Vehicle interaction scene analysis method and device, equipment, medium and product

Abstract

The invention provides a vehicle interaction scene analysis method and device, equipment, a medium and a product. According to one embodiment of the invention, the method comprises the following steps: carrying out feature extraction on time sequence position data to obtain relative features, the time sequence position data being position data of a target vehicle and a handcart in a plurality of time steps, and the relative features being used for reflecting relative motion data between the target vehicle and the handcart; weighting the relative features to obtain weighted features after weighting; performing maximum pooling on the weighted features to generate target features of the target vehicle and the handcart; according to the method, the target features are subjected to clustering analysis, the target interaction scene of the target vehicle and the handcart is determined, the weighted features are subjected to maximum pooling, and the most significant target features can be extracted, so that the target interaction scene of the target vehicle and the handcart is accurately analyzed.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment, medium and product for analyzing vehicle interaction scenarios. Background Art

[0002] With the rapid development of autonomous driving technology, especially in the early data processing stage of autonomous driving, accurately capturing and understanding the interactive scenarios such as overtaking and lane changing between the target vehicle and the opponent vehicle can provide more accurate data support for autonomous driving decision-making and planning.

[0003] At present, due to the complex interaction relationship between the target car and the opponent car, it is often impossible to accurately analyze the specific interaction scenario between the target car and the opponent car. Summary of the invention

[0004] In order to overcome the problems existing in the related art, the present application provides a method, device, equipment, medium and product for analyzing vehicle interaction scenarios.

[0005] According to a first aspect of any embodiment of the present application, a method for analyzing a vehicle interaction scenario is provided, the method comprising:

[0006] Extracting features from the time series position data to obtain relative features; wherein the time series position data is the position data of the target vehicle and the opponent vehicle within a plurality of time steps, and the relative features are used to reflect the relative motion data between the target vehicle and the opponent vehicle;

[0007] Weighting the relative features to obtain weighted features;

[0008] Performing maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle;

[0009] Cluster analysis is performed on the target features to determine a target interaction scenario between the target vehicle and the opponent vehicle.

[0010] According to a second aspect of any embodiment of the present application, a vehicle interaction scenario analysis device is provided, the device comprising:

[0011] An extraction module is used to extract features from the time series position data to obtain relative features; wherein the time series position data is the position data of the target vehicle and the opponent vehicle in multiple time steps, and the relative features are used to reflect the relative motion data between the target vehicle and the opponent vehicle;

[0012] A weighting module, used for weighting the relative features to obtain weighted features;

[0013] A pooling module, used for performing maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle;

[0014] The analysis module is used to perform cluster analysis on the target features to determine the target interaction scenario between the target vehicle and the opponent vehicle.

[0015] According to a third aspect of any embodiment of the present application, there is provided an electronic device, including:

[0016] processor;

[0017] a memory for storing processor-executable instructions;

[0018] The processor implements the method described in any embodiment of the present application by running the executable instructions.

[0019] According to a fourth aspect of any embodiment of the present application, a computer-readable storage medium is provided, on which computer instructions are stored. When the instructions are executed by a processor, the method described in any embodiment of the present application is implemented.

[0020] According to a fifth aspect of any embodiment of the present application, a computer program product is provided, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the method described in any embodiment of the present application is implemented.

[0021] The technical solution provided by this application may have the following beneficial effects:

[0022] According to the above embodiments, by performing feature extraction on the time series position data to obtain relative features, weighting the relative features to obtain weighted features, performing maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle, performing cluster analysis on the target features to determine the target interaction scenario between the target vehicle and the opponent vehicle, and performing maximum pooling on the weighted features to extract the most significant target features, thereby accurately analyzing the target interaction scenario between the target vehicle and the opponent vehicle.

[0023] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings, which are incorporated in the specification and constitute a part of this application, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.

[0025] Figure 1 is a flow chart of a vehicle interaction scenario analysis method shown in the present application according to an exemplary embodiment;

[0026] Figure 2 is a flow chart of a vehicle interaction scenario analysis method shown in the present application according to an exemplary embodiment;

[0027] Figure 3 is a schematic diagram of a vehicle interaction scenario analysis method shown in the present application according to an exemplary embodiment;

[0028] Figure 4 is a structural schematic diagram of an electronic device according to an exemplary embodiment of the present application;

[0029] Figure 5 It is a block diagram of a vehicle interaction scenario analysis device shown in the present application according to an exemplary embodiment. DETAILED DESCRIPTION

[0030] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0031] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0032] It should be understood that although the terms first, second, third, etc. may be used in the present application to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0033] Current vehicle interaction dynamic scene analysis methods mainly rely on implicit learning methods, which directly obtain and process vehicle trajectory data, allowing the model to explore and learn the interaction scenarios between vehicles on its own.

[0034] However, implicit learning methods have poor interpretability. Since the model learns the interaction pattern directly from the trajectory data and cannot reveal the relationship between the various features, the internal structure and decision-making process of the model are often opaque, making it difficult to intuitively understand and interpret the results of the model output, making it difficult for developers to optimize and adjust the model in a targeted manner, limiting the scalability and flexibility of the model.

[0035] In addition, implicit learning methods converge slowly when dealing with complex dynamic scenes, which often contain a large number of variable vehicle interaction behaviors that require the model to have strong learning and generalization capabilities to accurately capture.

[0036] Implicit learning methods usually also require a large amount of data and computing resources to train the model, which not only increases the time and cost of model training, but also may make it difficult for the model to converge effectively in complex scenarios, thus affecting its performance and stability in practical applications.

[0037] In order to solve the above problems, the present application proposes a method for analyzing vehicle interaction scenarios. To further illustrate the present application, the following embodiments are provided:

[0038] See also Figure 1 , Figure 1 This is a flow chart of an analysis method for a vehicle interaction scenario according to an exemplary embodiment of the present application. The analysis method for the vehicle interaction scenario can be executed by an analysis model in an analysis system, and the analysis system can be applied to a vehicle or to a server such as a cloud or a cluster server.

