer perception network model, and the multi-scale fusion high-dimensional feature vector is compressed into the same dimension as the number of classes in the class space to generate a class probability. The weighted multi-scale fusion high-dimensional feature map is input into the classifier, and a fully connected layer is used to compress the high-dimensional feature into the same dimension as the number of classes. Then the output of the fully connected layer is transmitted into the Softmax function to generate the probability distribution of the class. The Softmax function converts the output value of the network into a probability value, representing the predicted probability of each class.

[0100] A probability threshold is set, and the class whose probability reaches the set threshold is selected as the final prediction result of the hybrid activation multi-layer perception network model. A probability threshold is set, and if the probability value of a certain class is greater than or equal to the threshold, the class will be selected as the final prediction result. This threshold can ensure that the class output by the model has sufficient confidence and avoid selecting classes with low probability. It is set to a fixed value, usually between 0 and 1, for example, the threshold is set to 0.5, which means only the class whose probability is greater than or equal to 0.5 will be selected as the final prediction result. If no class probability reaches 0.5, other processing mechanisms (such as selecting the class with the largest probability or through further model correction processing) may be triggered.

[0101] S5.1, introduce channel attention mechanism and spatial attention mechanism, specifically,

[0102] Input the multi-scale fusion high-dimensional feature vector, and perform global average pooling on each channel of the multi-scale fusion high-dimensional feature vector to obtain the average pooling value of each channel.

[0103] Global maximum pooling is performed on each channel of the multi-scale fusion high-dimensional feature vector to obtain the maximum pooling value of each channel, which captures the most significant features in the channel.

[0104] The average pooling value and the maximum pooling value are spliced to obtain a fusion pooling feature vector; the vector contains global statistical information (average value and maximum value) of each channel, representing the importance of each channel;

[0105] The fusion pooling feature vector is input into a fully connected layer to generate a channel attention weight of each channel;

[0106] The channel attention weight is used to weight a multi-scale fusion high-dimensional feature vector to generate a weighted multi-scale fusion high-dimensional feature map;

[0107] The weighted multi-scale fusion high-dimensional feature map is input and spatial dimensions are performed;

[0108] Global average pooling is performed on the spatial dimensions of the weighted multi-scale fusion high-dimensional feature map to generate a global average pooling spatial map;

[0109] Global maximum pooling is performed on the spatial dimensions of the weighted multi-scale fusion high-dimensional feature map to generate a global maximum pooling spatial map;

[0110] The global average pooling spatial map and the global maximum pooling spatial map are spliced to generate a fusion feature map;

[0111] Convolution is performed on the fusion feature map to obtain a spatial attention weight map;

[0112] The spatial attention weight map is used to adjust (Hadamard product) the weighted multi-scale fusion high-dimensional feature map to obtain a final weighted multi-scale fusion high-dimensional feature map, so that the hybrid activation multi-layer perception network model pays more attention to important spatial regions.

[0113] It should be noted that by introducing the channel and spatial attention mechanism, the hybrid activation multi-layer perception network can automatically assign higher weights to important channels and key spatial regions; the channel attention mechanism captures the importance of each channel through global pooling, and the spatial attention mechanism focuses on key regions through convolution operation; this mechanism enhances the selective attention of the network to multi-scale features, enabling the classification model to more accurately identify and classify complex scenes, improving classification accuracy and robustness, and outputting more confident prediction results.

[0114] The embodiment also provides an intelligent classification system based on multi-dimensional feature extraction of point cloud data, comprising: a collection processing module, an adaptive fusion module, a construction module, a multi-scale processing module and a decision module; the collection processing module is used for collecting point cloud data and pre-processing the point cloud data; the adaptive fusion module is used for extracting features based on the pre-processed point cloud data and generating a fusion feature vector by adaptive fusion of the relative change rate of the features; the construction module is used for constructing a hybrid activation multi-layer perception network model by using the fused feature vector; the multi-scale processing module is used for multi-scale processing of original point cloud data, extracting features, and generating a multi-scale fusion high-dimensional feature vector by splicing and delivering to the hybrid activation multi-layer perception network model; and the decision module is used for introducing an attention mechanism for the hybrid activation multi-layer perception network model and making a classification decision.

