Intelligent vehicle collision risk trend prediction method and system

Abstract

The invention discloses an intelligent vehicle collision risk trend prediction method and system, and relates to the technical field of intelligent vehicle risk trend prediction, and the method comprises the steps: carrying out the modeling of collected trajectory data into an intelligent vehicle-surrounding vehicle-conceptual graph, carrying out the instance normalization and patch processing of the trajectory data, and carrying out the enhancement through employing complementary masking, designing comparative learning and patch reconstruction tasks to capture timestamp-level characterization of the trajectory sequence, and developing an instance comparative task to capture overall characterization of the sequence; inputting the extracted trajectory representation into a risk trend decoupling module to learn the risk trend representation of vehicle interaction, and performing soft clustering on the extracted trajectory concept representation to obtain the intention causing the risk trend so as to predict the collision risk of the intelligent vehicle; according to the method, risk trend reasoning in a complex traffic scene is promoted, and the accuracy of collision risk prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent vehicle risk trend prediction, and in particular to an intelligent vehicle collision risk trend prediction method and system. Background Art

[0002] The development of intelligent vehicles and the widespread adoption of advanced driver assistance systems (ADAS) have further improved driving safety. By collecting real-time trajectory data from surrounding vehicles, intelligent vehicles can predict risk trends and formulate appropriate driving strategies to avoid collisions.

[0003] Graph neural network (GNN)-based reasoning systems are currently a major trend. In risk trend reasoning, these systems can fully leverage high-order relationships to learn better representations of the trajectories of intelligent vehicles and surrounding vehicles. However, GNN-based risk trend reasoning systems cannot understand the decision-making intentions of intelligent vehicles from interaction data. Furthermore, when learning vehicle trajectory representations, proxy tasks are often required to extract timestamp-level or instance-level embeddings. Masking strategies and contrastive learning have been advanced proxy tasks in recent years. However, masking strategies can reduce model efficiency and robustness when capturing dependencies between patches, while contrastive learning often uses pooling methods to extract instance-level embeddings from timestamp-level embeddings during representation learning. This causes the learned vehicle trajectory embeddings to be mostly concentrated in a subspace of the vector space, resulting in the inference model being unable to effectively distinguish between different vehicle trajectories, leading to misjudgments of the collision risk between the intelligent vehicle and surrounding vehicles. Summary of the Invention

[0004] To address the shortcomings of existing technologies, such as their inability to effectively distinguish different vehicle trajectories and thus misjudging the collision risk between intelligent vehicles and surrounding vehicles, the present invention proposes a method and system for predicting collision risk trends of intelligent vehicles. By modeling the collected ego-vehicle trajectory and surrounding vehicle trajectories into an intelligent vehicle-surrounding vehicle-concept graph, the trajectory data is instance-normalized and patched and enhanced using complementary masking. Contrastive learning and patch reconstruction tasks are designed to capture the timestamp-level representation of the trajectory sequence. The extracted trajectory concept representations are soft-clustered to obtain the intention leading to the risk trend, thereby predicting the collision risk of the intelligent vehicle, thereby solving the problems existing in the existing technology.

[0005] A method for predicting collision risk trends of intelligent vehicles comprises the following steps: Collecting real-time trajectory data of the intelligent vehicle and its surrounding vehicles, and constructing a graph structure based on the trajectory data; the graph structure includes the intelligent vehicle trajectory, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts; the surrounding vehicle trajectory concepts include trajectory type, coverage, and complexity; The instance normalization function and patch processing function are used to extract the patch sequences of intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept in the graph structure respectively; the complementary masking strategy is used to generate two views of the patch sequence of intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept, and a [CLS] token is added at the beginning of each patch sequence. The added patch sequences are respectively input into the two-layer graph encoder LightGCN learning embedding function, and complementary contrast learning and patch reconstruction are performed according to the two views of each patch sequence to obtain timestamp-level representations respectively; through the [CLS] token strategy, the instance-level representation is obtained from each timestamp-level representation by using the instantiation contrast task; the intelligent vehicle trajectory representation, surrounding vehicle trajectory representation and surrounding vehicle trajectory concept representation are obtained by combining each generated timestamp-level representation and its corresponding instance-level representation; the decision intention of the intelligent vehicle is decoupled according to the surrounding vehicle trajectory concept representation, and all the generated intentions are combined to obtain the final risk trend representation; Based on the intelligent vehicle trajectory characterization, the surrounding vehicle trajectory characterization and the final risk trend characterization, the collision risk between the intelligent vehicle and surrounding vehicles is predicted.

[0006] Furthermore, the graph structure is represented by Represents; where the node set Contains the concepts of intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory, represents the trajectory sequence of intelligent vehicles, represents the set of A surrounding vehicle trajectory sequences, Represents the set of B surrounding vehicle trajectory concepts; edge set Indicates the relationship between the intelligent vehicle trajectory and the surrounding vehicle trajectories, Represents the trajectory of the intelligent vehicle and the surrounding vehicles Historical risk trends, Represents the trajectory of surrounding vehicles Belongs to the concept of surrounding vehicle trajectories .