[0039] like Figure 1 As shown, the analysis method may include the following steps:

[0040] Step 102: extract features from the time series position data to obtain relative features, wherein the time series position data is the position data of the target vehicle and the opponent vehicle within multiple time steps, and the relative features are used to reflect the relative motion data between the target vehicle and the opponent vehicle.

[0041] In this step, in the early data processing stage of autonomous driving, the vehicle operation data within a certain period of time in the vehicle driving scene pre-collected by the analysis system is analyzed. The vehicle operation data includes the time series position data of the target vehicle and the opponent vehicles around the target vehicle.

[0042] The analysis system inputs the time series position data within a preset time period into the analysis model, and the analysis model extracts features from the time series position data of the target vehicle and the opponent vehicle to obtain relative features within the preset time period.

[0043] The target vehicle is the subject in the traffic scene, that is, the vehicle to be analyzed for the interaction scene. The opponent vehicle is the vehicle with the highest probability of interaction with the target vehicle in the vehicle driving scene. There can only be the opponent vehicle with the highest probability of interaction with the target vehicle around the target vehicle, or there can be multiple vehicles around the target vehicle.

[0044] The time series position data is the position data of the target vehicle and the opponent vehicle in multiple time steps, that is, the position information recorded over time during the driving process of the target vehicle and the opponent vehicle, which can be the position coordinates expressed in the target vehicle coordinate system of the target vehicle. The multiple time steps can be multiple continuous time steps.

[0045] Relative features are used to reflect the relative motion data between the target car and the opponent car, which can be display features such as relative position difference, relative speed, relative acceleration, etc. Relative motion data refers to the motion information such as the relative position and speed between the target car and the opponent car.

[0046] If there are multiple vehicles around the target vehicle, the analysis system can input the time series position data into the analysis model, and the analysis model extracts features from the time series position data of the target vehicle and multiple vehicles to obtain multiple sets of relative features within a preset time period. Each set of relative features is used to reflect the relative motion data between the target vehicle and each vehicle.

[0047] Exemplarily, the interaction relationship of vehicles in the time series feature data may include: target vehicle and opponent vehicle a, target vehicle and opponent vehicle b, and target vehicle and opponent vehicle c. The analysis model extracts features from the time series position data of the target vehicle and opponent vehicle a, the time series position data of the target vehicle and opponent vehicle b, and the time series position data of the target vehicle and opponent vehicle c, respectively, to obtain three sets of relative features.

[0048] Step 104: weighting the relative features to obtain weighted features.

[0049] In this step, the analysis model weights the relative features, uses weighting techniques such as attention mechanism and gradient weighting method, calculates the weights of different feature dimensions of the relative features, and performs weighted summation of the feature dimensions according to these weights to generate weighted features. Among them, the weighted features are feature vectors that highlight the key features between the target car and the opponent car, which are used to enhance the accuracy and robustness of feature representation.

[0050] If there are multiple vehicles around the target vehicle, the analysis model can weight each set of relative features separately to obtain multiple sets of weighted features. Each set of weighted features retains the original feature information between the target vehicle and each vehicle, and further highlights the feature dimensions that are critical to analyzing the interaction scenario in each interaction relationship.

[0051] Step 106: Perform maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle.

[0052] In this step, the analysis model performs maximum pooling on the weighted features, selects the most significant interactive features from the weighted features, and generates target features for the target vehicle and the opponent vehicle.

[0053] Among them, the target feature is a feature representation that reflects the relative motion state and behavior pattern between the target vehicle and the opponent vehicle with the highest interaction probability, and is used to decide the target interaction scenario between the target vehicle and the opponent vehicle.

[0054] If there are multiple vehicles around the target vehicle, the analysis model can perform maximum pooling on multiple groups of weighted features, select the most significant interaction features from the multiple groups of weighted features, determine the most critical interaction relationship among the multiple interaction relationships in the time series position data, determine the opponent vehicle with the highest probability of interaction with the target vehicle in the vehicle driving scene, and generate target features for the target vehicle and the opponent vehicle.

[0055] Step 108: Perform cluster analysis on the target features to determine the target interaction scenario between the target vehicle and the opponent vehicle.

[0056] In this step, the analysis model uses clustering techniques such as density-based spatial clustering algorithm (DBSCAN) and K-Means algorithm to perform cluster analysis on target features. According to the clustering results of target features, the target interaction scenario between the target vehicle and the opponent vehicle is determined. The determined target interaction scenario can be used in scenarios such as traffic monitoring, accident warning learning, autonomous driving decision-making and path planning.

[0057] The target interaction scenario is the possible interaction between the target vehicle and the opponent vehicle, such as overtaking scenario, merging scenario, merging scenario, etc. The target interaction scenario can be pre-defined based on typical or key interaction modes in the actual traffic environment.

[0058] The vehicle interaction scenario analysis method of the present embodiment performs feature extraction on time series position data to obtain relative features, weights the relative features to obtain weighted features, performs maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle, performs cluster analysis on the target features, and determines the target interaction scenario between the target vehicle and the opponent vehicle. By performing maximum pooling on the weighted features, the most significant target features can be extracted, thereby accurately analyzing the target interaction scenario between the target vehicle and the opponent vehicle.

[0059] In addition, by directly obtaining the relative motion data of the interaction between the target vehicle and the opponent vehicle, the motion trajectory and direction changes of the vehicle can be intuitively reflected, thereby accelerating the convergence of the model and improving the interpretability of the model, avoiding reliance on the black box processing method of deep neural networks, and making the relationship between explicit relative features more transparent. At the same time, it also reduces the model's dependence on large-scale data and improves the analysis efficiency in complex vehicle interaction scenarios.

