Obstacle sensing method, device, electronic device and storage medium

Abstract

The embodiments of the present application provide an obstacle perception method, device, electronic device and storage medium, which relate to the field of intelligent driving technology. The method includes acquiring vehicle surrounding environment data from an on-board IoT sensor; using a gated graph neural network to extract features, using a trained autoencoder neural network model based on energy basis compression to reduce the feature dimension, and then using an extreme learning machine classifier to perform classification, so as to perceive and predict obstacles. The method uses a gated graph neural network to extract features, an autoencoder neural network model based on energy basis compression to reduce the feature dimension, and an extreme learning machine classification algorithm with an enhanced annealing strategy to classify, which effectively improves the accuracy and efficiency of the perception results, and has strong adaptability and robustness to abnormal environments, solving the problem of inaccurate obstacle perception and poor robustness of existing methods.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an obstacle perception method, device, electronic device and storage medium. Background Art

[0002] Intelligent driving systems rely on advanced perception technology, data processing capabilities, and automatic control technology to achieve automatic driving and intelligent decision-making of vehicles, significantly improve driving safety and efficiency, and reduce traffic accidents. However, the safety and efficiency of intelligent driving largely rely on accurate perception and understanding of the surrounding environment, especially the ability to identify and classify various obstacles. Therefore, obstacle perception technology has become a key and challenging area in the research of intelligent driving systems.

[0003] Existing technologies may not be able to fully capture the features of obstacles and the correlation between features in complex environments, resulting in inaccurate obstacle perception; there may be problems of inefficiency and insufficient accuracy in the feature dimensionality reduction and classification process, especially when processing large-scale data; in the face of environmental noise and abnormal data, the classifiers of existing technologies may lack sufficient robustness and be easily affected by outliers, resulting in degraded classification performance. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide an obstacle perception method, device, electronic device and storage medium, which utilize a gated graph neural network for feature extraction, an autoencoder neural network model based on energy basis compression for feature dimension reduction, and an extreme learning machine classification algorithm with an enhanced annealing strategy for classification, thereby effectively improving the accuracy and efficiency of the perception results, and having strong adaptability and robustness to abnormal environments, thereby solving the problems of inaccurate obstacle perception and poor robustness of existing methods.

[0005] The present application provides an obstacle perception method, the method comprising:

[0006] Obtain vehicle surrounding environment data from onboard IoT sensors;

[0007] The gated graph neural network is used for feature extraction, the trained energy-based compression autoencoder neural network model is used for feature dimensionality reduction, and then the extreme learning machine classifier is used for classification to perceive and predict obstacles.

[0008] In the above implementation process, the gated graph neural network is used for feature extraction, which improves the model's ability to identify and extract key features, and effectively improves the accuracy and efficiency of feature extraction; the autoencoder neural network model based on energy basis compression is used for feature dimensionality reduction, which can better consider the energy distribution of the data in the encoding process, obtain a more essential and compact feature representation, and ensure that the decoded data can accurately restore the original features; the extreme learning machine classification algorithm with enhanced annealing strategy is used for classification, and the learning parameters are dynamically adjusted to avoid falling into local optimality, accelerate convergence, and improve classification accuracy and robustness, thereby effectively improving the accuracy and efficiency of the perception results, and having strong adaptability and robustness to abnormal environments, solving the problems of inaccurate obstacle perception and poor robustness of existing methods.

[0009] Furthermore, the method further includes training a generative adversarial network to expand the training data:

[0010] Initialize the generator and discriminator according to the preset network structure parameters;

[0011] Calculate the optimal flow path between the new data generated by the generator and the real data to optimize the distribution of the new data;

[0012] Calculate the dynamic loss function for each new data generated;

[0013] Updating the discriminator using the dynamic loss function;

[0014] The generator is updated based on the update result of the discriminator, and the optimal flow path between the new data generated by the generator and the real data is recalculated until a preset number of iterations is reached.

[0015] In the above implementation process, the generative adversarial network is optimized through network flow theory. The global distribution characteristics of the data are taken into account in the process of generating data, the quality and diversity of the data are optimized, and the generalization ability and accuracy of the model when facing limited training samples are effectively improved.

[0016] Furthermore, the feature extraction using a gated graph neural network includes:

[0017] The features of the expanded training data are input into a fully connected neural network for processing to obtain the weight coefficient of each feature value;

[0018] Normalizing the weight coefficient to obtain an attention coefficient;

[0019] Select the first K largest attention coefficients and the corresponding index values, and obtain the corresponding eigenvalues ​​according to the index values, and use them as the key eigenvalues;

[0020] Constructing a graph neural network with the key feature values ​​as nodes;

[0021] A gated graph neural network is constructed using a gated cyclic update mechanism to interact with nodes and obtain the hidden states of the nodes.

[0022] In the above implementation process, graph neural networks are used to extract features, and the gating mechanism is used to enhance the model's ability to identify and extract key features, effectively improving the accuracy and efficiency of feature extraction.

