An intelligent vehicle collision risk trend prediction method and system

Abstract

The application discloses a kind of intelligent vehicle collision risk trend prediction method and system, it is related to intelligent vehicle risk trend prediction technical field, the method includes: the trajectory data collected is modeled as intelligent vehicle-surrounding vehicle-concept diagram, trajectory data is normalized and patching treatment and is enhanced using complementary masking, design contrast learning and patch reconstruction task to capture the timestamp level representation of trajectory sequence, develop instance contrast task to capture the overall representation of sequence;The trajectory representation extracted is input into the risk trend decoupling module to learn the risk trend representation of vehicle interaction, and the intention leading to the risk trend is obtained by soft clustering the trajectory concept representation extracted, so as to predict the collision risk of intelligent vehicle;The method promotes the risk trend reasoning under complex traffic scene, improves the accuracy of collision risk prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle risk trend prediction technology, specifically to an intelligent vehicle collision risk trend prediction method and system. Background Technology

[0002] The development of intelligent vehicles and the widespread adoption of advanced driver assistance systems have further improved vehicle driving safety. By collecting real-time data on the trajectories of surrounding vehicles, intelligent vehicles can predict the evolution of risks and thus formulate corresponding driving strategies in advance to avoid collisions.

[0003] Graph Neural Network (GNN)-based inference systems are a current mainstream trend. In risk trend inference, these systems can fully utilize higher-order relations to learn better representations of the trajectories of intelligent vehicles and surrounding vehicles. However, GNN-based risk trend inference systems cannot understand the decision-making intent of intelligent vehicles from interaction data. Furthermore, when learning vehicle trajectory representations, it is often necessary to rely on agent tasks to extract timestamp-level or instance-level embeddings. Masking strategies and contrastive learning are advanced agent tasks in recent years. However, masking strategies lead to reduced model efficiency and robustness when capturing dependencies between patches, while contrastive learning mostly uses pooling methods to extract instance-level embeddings from timestamp-level embeddings during representation learning. This results in the learned vehicle trajectory embeddings being mostly concentrated in a certain subspace of the vector space, causing the inference model to be unable to effectively distinguish different vehicle trajectories, thus leading to misjudgments of the collision risk between intelligent vehicles and surrounding vehicles. Summary of the Invention

[0004] To address the shortcomings of existing technologies that cannot effectively distinguish between different vehicle trajectories, thus leading to misjudgments of collision risks between intelligent vehicles and surrounding vehicles, this invention proposes a method and system for predicting collision risk trends of intelligent vehicles. The method involves modeling the collected vehicle trajectory and surrounding vehicle trajectories into an intelligent vehicle-surrounding vehicle-concept graph. Trajectory data undergoes instance normalization and patching, and is 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 then subjected to soft clustering to obtain the intent leading to risk trends, thereby predicting the collision risk of intelligent vehicles and solving the problems existing in the prior art.

[0005] A method for predicting collision risk trends in intelligent vehicles includes the following steps:

[0006] Real-time trajectory data of intelligent vehicles and surrounding vehicles are collected, and a graph structure is constructed based on the trajectory data; the graph structure includes intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concepts; the surrounding vehicle trajectory concept includes trajectory type, coverage area, and complexity.

[0007] Patch sequences of intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept are extracted from the graph structure using instance normalization and patching functions, respectively. A complementary masking strategy is used to generate two views of these patch sequences, and a [CLS] token is added to the beginning of each patch sequence. These patch sequences are then input into the LightGCN learning embedding function of a two-layer graph encoder. Complementary contrastive learning and patch reconstruction are performed based on the two views of each patch sequence to obtain timestamp-level representations. Instance-level representations are obtained from each timestamp-level representation using the [CLS] token strategy and an instance-level contrastive task. Combining each generated timestamp-level representation with its corresponding instance-level representation yields intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation. The decision-making intent of the intelligent vehicle is decoupled based on the surrounding vehicle trajectory concept representation, and all generated intents are combined to obtain the final risk trend representation.

[0008] Based on the trajectory representation of intelligent vehicles, the trajectory representation of surrounding vehicles, and the final risk trend representation, the collision risk between intelligent vehicles and surrounding vehicles is predicted.

[0009] Furthermore, the graph structure uses a graph. Representation; where, node set It includes the concepts of intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory. This represents the trajectory sequence of intelligent vehicles. Let A represent the set of A surrounding vehicle trajectory sequences. Represents a set of B surrounding vehicle trajectory concepts; edge set This indicates the correlation between the trajectory of the intelligent vehicle and the trajectories of surrounding vehicles. Represents the trajectory of intelligent vehicles and the trajectories of surrounding vehicles. A historical risk trend, Indicates the trajectory of surrounding vehicles Belongs to the concept of surrounding vehicle trajectory .