[0060] In the above-mentioned embodiment, it is introduced that the target interaction scene between the target vehicle and the opponent vehicle in a complex traffic scene is determined by performing feature extraction, weighting, maximum pooling and cluster analysis on the time series position data of the target vehicle and the opponent vehicle. In the following embodiment, the analysis process of the target interaction scene will be described in more detail and can be applied to any of the above embodiments.

[0061] In one embodiment, the time series position data may include the position of the target vehicle and the position of the opponent vehicle. The analysis model uses the median difference method to determine the position difference between the target vehicle and the opponent vehicle based on the position of the target vehicle and the opponent vehicle in the time series position data, and calculate the relative characteristics between the target vehicle and the opponent vehicle.

[0062] The position difference is the relative position change between the target vehicle and the opponent vehicle at adjacent time points. The relative feature may include: at least one of relative position difference data, relative speed data, relative acceleration data, the opponent vehicle heading angle, and the target vehicle heading angle.

[0063] The relative position difference data is used to reflect the distance change between the target car and the opponent car, for example: the modulus of the relative position difference. The relative speed data is used to reflect the speed difference between the target car and the opponent car, and can capture the interaction dynamics, for example: the modulus of the relative speed.

[0064] Relative acceleration data is used to reflect the acceleration difference between the target car and the opponent car, and can capture sudden acceleration or deceleration behavior, such as the modulus of relative acceleration.

[0065] The opponent car's heading angle is used to reflect the direction of movement of the opponent car relative to the target car, for example: the heading angle of the opponent car in the target car's coordinate system. The target car's heading angle is used to reflect the direction of movement of the target car relative to the opponent car, for example: the heading angle of the target car in the opponent car's coordinate system.

[0066] Exemplarily, the position of the target vehicle at time t is in, represents the position of the target vehicle at time t, x i (t) represents the x-axis coordinate of the target vehicle in the world coordinate system, y i (t) represents the y-axis coordinate of the target vehicle in the world coordinate system.

[0067] The position of the opponent car a is in, represents the position of the opponent car a at time t, x j (t) represents the x-axis coordinate of the opponent car a in the world coordinate system, y j (t) represents the y-axis coordinate of the opponent car a in the world coordinate system.

[0068] The analysis model can calculate the vector of the relative position difference in the world coordinate system according to the following formula 1:

[0069]

[0070]

[0071] The analytical model can calculate the modulus of the vector of the relative position difference in the world coordinate system according to the following formula 2:

[0072]

[0073] The analytical model can calculate the speed of the target vehicle at time t according to the following formula 3:

[0074]

[0075] in, represents the position of the target vehicle at time t+1, represents the position of the target vehicle at time t-1, t i (t+1) represents the time point t+1, t i (t-1) represents the time point at time t-1.

[0076] The analytical model can calculate the speed of the opponent car a at time t according to the following formula 4:

[0077]

[0078] in, represents the position of the opponent car a at time t+1, represents the position of the opponent car a at time t-1, t j (t+1) represents the time point at time t+1, t j (t-1) represents the time point at time t-1.

[0079] The analytical model can calculate the relative velocity vector between the target vehicle and the opponent vehicle a according to the following formula 5:

[0080]

[0081] The analytical model can calculate the modulus of the relative speed between the target car and the opponent car a according to the following formula 6:

[0082]

[0083] The analytical model can calculate the acceleration of the target vehicle at time t according to the following formula 7:

[0084]

[0085] in, represents the speed of the target vehicle at time t+1, represents the speed of the target vehicle at time t-1.

[0086] The analytical model can calculate the acceleration of the opponent car a at time t according to the following formula 8:

[0087]

[0088] in, represents the speed of the opponent car a at time t+1, Represents the speed of the opponent car a at time t-1.

[0089] The analytical model can calculate the relative acceleration vector between the target vehicle and the opponent vehicle a according to the following formula 9:

[0090]

[0091] The analytical model can calculate the modulus of the relative acceleration vector between the target vehicle and the opponent vehicle a according to the following formula 10:

[0092]

[0093] The analytical model can calculate the heading angle of the target vehicle and the heading angle of the opponent vehicle through the inverse tangent function. The opponent vehicle heading angle θ in the target vehicle coordinate system is calculated according to the following formula 11: j (t):

[0094] θ j (t) = atan2(y j (t)-y i (t),x j (t)-x i (t)) Formula 11

[0095] Among them, atan2 is the inverse tangent function.

[0096] The analytical model can calculate the target vehicle heading angle θ in the opponent vehicle coordinate system according to the following formula 12: i (t):

[0097] θ i (t) = atan2(y i (t)-y j (t),x i (t)-x j (t)) Formula 12

[0098] As described above, by calculating the relative position difference data, relative speed data, relative acceleration data, opponent vehicle heading angle and target vehicle heading angle of the target vehicle and opponent vehicle based on the positions of the target vehicle and opponent vehicle in the time series position data, it is ensured that even in the case of only position information, the motion trajectory and direction change of the target vehicle and opponent vehicle can be accurately captured, and the intuitive relative features can accurately capture the relative motion and interaction between vehicles, making the interaction mode between the target vehicle and the opponent vehicle clearer and more explicit.

[0099] In one embodiment, the analysis model may use dimensionality reduction techniques such as Linear Discriminant Analysis (LDA) and autoencoders to compress relative features and generate compressed relative features.

[0100] Among them, the compressed relative features are low-dimensional potential representations of the relative features between the target car and the opponent car, and the compressed relative features retain the key information in the original relative features.

[0101] By reducing the dimension of relative features, the computational complexity of subsequent processing and analysis can be significantly reduced, less storage space is occupied, and the real-time performance of the analysis model is improved; in addition, low-dimensional relative features are easier to be learned and generalized by the analysis model, which helps to improve the accuracy and robustness of the analysis model.

[0102] Taking the use of LSTM autoencoder for compression as an example, the analysis model can input the extracted relative features into the LSTM autoencoder. The LSTM autoencoder processes the relative features of the time series, compresses the high-dimensional relative features into low-dimensional potential representations, and obtains the relative features compressed by the LSTM autoencoder.