[0023] Furthermore, the gated graph neural network is constructed by using a gated cyclic update mechanism to perform information interaction of nodes to obtain the hidden state of the nodes, including:

[0024] Using characteristic key values initialization t =0 when the hidden state;

[0025] At time t, a gated loop update mechanism is used to enable each node to aggregate messages from its neighbors and obtain an aggregated feature vector;

[0026] The aggregated feature vector and the previous time step are used to update the hidden state of each node through a gating mechanism, and the final hidden state is obtained after multiple update iterations.

[0027] In the above implementation process, a gated recurrent update mechanism is used to pass information through the graph and learn the associated features of the context.

[0028] Furthermore, before the step of performing feature dimensionality reduction using the trained autoencoder neural network model based on energy basis compression, the method further comprises training the autoencoder neural network model based on energy basis compression:

[0029] Randomly initialize the encoder parameters and decoder parameters of the autoencoder;

[0030] Convert the input data into a low-dimensional feature representation through an encoder;

[0031] Calculating the energy function of the low-dimensional feature;

[0032] Inputting the low-dimensional features into a decoder to obtain reconstructed data of the input data;

[0033] Minimize the reconstruction error between input data and reconstructed data using the optimal mapping strategy;

[0034] A loss function is constructed based on the reconstruction error and the energy machine function, and the encoder parameters and decoder parameters are updated using the gradients of the loss function on the encoder parameters and the decoder parameters, and then reconverted into low-dimensional feature representations, and the above steps are repeated until the preset number of iterations is reached.

[0035] In the above implementation process, the energy basis function is used so that the autoencoder not only reconstructs the data, but also takes into account the energy distribution of the data during the encoding process, thereby obtaining a more essential and compact feature representation.

[0036] Furthermore, the method also includes training the extreme learning machine classifier using an enhanced annealing strategy:

[0037] Initialize the weights and bias parameters of the extreme learning machine classifier;

[0038] Based on the weight and bias parameters, input the training samples into the extreme learning machine classifier to obtain a prediction result;

[0039] The loss function is calculated based on the actual labels and prediction results of the training samples:

[0040] In each iteration, the learning parameters are dynamically adjusted through iterative updates of the simulated annealing algorithm;

[0041] The weight and bias parameters of the extreme learning machine classifier are updated according to the gradient of the loss function, and the prediction results are obtained again according to the updated weight and bias parameters. The above steps are repeated until the preset number of iterations is reached.

[0042] In the above implementation process, an enhanced annealing strategy is adopted to simulate the annealing process in the physical process and find the global optimal solution by gradually lowering the temperature of the system; during the training process of the classifier, this strategy is used to dynamically adjust the learning parameters to avoid falling into the local optimum.

[0043] The present application also provides an obstacle sensing device, the device comprising:

[0044] A data acquisition module, used to acquire vehicle surrounding environment data from vehicle-mounted IoT sensors;

[0045] The perception prediction module is used to extract features using a gated graph neural network, reduce the feature dimension using a trained autoencoder neural network model based on energy basis compression, and then classify using an extreme learning machine classifier to predict the perception of obstacles.

[0046] In the above implementation process, the gated graph neural network is used for feature extraction, which improves the model's ability to identify and extract key features, and effectively improves the accuracy and efficiency of feature extraction; the autoencoder neural network model based on energy basis compression is used for feature dimensionality reduction, which can better consider the energy distribution of the data in the encoding process, obtain a more essential and compact feature representation, and ensure that the decoded data can accurately restore the original features; the extreme learning machine classification algorithm with enhanced annealing strategy is used for classification, and the learning parameters are dynamically adjusted to avoid falling into local optimality, accelerate convergence, and improve classification accuracy and robustness, thereby effectively improving the accuracy and efficiency of the perception results, and having strong adaptability and robustness to abnormal environments, solving the problems of inaccurate obstacle perception and poor robustness of existing methods.

[0047] Furthermore, the perception prediction module includes a feature extraction module, which is used to:

[0048] The features of the expanded training data are input into a fully connected neural network for processing to obtain the weight coefficient of each feature value;

[0049] Normalizing the weight coefficient to obtain an attention coefficient;

[0050] Select the first K largest attention coefficients and the corresponding index values, and obtain the corresponding eigenvalues ​​according to the index values, and use them as the key eigenvalues;

[0051] Constructing a graph neural network with the key feature values ​​as nodes;

[0052] A gated graph neural network is constructed using a gated cyclic update mechanism to interact with nodes and obtain the hidden states of the nodes.

[0053] In the above implementation process, graph neural networks are used to extract features, and the gating mechanism is used to enhance the model's ability to identify and extract key features, effectively improving the accuracy and efficiency of feature extraction.

[0054] An embodiment of the present application further provides an electronic device, which includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform any of the above-mentioned obstacle perception methods.