[0010] Furthermore, the generation process of the timestamp-level representation specifically includes the following steps:

[0011] intelligent vehicle trajectory sequences Using instance normalization function and patching processing functions A series of patches were obtained. The patching process transforms the time dimension of a single vehicle trajectory sequence from... Reduce to Feature dimensions from Expand to ; and These represent the time dimension and feature dimension of the sequence, respectively. This indicates the number of patches in a single vehicle trajectory sequence. Indicates the patch length;

[0012] Two views are generated using a complementary masking strategy, and a [CLS] token is added at the beginning of the patch sequence to obtain... Similarly, the trajectory patch sequence of surrounding vehicles is obtained. and surrounding vehicle trajectory concept patch sequence , This represents the sequence of trajectories of the nth surrounding vehicle that interacted with the intelligent vehicle. This represents the concept sequence of the b-th trajectory to which the a-th surrounding vehicle trajectory belongs;

[0013] By , The input is the embedding function learned by the two-layer graph encoder LightGCN. In this process, based on a graph-structured message passing mechanism, neighborhood information of the intelligent vehicle trajectory sequence is aggregated and combined with the intelligent vehicle trajectory sequence representation from the previous iteration to obtain... Representation 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 a timestamp-level representation of the intelligent vehicle trajectory sequence;

[0014] Similarly, by using the surrounding vehicle trajectory patch sequence and its adjacent nodes... and The input is the embedding function learned by the two-layer graph encoder LightGCN. In the process, a timestamp-level representation of the surrounding vehicle trajectory sequence is obtained; the surrounding vehicle trajectory concept patch sequence is then processed. and its adjacent nodes The input is the embedding function learned by the two-layer graph encoder LightGCN. In this process, a timestamp-level representation of the concept sequence of surrounding vehicle trajectories is obtained.

[0015] Furthermore, the step of obtaining instance-level representations from timestamp-level representations using the [CLS] token strategy and an instance comparison task specifically includes the following steps:

[0016] Will and As two views representing the timestamp level of the trajectory sequence of intelligent vehicles, for each view, the embedding corresponding to the [CLS] token is extracted as an instance-level embedding;

[0017] 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;

[0018] Similarly, the [CLS] token strategy is used to obtain instance-level representations of the surrounding vehicle trajectory sequence and the surrounding vehicle trajectory concept sequence.

[0019] Furthermore, it also includes obtaining the total loss through training complementary contrastive learning, patch reconstruction, and instantiated contrastive tasks, and obtaining the best intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation by optimizing the total loss.

[0020] The total loss includes the loss from the contrastive learning process, the loss from the patch reconstruction process, and the instantiation contrastive loss; the loss from the contrastive learning is obtained through the interaction between the two views. The probability calculation yielded that the 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 calculating the loss between the instance-level prediction result of the first view and the instance-level embedding corresponding to the second view using negative cosine similarity and calculating the loss between the instance-level prediction result of the second view and the instance-level embedding corresponding to the first view.

[0021] Furthermore, the step of decoupling the decision-making intent of the intelligent vehicle based on the concept of surrounding vehicle trajectories, and combining all generated intents to obtain the final risk trend representation, specifically includes the following steps:

[0022] Assuming the underlying intentions leading to the interactive risk trend are: One, extract from the trajectory representation of surrounding vehicles through soft clustering. A high-level semantic base ;

[0023] Learning a probability assignment matrix as:

[0024] ;

[0025] in, For activation function, For model hyperparameters, C Each row provides a soft assignment of probability nodes to different intentions, based on C for probability nodes. Aggregate to obtain One aggregate embedding;

[0026] Use a semantic projection head right The aggregated embeddings are used for feature transformation, outputting a set of concept-aware semantic bases:

[0027] ;

[0028] in, , , Corresponding to different semantic spaces, it is jointly decoupled from the risk trend embedding of the intelligent vehicle's decision-making intent. Dimensions representing the concept of trajectory:

[0029] ;

[0030] in, This indicates a connection between two embedded elements. This indicates that the combination is mapped to the first... Risk trend projection head in the intent space, Indicates the first One intention;

[0031] Combining all intentions yields the final decoupling risk trend characterization as follows:

[0032] .

[0033] Furthermore, it also includes enhancing the decision-making intent of decoupled intelligent vehicles through intent-based contrastive learning and encoding decoding regularization, specifically including the following steps:

[0034] Using an edge-dropping strategy for the original graph Constructing an augmented graph This leads to the augmented decomposition of risk trend representation:

[0035] ;

[0036] The intention-based contrastive learning loss is:

[0037] ;

[0038] in, Indicates risk trend By the The probability caused by an intention, For the first The risk trend comparison learning subtask corresponding to each intention; for learning An optimal intent that maximizes the expectations of subtasks, based on a concept-aware semantic base. , obtained the Trend confidence of an intention:

[0039] ;

[0040] in, It is related to temperature and The cosine similarity; therefore, the first k The comparative learning subtask for each intention is:

[0041] ;

[0042] in, and For the first The intention is exactly the opposite;

[0043] Encoding reduction 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:

[0044] ;

[0045] in, The coding rate calculation function, is a hyperparameter representing the threshold, indicating the expected decoding error is less than . , Let T be the identity matrix, and let T denote the matrix transpose.

[0046] Define a set of membership matrices This orthogonally maps the risk trends corresponding to different intentions to different subspaces; among them, It is a diagonal matrix, and its diagonal elements correspond to the first... The probability of risk trends under each intention; if each risk trend group is encoded separately, then the probability of risk trends under each intention is... The group has the expected number Vectors; then for matrices The total tightness of each risk trend group is the sum of the coding rates of all risk trend groups:

[0047] ;

[0048] in, For compactness calculation functions, This indicates taking the determinant of a matrix. Represents the trace of a matrix;

[0049] The difference in coding rate between the overall risk trend representation and each group of risk trend representations for intelligent vehicles is expressed as:

[0050] .