[0103] As mentioned above, by inputting relative features into the LSTM autoencoder, the relative features compressed by the LSTM autoencoder are obtained. The LSTM autoencoder can be used to understand the changes in the time step and avoid relying only on instantaneous features, thereby capturing the long-term dependent features in vehicle interactions. It can not only compress the relative features of high-dimensional time series, but also make the compressed relative features retain the key dynamic information in vehicle interactions.

[0104] In one embodiment, the LSTM autoencoder may include: a first encoding module and a second encoding module. The analysis model may use the first encoding module to compress the relative features to obtain the compressed features of the hidden layer dimension. The compressed features are further compressed using the second encoding module to compress the compressed features to the dimension of the potential representation to obtain the compressed relative features. The first encoding module and the second encoding module may be two layers of connected LSTM.

[0105] Exemplarily, the analysis model can input the relative features into the first encoding module, use the first encoding module to compress the relative features, and obtain the compressed features of the hidden layer dimension hidden_dim output by the first encoding module. The compressed features are features of shape (batch_size, seq_len, hidden_dim), where batch_size is the batch size of the input data, seq_len is the time step of the input data, and n_features is the feature dimension of the input data.

[0106] For example, the relative feature has 100 time steps, and each time step has 5 features, that is, seq_len = 100, n_features = 5. Assuming hidden_dim = 64, the first encoding module compresses the input features from n_features to hidden_dim, and the shape of the output compressed features is (batch_size, seq_len, 64).

[0107] The compressed features of the hidden layer dimension are input into the second encoding module, and the compressed features are further compressed using the second encoding module to compress the compressed features to the dimension embedding_dim of the potential representation to obtain the compressed relative features. The shape of the compressed relative features is (batch_size, seq_len, embedding_dim). Assuming embedding_dim = 32, the shape of the compressed relative features is (batch_size, seq_len, 32).

[0108] As described above, the relative features are compressed using the first encoding module to obtain compressed features of the hidden layer dimension, and the compressed features are compressed to the dimension of the potential representation using the second encoding module to obtain the compressed relative features. Through the hierarchical compression of the two encoding modules, the relative features can be effectively reduced in dimensionality, which not only reduces the complexity of the relative features, but also helps to remove noise information in the original features, thereby improving the accuracy and efficiency of subsequent analysis.

[0109] In one embodiment, the LSTM autoencoder may include: an encoding unit and a decoding unit. The LSTM autoencoder may be pre-trained using training compression features, the training compression features may be compressed using the encoding unit to obtain predicted compression features, and the predicted compression features may be reconstructed using the decoding unit to generate reconstructed compression features.

[0110] Based on the predicted compression features and the reconstructed compression features, the difference between the original training compression features and the reconstructed compression features is calculated, and the compression loss value of the compression loss function is calculated. According to the compression loss value, the LSTM autoencoder is adjusted, and the weights of the LSTM autoencoder are updated using back propagation algorithms, gradient descent and other methods to minimize the compression loss value, and the trained LSTM autoencoder is obtained.

[0111] During the training process, as the weights of the LSTM autoencoder are updated, the compression loss value will gradually decrease, indicating that the reconstruction ability of the LSTM autoencoder is constantly improving. By minimizing the compression loss function, the model can learn an effective low-dimensional representation of the input data while maintaining the ability to reconstruct the original data.

[0112] The training compressed features are relative features extracted from simulated or real time series position data, which are used as training input data for the LSTM autoencoder. The predicted compressed features are low-dimensional potential representations of the training compressed features, which are compact representations of the training compressed features. The training compressed features include the key information required to reconstruct the original input.

[0113] The reconstructed compressed feature is a feature vector obtained after the decoding unit decodes the predicted compressed feature, which is used to compare the difference between the reconstructed compressed feature and the original training compressed feature. The reconstructed compressed feature needs to be as close to the original training compressed feature as possible.

[0114] The compression loss function is a function that measures the difference between the reconstructed compressed features and the original trained compressed features. For example, loss functions such as Mean Square Error (MSE) and Mean Absolute Error (MAE). The compression loss value is the specific value of the compression loss function on specific input data, which indicates the performance of the LSTM autoencoder in reconstructing the input data. The compression loss value is used to evaluate the training effect of the model.

[0115] The encoding unit may include a first encoding module and a second encoding module, and the encoding unit may include a first decoding module and a second decoding module, wherein the first decoding module and the second decoding module may be two layers of LSTM connected.

[0116] During the training process of the LSTM autoencoder, the encoding unit uses the first encoding module and the second encoding module to gradually compress the training features to obtain the prediction features. Continuing with the example of the input training compressed features with the shape of (batch_size, seq_len, n_features), the training compressed features are input to the first encoding module, and the first encoding module compresses the training compressed features to the hidden layer dimension of the LSTM, and outputs the hidden layer dimension features with the shape of (batch_size, seq_len, hidden_dim).

[0117] The features output by the first encoding module are input to the second encoding module, which further compresses the features to the dimension embedding_dim of the potential representation and outputs the predicted compressed features with a shape of (batch_size, seq_len, embedding_dim).

[0118] After the predicted compressed features are gradually compressed, the predicted compressed features finally output by the encoding unit are expressed as (batch_size, seq_len, embedding_dim). The hidden state and cell state (batch_size, embedding_dim) of the last time step in the second encoding module are retained to initialize the hidden state and cell state of the decoding unit, which are also the potential compressed features required for subsequent weighting and clustering.

[0119] The decoding unit uses the first decoding module and the second decoding module to reconstruct the predicted features and generate reconstructed features. First, the hidden state and cell state are initialized, and the hidden state and cell state (batch_size, embedding_dim) of the last time step in the second encoding module are used as the initial hidden state and cell state of the decoding unit, allowing the decoding unit to use the representation learned by the encoding unit for decoding.

[0120] The predicted compressed features of the shape (batch_size, seq_len, embedding_dim) output by the second encoding module are input into the first decoding module. Taking embedding_dim=32 as an example, the shape of the input predicted features is (batch_size, seq_len, 32).