[0055] An embodiment of the present application further provides a readable storage medium, wherein the readable storage medium stores computer program instructions, and when the computer program instructions are read and executed by a processor, any of the above-mentioned obstacle perception methods is executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 A flowchart of an obstacle perception method provided in an embodiment of the present application;

[0058] Figure 2 A flowchart for training a generative adversarial network provided in an embodiment of the present application;

[0059] Figure 3 A schematic diagram of the structure of a gated graph neural network provided in an embodiment of the present application;

[0060] Figure 4 A flow chart of feature extraction using a gated graph neural network provided in an embodiment of the present application;

[0061] Figure 5 A flowchart for obtaining the hidden state of a node provided in an embodiment of the present application;

[0062] Figure 6 A flowchart for training an autoencoder neural network model based on energy basis compression provided in an embodiment of the present application;

[0063] Figure 7 A training flow chart of an extreme learning machine classifier provided in an embodiment of the present application;

[0064] Figure 8 A structural block diagram of an obstacle sensing device provided in an embodiment of the present application;

[0065] Fig. 9 A structural block diagram of another obstacle sensing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0067] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0068] Example 1

[0069] Please see Figure 1 , Figure 1 A flowchart of an obstacle perception method provided in an embodiment of the present application. The method specifically comprises the following steps:

[0070] Step S100: Acquire vehicle surrounding environment data from the vehicle-mounted IoT sensor;

[0071] The vehicle-mounted IoT sensor can capture comprehensive information about the vehicle's surroundings. In one embodiment, the attributes of the data include but are not limited to:

[0072] : Longitude information of the vehicle's current location; : Latitude information of the vehicle’s current location; : The relative height between the vehicle and the ground; : The current speed of the vehicle; : The angle between the vehicle's forward direction and the north direction; : vehicle tilt angle; : Timestamp of data collection; : Data from the acoustic wave sensor, reflecting the distance of surrounding obstacles; : Data from LIDAR (laser radar) in the form of a point cloud that details the 3D structure around the vehicle; : Ambient temperature and humidity information helps determine the impact of weather conditions on obstacle perception.

[0073] It should be noted that the above are only examples of data formats and types. In actual applications, the attributes of data are usually more than 10 attributes, and may even reach dozens or even hundreds. No limitation is made on the types of data here.

[0074] During training, the collected data is labeled. For example, the labeled categories include pedestrians, other vehicles, roadblocks, animals, and other obstacles.

[0075] Step S200: Use a gated graph neural network to extract features, use a trained energy-based compression autoencoder neural network model to reduce feature dimensions, and then use an extreme learning machine classifier to classify, so as to perceive and predict obstacles.

[0076] The trained model is used to process new samples to complete the task of intelligent driving obstacle perception. In one embodiment, the trained model is used to extract features and reduce the dimension of new samples, and then input them into the classifier for classification. In this embodiment, the labeled categories include one of the five categories of pedestrians, other vehicles, roadblocks, animals, and other obstacles.

[0077] Insufficient training data can easily lead to poor model generalization ability and affect the accuracy of the model. Therefore, this application adopts a network flow optimized generative adversarial network. The network flow strategy takes into account the global characteristics of data distribution, making the generated data more diverse and closer to the real data distribution. Compared with traditional generative adversarial networks, this application takes into account the fluidity of data in the feature space during the generation process, thereby optimizing the quality and diversity of generated data, and effectively improving the generalization ability and accuracy of the model when facing limited training samples.

[0078] The method also includes training a generative adversarial network to expand the training data, such as Figure 2 As shown in FIG. 1 , it is a flowchart for training a generative adversarial network, which specifically includes the following steps:

[0079] Step S211: Initialize the generator and discriminator according to preset network structure parameters;

[0080] The generator is used to generate data, while the discriminator’s task is to distinguish between generated data and real data. G Parameters Initialize by random initialization, the discriminator D Parameters Initialization is performed using random initialization.

[0081] Step S212: Calculate the optimal flow path between the new data generated by the generator and the real data to optimize the distribution of the new data;

[0082] In each iteration, the generator generates new data samples and uses network flow theory to calculate the optimal flow path between the generated new data and the real data, so as to optimize the distribution of the generated data and make it closer to the distribution of the real data. Specifically, let the generator G Noise at the input The new data generated is:

[0083] ;

[0084] Furthermore, let the real data set be , where each data item , designing network flow models to optimize generators G , so that the generated new data The distribution is as close to the real data as possible The distribution of . Using Wasserstein distance As an optimization objective, it can be expressed as:

[0085] ;

[0086] in, express( x , y ) is subject to Y Distribution of x , y Represent two data points, for example x represents the original sample, y Represents the generated new sample; Represents the set of all possible joint distributions that transform real data into new data;

[0087] Furthermore, the Wasserstein distance is calculated by solving an optimal transmission problem, that is, the method to find the lowest cost is to transform the distribution of generated data into the distribution of real data. The calculation method of the cost function can be expressed as:

[0088] ;

[0089] in, Indicates that from the real data point To new data point The flow rate must satisfy the flow conservation and non-negativity constraints.