[0007] Furthermore, the generation process of the timestamp-level representation specifically includes the following steps: Intelligent vehicle trajectory sequence Using instance normalization function and patching functions , get a series of patches ; Among them, the patching process changes the time dimension of a single vehicle trajectory sequence from Reduce to , the feature dimension changes from Expand to ; and Represent the time dimension and feature dimension of the sequence respectively, The number of patches representing a single vehicle trajectory sequence, Indicates the patch length; A complementary masking strategy is adopted to generate two views and a [CLS] token is added at the beginning of the patch sequence to obtain ; Similarly, we get the surrounding vehicle trajectory patch sequence and surrounding vehicle trajectory concept patch sequence , Indicates the trajectory sequence of the number of surrounding vehicles that interact with the smart vehicle. represents the bth trajectory concept sequence to which the ath surrounding vehicle trajectory belongs; By 、 Input two-layer graph encoder LightGCN learned embedding function In the process, based on the message passing mechanism of the graph structure, the neighborhood information of the intelligent vehicle trajectory sequence is aggregated and combined with the intelligent vehicle trajectory sequence representation of the previous iteration to obtain Representations in the first and second layers of the encoder; using a readout function Integrate the representations output by the two-layer graph encoder to obtain the timestamp-level representation of the intelligent vehicle trajectory sequence; Similarly, by combining the surrounding vehicle trajectory patch sequence and its adjacent nodes and Input two-layer graph encoder LightGCN learned embedding function In the process, the timestamp level representation of the surrounding vehicle trajectory sequence is obtained; the surrounding vehicle trajectory concept patch sequence is and its adjacent nodes Input two-layer graph encoder LightGCN learned embedding function In this paper, we obtain the timestamp-level representation of the concept sequence of surrounding vehicle trajectories.

[0008] Furthermore, the instantiation comparison task is used to obtain the instance-level representation from the timestamp-level representation through the [CLS] token strategy, which specifically includes the following steps: Will and As two views of the timestamp-level representation of intelligent vehicle trajectory sequences, for each view, the embedding corresponding to the [CLS] token is extracted as the instance-level embedding; Through a two-layer bottleneck MLP encoder with BatchNorm and ReLU Each instance-level embedding is processed to generate an instance-level representation of the intelligent vehicle trajectory sequence; Similarly, the [CLS] token strategy is adopted to obtain instance-level representations of surrounding vehicle trajectory sequences and surrounding vehicle trajectory concept sequences.

[0009] Furthermore, the method further includes obtaining a total loss by training complementary contrastive learning, patch reconstruction, and instantiation contrast tasks, and obtaining an optimal intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation by optimizing the total loss; The total loss includes the loss of contrastive learning process, the loss of patch reconstruction process and instantiation contrast loss; the loss of contrastive learning is calculated by comparing the two views. The probability is calculated, The probability is used to learn the missing patch information in the complementary view; the loss of the patch reconstruction process is calculated based on the patch sequence after adding the [CLS] token and the result of patch reconstruction; the instantiation contrast loss is obtained by using negative cosine similarity to calculate the loss between the instance-level prediction result of the first view and the instance-level embedding corresponding to the second view, and the loss between the instance-level prediction result of the second view and the instance-level embedding corresponding to the first view.

[0010] Furthermore, the decision intention of the intelligent vehicle is decoupled based on the concept representation of the surrounding vehicle trajectory, and all the generated intentions are combined to obtain the final risk trend representation, which specifically includes the following steps: Assume that the underlying intentions that lead to the risk trend of interaction are Extracted from the trajectory representation of surrounding vehicles through soft clustering High-level semantic base ; A probability assignment matrix is ​​learned as: ; in, is the activation function, is the model hyperparameter, C Each row of provides a soft assignment of probability nodes to different intentions, based on C Perform polymerization to obtain aggregate embeddings; Using a semantic projection head right Aggregate embedding is used to transform features and output a set of concept-aware semantic bases: ; in, , , Corresponding to different semantic spaces, it is embedded with risk trends to jointly decouple intelligent vehicle decision intentions. Dimensions representing trajectory concepts: ; in, represents the connection between two embedded Indicates mapping the combination pair to the Risk trend projection head of the intention space, Indicates the intention; Combining all intentions together, the final decoupling risk trend is characterized as follows: .

[0011] Furthermore, the method also includes enhancing the decoupled intelligent vehicle decision intention through intent-based contrastive learning and encoding-downcoding regularization, which specifically includes the following steps: The edge discarding strategy is used for the original graph Constructing the augmented graph , and then we get the augmented decomposition risk trend representation: ; Then the intent-based contrastive learning loss is: ; in, Indicates risk trends By The probability of an intention leading to For the The risk trend comparison learning subtask corresponding to the intention; The optimal intention that can maximize the subtask expectation is based on the concept perception semantic basis. , get the Trend confidence for each intent: ; in, Is related to temperature and The cosine similarity of k The contrastive learning subtask for each intent is: ; in, and For the The right intention; Encoding and decoding regularization is used to further separate the risk trends corresponding to different intentions. For a given risk trend representation , the coding rate of the overall risk trend is defined as the average coding length of each risk trend: ; in, is the coding rate calculation function, is a hyperparameter representing the threshold, which indicates that the expected decoding error is less than , is the identity matrix, T represents the matrix transpose; Define a set of membership matrices , orthogonally mapping the risk trends corresponding to different intentions to different subspaces; among them, is a diagonal matrix whose diagonal elements correspond to the If each risk trend group is coded separately, the probability of risk trend under the intention is Groups have an expected number of vector; then for the matrix , the total compactness of each risk trend group is the sum of the coding rates of all risk trend groups: ; in, is the compactness calculation function, represents the matrix determinant, represents the trace of the matrix; The difference in coding rate between the overall risk trend representation of intelligent vehicles and each group of risk trend representations is expressed as: .