[0121] The first decoding module decodes the predicted compressed features of the potential representation dimension back to the feature representation of the hidden layer dimension, and outputs features of shape (batch_size, seq_len, hidden_dim). Taking hidden_dim=64 as an example, the output shape of the first decoding module is (batch_size, seq_len, 64).

[0122] The features of the hidden layer dimension output by the first decoding module are input into the second encoding module, which further decodes the features of the hidden layer dimension back to the original feature dimension of the training compressed feature, and the second encoding module outputs the reconstructed features with a shape of (batch_size, seq_len, n_features). The decoding unit finally outputs the reconstructed features with a shape of (batch_size, seq_len, n_features), which represents the original feature sequence recovered from the potential representation.

[0123] The encoding unit gradually compresses the input time series features (batch_size, seq_len, n_features) through two layers of LSTM, outputs the predicted features of the dimension of the potential representation (batch_size, seq_len, embedding_dim), and retains the hidden state and cell state of the last time step to facilitate the subsequent initialization of the decoding unit.

[0124] The decoding unit uses the output sequence of the encoding unit as input. The decoding unit gradually decodes the predicted compressed features of the potential representation dimension (batch_size, seq_len, embedding_dim) back to the original feature dimension (batch_size, seq_len, n_features) through two layers of LSTM, and outputs the restored input time series.

[0125] As described above, the training compression features are compressed by using the encoding unit to obtain the predicted compression features, and the predicted compression features are reconstructed by using the decoding unit to generate reconstructed compression features. The compression loss value of the compression loss function is calculated based on the training compression features and the reconstructed compression features. According to the compression loss value, the LSTM autoencoder is adjusted to obtain the trained LSTM autoencoder, so that the LSTM autoencoder can learn how to extract the most useful information from the original features and compress it into a smaller feature vector to ensure the accuracy of the relative features after compression.

[0126] In one embodiment, the analysis model uses an attention mechanism when weighting relative features to generate attention scores of the relative features in different feature dimensions.

[0127] The analysis model normalizes the attention scores to obtain the attention weights of different feature dimensions, and highlights the most important feature dimensions in the current interaction relationship according to the specific situation of the interaction relationship between the target vehicle and each vehicle. The softmax function is used to convert the attention scores into a probability distribution form, that is, the attention weight of each feature dimension is between 0 and 1, and the sum of the attention weights of all feature dimensions is 1.

[0128] After obtaining the normalized attention score, it is weighted according to the attention weight, and the value of each feature dimension in the relative feature is multiplied by the corresponding attention weight to generate a weighted feature. The analysis model can also dynamically adjust the attention weight of each feature dimension according to the different contextual information of the relative feature to enhance the focus on key feature dimensions.

[0129] The weighted features have higher weights on key feature dimensions and can therefore more accurately reflect the interaction between the target car and the opponent car.

[0130] Among them, the attention score is a quantitative indicator of the importance of different feature dimensions in the relative features to the analysis of the interaction scene, which is used to identify the importance of different feature dimensions in the interaction relationship between the target car and the opponent car.

[0131] Feature dimensions refer to the elements in the relative features, and feature dimensions are used to reflect the key information in the interaction between the target car and the opponent car. Attention weights are obtained after normalization of attention scores, and are used to indicate the contribution of each feature dimension in the weighted features.

[0132] As described above, by using the attention mechanism, generating attention scores for relative features, normalizing the attention scores, obtaining attention weights for different feature dimensions, and weighting them according to the attention weights, weighted features are generated. This can highlight the most important feature dimensions in the interaction between the target car and the opponent car, so that the weighted features retain the most important feature dimensions between the target car and the opponent car, while reducing the influence of minor feature dimensions, further optimizing the results of subsequent maximum pooling, and enhancing the recognition ability and accuracy of key target features, thereby effectively coping with complex dynamic interaction environments.

[0133] In one embodiment, the pre-collected training weighted features are input into the initial model, and the initial model uses the attention mechanism to calculate the attention scores and attention weights of different feature dimensions in the training weighted features, and predict the weighted predicted weighted features.

[0134] For the training weighted features, there is a corresponding binary label data, which represents the interaction relationship between the training target car and the training opponent car (positive class) and the non-training opponent car (negative class). The training target car is the vehicle to be analyzed for the interaction scene during the training process, and the training opponent car is the opponent car with the highest interaction probability with the training target car in the training weighted features.

[0135] The predicted weighted features are input to the output head of the initial model to generate a binary classification result. Based on the binary classification result and the label data, the weighted loss value of the weighted loss function is calculated.

[0136] According to the weighted loss value, the back propagation algorithm is used to adjust the parameters of the initial model. After multiple iterative training, the training termination conditions such as the initial model's performance no longer significantly improves on the validation set and the preset number of training rounds are reached are met, and the trained feature weighted model is obtained.

[0137] The analysis model extracts features from the time series position data, and inputs the extracted relative features into the trained feature weighting model. The feature weighting model weights the relative features, and the analysis model obtains the weighted features generated by the feature weighting model.

[0138] The initial model refers to a model structure before training that has not learned any specific task knowledge, which can be a neural network model. The initial model can include: an attention layer for feature weighting and an output head for classification.

[0139] The feature weighted model is a model that introduces a feature weighting mechanism based on the initial model. The feature weighted model is used to assign different weights to each feature dimension in the input features, thereby highlighting the features that are more important to the classification task.

[0140] The training weighted features can be relative features extracted from simulated or real time series position data, or feature vectors obtained by compressing simulated or real relative features, which are used as training input data for the initial model. The prediction weighted features are feature information after feature weighting processing, and the prediction weighted features are used for subsequent classification tasks.

[0141] The binary classification result refers to the two possible results obtained after the model classifies the training weighted features. The binary classification result may include: the interaction relationship between the training target car and the training opponent car, and the interaction relationship between the training target car and the non-training opponent car.