[0051] Furthermore, the prediction of 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:

[0052] intelligent vehicles and surrounding vehicles The collision risk of interaction is represented as:

[0053] ;

[0054] Employing paired Bayesian personalized ranking loss, the observed positive pair... The score was higher than the unobserved pair To identify the surrounding vehicles with the highest risk of interacting with intelligent vehicles:

[0055] ;

[0056] The total loss from the predicted interaction risk between intelligent vehicles and surrounding vehicles is expressed as:

[0057] ;

[0058] in, For the model parameter set, , and Hyperparameters for controlling the intensity of each component.

[0059] The present invention also includes an intelligent vehicle collision risk trend prediction system, comprising:

[0060] The data acquisition module is used to collect real-time trajectory data of the intelligent vehicle and its surrounding vehicles, and to construct a graph structure based on the trajectory data. The graph structure includes the intelligent vehicle trajectory, the surrounding vehicle trajectory, and the concept of surrounding vehicle trajectory. The concept of surrounding vehicle trajectory includes the type, coverage area, and complexity of the trajectory.

[0061] The trajectory representation generation module extracts patch sequences of intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept from the graph structure using instance normalization and patching processing functions, respectively. It generates two views of these patch sequences using a complementary masking strategy, adding a [CLS] token to the beginning of each patch sequence. These sequences are then input into the LightGCN learning embedding function of a two-layer graph encoder. Complementary contrastive learning and patch reconstruction are performed based on the two views of each patch sequence to obtain timestamp-level representations. Instance-level representations are then obtained from each timestamp-level representation using the [CLS] token strategy and an instance-level contrastive task. Combining each generated timestamp-level representation with its corresponding instance-level representation yields intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation. Finally, the intelligent vehicle's decision-making intent is decoupled based on the surrounding vehicle trajectory concept representation, and all generated intents are combined to obtain the final risk trend representation.

[0062] The prediction module is used to predict the collision risk between the intelligent vehicle and surrounding vehicles based on the trajectory representation of the intelligent vehicle, the trajectory representation of surrounding vehicles, and the final risk trend representation.

[0063] This invention provides a method for predicting the trend of collision risk in intelligent vehicles, which has the following beneficial effects:

[0064] This invention constructs a smart vehicle-surrounding vehicle-concept graph and extracts patch sequences of smart vehicle trajectories, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts using instance normalization and patching functions, respectively. The patch independence strategy employed is a pre-training task, unlike masking modeling which predicts masked patches; instead, it reconstructs unmasked patches. This approach does not require capturing dependencies between patches, resulting in stronger robustness and higher efficiency in various real-world road scenarios. Furthermore, a complementary masking strategy generates two views of the patch sequences of smart vehicle trajectories, surrounding vehicle trajectories, and surrounding vehicle trajectory concepts, with a [CLS] token added at the beginning of each patch sequence to identify different trajectories. The sequence is used to extract instance-level representations in a decoupled manner. To improve the performance of trajectory representation learning, a two-layer graph encoder is further employed in the patch-independent strategy to make the model focus more on extracting patch representations, and a contrastive learning strategy is used to learn the timestamp-level representation of the trajectory. In risk trend inference of intelligent vehicles, the patch-independent strategy can comprehensively learn the rich representations of trajectory data, thus providing valuable information for subsequent risk trend inference. This invention is expected to provide interpretable, accurate, and efficient risk trend inference services for intelligent vehicles, promote risk trend inference in complex traffic scenarios, improve the accuracy of collision risk prediction, enhance the driving safety of intelligent vehicles, and contribute to building a safe road traffic environment. Attached Figure Description

[0065] 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;

[0066] Figure 2 This is a flowchart of the intelligent vehicle collision risk trend prediction method based on patch-independent strategy and decoupled graph representation learning in an embodiment of the present invention.

[0067] Figure 3 This is a flowchart illustrating the model training and validation process in an embodiment of the present invention.

[0068] Figure 4 This is a structural framework diagram of trajectory representation extraction in an embodiment of the present invention;

[0069] Figure 5 This is a flowchart of the trajectory representation learning method in an embodiment of the present invention;

[0070] Figure 6 This is a scenario example diagram of intelligent vehicle collision risk trend prediction in an embodiment of the present invention;

[0071] Figure 7 This is a diagram illustrating the risk trend results of intelligent vehicles in an embodiment of the present invention. Detailed Implementation

[0072] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0073] This invention proposes a method for predicting the trend of collision risk in intelligent vehicles, referring to... Figure 1 , Figure 2 The method specifically includes the following steps:

[0074] S1. Data Acquisition; Data acquisition devices include roadside equipment (cameras, lidar, millimeter-wave radar, etc.) and communication technologies (GPS technology, vehicle-to-vehicle communication technology, vehicle-to-infrastructure communication technology, etc.) used to collect trajectory data of intelligent vehicles and surrounding vehicles and transmit it to intelligent vehicles; the collected trajectory data includes sampling point location, sampling time, speed, etc.

[0075] S2. Data Preprocessing: The collected raw vehicle trajectory data is cleaned, normalized, missing value handled, feature selected, and the dataset partitioned. This aims to eliminate various problems in the data, enhance its quality and reliability, and thus improve the effectiveness and generalization ability of the model. The preprocessed data is then stored in the trajectory database. The specific preprocessing steps are as follows:

[0076] (1) Data cleaning: Remove invalid, duplicate and irrelevant data from the original trajectory data and perform time error correction.