[0012] Furthermore, the method of predicting the collision risk between the intelligent vehicle and surrounding vehicles based on the learned intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and decoupling risk trend representation specifically includes the following steps: Smart Vehicles With surrounding vehicles The collision risk of an interaction is expressed as: ; Adopt pairwise Bayesian personalized ranking loss to make the observed positive pairs The score is higher than the unobserved , to obtain the surrounding vehicles with the highest interaction risk with the smart vehicle: ; The total loss of the interaction risk between the intelligent vehicle and surrounding vehicles is predicted to be: ; in, is the model parameter set, 、 and is a hyperparameter that controls the strength of each component.

[0013] The present invention also includes an intelligent vehicle collision risk trend prediction system, comprising: The acquisition module is used to collect real-time trajectory data of the intelligent vehicle and its surrounding vehicles, and construct a graph structure based on the trajectory data; the graph structure includes the intelligent vehicle trajectory, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts; the surrounding vehicle trajectory concepts include trajectory type, coverage, and complexity; The trajectory representation generation module is used to extract the patch sequences of the intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept in the graph structure using the instance normalization function and the patch processing function respectively; the complementary masking strategy is used to generate two views of the patch sequences of the intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept, and a [CLS] token is added to the beginning of each patch sequence. The added patch sequences are respectively input into the two-layer graph encoder LightGCN learning embedding function, and complementary contrast learning and patch reconstruction are performed according to the two views of each patch sequence to obtain timestamp-level representations respectively; through the [CLS] token strategy, the instance-level representation is obtained from each timestamp-level representation using the instantiation contrast task; the intelligent vehicle trajectory representation, the surrounding vehicle trajectory representation and the surrounding vehicle trajectory concept representation are obtained by combining each generated timestamp-level representation and its corresponding instance-level representation; the decision intention of the intelligent vehicle is decoupled according to the surrounding vehicle trajectory concept representation, and all the generated intentions are combined to obtain the final risk trend representation; The prediction module is used to predict the collision risk between the intelligent vehicle and surrounding vehicles based on the intelligent vehicle trajectory representation, the surrounding vehicle trajectory representation and the final risk trend representation.

[0014] The present invention provides a method for predicting the collision risk trend of intelligent vehicles, which has the following beneficial effects: The present invention constructs an intelligent vehicle-surrounding vehicle-concept graph, and uses an instance normalization function and a patch processing function to extract patch sequences of the intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept, respectively. The patch independence strategy adopted is a pre-training task. Unlike mask modeling, which predicts masked patches, it reconstructs unmasked patches. It does not need to capture the dependencies between patches and has stronger robustness and higher efficiency in various actual road scenarios. The complementary masking strategy is used to generate two views of the patch sequence of the intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept, and a [CLS] token is added at the beginning of each patch sequence to mark different trajectories. sequence, thereby extracting instance-level representations in a decoupled manner; at the same time, to improve the performance of trajectory representation learning, a two-layer graph encoder is further adopted in the patch-independent strategy to make the model more focused on the extraction of patch representations, and the timestamp-level representation of the trajectory is learned through a comparative learning strategy; in the risk trend reasoning of intelligent vehicles, the patch-independent strategy can comprehensively learn the rich representations of trajectory data, thereby providing valuable information for subsequent risk trend reasoning; the present invention is expected to provide interpretable, accurate and efficient risk trend reasoning services for intelligent vehicles, promote risk trend reasoning in complex traffic scenarios, improve the accuracy of collision risk prediction, enhance the driving safety of intelligent vehicles, and help build a safe road traffic environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a model framework diagram based on patch independence strategy and decoupled graph representation learning in an embodiment of the present invention; Figure 2 This is a flow chart of a method for predicting collision risk trends of intelligent vehicles based on a patch independence strategy and decoupled graph representation learning in an embodiment of the present invention; Figure 3 This is a flow chart of model training and verification in an embodiment of the present invention; Figure 4 This is a structural framework diagram of trajectory characterization extraction in an embodiment of the present invention; Figure 5 This is a flow chart of a method for trajectory representation learning in an embodiment of the present invention; Figure 6 This is an example diagram of a scenario for predicting collision risk trends of intelligent vehicles in an embodiment of the present invention; Figure 7 This is a diagram showing the results of the smart vehicle risk trend diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0017] The present invention proposes a method for predicting the collision risk trend of intelligent vehicles, referring to Figure 1 、 Figure 2 , the method specifically comprises the following steps: S1. Data collection. Data collection devices include roadside equipment (cameras, lidar, millimeter-wave radar, etc.) and communication technologies (GPS, vehicle-to-vehicle communication, and vehicle-to-infrastructure communication), which are used to collect trajectory data of the intelligent vehicle and its surrounding vehicles and transmit it to the intelligent vehicle. The collected trajectory data includes sampling point location, sampling time, and speed.

[0018] S2. Data preprocessing: The collected raw vehicle trajectory data is cleaned, normalized, missing value processed, feature selected, and dataset partitioned. This aims to eliminate data issues, enhance data quality and reliability, and thereby improve the effectiveness and generalization capabilities of the model. The preprocessed data is stored in the trajectory database. The specific steps of preprocessing are as follows:

[0019] (1) Data cleaning: remove invalid data, duplicate data, and irrelevant data from the original trajectory data, and perform time error correction.