[0142] The weighted loss function is a loss function that applies different weights to different feature dimensions when calculating the weighted loss value. It is used to help the model pay more attention to those features that are more important for the classification task. It can be a cross entropy loss function, etc.

[0143] The weighted loss value is the loss value calculated according to the weighted loss function. It is used to reflect the degree of fit of the model to the training data under the current parameters and to guide the parameter update of the model.

[0144] As described above, by inputting the training weighted features into the initial model, the initial model predicts the weighted predicted weighted features, and the predicted weighted features are input into the output head of the initial model to generate a binary classification result. Based on the binary classification result and the label data, the weighted loss value of the weighted loss function is calculated. According to the weighted loss value, the parameters of the initial model are adjusted using the back propagation algorithm to obtain the trained feature weighted model, so that the feature weighted model can more accurately identify the interaction relationship between the training target vehicle and the training opponent vehicle, as well as the interaction relationship between the training target vehicle and the non-training opponent vehicle, and learn the importance of feature dimensions in different interaction relationships. While making the generated weighted features more accurate, it enhances the consistency between the model and human subjective judgment.

[0145] In one embodiment, the analysis model performs global maximum pooling on the weighted features and calculates the Euclidean norm (L2 norm) of the weighted features, that is, the length or modulus of the vector of the weighted features. The L2 norm is used to reflect the comprehensive strength of all feature dimensions in the weighted features. The most significant target feature is generated based on the feature dimension with the largest Euclidean norm.

[0146] If there are multiple vehicles around the target vehicle, all N groups of weighted features are traversed along the different interaction relationships between the target vehicle and multiple vehicles. In each feature dimension, the maximum values ​​of multiple Euclidean norms, that is, the corresponding dimension value of the weighted feature with the largest L2 norm, are selected. Based on the maximum values ​​from different weighted features and feature dimensions, the target feature between the target vehicle and the opponent vehicle with the highest interaction probability is generated.

[0147] As described above, by calculating the Euclidean norm corresponding to the weighted feature and generating the target feature according to the feature dimension with the largest Euclidean norm, the most representative target feature can be selected, so that the subsequent clustering analysis can analyze the interaction scenario more accurately and stably, thereby improving the efficiency and accuracy of the entire system.

[0148] In one embodiment, the analysis model may use a density-based clustering algorithm to perform cluster analysis on the target features, and determine the core points in the target features according to a preset neighborhood radius and a minimum number of samples.

[0149] A density-based clustering algorithm is used. According to the density connectivity of core points, if two sample points are in the neighborhood of a core point or are connected through a series of core points, it can be determined that the two sample points are density connected.

[0150] The multiple sample points of the target feature are divided into multiple clusters, each cluster represents an interaction scenario. By traversing all sample points, the core point and its density-connected sample points are divided into a cluster according to density connectivity, and the sample points that do not belong to any cluster are regarded as noise points.

[0151] According to the density distribution of multiple clusters, the density of each cluster is calculated by calculating the number of sample points or the average distance within the cluster, and the target interaction scene is determined. For example, a cluster with a higher density represents a more important interaction scene.

[0152] The neighborhood radius (ε) is used to define the size of the neighborhood around the sample point. The minimum number of samples (MinPts) is used to define the minimum number of sample points in the neighborhood required to become a core point.

[0153] A core point is a sample point whose number of sample points within the neighborhood radius is greater than or equal to the minimum number of sample points. Sample points are used to represent data of different feature dimensions in the target feature, and each sample point corresponds to a label of an interaction scene.

[0154] As described above, by determining the core points in the target features according to the preset neighborhood radius and the minimum number of samples, a density-based clustering algorithm is used to divide multiple sample points of the target features into multiple clustering clusters according to the density connectivity of the core points, and the target interaction scenario is determined according to the density distribution of the multiple clustering clusters. By utilizing the density-adaptive feature processing, complex vehicle interaction scenarios of different densities can be responded to more accurately and stably, ensuring the stability and accuracy of the clustering results.

[0155] To further introduce the analysis process of vehicle interaction scenarios, Figure 2A flowchart of another vehicle interaction scenario analysis method is shown. The analysis method can be applied to the analysis model. Taking the case where there are multiple vehicles around the target vehicle as an example, the analysis method may include the following steps:

[0156] Step 202: Calculate multiple sets of relative features based on the positions of the target vehicle and multiple vehicles in the time series position data.

[0157] See also Figure 3 , Figure 3 A schematic diagram of an analysis method for a vehicle interaction scenario is shown. The analysis model 30 may include: a feature extraction model 301 , an LSTM autoencoder module 302 , an attention mechanism module 303 , a maximum pooling module 304 and a DBSCAN clustering module 305 .

[0158] Exemplarily, the time series position data of the target vehicle and multiple vehicles are the position data of n time steps from scene (Scene)-0 to scene-n. The time series position data is input into the analysis model 30, and the feature extraction model 301 uses the median difference method to calculate the relative features such as the modulus length of the relative speed and the modulus length of the relative acceleration between the target vehicle and each vehicle in each time step t0-tn based on the positions of the target vehicle and multiple vehicles in the time series position data, and obtain multiple sets of relative features.

[0159] Step 204: compress the multiple groups of relative features respectively to generate multiple groups of compressed relative features.

[0160] The LSTM autoencoder module 302 may include: an encoding unit 3020, a bottleneck unit 3021, and a decoding unit 3022. The encoding unit 3020 compresses multiple sets of relative features respectively, and inputs the generated low-dimensional multiple sets of relative features into the bottleneck unit 3021. The decoding unit 3022 reconstructs the multiple sets of relative features to generate reconstructed compressed features for training the LSTM autoencoder module 302.

[0161] Step 206: Use the attention mechanism to weight each group of compressed relative features to obtain multiple groups of weighted features.