[0077] (2) Data normalization: Normalize the values ​​of different orders of magnitude in the trajectory data so that they are within the same order of magnitude; at the same time, ensure that all trajectory data use a consistent coordinate system.

[0078] (3) Missing value handling: interpolate, extrapolate, or delete missing data.

[0079] (4) Feature selection: Highly relevant features are screened out by means of correlation analysis and feature importance assessment, and the most representative features are selected.

[0080] (5) Dataset partitioning: 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's hyperparameters and select the model structure. 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%, respectively.

[0081] S3. Construct a predictive 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-conceptual map, and extracts the timestamp-level representation of the trajectory data through complementary contrastive learning and patch independence strategies. It captures the overall representation of the trajectory data through instantiated contrastive learning. The risk trend inference module consists of three parts: risk trend decoupling, intent contrastive learning, and encoding rate reduction regularization. Among them, the risk trend decoupling module is used to learn the different decision intentions of the intelligent vehicle, the intent contrastive learning is used to enhance intent decoupling and infer the risk trend distribution, and the encoding rate reduction regularization ensures that the risk trends corresponding to different intentions are orthogonal, thereby making the learned risk trend representation more discriminative among different intentions.

[0082] The intelligent vehicle collision risk trend prediction method based on patch-independent policies and decoupled graph representation learning mainly relies on trajectory representation extraction and decoupled graph comparative learning to achieve accurate and comprehensive risk trend inference and collision risk prediction. The specific construction process of this model framework is as follows:

[0083] (1) Intelligent vehicle-surrounding vehicles-concept map construction: vehicle trajectory data (refer to Figure 6 (Use diagram) Indicate (refer to) Figure 1 ), where the node set It includes concepts such as intelligent vehicle trajectory, surrounding vehicle trajectory, and related concepts, edge set. This indicates the correlation between the trajectory of the intelligent vehicle and the trajectories of surrounding vehicles. Among them, This represents the trajectory sequence of intelligent vehicles. Let A represent the set of A surrounding vehicle trajectory sequences. This represents a set of B surrounding vehicle trajectory-related concepts, including trajectory type, coverage area, complexity, etc. Represents the trajectory of intelligent vehicles and the trajectories of surrounding vehicles. A historical risk trend, Indicates the trajectory of surrounding vehicles Belongs to the concept .

[0084] Given a trajectory of an intelligent vehicle and surrounding vehicle trajectories Candidate pairs The goal is to learn the decoupled intent of intelligent vehicle trajectories and the risk trend distribution of those intents, and then predict... This indicates the risk of collision between the intelligent vehicle and surrounding vehicles.

[0085] (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 computational costs. Trajectory representation learning can extract the most critical features of trajectory data, such as trajectory direction changes, speed patterns, and path regularities, thereby enhancing the model's understanding and generalization abilities and improving the efficiency and accuracy of risk trend inference. Therefore, before performing risk trend inference, it is necessary to extract trajectory representations of intelligent vehicles and surrounding vehicles (refer to...). Figure 4 ).

[0086] Specifically, firstly, the trajectory sequence of intelligent vehicles Using instance normalization function and patching processing functions A series of patches were obtained. .

[0087] ;

[0088] Among them, the patching process transforms the time dimension of a single vehicle trajectory sequence from... Reduce to Feature dimensions from Expand to .in, and These represent the time dimension and feature dimension of the sequence, respectively. This indicates the number of patches in a single vehicle trajectory sequence. This indicates the patch length. Two views are generated using a complementary masking strategy (for complementary contrastive learning and instance-based contrastive learning), and a [CLS] token is added to the beginning of the patch sequence (to label different trajectory sequences, thereby extracting instance-level representations in a decoupled manner). :

[0089] ;

[0090] Similarly, the trajectory patch sequence of surrounding vehicles can be obtained. and surrounding vehicle trajectory concept patch sequence , This represents the sequence of trajectories of the nth surrounding vehicle that interacted with the intelligent vehicle. This represents the concept sequence of the b-th trajectory to which the a-th surrounding vehicle trajectory belongs.

[0091] Secondly, in order to extract the timestamp-level representation of the trajectory sequence of intelligent vehicles, , Input a simple two-layer graph encoder, LightGCN, to learn embedding functions. Based on a graph-structured message passing mechanism, neighborhood information of the intelligent vehicle trajectory sequence is aggregated and combined with the intelligent vehicle trajectory sequence representation from the previous iteration to obtain... Representation in the first and second layers of the encoder and They are used for complementary contrastive learning and patch reconstruction, respectively.

[0092] ;

[0093] ;

[0094] In the formula, , , , and , , L represents the number of trajectory sequences of the intelligent vehicle, M represents the number of channels in a single vehicle trajectory sequence, and N represents the number of patches in a single channel of a single vehicle trajectory sequence. for The input dimension, O is The output dimensions are the patch size and the embedding dimension, respectively. and These are aggregation and combination functions, respectively. yes The set of adjacent nodes represents, in this case, the set of trajectories of surrounding moving vehicles interacting with the intelligent vehicle. A readout function is used. To integrate the representations output by the two-layer graph encoder, a timestamp-level representation is obtained:

[0095] ;

[0096] Similarly, by using the surrounding vehicle trajectory patch sequence and its adjacent nodes... and Inputting a two-layer graph encoder (LightGCN) yields a timestamp-level representation of the surrounding vehicle trajectory sequence. This is achieved by patching the surrounding vehicle trajectory concept sequence with its adjacent nodes. Inputting a two-layer graph encoder LightGCN yields a timestamp-level representation of the surrounding vehicle trajectory concept sequence.