[0090] Step S213: Calculate the dynamic loss function of the new data generated each time;

[0091] After each data generation, a dynamic loss function is calculated, which automatically adjusts the loss weight according to the distribution difference between the current new data and the real data, so that the model can flexibly adapt to the changes in data distribution, thereby more effectively optimizing the generation process. Specifically, for the discriminator D , the loss function is defined as:

[0092] ;

[0093] Among them, E represents expectation;

[0094] And, for the generator G , the loss function is defined as:

[0095] ;

[0096] Furthermore, the calculation method of the dynamic adjustment factor can be expressed as:

[0097] ;

[0098] in, t is the number of iterations, k and t 0 is a preset constant. For example, in one embodiment, k is set to 0.01, t0 is set to 50.

[0099] Furthermore, the calculation method of the dynamic loss function can be expressed as:

[0100] ;

[0101] ;

[0102] in, and They represent the loss functions of the discriminator and generator that are dynamically adjusted during the training process.

[0103] Step S214: updating the discriminator using the dynamic loss function;

[0104] Use the dynamic loss function to update the discriminator to improve its ability to distinguish between generated new data and real data. Specifically, the discriminator update method can be expressed as:

[0105] ;

[0106] in, represents the parameter gradient of the discriminator;

[0107] And, the generator update method can be expressed as:

[0108] ;

[0109] in, and are the learning rates of the discriminator and the generator, respectively. In one embodiment, the learning rate and Set to 0.0002 and 0.0002 respectively.

[0110] Furthermore, the calculation method of the discriminator parameter gradient can be expressed as:

[0111] ;

[0112] Moreover, the calculation method of the generator parameter gradient can be expressed as:

[0113] .

[0114] Step S215: updating the generator based on the update result of the discriminator, and recalculating the optimal flow path between the new data generated by the generator and the real data until a preset number of iterations is reached.

[0115] According to the feedback from the discriminator, the generator is updated to generate new data that is closer to the real data distribution. Specifically, steps S212 to S214 are repeated until a preset number of iterations is reached. In one embodiment, the number of iterations is set to 100 times.

[0116] The graph neural network is used to extract features, and the gating mechanism is used to enhance the model's ability to identify and extract key features, effectively improving the accuracy and efficiency of feature extraction. Figure 3 As shown in the figure, it is a schematic diagram of the structure of the gated graph neural network. Figure 4 As shown in FIG, it is a flowchart of feature extraction using a gated graph neural network, which specifically includes the following steps:

[0117] Step S220: input the features of the expanded training data into a fully connected neural network for processing to obtain a weight coefficient for each feature value;

[0118] The characteristics of the expanded training data are consistent with the characteristics of the collected training data. The characteristics of the expanded training data here refer to the expanded data set. For one data, when it is input into the fully connected neural network, it is input in the form of feature values. This is a common description in this field and will not be described in detail here.

[0119] First, the features of the expanded training data input into the feature extraction module are sent to the fully connected neural network to learn the attention coefficient of each eigenvalue, and the first K eigenvalues ​​with larger attention coefficients are selected as key eigenvalues; secondly, the key eigenvalues ​​are used as nodes to construct a graph neural network, and the attention coefficients are used to construct an adjacency matrix. The graph neural network is used to interact with the features to obtain updated features, that is, the features after feature extraction of the expanded training data.

[0120] Specifically, in the node generation module, the input features are set as:

[0121] ;

[0122] in, B , R Represent the feature dimension and the number of eigenvalues ​​respectively. After being processed by the fully connected neural network, the weight coefficient of each eigenvalue is obtained, which can be expressed as:

[0123] ;

[0124] And the processing method of the fully connected neural network can be expressed as:

[0125] ;

[0126] in, W is the weight, b is the bias,x is the input feature.

[0127] Step S230: normalizing the weight coefficient to obtain an attention coefficient;

[0128] The weight coefficient C The attention coefficient is obtained after normalization using the softmax function and can be expressed as:

[0129] ;

[0130] in, is the attention coefficient, C r For the r The weight coefficient of the eigenvalue.

[0131] Step S240: Select the first K maximum attention coefficients and the corresponding index values, and obtain the corresponding eigenvalues ​​according to the index values, and use them as key eigenvalues;

[0132] Furthermore, we select the first K largest attention coefficients and record them as , and record the index value of the corresponding eigenvalue , get the corresponding eigenvalue according to the index value These K eigenvalues ​​have larger attention coefficients and receive more attention from the model, so they are considered key eigenvalues.

[0133] Step S250: constructing a graph neural network with the key feature values ​​as nodes;

[0134] Furthermore, there will be K The data of each feature is input into the feature interaction module, and a graph neural network is constructed with key feature values ​​as nodes. A gating mechanism is introduced into the graph neural network to realize information interaction of node features through gated graph neural networks (GGNN). Specifically, the constructed graph neural network is , where the node Indicates the node associated with the key eigenvalue, that is, k Key eigenvalues ​​and a node Associated. A is a size of The adjacency matrix of ,and Represents known eigenvalues The characteristic value under the condition k The probability of occurrence is calculated as follows:

[0135] ;

[0136] in, Indicates ks The attention coefficient of key eigenvalues, Indicates j The attention coefficient of the feature value, Indicates The attention coefficient of the feature value.