[0162] In this step, the attention mechanism module 303 weights each group of relative features in the bottleneck unit 3021, and uses the attention mechanism to normalize the attention scores of each group of relative features in different feature dimensions to obtain the attention weights of different feature dimensions. The corresponding weighted features are weighted according to the attention weights to obtain multiple groups of weighted features.

[0163] The attention mechanism module 303 can map multiple groups of weighted features to binary classification results, and train the feature weighting model in the attention mechanism module 303 based on the binary classification results and corresponding label data.

[0164] Step 208: Perform global maximum pooling on multiple groups of weighted features to generate target features.

[0165] In this step, the maximum pooling module 304 performs global maximum pooling on multiple groups of weighted features, calculates the L2 norm of each group of weighted features, selects the maximum value of multiple Euclidean norms in each feature dimension, and generates a target feature between the target vehicle and the opponent vehicle with the highest interaction probability.

[0166] Step 210: Using a density-based clustering algorithm, cluster analysis is performed on target features to determine target interaction scenarios.

[0167] In this step, the DBSCAN clustering module 305 uses a density-based clustering algorithm to perform cluster analysis on the target features, divides multiple sample points of the target features into multiple clusters according to the density connectivity of the core points, and determines the target interaction scene between the target vehicle and the opponent vehicle according to the density distribution of the multiple clusters.

[0168] Figure 4 1 is a schematic diagram of the structure of an electronic device according to an exemplary embodiment of the present application. The electronic device may be, for example, a mobile phone, a computer, a digital broadcast terminal, a message transceiver, a game console, a tablet device, a personal digital assistant, a server, a smart home appliance, a car computer, etc. Figure 4 At the hardware level, the electronic device includes a processor 402, an internal bus 404, a network interface 406, a memory 408, and a non-volatile memory 410, and may also include hardware required for other services. The processor 402 reads the corresponding computer program from the non-volatile memory 410 into the memory 408 and then runs it, forming an analysis device for the vehicle interaction scene at the logical level. Of course, in addition to the software implementation method, this application does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0169] Figure 5 is a block diagram of a vehicle interaction scenario analysis device according to an exemplary embodiment of the present application. Figure 5 , the apparatus may include: an extraction module 502, a weighting module 504, a pooling module 506 and an analysis module 508, wherein:

[0170] The extraction module 502 is used to extract features from the time series position data to obtain relative features; wherein the time series position data is the position data of the target vehicle and the opponent vehicle in multiple time steps, and the relative features are used to reflect the relative motion data between the target vehicle and the opponent vehicle;

[0171] The weighting module 504 is used to weight the relative features to obtain weighted features;

[0172] The pooling module 506 is used to perform maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle;

[0173] The analysis module 508 is used to perform cluster analysis on the target features to determine the target interaction scenario between the target vehicle and the opponent vehicle.

[0174] In one example, the extraction module 502, the time series position data includes: the position of the target vehicle and the position of the opponent vehicle; when used to extract features from the time series position data to obtain relative features, it includes: calculating the relative features based on the position of the target vehicle and the position of the opponent vehicle; wherein the relative features include: at least one of relative position difference data, relative speed data, relative acceleration data, the heading angle of the opponent vehicle and the heading angle of the target vehicle.

[0175] In one example, the weighting module 504, before being used to weight the relative features to obtain the weighted weighted features, the method also includes: inputting the relative features into a long short-term memory network LSTM autoencoder to obtain the relative features compressed by the LSTM autoencoder.

[0176] In one example, the LSTM autoencoder includes: a first encoding module and a second encoding module; the weighting module 504, when used to input the multiple groups of relative features into the long short-term memory network LSTM autoencoder to obtain the multiple groups of relative features compressed by the LSTM autoencoder, includes: using the first encoding module to compress the relative features to obtain compressed features of the hidden layer dimension; using the second encoding module to compress the compressed features to the dimension of the potential representation to obtain the compressed relative features.

[0177] In one example, the LSTM autoencoder includes: an encoding unit and a decoding unit; the weighting module 504, before being used to input the multiple groups of relative features into the long short-term memory network LSTM autoencoder to obtain the multiple groups of relative features compressed by the LSTM autoencoder, also includes: using the encoding unit to compress the training compression features to obtain predicted compression features; using the decoding unit to reconstruct the predicted compression features to generate reconstructed compression features; based on the training compression features and the reconstructed compression features, calculating the compression loss value of the compression loss function; according to the compression loss value, adjusting the LSTM autoencoder to obtain the trained LSTM autoencoder.

[0178] In one example, the weighting module 504, when weighting the relative feature to obtain the weighted weighted feature, includes: using an attention mechanism to generate an attention score for the relative feature, wherein the attention score is used to identify the importance of different feature dimensions between the target vehicle and the opponent vehicle; normalizing the attention score to obtain the attention weights of the different feature dimensions; and weighting according to the attention weights to generate the weighted feature.

[0179] In one example, the weighting module 504, when weighting the relative features to obtain weighted features, includes: inputting the relative features into a feature weighting model to obtain the weighted features generated by the feature weighting model; during the training process of the feature weighting model, the weighting module 504 is used to: input the training weighted features into an initial model, and the initial model predicts the weighted predicted weighted features; input the predicted weighted features into the output head of the initial model to generate a binary classification result; the binary classification result includes: the interaction relationship between the target vehicle and the training opponent vehicle, and the interaction relationship between the target vehicle and the non-training opponent vehicle; based on the binary classification results and label data, calculating the weighted loss value of the weighted loss function; according to the weighted loss value, using a back propagation algorithm to adjust the parameters of the initial model to obtain a trained feature weighted model.

[0180] In one example, the pooling module 506, when used to perform maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle, includes: calculating the Euclidean norm corresponding to the weighted features; and generating the target features according to the feature dimension with the largest Euclidean norm.