[0097] Reference Figure 5 The specific process of complementary contrastive learning and patch reconstruction tasks is as follows:

[0098] ① Complementary and comparative learning: A system was established The system uses probability to learn missing patch information in complementary views, thereby capturing the relationship between adjacent timestamps in a single vehicle trajectory sequence in a hierarchical manner. and The representations of the two views generated by the complementary masking strategy at the first layer, and the relationship between these two views. Probability can be defined as follows.

[0099] ;

[0100] In the formula, the dot product This represents the computation of similarity. Therefore, the loss for contrastive learning can be expressed as:

[0101] ;

[0102] right Perform max pooling and contrastive learning, and iteratively calculate the hierarchical loss until... :

[0103] ;

[0104] ② Patch Reconstruction: Represent the vehicle trajectory from the second layer output. Input a block-by-block linear projection head to obtain the patch reconstruction results:

[0105] ;

[0106] In the formula, This is the weight matrix of the linear layer.

[0107] The loss from patch rebuilding can be expressed as:

[0108] ;

[0109] Finally, an instantiation comparison task is used to capture the overall information of the entire sequence (see reference). Figure 5 Specifically, let's assume... and Two views are used to represent the timestamp-level representation of the trajectory sequence of intelligent vehicles. For each view, the embedding corresponding to the first position ([CLS] token) is extracted as an instance-level embedding.

[0110] ;

[0111] ;

[0112] Each instance-level embedding is passed through a two-layer bottleneck MLP encoder with BatchNorm and ReLU. Processing is performed to obtain and Used for instantiating contrastive learning:

[0113] ;

[0114] ;

[0115] The instantiation contrast loss is calculated using negative cosine similarity. First, calculate... and Losses between:

[0116] ;

[0117] In the formula, This represents the function for calculating cosine similarity. This indicates stopping the gradient operation, used here to address sampling bias in instanced contrastive learning. Further calculations are performed to symmetrically optimize the network. and Losses between:

[0118] ;

[0119] Therefore, the overall instantiation contrast loss for and Average value:

[0120] ;

[0121] Finally, the complementary contrastive learning task, patch reconstruction task, and instantiated contrastive task are trained jointly, and then... and Adjustments are made among the three losses to obtain the total loss for character extraction. By optimizing this loss, the optimal trajectory representation of autonomous vehicles can be obtained. .

[0122] ;

[0123] Similarly, the optimal trajectory of surrounding vehicles can be obtained. The representation and concept The representation Then, the interactive risk trend representation combines the trajectory representation of intelligent vehicles and the trajectory representation of surrounding vehicles. for:

[0124] .

[0125] ③ Risk Trend Inference: The interaction risk between intelligent vehicles and surrounding vehicles may be influenced by multiple decision-making intentions. These complex intentions are usually closely related to the decision-making mechanism of the intelligent vehicle and the trajectory characteristics of surrounding vehicles. Ignoring information from any one aspect may lead to insufficient and inaccurate risk trend inference. 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 representations. Furthermore, intention-based risk trend distribution can further reflect the different decision-making intentions of intelligent vehicles, improving the interpretability of collision risk prediction.

[0126] Specifically, the first step is to decouple the risk trend representation and extract the decision-making intent of the intelligent vehicle. It is assumed that the potential intents leading to the interactive risk trends include... One. Extracting from the trajectory concept representation of surrounding vehicles through soft clustering. A high-level semantic base First, a probability assignment matrix is ​​learned as follows:

[0127] ;

[0128] In the formula, For activation function, and For the model hyperparameters, each row of C provides a soft assignment of probability nodes to different intentions, based on C to the probability nodes. Aggregate to obtain An aggregate embedding. A semantic projection head is used. Perform feature transformation on these clustering embeddings to output a set of concept-aware semantic bases:

[0129] ;

[0130] In the formula, , , Corresponding to different semantic spaces, it is jointly decoupled from the risk trend embedding of the intelligent vehicle's decision-making intent. Dimensions representing the concept of trajectory:

[0131] ;

[0132] In the formula, This indicates a connection between two embedded elements. This indicates that the combination is mapped to the first... Risk trend projection head in the intent space, Indicates the first Each intention, when combined, yields a decoupled risk trend representation:

[0133] ;

[0134] Secondly, intention-based contrastive learning is employed to enhance risk trend decoupling. Specifically, an edge-dropping strategy is first used for the original graph. Constructing an augmented graph This leads to the augmented decomposition of risk trend representation:

[0135] ;

[0136] Therefore, the intention-based contrastive learning loss is:

[0137] ;

[0138] In the formula, Indicates risk trend By the The probability caused by an intention, For the first The subtask involves comparing and learning risk trends corresponding to each intention. This is for learning... An optimal intent that maximizes the expectations of subtasks, based on a concept-aware semantic base. , can obtain the first Trend confidence of an intention:

[0139] ;

[0140] In the formula, It is related to temperature and The cosine similarity is used to assign a greater weight to the decision-making intent of intelligent vehicles with high confidence in contrastive learning. Therefore, the first... k The comparative learning subtask for each intention is:

[0141] ;

[0142] In the formula, and For the first The intention is exactly the right one.