[0020] (2) Data normalization: Normalize the values ​​of different magnitudes in the trajectory data so that they are in the same magnitude; at the same time, ensure that all trajectory data use a consistent coordinate system.

[0021] (3) Missing value processing: interpolation, extrapolation estimation or direct deletion of missing data.

[0022] (4) Feature selection: Screen out highly correlated features through correlation analysis, feature importance assessment, and other methods, and select the most representative features.

[0023] (5) Dataset division: The dataset is divided into a training set, a validation set, and a test set. The training set is used to train the model so that it can fit the trajectory training data as well as possible; the validation set is used to adjust the model hyperparameters and select the model structure; and the test set is used to evaluate the model's performance in trajectory feature extraction and risk trend inference. The training set, validation set, and test set are divided in a ratio of 70%, 15%, and 15%.

[0024] S3. Construct a prediction model, including a trajectory representation extraction module and a risk trend inference module. The trajectory representation extraction module models the trajectory data of the intelligent vehicle and surrounding vehicles as an intelligent vehicle-surrounding vehicle-concept map, extracts timestamp-level representations of the trajectory data through complementary contrastive learning and patch-independent strategies, and captures the overall representation of the trajectory data through instanced contrastive learning. The risk trend inference module consists of three parts: risk trend decoupling, intent contrastive learning, and coding rate reduction regularization. The risk trend decoupling module is used to learn the different decision intentions of the intelligent vehicle, intent contrastive learning is used to enhance intent decoupling and infer risk trend distribution, and coding rate reduction regularization is used to orthogonalize the risk trends corresponding to different intentions, making the learned risk trend representation more discriminative between different intentions.

[0025] The intelligent vehicle collision risk trend prediction method based on a patch-independent strategy and decoupled graph representation learning relies primarily on trajectory representation extraction and decoupled graph comparative learning to achieve accurate and comprehensive risk trend reasoning and collision risk prediction. The construction process of this model framework is as follows:

[0026] (1) Intelligent vehicle-surrounding vehicle-concept map construction: Vehicle trajectory data (refer to Figure 6 ) Use pictures Indicates (refer to Figure 1 ), where the node set Contains the concepts of intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory, edge set Represents the correlation between the intelligent vehicle trajectory and the surrounding vehicle trajectories. represents the trajectory sequence of intelligent vehicles, represents the set of A surrounding vehicle trajectory sequences, Represents a set of concepts related to B surrounding vehicle trajectories, including trajectory type, coverage, complexity, etc. Represents the trajectory of the intelligent vehicle and the surrounding vehicles Historical risk trends, Represents the trajectory of surrounding vehicles Belong to the concept .

[0027] Given a trajectory of an intelligent vehicle and surrounding vehicle trajectories Candidate pairs , the goal is to learn the decoupled intentions of intelligent vehicle trajectories and the risk trend distribution on the intentions, and then predict , which represents the interactive collision risk between the smart vehicle and the surrounding vehicles.

[0028] (2) Trajectory representation learning: Vehicle trajectory datasets contain a large amount of high-dimensional information, which may lead to overfitting of the prediction model and increase the computational cost. Trajectory representation learning can extract the most critical features of trajectory data, such as trajectory direction changes, speed patterns, path regularity, etc., thereby enhancing the model's understanding and generalization capabilities and improving the efficiency and accuracy of risk trend reasoning. Therefore, before performing risk trend reasoning, it is necessary to extract trajectory representations of the intelligent vehicle and surrounding vehicles (refer to Figure 4 ).

[0029] Specifically, firstly, the intelligent vehicle trajectory sequence Using instance normalization function and patching functions , get a series of patches .

[0030] ; Among them, the patching process changes the time dimension of a single vehicle trajectory sequence from Reduce to , the feature dimension changes from Expand to .in, and Represent the time dimension and feature dimension of the sequence respectively, The number of patches representing a single vehicle trajectory sequence, Represents the patch length. A complementary masking strategy is used to generate two views (for complementary contrastive learning and instance-wise contrastive learning), and a [CLS] token is added at the beginning of the patch sequence (for marking different trajectory sequences, thereby extracting instance-level representations in a decoupled manner), resulting in :

[0031] ; Similarly, we can get the surrounding vehicle trajectory patch sequence and surrounding vehicle trajectory concept patch sequence , Indicates the trajectory sequence of the number of surrounding vehicles that interact with the smart vehicle. Represents the bth trajectory concept sequence to which the ath surrounding vehicle trajectory belongs.

[0032] Secondly, in order to extract the timestamp-level representation of the trajectory sequence of the intelligent vehicle, 、 Input a simple two-layer graph encoder LightGCN to learn the embedding function Based on the message passing mechanism of the graph structure, the neighborhood information of the intelligent vehicle trajectory sequence is aggregated and combined with the intelligent vehicle trajectory sequence representation of the previous iteration to obtain Representation in the first and second layers of the encoder and , for complementary contrastive learning and patch reconstruction, respectively:

[0033] ; ; Where, , , , and , , L represents the number of intelligent vehicle trajectory sequences, M represents the number of channels of a single vehicle trajectory sequence, and N represents the number of patches of a single channel of a single vehicle trajectory sequence. for The input dimension, O is The output dimensions of , which represent the patch size and embedding dimension respectively. and are aggregation and composition functions, yes The set of adjacent nodes of , here represents the set of trajectories of surrounding moving vehicles interacting with the smart vehicle. Using a readout function To integrate the representations output by the two-layer graph encoder to obtain timestamp-level representations:

[0034] ; Similarly, by combining the surrounding vehicle trajectory patch sequence and its adjacent nodes and Input the two-layer graph encoder LightGCN to obtain the timestamp-level representation of the surrounding vehicle trajectory sequence. Input the two-layer graph encoder LightGCN to obtain the timestamp-level representation of the surrounding vehicle trajectory concept sequence.