[0181] In one example, the analysis module 508, when used to perform cluster analysis on the target feature to determine the target interaction scenario between the target vehicle and the opponent vehicle, includes: determining the core point in the target feature according to a preset neighborhood radius and a minimum number of samples; the core point is a sample point whose number of sample points within the neighborhood radius is greater than or equal to the minimum number of samples; the sample point is used to represent data of different feature dimensions in the target feature; using a density-based clustering algorithm, multiple sample points of the target feature are divided into multiple cluster clusters according to the density connectivity of the core points, each cluster cluster represents an interaction scenario; and determining the target interaction scenario according to the density distribution of the multiple cluster clusters.

[0182] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0183] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application scheme. A person of ordinary skill in the art can understand and implement it without paying creative labor.

[0184] In an exemplary embodiment, a non-temporary computer-readable storage medium including instructions is also provided, such as a memory including instructions. The above instructions can be executed by a processor of an analysis device for a vehicle interaction scenario to implement a method as described in any of the above embodiments.

[0185] The non-temporary computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc., and the present application does not limit this.

[0186] In an exemplary embodiment, a computer program product including a computer program / instruction is also provided. The computer program / instruction can be executed by a processor of a vehicle interaction scenario analysis device to implement a method as described in any of the above embodiments.

[0187] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0188] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the inventions claimed herein. The present application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

[0189] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for analyzing vehicle interaction scenarios, characterized in that: The method comprises: Extracting features from the time series position data to obtain relative features; wherein the time series position data is the position data of the target vehicle and the opponent vehicle within a plurality of time steps, and the relative features are used to reflect the relative motion data between the target vehicle and the opponent vehicle; Weighting the relative features to obtain weighted features; Performing maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle; Cluster analysis is performed on the target features to determine a target interaction scenario between the target vehicle and the opponent vehicle.

2. The method according to claim 1, characterized in that The time series position data includes: the position of the target vehicle and the position of the opponent vehicle; The feature extraction of the time series position data to obtain relative features includes: Calculating the relative feature based on the position of the target vehicle and the position of the opponent vehicle; The relative features include: at least one of relative position difference data, relative speed data, relative acceleration data, and the heading angle of the opponent vehicle and the heading angle of the target vehicle.

3. The method according to claim 1, characterized in that Before weighting the relative features to obtain weighted features, the method further includes: The relative features are input into a long short-term memory network (LSTM) autoencoder to obtain the relative features compressed by the LSTM autoencoder.

4. The method according to claim 3, characterized in that: The LSTM autoencoder includes: a first encoding module and a second encoding module; The step of inputting the relative features into a long short-term memory network (LSTM) autoencoder to obtain the relative features compressed by the LSTM autoencoder comprises: Using the first encoding module to compress the relative features to obtain compressed features of hidden layer dimensions; The compressed features are compressed to the dimension of the potential representation using the second encoding module to obtain the compressed relative features.

5. The method according to claim 3, characterized in that: The LSTM autoencoder includes: an encoding unit and a decoding unit; Before inputting the relative features into a long short-term memory network LSTM autoencoder to obtain the relative features compressed by the LSTM autoencoder, the method further includes: Using the encoding unit, compressing the training compression feature to obtain the predicted compression feature; Using the decoding unit, reconstructing the predicted compressed features to generate reconstructed compressed features; Calculating a compression loss value of a compression loss function based on the training compression feature and the reconstructed compression feature; The LSTM autoencoder is adjusted according to the compression loss value to obtain a trained LSTM autoencoder.

6. The method according to claim 1, characterized in that The step of weighting the relative features to obtain weighted features includes: Using an attention mechanism, generating an attention score of the relative feature, wherein the attention score is used to identify the importance of different feature dimensions between the target vehicle and the opponent vehicle; Normalizing the attention scores to obtain attention weights of the different feature dimensions; The weighted features are generated by weighting according to the attention weights.

7. The method according to claim 1, characterized in that The step of weighting the relative features to obtain weighted features includes: Inputting the relative features into a feature weighting model to obtain the weighted features generated by the feature weighting model; The training process of the feature weighted model includes: Inputting the training weighted features into an initial model, wherein the initial model predicts weighted predicted weighted features; Inputting the predicted weighted features into the output head of the initial model to generate a binary classification result; the binary classification result includes: the interaction relationship between the training target vehicle and the training opponent vehicle, and the interaction relationship between the training target vehicle and the non-training opponent vehicle; Based on the binary classification result and the label data, calculating a weighted loss value of a weighted loss function; According to the weighted loss value, the parameters of the initial model are adjusted using a back propagation algorithm to obtain a trained feature weighted model.

8. The method according to claim 1, characterized in that The step of performing maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle includes: Calculating the Euclidean norm corresponding to the weighted feature; The target feature is generated according to the feature dimension with the largest Euclidean norm.

9. The method according to claim 1, characterized in that: The performing cluster analysis on the target features to determine the target interaction scene between the target vehicle and the opponent vehicle includes: Determine the core point in the target feature according to a preset neighborhood radius and a minimum number of samples; the core point is a sample point whose number of sample points within the neighborhood radius is greater than or equal to the minimum number of samples; the sample point is used to represent data of different feature dimensions in the target feature; Using a density-based clustering algorithm, according to the density connectivity of the core points, multiple sample points of the target feature are divided into multiple clusters, each cluster representing an interaction scenario; The target interaction scenario is determined according to the density distribution of the multiple clusters.

10. A vehicle interaction scene analysis device, characterized in that: The device comprises: An extraction module is used to extract features from the time series position data to obtain relative features; wherein the time series position data is the position data of the target vehicle and the opponent vehicle in multiple time steps, and the relative features are used to reflect the relative motion data between the target vehicle and the opponent vehicle; A weighting module, used for weighting the relative features to obtain weighted features; A pooling module, used for performing maximum pooling on the weighted features to generate target features of the target vehicle and the opponent vehicle; The analysis module is used to perform cluster analysis on the target features to determine the target interaction scenario between the target vehicle and the opponent vehicle.

11. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor implements the method according to any one of claims 1 to 9 by running the executable instructions.

12. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

13. A computer program product having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images