[0143] Furthermore, since risk trends corresponding to different intentions may be distributed in different subspaces, this makes the learned risk trend representations discriminative. Therefore, maximizing encoding rate decay is employed to enhance the diversity of risk trend representations. 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:

[0144] ;

[0145] In the formula, The coding rate calculation function, is a hyperparameter representing the threshold, indicating the expected decoding error is less than . , It is the identity matrix. T This indicates the matrix transpose.

[0146] To ensure that risk trends corresponding to different intentions are orthogonal, they need to be mapped to different subspaces. Therefore, a set of membership matrices is defined. In the formula, It is a diagonal matrix, and its diagonal elements correspond to the first... The probability of risk trends under each intention. If each risk trend group is encoded separately, then the probability of risk trends under each intention is... The group has the expected number Vectors. Therefore, for matrices The total tightness of each risk trend group is the sum of the coding rates of all risk trend groups:

[0147] ;

[0148] In the formula, For compactness calculation functions, This indicates taking the determinant of a matrix. Represents the trace of a matrix.

[0149] Therefore, the difference in coding rate between the overall risk trend representation of intelligent vehicles and the risk trend representation of each group is expressed as follows, and a better trend representation can be obtained by maximizing this value:

[0150] ;

[0151] Finally, based on the learned trajectory representations of the intelligent vehicle and surrounding vehicles, the interaction risk between the intelligent vehicle and surrounding vehicles is predicted (see reference). Figure 7 Specifically, intelligent vehicles and surrounding vehicles The collision risk of interaction can be represented as:

[0152] ;

[0153] Employing paired Bayesian personalized ranking loss, the observed positive pair... The score was higher than the unobserved pair To identify the surrounding vehicles with the highest risk of interacting with intelligent vehicles:

[0154] ;

[0155] Therefore, the total loss from predicting the interaction risks between intelligent vehicles and surrounding vehicles can be expressed as:

[0156] ;

[0157] In the formula, For the model parameter set, , and Hyperparameters for controlling the intensity of each component.

[0158] 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 achieve optimal performance. Then, the optimal model is tested using test set data to evaluate its predictive performance, and finally, the model prediction results are output.

[0159] The intelligent vehicle collision risk trend prediction system proposed in this invention, based on a patch-independent strategy and decoupled graph representation learning, employs a patch-independent strategy as a pre-training task. Unlike mask modeling, which predicts masked patches, this system reconstructs unmasked patches. Therefore, it does not need to capture dependencies between patches, exhibiting stronger robustness and higher efficiency in various real-world road scenarios.

[0160] To improve the performance of trajectory representation learning, this invention further employs a simplified two-layer graph encoder in the patch-independent strategy, allowing the model to focus more on extracting patch representations. A contrastive learning strategy is used to learn both the timestamp-level and overall representations of the trajectory. In risk trend inference for intelligent vehicles, the patch-independent strategy can comprehensively learn rich representations of trajectory data, thus providing valuable information for subsequent risk trend inference.

[0161] The decoupling technique employed in this invention is a method for handling highly correlated data in deep learning. By reducing feature correlation, it enables the model to better understand and utilize features, thereby improving the model's inference and prediction performance and robustness. This invention further extends this to decoupling graph contrastive learning to enhance decoupling. In risk trend inference for intelligent vehicles, decoupling graph contrastive learning can achieve better, more interpretable risk trend inference by learning the complex and diverse decision intentions of intelligent vehicles. Therefore, this invention is expected to provide interpretable, accurate, and efficient risk trend inference services for intelligent vehicles, promote risk trend inference in complex traffic scenarios, improve the accuracy of collision risk prediction, enhance the driving safety of intelligent vehicles, and contribute to building a safe road traffic environment.

[0162] This invention proposes a method for predicting collision risk trends in intelligent vehicles based on a patch-independent strategy and decoupled graph contrastive learning. It models the collected vehicle trajectory and surrounding vehicle trajectories into a concept graph, which is then used to learn the interaction relationships between the intelligent vehicle and surrounding vehicles and infer the intelligent vehicle's decision-making intent. Next, the intelligent vehicle trajectory data and surrounding vehicle trajectory data from the test set are input into the trained model to learn important representations of the trajectory data, infer the intelligent vehicle's decision-making intent, and predict the collision risk. This invention can provide intelligent vehicles with more accurate and reliable risk trend inference and collision risk prediction, enhancing their risk assessment capabilities in complex traffic scenarios, thereby promoting the development and upgrading of intelligent vehicles.

[0163] Based on the same inventive concept, this invention also proposes an intelligent vehicle collision risk trend prediction system, comprising:

[0164] The data acquisition module is used to collect real-time trajectory data of intelligent vehicles and surrounding vehicles, and to construct a concept map of intelligent vehicles, surrounding vehicles, and based on the trajectory data.