[0035] Reference Figure 5 ,The specific process of complementary contrastive learning and patch reconstruction tasks is as follows: ① Complementary contrastive learning: established a We use the probability to learn the missing patch information in the complementary view, and then hierarchically capture the relationship between adjacent timestamps in a single vehicle trajectory sequence. and The representation of the first layer of the two views generated by the complementary masking strategy is The probability can be defined as .

[0036] ; Where, the dot product Represents the calculation similarity. Therefore, the loss of contrastive learning can be expressed as:

[0037] ; right Perform maximum pooling and contrastive learning, and repeatedly calculate the layer loss until : ; ② Patch reconstruction: Represent the vehicle trajectory output by the second layer Input the block-wise linear projection head to obtain the result of patch reconstruction: ; Where, is the weight matrix of the linear layer.

[0038] The loss of patch reconstruction can be expressed as: ; Finally, an instance contrast task is used to capture the overall information of the entire sequence (refer to Figure 5 Specifically, suppose and Two views of timestamp-level representation for intelligent vehicle trajectory sequences. For each view, the embedding corresponding to the first position ([CLS] token) is extracted as the instance-level embedding.

[0039] ; ; Each instance-level embedding is passed through a two-layer bottleneck MLP encoder with BatchNorm and ReLU Processing to obtain and , used to instantiate contrastive learning: ; ; Using negative cosine similarity to calculate the instantiation contrast loss, first calculate and The loss between: ; Where, Represents the cosine similarity calculation function, Indicates stopping the gradient operation, which is used here to solve the sampling bias of instantiated contrastive learning. In order to symmetrically optimize the network, further calculation and The loss between:

[0040] ; Therefore, the total instantiation contrast loss for and The average value of: ; Finally, we jointly train the complementary contrastive learning task, the patch reconstruction task, and the instance contrast task, and use and Adjust between the three losses to get the total loss of representation extraction By optimizing this loss, the best intelligent driving vehicle trajectory representation can be obtained. .

[0041] ; Similarly, the best surrounding vehicle trajectory can be obtained Characterization and concepts Characterization , then the interactive risk trend representation is combined with the intelligent vehicle trajectory representation and the surrounding vehicle trajectory representation for: .

[0042] ③ Risk Trend Reasoning: The interaction risk between an intelligent vehicle and its surrounding vehicles may be influenced by multiple decision-making intentions. These complex intentions are often closely related to the intelligent vehicle's decision-making mechanism and the trajectory characteristics of surrounding vehicles. Ignoring any one aspect of this information can lead to insufficient and inaccurate risk trend reasoning. Therefore, the key to achieving intelligent vehicle collision risk trend prediction is to learn the intelligent vehicle's decision-making intentions from the extracted trajectory representation. In addition, the intention-based risk trend distribution can further reflect the different decision-making intentions of intelligent vehicles and improve the interpretability of collision risk prediction.

[0043] Specifically, we first decouple the risk trend representation and extract the decision intention of the intelligent vehicle. Assuming that the potential intentions that lead to the interactive risk trend are Extract the concept representation of the trajectory of surrounding vehicles through soft clustering High-level semantic base First, a probability distribution matrix is ​​learned as:

[0044] ; Where, is the activation function, and For model hyperparameters, each row of C provides a soft assignment of probability nodes to different intents, based on C to probability nodes Perform polymerization to obtain Aggregate embedding. Using a semantic projection head Perform feature transformation on these cluster embeddings to output a set of concept-aware semantic bases:

[0045] ; Where, , , Corresponding to different semantic spaces, it is embedded with risk trends to jointly decouple intelligent vehicle decision intentions. Dimensions representing trajectory concepts: ; Where, represents the connection between two embedded Indicates mapping the combination pair to the Risk trend projection head of the intention space, Indicates the Combining all the intentions together yields a decoupled risk trend representation:

[0046] ; Secondly, intent-based contrastive learning is used to enhance risk trend decoupling. Specifically, an edge discarding strategy is first used to Constructing the augmented graph , and then we get the augmented decomposition risk trend representation:

[0047] ; Therefore, the intent-based contrastive learning loss is: ; Where, Indicates risk trends By The probability of an intention leading to For the The risk trend comparison learning subtask corresponding to the intention. The optimal intention that can maximize the subtask expectation is based on the concept perception semantic basis. , we can get the Trend confidence for each intent:

[0048] ; Where, Is related to temperature and The cosine similarity of , which gives a greater weight coefficient to the intelligent vehicle decision intention with high confidence in contrastive learning. k The contrastive learning subtask for each intent is:

[0049] ; Where, and For the The intention is right.

[0050] In addition, since the risk trends corresponding to different intentions may be distributed in different subspaces, the learned risk trend representation has discriminativeness. Therefore, the maximum coding rate attenuation is adopted to enhance the diversity of risk trend representation. For a given risk trend representation , the coding rate of the overall risk trend is defined as the average coding length of each risk trend:

[0051] ; Where, is the coding rate calculation function, is a hyperparameter representing the threshold, which indicates that the expected decoding error is less than , is the identity matrix, T Represents matrix transpose.