[0115] The embodiment also provides a computer device suitable for the intelligent classification method based on multi-dimensional feature extraction of point cloud data, comprising: a memory and a processor; the memory is used for storing computer executable instructions, and the processor is used for executing the computer executable instructions to realize the intelligent classification method based on multi-dimensional feature extraction of point cloud data as proposed in the above embodiment.

[0116] The computer device can be a terminal, and the computer device comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used for providing computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals. The wireless communication can be realized through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0117] The embodiment also provides a storage medium on which a computer program is stored, the program being executed by a processor to implement the intelligent classification method based on multi-dimensional feature extraction of point cloud data proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk.

[0118] To sum up, the application combines the advantages of ReLU and Tanh activation functions, enhances the nonlinear expression ability of different features, and improves the generalization ability of the hybrid activation multi-layer perception network model. In addition, the combination of the cross-entropy loss function and the Adam optimization algorithm enables the hybrid activation multi-layer perception network model to converge quickly and improve the classification accuracy during the training process. The hybrid activation mechanism effectively solves the limitations of a single activation function and exhibits better classification performance when processing high-dimensional fusion feature vectors, thereby improving the overall classification effect and robustness.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the application rather than limit the application. Although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the application, and all modifications or equivalent replacements should be included in the scope of the claims of the application.

Claims

1. An intelligent classification method based on multidimensional feature extraction from point cloud data, characterized in that: include, Collect point cloud data and preprocess the point cloud data; Based on the preprocessed point cloud data, features are extracted and adaptively fused using the relative rate of change of the features to generate a fused feature vector. A hybrid activation multilayer perceptron model is constructed using the fused feature vectors. The original point cloud data is processed at multiple scales to extract features, and the resulting multi-scale fused high-dimensional feature vector is fed into a hybrid activation multilayer perceptron model. An attention mechanism is introduced into the hybrid activation multilayer perceptron model to make classification decisions; The specific steps for constructing the hybrid activation multilayer perceptron model are as follows: The input layer is defined as the fused feature vector, and the output layer is the prediction result. The fused feature vectors from multiple hidden layers are input and subjected to a linear transformation to introduce nonlinearity. The resulting activation vectors are then processed using a hybrid activation function. The expression is: ； in, It is an adaptive parameter. It is an activation function, Tanh(x) is the hyperbolic tangent activation function. It is the input data; Based on the activation results, a linear transformation is used to map the activation results to the category space of the classification task; Set a classification threshold, and select the category whose category score in the category space reaches the set threshold as the prediction result of the hybrid activation multilayer perceptron model; The cross-entropy loss function is used to measure the difference between the prediction results of the hybrid activation multilayer perceptron model and the true class label. The parameters of the hybrid activation multilayer perceptron model are optimized using the Adam optimization algorithm.

2. The intelligent classification method based on multi-dimensional feature extraction of point cloud data as described in claim 1, characterized in that: The point cloud data is collected, and the point cloud data is preprocessed. The specific steps are as follows: Collect 3D point cloud data; Statistical filtering is used for noise reduction, and voxel grid filters are used for downsampling.

3. The intelligent classification method based on multi-dimensional feature extraction of point cloud data as described in claim 2, characterized in that: The specific steps for extracting features and adaptively fusing them using their relative rates of change to generate a fused feature vector are as follows: Principal component analysis (PCA) is used to extract geometric features from 3D point cloud data. Physical features are obtained based on the reflection intensity in 3D point cloud data; Based on the adaptive principle of the relative rate of change of features, the extracted geometric and physical features are fused to obtain a fused feature vector. The expression is: ； in, It is a normalization factor. It is the relative rate of change of geometric features. It is the relative rate of change of physical characteristics. It is the first Geometric features, It is the first Each physical feature is an index of its geometric features. It is an index of physical characteristics. It is the total number of physical characteristics. It is the total number of geometric features.