[0165] The trajectory representation generation module extracts patch sequences of intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept from the intelligent vehicle-surrounding vehicle-concept graph using instance normalization and patching functions, respectively. It generates two views of these patch sequences using a complementary masking strategy, adding a [CLS] token to the beginning of each sequence. These sequences are then input into the LightGCN learning embedding function of a two-layer graph encoder. Complementary contrastive learning and patch reconstruction are performed based on the two views of each sequence, generating timestamp-level representations. Instance-level representations are obtained from the timestamp-level representations using the [CLS] token strategy and an instance-level contrastive task. Combining each generated timestamp-level representation with its corresponding instance-level representation yields the intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation. The intelligent vehicle's decision-making intent is decoupled based on the surrounding vehicle trajectory concept representation, and all generated intents are combined to obtain the final risk trend representation.

[0166] The prediction module is used to predict the collision risk between the intelligent vehicle and surrounding vehicles based on the trajectory representation of the intelligent vehicle, the trajectory representation of surrounding vehicles, and the final risk trend representation.

[0167] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the trend of collision risk in intelligent vehicles, characterized in that, Includes the following steps: Real-time trajectory data of intelligent vehicles and surrounding vehicles are collected, and a graph structure is constructed based on the trajectory data; the graph structure includes intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concepts; the surrounding vehicle trajectory concept includes trajectory type, coverage area, and complexity. Patch sequences for intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept are extracted from the graph structure using instance normalization and patching functions, respectively. A complementary masking strategy is employed to generate two views of these patch sequences, and a [CLS] token is added to the beginning of each patch sequence. These patch sequences are then input into the LightGCN learning embedding function of a two-layer graph encoder. Complementary contrastive learning and patch reconstruction are performed based on the two views of each patch sequence to obtain timestamp-level representations. Instance-level representations are then obtained from each timestamp-level representation using the [CLS] token strategy and an instance-level contrastive task. Finally, by combining each generated timestamp-level representation with its corresponding instance-level representation, intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation are obtained. Based on the concept of surrounding vehicle trajectories, the decision-making intent of intelligent vehicles is decoupled, and all generated intents are combined to obtain the final risk trend representation. Based on the trajectory representation of intelligent vehicles, the trajectory representation of surrounding vehicles, and the final risk trend representation, the collision risk between intelligent vehicles and surrounding vehicles is predicted.

2. The intelligent vehicle collision risk trend prediction method according to claim 1, characterized in that, The graph structure is a diagram. Representation; where, node set It includes the concepts of intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory. This represents the trajectory sequence of intelligent vehicles. Let A represent the set of A surrounding vehicle trajectory sequences. Represents a set of B surrounding vehicle trajectory concepts; edge set This indicates the correlation between the trajectory of the intelligent vehicle and the trajectories of surrounding vehicles. Represents the trajectory of intelligent vehicles and the trajectories of surrounding vehicles. A historical risk trend, Indicates the trajectory of surrounding vehicles Belongs to the concept of surrounding vehicle trajectory .

3. The intelligent vehicle collision risk trend prediction method according to claim 2, characterized in that, The generation process of the timestamp-level representation specifically includes the following steps: intelligent vehicle trajectory sequences Using instance normalization function and patching processing functions A series of patches were obtained. The patching process transforms the time dimension of a single vehicle trajectory sequence from... Reduce to Feature dimensions from Expand to ; and These represent the time dimension and feature dimension of the sequence, respectively. This indicates the number of patches in a single vehicle trajectory sequence. Indicates the patch length; Two views are generated using a complementary masking strategy, and a [CLS] token is added at the beginning of the patch sequence to obtain... Similarly, the trajectory patch sequence of surrounding vehicles is obtained. and surrounding vehicle trajectory concept patch sequence , This represents the sequence of trajectories of the nth surrounding vehicle that interacted with the intelligent vehicle. This represents the concept sequence of the b-th trajectory to which the a-th surrounding vehicle trajectory belongs; By , The input is the embedding function learned by the two-layer graph encoder LightGCN. In this process, based on a graph-structured message passing mechanism, neighborhood information of the intelligent vehicle trajectory sequence is aggregated and combined with the intelligent vehicle trajectory sequence representation from the previous iteration to obtain... Representation 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 a timestamp-level representation of the intelligent vehicle trajectory sequence; Similarly, by inputting the surrounding vehicle trajectory patch sequence and its adjacent nodes into the embedding function learned by the two-layer graph encoder LightGCN, In the process, a timestamp-level representation of the surrounding vehicle trajectory sequence is obtained; the surrounding vehicle trajectory concept patch sequence is then processed. The embedding function learned by the LightGCN two-layer graph encoder, along with its adjacent nodes, is input to the graph. In this process, a timestamp-level representation of the surrounding vehicle trajectory concept sequence is obtained; wherein, the surrounding vehicle trajectory patch sequence The adjacent nodes include and The surrounding vehicle trajectory concept patch sequence The adjacent nodes are .

4. The intelligent vehicle collision risk trend prediction method according to claim 3, characterized in that, The process of obtaining instance-level representations from timestamp-level representations using the [CLS] token strategy and an instance comparison task specifically includes the following steps: Will and As two views representing the timestamp level of the trajectory sequence of intelligent vehicles, for each view, the embedding corresponding to the [CLS] token is extracted as an 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 used to obtain instance-level representations of the surrounding vehicle trajectory sequence and the surrounding vehicle trajectory concept sequence.