[0052] In order to make the risk trends corresponding to different intentions orthogonal, they need to be mapped to different subspaces. Therefore, a set of membership matrices are defined: , where is a diagonal matrix whose diagonal elements correspond to the If each risk trend group is coded separately, then Groups have an expected number of vector. Therefore, for the matrix , the total compactness of each risk trend group is the sum of the coding rates of all risk trend groups:

[0053] ; Where, is the compactness calculation function, represents the matrix determinant, represents the trace of the matrix.

[0054] Therefore, the difference in coding rate between the overall risk trend representation of the intelligent vehicle and the risk trend representation of each group is expressed as follows. By maximizing this value, a better trend representation can be further obtained: ; Finally, the interaction risk between the intelligent vehicle and surrounding vehicles is predicted based on the learned intelligent vehicle trajectory representation and the surrounding vehicle trajectory representation (refer to Figure 7 ). Specifically, smart vehicles With surrounding vehicles The collision risk of an interaction can be expressed as:

[0055] ; Adopt pairwise Bayesian personalized ranking loss to make the observed positive pairs The score is higher than the unobserved , to obtain the surrounding vehicles with the highest risk of interaction with the smart vehicle: ; Therefore, the total loss of the interaction risk between the intelligent vehicle and surrounding vehicles can be expressed as: ; Where, is the model parameter set, 、 and is a hyperparameter that controls the strength of each component.

[0056] Reference Figure 3 Before conducting actual risk trend inference, the model needs to be trained and validated, and the model parameters need to be continuously adjusted to optimize. Then, the optimal model is tested using the test set data to evaluate its predictive performance and ultimately output the model prediction results.

[0057] This paper proposes an intelligent vehicle collision risk trend prediction system based on a patch-independent strategy and decoupled graph representation learning. The patch-independent strategy employed is a pre-training task that, unlike mask modeling, reconstructs unmasked patches instead of predicting masked patches. Therefore, it eliminates the need to capture dependencies between patches, resulting in greater robustness and efficiency across various real-world road scenarios.

[0058] To improve the performance of trajectory representation learning, this paper further employs a simplified two-layer graph encoder within the patch-independent strategy, allowing the model to focus more on extracting patch representations. A comparative learning strategy is then used to learn both timestamp-level and overall trajectory representations. For intelligent vehicle risk trend inference, the patch-independent strategy can comprehensively learn rich representations of trajectory data, providing valuable information for subsequent risk trend inference.

[0059] The decoupling adopted by the present invention is a technology for processing highly correlated data in deep learning. It reduces feature correlation, enables the model to better understand and utilize features, and improves the model's reasoning and prediction performance and robustness. The present invention further expands it into decoupled graph comparative learning to enhance decoupling. In the risk trend reasoning of intelligent vehicles, decoupled graph comparative learning can achieve better and explainable risk trend reasoning of intelligent vehicles by learning complex and diverse intelligent vehicle decision intentions. Therefore, the present invention is expected to provide explainable, accurate and efficient risk trend reasoning services for intelligent vehicles, promote risk trend reasoning in complex traffic scenarios, improve the accuracy of collision risk prediction, enhance the driving safety of intelligent vehicles, and help build a safe road traffic environment.

[0060] This paper proposes a method for predicting collision risk trends for intelligent vehicles based on a patch-independent strategy and decoupled graph comparative learning. The method models the collected trajectories of the vehicle and surrounding vehicles as an intelligent vehicle-surrounding vehicle-conceptual graph, which is then used to learn the interactive relationships between the intelligent vehicle and surrounding vehicles and infer the intelligent vehicle's decision-making intentions. The method then inputs the intelligent vehicle trajectory data and surrounding vehicle trajectory data from the test set into the trained model to learn important representations of the trajectory data, infer the intelligent vehicle's decision-making intentions, and predict the intelligent vehicle's collision risk. This method can provide more accurate and reliable risk trend reasoning and collision risk prediction for intelligent vehicles, enhancing the risk assessment capabilities of intelligent vehicles in complex traffic scenarios, thereby promoting the development and upgrading of intelligent vehicles.

[0061] Based on the same inventive concept, the present invention also proposes an intelligent vehicle collision risk trend prediction system, comprising: The acquisition module is used to collect real-time trajectory data of the intelligent vehicle and its surrounding vehicles, and to construct an intelligent vehicle-surrounding vehicle-concept map based on the trajectory data.

[0062] The trajectory representation generation module is used to extract the patch sequences of the intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept in the intelligent vehicle-surrounding vehicle-concept graph by using instance normalization function and patch processing function respectively; a complementary masking strategy is used to generate two views of the patch sequence of the intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept, and a [CLS] token is added at the beginning of each patch sequence. The added patch sequences are respectively input into the two-layer graph encoder LightGCN learning embedding function, and complementary contrast learning and patch reconstruction are performed according to the two views of each patch sequence to generate timestamp-level representations respectively; through the [CLS] token strategy, the instance-level representation is obtained from the timestamp-level representation by using the instantiation contrast task; the intelligent vehicle trajectory representation, the surrounding vehicle trajectory representation and the surrounding vehicle trajectory concept representation are obtained by combining each generated timestamp-level representation and its corresponding instance-level representation; the decision intention of the intelligent vehicle is decoupled according to the surrounding vehicle trajectory concept representation, and all the generated intentions are combined to obtain the final risk trend representation.