[0137] Step S260: construct a gated graph neural network using a gated loop update mechanism to perform information interaction of nodes to obtain the hidden state of the nodes.

[0138] Furthermore, the present application adopts a gated loop update mechanism to pass information through the graph and learn the associated features of the context.

[0139] like Figure 5 As shown, it is a flowchart of obtaining the hidden state of a node, and step S260 specifically includes the following steps:

[0140] Step S261: Using feature key values initialization t =0 when the hidden state:

[0141] ;

[0142] For each node , assuming that t The hidden state at time , initialized in the above manner.

[0143] Step S262: t At this moment, the gated loop update mechanism is used to make each node aggregate messages from its neighbors and obtain the aggregated feature vector:

[0144] ;

[0145] in, express t The hidden state at time -1.

[0146] Step S263: using the aggregated feature vector and the previous time step and updating the hidden state of each node through a gating mechanism, and obtaining the final hidden state after multiple update iterations.

[0147] Based on the aggregated feature vector and the hidden state of the previous time step , each node updates its hidden state through the gating mechanism, which can be expressed as:

[0148] ;

[0149] ;

[0150] ;

[0151] ;

[0152] in, represents the sigmoid function, is the hyperbolic tangent function, is the element-wise multiplication operation, , ,W,U, , , are the network parameters to be learned, , , , All are intermediate variables.

[0153] When the characteristics With features When the correlation is high, the graph neural network will transfer information between the two, otherwise it will inhibit information transfer. T Second, generate the final hidden state . That is, The output of the feature extraction module not only encodes the features It not only combines the information of feature sequences, but also integrates the context information of feature sequences.

[0154] In the process of feature dimensionality reduction, by introducing the energy basis model, the present invention can better consider the energy distribution of data in the encoding process, obtain a more essential and compact feature representation, and ensure that the decoded data can accurately restore the original features.

[0155] Feature dimension reduction model training: Input the data after feature extraction into the feature dimension reduction model to train the feature dimension reduction model. The present application proposes an autoencoder neural network algorithm based on energy basis compression. The traditional autoencoder learns the compressed representation of data by minimizing the reconstruction error. On this basis, the present invention introduces an energy basis model and defines an energy function to evaluate the compression efficiency and quality of the encoding, so that the autoencoder not only reconstructs the data, but also takes into account the energy distribution of the data during the encoding process, thereby obtaining a more essential and compact feature representation. In addition, during the decoding process, the present invention ensures that the decoded data can more accurately restore the characteristics of the original data through the optimal mapping strategy. By establishing the optimal mapping relationship between encoding and decoding, the feature loss is further reduced and the quality of the data after dimensionality reduction is improved.

[0156] like Figure 6As shown, it is a flow chart for training the autoencoder neural network model based on energy basis compression. Before the step of performing feature dimension reduction using the trained autoencoder neural network model based on energy basis compression, the method further includes:

[0157] Step S271: Randomly initialize the encoder parameters of the autoencoder and decoder parameters ;

[0158] Initialize the encoder and decoder parameters of the autoencoder, including network weights and biases. Specifically, set the encoder of the autoencoder E and decoder D The parameters are and , the way to initialize these parameters is random initialization.

[0159] Step S272: converting the input data into a low-dimensional feature representation through an encoder;

[0160] Perform forward propagation, during which the input data is converted into low-dimensional feature representations by the encoder, and then these features are reconstructed back to the original data space by the decoder. Consider the energy basis model to ensure that the feature representation has low energy, thereby achieving effective compression. Specifically, for a given input data x , encoder E Map it to a low-dimensional feature space z can be expressed as:

[0161] ;

[0162] in, is the input data, It is a feature representation.

[0163] Step S273: Calculate the energy function of the low-dimensional feature;

[0164] For the encoded features, calculate its energy function value to evaluate its compression efficiency and quality. Specifically, define the energy function To evaluate the feature representation z The compression efficiency and quality can be expressed as:

[0165] ;

[0166] in, is the first i elements, yes z The average value of n is the dimension of feature representation.

[0167] Furthermore, the feature representation z The average The calculation method can be expressed as:

[0168] ;

[0169] Step S274: inputting the low-dimensional features into a decoder to obtain reconstructed data of the input data;

[0170] In the decoding stage, the decoder parameters are optimized through the optimal mapping strategy to ensure that the reconstructed data is as close to the original data as possible while retaining the essential information of the low-dimensional features. D Try to get the feature representation from compression z Reconstructing the original input , which can be expressed as:

[0171] .