5. The intelligent vehicle collision risk trend prediction method according to claim 4, characterized in that, It also includes obtaining the total loss through training complementary contrastive learning, patch reconstruction and instantiation contrastive 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 from the contrastive learning process, the loss from the patch reconstruction process, and the instantiation contrastive loss; the loss from the contrastive learning is obtained through the interaction between the two views. The probability calculation yielded that the 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 calculating the loss between the instance-level prediction result of the first view and the instance-level embedding corresponding to the second view using negative cosine similarity and calculating the loss between the instance-level prediction result of the second view and the instance-level embedding corresponding to the first view.

6. The intelligent vehicle collision risk trend prediction method according to claim 2, characterized in that, The process of decoupling the decision-making intent of the intelligent vehicle based on the concept of surrounding vehicle trajectories, and combining all generated intents to obtain the final risk trend representation, specifically includes the following steps: Assuming the underlying intentions leading to the interactive risk trend are: One, extract from the trajectory representation of surrounding vehicles through soft clustering. A high-level semantic base ; Learning a probability assignment matrix as: ； in, For activation function, For model hyperparameters, C Each row provides a soft assignment of probability nodes to different intentions, based on C for probability nodes. Aggregate to obtain One aggregate embedding; Use a semantic projection head right The aggregated embeddings are used for feature transformation, outputting a set of concept-aware semantic bases: ； in, , , Corresponding to different semantic spaces, it is jointly decoupled from the intelligent vehicle's decision-making intent by embedding the risk trend given by the intelligent vehicle. Dimensions representing the concept of trajectory: ； in, This indicates a connection between two embedded elements. This indicates that the combination is mapped to the first... Risk trend projection head in the intent space, Indicates the first One intention; Combining all intentions yields the final risk trend characterization as follows: 。 7. The intelligent vehicle collision risk trend prediction method according to claim 6, characterized in that, It also includes enhancing the decision-making intent of decoupled intelligent vehicles through intent-based contrastive learning and encoding decoding regularization, specifically including the following steps: Using an edge-dropping strategy for the original graph Constructing an augmented graph This leads to the augmented decomposition of risk trend representation: ； The intention-based contrastive learning loss is: ； in, Indicates risk trend By the The probability caused by an intention, For the first The risk trend comparison learning subtask corresponding to each intention; for learning An optimal intent that maximizes the expectations of subtasks, based on a concept-aware semantic base. , obtained the Trend confidence of an intention: ； in, It is related to temperature and The cosine similarity; therefore, the first k The comparative learning subtask for each intention is: ； in, and For the first The intention is exactly the opposite; Encoding reduction 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, The coding rate calculation function, is a hyperparameter representing the threshold, indicating the expected decoding error is less than . , Let T be the identity matrix, and let T denote the matrix transpose. Define a set of membership matrices This orthogonally maps the risk trends corresponding to different intentions to different subspaces; among them, It is a diagonal matrix, and its diagonal elements correspond to the first... The probability of risk trends under each intention; if each risk trend group is encoded separately, then the probability of risk trends under each intention is... The group has the expected number Vectors; then for matrices The total tightness of each risk trend group is the sum of the coding rates of all risk trend groups: ； in, For compactness calculation functions, This indicates taking the determinant of a matrix. Represents the trace of a matrix; The difference in coding rate between the overall risk trend representation and each group of risk trend representations for intelligent vehicles is expressed as: 。 8. The intelligent vehicle collision risk trend prediction method according to claim 7, characterized in that, The method for 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 includes the following steps: intelligent vehicles and surrounding vehicles The collision risk of interaction is represented as: ； Employing paired Bayesian personalized ranking loss, the observed positive pair... The score was higher than the unobserved pair To identify the surrounding vehicles with the highest risk of interacting with intelligent vehicles: ； The total loss from the predicted interaction risk between intelligent vehicles and surrounding vehicles is expressed as: ； in, For the model parameter set, , and Hyperparameters for controlling the intensity of each component.

9. An intelligent vehicle collision risk trend prediction system, characterized in that, include: The data acquisition module is used to collect real-time trajectory data of the intelligent vehicle and its surrounding vehicles, and to construct a graph structure based on the trajectory data. The graph structure includes the intelligent vehicle trajectory, the surrounding vehicle trajectory, and the concept of surrounding vehicle trajectory. The concept of surrounding vehicle trajectory includes the type, coverage area, and complexity of the trajectory. The trajectory representation generation module extracts patch sequences of intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept from the graph structure using instance normalization and patching processing functions, respectively. It generates two views of the patch sequences of intelligent vehicle trajectory, surrounding vehicle trajectory, and surrounding vehicle trajectory concept using a complementary masking strategy, and adds a [CLS] token to the beginning of each patch sequence. The resulting patch sequences are then input into the LightGCN learning embedding function of a two-layer graph encoder. Complementary contrastive learning and patch reconstruction are performed based on the two views of each patch sequence to obtain timestamp-level representations. Using the [CLS] token strategy, an instance-level representation is obtained from each timestamp-level representation through an instance-level contrastive task. Finally, by combining each generated timestamp-level representation with its corresponding instance-level representation, the intelligent vehicle trajectory representation, surrounding vehicle trajectory representation, and surrounding vehicle trajectory concept representation are obtained. Based on the concept of surrounding vehicle trajectories, the decision-making intent of intelligent vehicles is decoupled, and all generated intents 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 trajectory representation of the intelligent vehicle, the trajectory representation of surrounding vehicles, and the final risk trend representation.

Images