[0063] The prediction module is used to predict the collision risk between the intelligent vehicle and surrounding vehicles based on the intelligent vehicle trajectory representation, the surrounding vehicle trajectory representation and the final risk trend representation.

[0064] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting collision risk trends of intelligent vehicles, characterized in that: The following steps are involved: Collecting real-time trajectory data of the intelligent vehicle and its surrounding vehicles, and constructing a graph structure based on the trajectory data; the graph structure includes the intelligent vehicle trajectory, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts; the surrounding vehicle trajectory concepts include trajectory type, coverage, and complexity; The instance normalization function and patch processing function are used to extract the patch sequences of intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept in the graph structure respectively; the complementary masking strategy is used to generate two views of the patch sequence of intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory concept, and a [CLS] token is added to the beginning of each patch sequence. The added patch sequences are respectively input into the two-layer graph encoder LightGCN learning embedding function, and complementary contrast learning and patch reconstruction are performed according to the two views of each patch sequence to obtain timestamp-level representations respectively; through the [CLS] token strategy, the instance-level representation is obtained from each timestamp-level representation by using the instance-level contrast task; the intelligent vehicle trajectory representation, surrounding vehicle trajectory representation and surrounding vehicle trajectory concept representation are obtained by combining each generated timestamp-level representation with its corresponding instance-level representation; Decouple the decision intentions of intelligent vehicles based on the concept of surrounding vehicle trajectories, and combine all generated intentions to obtain the final risk trend representation; Based on the intelligent vehicle trajectory characterization, the surrounding vehicle trajectory characterization and the final risk trend characterization, the collision risk between the intelligent vehicle and surrounding vehicles is predicted.

2. The method for predicting collision risk trend of intelligent vehicles according to claim 1, characterized in that: The graph structure is shown in FIG Represents; where the node set Contains the concepts of intelligent vehicle trajectory, surrounding vehicle trajectory and surrounding vehicle trajectory, represents the trajectory sequence of intelligent vehicles, represents the set of A surrounding vehicle trajectory sequences, Represents the set of B surrounding vehicle trajectory concepts; edge set Indicates the relationship between the intelligent vehicle trajectory and the surrounding vehicle trajectories, Represents the trajectory of the intelligent vehicle and the surrounding vehicles Historical risk trends, Represents the trajectory of surrounding vehicles Belongs to the concept of surrounding vehicle trajectories .

3. The method for predicting collision risk trend of intelligent vehicles according to claim 2, characterized in that: The generation process of the timestamp-level representation specifically includes the following steps: Intelligent vehicle trajectory sequence Use instance normalization function and patching functions , get a series of patches ; Among them, the patching process changes the time dimension of a single vehicle trajectory sequence from Reduce to , the feature dimension changes from Expand to ; and Represent the time dimension and feature dimension of the sequence respectively, The number of patches representing a single vehicle trajectory sequence, Indicates the patch length; A complementary masking strategy is adopted to generate two views and a [CLS] token is added at the beginning of the patch sequence to obtain ; Similarly, we get the surrounding vehicle trajectory patch sequence and surrounding vehicle trajectory concept patch sequence , Indicates the trajectory sequence of the number of surrounding vehicles that interact with the smart vehicle. represents the bth trajectory concept sequence to which the ath surrounding vehicle trajectory belongs; By 、 Input two-layer graph encoder LightGCN learned embedding function In the process, based on the message passing mechanism of the graph structure, the neighborhood information of the intelligent vehicle trajectory sequence is aggregated and combined with the intelligent vehicle trajectory sequence representation of the previous iteration to obtain Representations in the first and second layers of the encoder; using a readout function Integrate the representations output by the two-layer graph encoder to obtain the timestamp-level representation of the intelligent vehicle trajectory sequence; Similarly, by combining the surrounding vehicle trajectory patch sequence and its adjacent nodes and Input two-layer graph encoder LightGCN learned embedding function In the process, the timestamp level representation of the surrounding vehicle trajectory sequence is obtained; the surrounding vehicle trajectory concept patch sequence is and its adjacent nodes Input two-layer graph encoder LightGCN learned embedding function In the above example, we obtain the timestamp-level representation of the concept sequence of surrounding vehicle trajectories.

4. The method for predicting collision risk trend of intelligent vehicles according to claim 3, characterized in that: The [CLS] token strategy is used to obtain instance-level representations from timestamp-level representations using an instantiation comparison task, specifically including the following steps: Will and As two views of the timestamp-level representation of intelligent vehicle trajectory sequences, for each view, the embedding corresponding to the [CLS] token is extracted as the instance-level embedding; Through a two-layer bottleneck MLP encoder with BatchNorm and ReLU Each instance-level embedding is processed to generate an instance-level representation of the intelligent vehicle trajectory sequence; Similarly, the [CLS] token strategy is adopted to obtain instance-level representations of surrounding vehicle trajectory sequences and surrounding vehicle trajectory concept sequences.