4. The intelligent classification method based on multidimensional feature extraction of point cloud data as described in claim 1, characterized in that: The steps involve multi-scale processing of the original point cloud data, feature extraction, and concatenation to generate a multi-scale fused high-dimensional feature vector, which is then fed into a hybrid activation multilayer perceptron model. The original point cloud data is processed at multiple scales, and geometric and physical features are extracted from the point cloud data at each scale to generate feature vectors for each scale. Feature vectors of different scales are concatenated to generate a multi-scale fused high-dimensional feature vector; The multi-scale fused high-dimensional feature vector is input into the hybrid activation multilayer perceptron model for multi-dimensional feature processing, compressing the multi-scale fused high-dimensional feature vector, and outputting the classification result.

5. The intelligent classification method based on multi-dimensional feature extraction of point cloud data as described in claim 1, characterized in that: The specific steps for introducing an attention mechanism to make classification decisions are as follows: Introducing channel attention and spatial attention mechanisms into a hybrid activation multilayer perceptron model; By analyzing different channels of multi-scale fused high-dimensional feature vectors through channel attention mechanism, and using spatial attention mechanism to highlight the most important spatial regions of multi-scale fused high-dimensional feature vectors, feature maps are obtained. The feature map is input into the hybrid activation multilayer perceptron model, and the multi-scale fused high-dimensional feature vector is compressed into a dimension with the same number of categories in the category space to generate category probabilities. Set a probability threshold, and select the category whose category probability reaches the set threshold as the final prediction result of the hybrid activation multilayer perceptron model.

6. The intelligent classification method based on multidimensional feature extraction of point cloud data as described in claim 5, characterized in that: The specific steps for introducing the channel attention mechanism and the spatial attention mechanism are as follows: Input a multi-scale fused high-dimensional feature vector, and perform global average pooling on each channel of the multi-scale fused high-dimensional feature vector to obtain the average pooling value of each channel. Global max pooling is performed on each channel of the multi-scale fused high-dimensional feature vector to obtain the max pooling value of each channel; The average pooling value and the max pooling value are concatenated to obtain the fused pooling feature vector; The fused pooled feature vectors are input into the fully connected layer to generate channel attention weights; The multi-scale fusion high-dimensional feature vector is weighted using channel attention weights to generate a weighted multi-scale fusion high-dimensional feature map; Input a weighted, multi-scale fused high-dimensional feature map and perform spatial dimension calculation; Global average pooling is applied to the spatial dimension of the weighted multi-scale fused high-dimensional feature map to generate a global average pooling spatial map. Global max pooling is performed on the spatial dimension of the weighted multi-scale fused high-dimensional feature map to generate a global max pooling spatial map; The global average pooling space map and the global max pooling space map are concatenated to generate a fused feature map; The fused feature map is convolved to obtain the spatial attention weight map; The weighted multi-scale fusion high-dimensional feature map is adjusted using the spatial attention weight map to obtain the final weighted multi-scale fusion high-dimensional feature map.

7. An intelligent classification system based on multidimensional feature extraction of point cloud data, based on the intelligent classification method based on multidimensional feature extraction of point cloud data according to any one of claims 1 to 6, characterized in that: It includes an acquisition and processing module, an adaptive fusion module, a construction module, a multi-scale processing module, and a decision-making module; The acquisition and processing module is used to acquire point cloud data and preprocess the point cloud data. The adaptive fusion module is used to extract features based on preprocessed point cloud data and adaptively fuse them using the relative rate of change of the features to generate a fused feature vector. The building module is used to construct a hybrid activation multilayer perceptron model using the fused feature vectors. The multi-scale processing module is used to perform multi-scale processing on the original point cloud data, extract features, and generate a multi-scale fused high-dimensional feature vector by splicing, which is then fed into the hybrid activation multilayer perceptron model. The decision module is used to introduce an attention mechanism into the hybrid activation multilayer perceptron model to make classification decisions.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent classification method based on multidimensional feature extraction of point cloud data as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent classification method based on multidimensional feature extraction of point cloud data as described in any one of claims 1 to 6.

Images