[0172] Step S275: Minimizing the reconstruction error between the input data and the reconstructed data using the optimal mapping strategy;

[0173] Furthermore, the optimal mapping strategy aims to minimize the original data x and reconstructing data The difference between them is usually expressed as:

[0174] ;

[0175] in, m is the number of data samples.

[0176] Step S276: construct a loss function based on the reconstruction error and the energy machine function and use the loss function to adjust the encoder parameters and decoder parameters Update the encoder parameters with the gradient and decoder parameters , and reconvert it into a low-dimensional feature representation, repeating the above steps until the preset number of iterations is reached.

[0177] Back propagation and parameter update are performed, and the parameters of the autoencoder are updated according to the gradient descent method by calculating the reconstruction error and its gradient to the parameters. The influence of the energy basis model and the optimal mapping strategy on the gradient is taken into account to guide the parameter update. Specifically, the loss function of the entire network combines the reconstruction error and the energy basis model, which can be expressed as:

[0178] ;

[0179] in, is a preset hyperparameter used to balance the impact of reconstruction error and energy function. In one embodiment, Set to 0.05.

[0180] Furthermore, by calculating the loss function L right and The gradient update parameter can be expressed as:

[0181] ;

[0182] ;

[0183] in, is the learning rate. In one embodiment, the learning rate is set to 0.1.

[0184] Furthermore, the loss function L Encoder parameters The gradient of can be expressed as:

[0185] ;

[0186] And, for the decoder parameters , the gradient is calculated as:

[0187] ;

[0188] in, and Respectively represent the impact of reconstruction error on encoder and decoder parameters, Represents the influence of the energy function on the encoder parameters.

[0189] Repeat step S272 to step S276 until a preset number of iterations is reached. In one embodiment, the preset number of iterations is 1000.

[0190] An enhanced annealing strategy is introduced during classifier training to dynamically adjust learning parameters, avoid falling into local optimality, accelerate convergence, and improve classification accuracy and robustness.

[0191] Specifically, training the classifier: inputting the data after feature dimension reduction into the classifier to train the classifier. The present invention proposes an extreme learning machine classification algorithm based on an enhanced annealing strategy, which uses an enhanced annealing strategy to simulate the annealing process in the physical process and gradually lower the temperature of the system to find the global optimal solution. During the training process of the classifier, this strategy is used to dynamically adjust the learning parameters to avoid falling into the local optimum and accelerate convergence. In addition, in order to improve the robustness of the classifier in the face of noisy data, the present application adopts a robust optimization strategy, by adding a regularization term to the loss function, reducing the impact of outliers and enhancing the model's ability to handle abnormal data.

[0192] like Figure 7 As shown, it is a training flow chart of the extreme learning machine classifier, and the method also includes training the extreme learning machine classifier using an enhanced annealing strategy:

[0193] Step S281: Initialize the weight and bias parameters of the extreme learning machine classifier;

[0194] Initialize the weights and bias parameters of the extreme learning machine classifier, weights w The initialization method can be expressed as:

[0195] ;

[0196] Bias b The initialization method can be expressed as:

[0197] ;

[0198] in, represents the number of neurons in the input layer, represents the number of neurons in the output layer, randn ()and rand () represent random sampling of initial values ​​from normal distribution and uniform distribution respectively.

[0199] Step S282: Based on the weight and bias parameters, input the training sample into the extreme learning machine classifier to obtain a prediction result;

[0200] The input data is passed through the extreme learning machine classifier to calculate the probability of each sample belonging to each category. Specifically, the input data The process of forward propagation through the extreme learning machine can be expressed as:

[0201] ;

[0202] in, H represents the output of the hidden layer, is an activation function. In one embodiment, a Sigmoid activation function is used.

[0203] Step S283: Calculate the loss function according to the actual label and prediction result of the training sample. The loss function combines the cross entropy loss and the regularization term of robust optimization, which can be expressed as:

[0204] ;

[0205] in, N represents the number of training samples, Representation sample i The actual label, Representation sample i The predicted label of represents the regularization coefficient, Represents the weight w Regularization term for robust optimization;

[0206] In one embodiment, the regularization coefficient is set to 0.001.

[0207] Regularization term R ( w ) is used to reduce the overfitting of the model, and its calculation formula can be expressed as:

[0208] ;

[0209] in, w i Represents the weight matrix w The i elements, n is the total number of elements in the weight matrix. L2 regularization encourages weight values ​​to approach 0, but not completely 0, which helps control model complexity and avoid overfitting.

[0210] Step S284: In each iteration process, dynamically adjusting the learning parameters through iterative updates of the simulated annealing algorithm;

[0211] In each iteration, the learning parameters, including the learning rate and the coefficient of the regularization term, are dynamically adjusted according to the enhanced annealing strategy. This process simulates the annealing process, gradually lowering the temperature of the system to find the optimal solution. Specifically, the process of adjusting the learning parameters by the enhanced annealing strategy is achieved through iterative updates of the simulated annealing algorithm, where the parameter update at each step can be expressed as:

[0212] ;

[0213] ;

[0214] in, It is in tThe learning rate is dynamically adjusted according to the annealing strategy in the iteration. and Represent the loss function L Weight w and bias b gradient.