5. The method for predicting collision risk trend of intelligent vehicles according to claim 4, characterized in that: It also includes obtaining the total loss by training complementary contrast learning, patch reconstruction and instantiation contrast tasks, and obtaining the best intelligent vehicle trajectory representation, surrounding vehicle trajectory representation and surrounding vehicle trajectory concept representation by optimizing the total loss; The total loss includes the loss of contrastive learning process, the loss of patch reconstruction process and instantiation contrast loss; the loss of contrastive learning is calculated by comparing the two views. The probability is calculated, The probability is used to learn the missing patch information in the complementary view; the loss of the patch reconstruction process is calculated based on the patch sequence after adding the [CLS] token and the result of patch reconstruction; the instantiation contrast loss is obtained by using negative cosine similarity to calculate the loss between the instance-level prediction result of the first view and the instance-level embedding corresponding to the second view, and the loss between the instance-level prediction result of the second view and the instance-level embedding corresponding to the first view.

6. The method for predicting collision risk trend of intelligent vehicles according to claim 2, characterized in that: The decoupling of the intelligent vehicle's decision intentions based on the surrounding vehicle trajectory concept characterization and combining all generated intentions to obtain the final risk trend characterization specifically include the following steps: Assume that the underlying intentions that lead to the risk of interaction are Extracted from the trajectory representation of surrounding vehicles through soft clustering High-level semantic base ; A probability assignment matrix is ​​learned as: ； in, is the activation function, is the model hyperparameter, C Each row of provides a soft assignment of probability nodes to different intentions, based on C Perform polymerization to obtain aggregate embeddings; Using a semantic projection head right Aggregate embedding is used to transform features and output a set of concept-aware semantic bases: ； in, , , Corresponding to different semantic spaces, it is embedded in the risk trend of the intelligent vehicle and jointly decouples the intelligent vehicle's decision intention. Dimensions representing trajectory concepts: ； in, represents the connection between two embedded Indicates mapping the combination pair to the Risk trend projection head of the intention space, Indicates the intention; Combining all intentions together, the final risk trend is characterized as follows: 。 7. The method for predicting collision risk trend of intelligent vehicles according to claim 6, characterized in that: It also includes enhancing the decoupled intelligent vehicle decision intention through intent-based contrastive learning and encoding-downcoding regularization, which specifically includes the following steps: Adopt edge discarding strategy for the original graph Constructing the augmented graph , and then we get the augmented decomposition risk trend representation: ； Then the intent-based contrastive learning loss is: ； in, Indicates risk trends By The probability of an intention leading to For the The risk trend comparison learning subtask corresponding to the intention; The optimal intention that can maximize the subtask expectation is based on the concept perception semantic basis. , get the Trend confidence for each intent: ； in, Is related to temperature and The cosine similarity of k The contrastive learning subtask for each intent is: ； in, and For the The right intention; Encoding and decoding regularization is used to further separate the risk trends corresponding to different intentions. For a given risk trend representation , the coding rate of the overall risk trend is defined as the average coding length of each risk trend: ； in, is the coding rate calculation function, is a hyperparameter representing the threshold, which indicates that the expected decoding error is less than , is the identity matrix, T represents the matrix transpose; Define a set of membership matrices , orthogonally mapping the risk trends corresponding to different intentions to different subspaces; among them, is a diagonal matrix whose diagonal elements correspond to the If each risk trend group is coded separately, the probability of risk trend under the intention is Groups have an expected number of vector; then for the matrix , the total compactness of each risk trend group is the sum of the coding rates of all risk trend groups: ； in, is the compactness calculation function, represents the matrix determinant, represents the trace of the matrix; The difference in coding rate between the overall risk trend representation of intelligent vehicles and each group of risk trend representations is expressed as: 。 8. The method for predicting collision risk trend of intelligent vehicles according to claim 7, characterized in that: The method of predicting the collision risk between the intelligent vehicle and surrounding vehicles based on the learned intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and decoupling risk trend representation specifically includes the following steps: Smart Vehicles With surrounding vehicles The collision risk of an interaction is expressed as: ； Adopt pairwise Bayesian personalized ranking loss to make the observed positive pairs The score is higher than the unobserved , to obtain the surrounding vehicles with the highest interaction risk with the smart vehicle: ； The total loss of the interaction risk between the intelligent vehicle and surrounding vehicles is predicted to be: ； in, is the model parameter set, 、 and is a hyperparameter that controls the strength of each component.

9. An intelligent vehicle collision risk trend prediction system, characterized in that: include: The acquisition module is used to collect real-time trajectory data of the intelligent vehicle and its surrounding vehicles, and construct a graph structure based on the trajectory data; the graph structure includes the intelligent vehicle trajectory, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts; the surrounding vehicle trajectory concepts include trajectory type, coverage, and complexity; The trajectory representation generation module is used to extract patch sequences of intelligent vehicle trajectories, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts from the graph structure using instance normalization functions and patch processing functions. A complementary masking strategy is used to generate two views of the patch sequences of intelligent vehicle trajectories, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts, and a [CLS] token is added to the beginning of each patch sequence. The added patch sequences are input into the two-layer graph encoder LightGCN learning embedding function, and complementary contrast learning and patch reconstruction are performed based on the two views of each patch sequence to obtain timestamp-level representations. Through the [CLS] token strategy, an instance-level representation is obtained from each timestamp-level representation using an instance-level contrast task. The intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation are obtained by combining each generated timestamp-level representation with its corresponding instance-level representation. Decouple the decision intentions of intelligent vehicles based on the concept of surrounding vehicle trajectories, and combine all generated intentions to obtain the final risk trend representation; The prediction module is used to predict the collision risk between the intelligent vehicle and surrounding vehicles based on the intelligent vehicle trajectory representation, the surrounding vehicle trajectory representation and the final risk trend representation.

Images