[0215] Furthermore, the learning rate The calculation method can be expressed as:

[0216] ;

[0217] in, is the initial learning rate, t is the current iteration number, is a preset time constant. In one embodiment, the initial learning rate is set to 0.01, the time constant is set to 50.

[0218] Step S285: Update the weight and bias parameters of the extreme learning machine classifier according to the gradient of the loss function, and re-obtain the prediction result according to the updated weight and bias parameters, and repeat the above steps until the preset number of iterations is reached.

[0219] According to the gradient of the loss function, the weight and bias parameters of the extreme learning machine classifier are updated. Specifically, based on the back propagation algorithm, the update method of weights and biases can be expressed as:

[0220] ;

[0221] ;

[0222] in, is the global learning rate, and They represent the gradients of the loss function with respect to weights and biases, respectively.

[0223] Furthermore, the loss function L has a certain effect on the weights w and bias b The gradient calculation is implemented through the back propagation algorithm and can be expressed as:

[0224] ;

[0225] in, It represents the error between the predicted value and the true value. is the corresponding input feature, From the L2 regularization term w s The derivative of N Indicates the sample size.

[0226] Bias b Update method and weight of w same.

[0227] Iterative optimization: Steps S282 to S285 are repeatedly executed until a preset maximum number of iterations is met. In one embodiment, the preset maximum number of iterations is 1000.

[0228] Through efficient data collection and processing methods, the present invention can more accurately identify and classify various obstacles in the intelligent driving environment, including pedestrians, vehicles, etc., and improve the safety of the intelligent driving system. Using the network flow optimized generative adversarial network and the feature extraction method based on the graph neural network, the present invention can effectively improve the generalization ability of the model in different environments and conditions, and enhance the adaptability and flexibility of the system.

[0229] Through advanced algorithm design, the present invention improves the efficiency of data processing and model training and reduces resource consumption while ensuring high data processing quality.

[0230] The present invention effectively improves the quality of feature dimensionality reduction and classification accuracy through the extreme learning machine classification algorithm of the energy-based compression autoencoder and the enhanced annealing strategy, providing more accurate decision support for intelligent driving.

[0231] Example 2

[0232] The present application also provides an obstacle sensing device, such as Figure 8 The figure is a structural block diagram of an obstacle sensing device, which includes but is not limited to:

[0233] The data acquisition module 100 is used to acquire vehicle surrounding environment data from the vehicle-mounted IoT sensor;

[0234] The perception prediction module 200 is used to perform feature extraction using a gated graph neural network, perform feature dimension reduction using a trained autoencoder neural network model based on energy basis compression, and then perform classification using an extreme learning machine classifier to perform perception prediction of obstacles.

[0235] like Fig. 9 FIG. 2 is a block diagram of another obstacle sensing device, wherein the sensing prediction module 200 includes a feature extraction module 201, which is used to:

[0236] The features of the expanded training data are input into a fully connected neural network for processing to obtain the weight coefficient of each feature value;

[0237] Normalizing the weight coefficient to obtain an attention coefficient;

[0238] Select the first K largest attention coefficients and the corresponding index values, and obtain the corresponding eigenvalues ​​according to the index values, and use them as the key eigenvalues;

[0239] Constructing a graph neural network with the key feature values ​​as nodes;

[0240] A gated graph neural network is constructed using a gated cyclic update mechanism to interact with nodes and obtain the hidden states of the nodes.

[0241] The device also includes a generative adversarial network training module 301, an autoencoder neural network model training module 302, and a classifier training module 303. The specific training process has been specifically described in Example 1 and will not be repeated here.

[0242] An embodiment of the present application further provides an electronic device, which includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the obstacle perception method described in Example 1.

[0243] The embodiment of the present application further provides a readable storage medium, wherein the readable storage medium stores computer program instructions, and when the computer program instructions are read and executed by a processor, the obstacle perception method described in Embodiment 1 is executed.

[0244] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0245] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0246] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0247] The above description is only an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0248] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0249] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

Claims

1. An obstacle perception method, characterized in that: The method comprises: Obtain vehicle surrounding environment data from onboard IoT sensors; The gated graph neural network is used for feature extraction, the trained energy-based compression autoencoder neural network model is used for feature dimension reduction, and then the extreme learning machine classifier is used for classification to perceive and predict obstacles. The method further includes training a generative adversarial network to expand the training data: Initialize the generator and discriminator according to the preset network structure parameters; Calculate the optimal flow path between the new data generated by the generator and the real data to optimize the distribution of the new data; Calculate the dynamic loss function for each new data generated; Updating the discriminator using the dynamic loss function; Based on the update result of the discriminator, the generator is updated, and the optimal flow path between the new data generated by the generator and the real data is recalculated until a preset number of iterations is reached; Before the step of using the trained autoencoder neural network model based on energy basis compression to perform feature dimension reduction, the method further includes training the autoencoder neural network model based on energy basis compression: Randomly initialize the encoder parameters and decoder parameters of the autoencoder; Convert the input data into a low-dimensional feature representation through an encoder; Calculating the energy function of the low-dimensional feature; Inputting the low-dimensional features into a decoder to obtain reconstructed data of the input data; Minimize the reconstruction error between input data and reconstructed data using the optimal mapping strategy; A loss function is constructed based on the reconstruction error and the energy basis function, and the encoder parameters and decoder parameters are updated using the gradient of the loss function to reconvert them into low-dimensional feature representations, and the above steps are repeated until the preset number of iterations is reached.

2. The obstacle sensing method according to claim 1, characterized in that: The feature extraction using a gated graph neural network includes: The features of the expanded training data are input into a fully connected neural network for processing to obtain the weight coefficient of each feature value; Normalizing the weight coefficient to obtain an attention coefficient; Select the first K largest attention coefficients and the corresponding index values, and obtain the corresponding eigenvalues ​​according to the index values, and use them as the key eigenvalues; Constructing a graph neural network with the key feature values ​​as nodes; A gated graph neural network is constructed using a gated cyclic update mechanism to interact with nodes and obtain the hidden states of the nodes.

3. The obstacle sensing method according to claim 2, characterized in that: The gated graph neural network is constructed by using a gated cyclic update mechanism to perform information interaction of nodes to obtain the hidden state of the nodes, including: Using the feature key value f k Initialize the hidden state at t = 0; At time t, a gated loop update mechanism is used to enable each node to aggregate messages from its neighbors and obtain an aggregated feature vector; The aggregated feature vector and the previous time step are used to update the hidden state of each node through a gating mechanism, and the final hidden state is obtained after multiple update iterations.

4. The obstacle sensing method according to claim 1, characterized in that: The method also includes training an extreme learning machine classifier using an enhanced annealing strategy: Initialize the weights and bias parameters of the extreme learning machine classifier; Based on the weight and bias parameters, input the training samples into the extreme learning machine classifier to obtain a prediction result; The loss function is calculated based on the actual labels and prediction results of the training samples: In each iteration, the learning parameters are dynamically adjusted through iterative updates of the simulated annealing algorithm; The weight and bias parameters of the extreme learning machine classifier are updated according to the gradient of the loss function, and the prediction results are regained according to the updated weight and bias parameters. The above steps are repeated until the preset number of iterations is reached.

5. An obstacle sensing device, characterized in that: The device comprises: A data acquisition module, used to acquire vehicle surrounding environment data from vehicle-mounted IoT sensors; The perception prediction module is used to extract features using a gated graph neural network, reduce the feature dimension using a trained autoencoder neural network model based on energy basis compression, and then classify using an extreme learning machine classifier to predict the perception of obstacles; Wherein, the device further comprises: Generate adversarial network training module, used to initialize the generator and discriminator according to the preset network structure parameters; Calculate the optimal flow path between the new data generated by the generator and the real data to optimize the distribution of the new data; Calculate the dynamic loss function for each new data generated; Updating the discriminator using the dynamic loss function; Based on the update result of the discriminator, the generator is updated, and the optimal flow path between the new data generated by the generator and the real data is recalculated until a preset number of iterations is reached; Wherein, the device further comprises: An autoencoder neural network model training module is used to randomly initialize the encoder parameters and decoder parameters of the autoencoder before the step of using the trained energy-based compression-based autoencoder neural network model to perform feature dimensionality reduction; Convert the input data into a low-dimensional feature representation through an encoder; Calculating the energy function of the low-dimensional feature; Inputting the low-dimensional features into a decoder to obtain reconstructed data of the input data; Minimize the reconstruction error between input data and reconstructed data using the optimal mapping strategy; A loss function is constructed based on the reconstruction error and the energy basis function, and the encoder parameters and decoder parameters are updated using the gradient of the loss function to reconvert them into low-dimensional feature representations, and the above steps are repeated until the preset number of iterations is reached.

6. The obstacle sensing device according to claim 5, characterized in that: The perception prediction module includes a feature extraction module, which is used to: The features of the expanded training data are input into a fully connected neural network for processing to obtain the weight coefficient of each feature value; Normalizing the weight coefficient to obtain an attention coefficient; Select the first K largest attention coefficients and the corresponding index values, and obtain the corresponding eigenvalues ​​according to the index values, and use them as the key eigenvalues; Constructing a graph neural network with the key feature values ​​as nodes; A gated graph neural network is constructed using a gated cyclic update mechanism to interact with nodes and obtain the hidden states of the nodes.

7. An electronic device, characterized in that: The electronic device comprises a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the obstacle perception method according to any one of claims 1 to 4.

8. A readable storage medium, characterized in that: The readable storage medium stores computer program instructions, and when the computer program instructions are read and executed by a processor, the obstacle perception method according to any one of claims 1 to 4 is executed.

Images