A port IGV lane-changing risk prediction method based on multi-vehicle interaction graph modeling

Abstract

The application discloses a port IGV lane-changing risk prediction method based on multi-vehicle interaction graph modeling, and belongs to the technical field of traffic safety evaluation. First, a lane-changing vehicle group under a network connection environment is constructed, a multi-vehicle interaction graph and a time sequence multi-vehicle interaction graph are established by using graph theory, and modeling is completed by fusing vehicle individual characteristics, interaction characteristics and roadside environment characteristics. Second, the comprehensive risk of lane changing is calculated through the safe stopping distance, the stopping distance index, the risk exposure level and the risk severity level. Third, a space-time attention graph neural network is constructed by using a graph attention neural network and a gated recurrent unit, and accurate lane-changing risk prediction is realized. Finally, the IGV lane-changing strategy is triggered according to the risk threshold, and the problems that the traditional port vehicle risk evaluation method is dependent on static safety rules or single-vehicle state calculation, is difficult to depict the influence of multi-vehicle dynamic interaction, and is unable to support real-time safety decision-making are solved.

Description

Technical Field

[0001] This invention belongs to the field of traffic safety evaluation technology, specifically relating to a method for predicting IGV lane-changing risks in port areas based on multi-vehicle interaction graph modeling. Background Technology

[0002] Against the backdrop of the deep integration of smart port construction and intelligent connected vehicle technology, automated terminals are gradually realizing large-scale cluster operations of Intelligent Guided Vehicles (IGVs) and vehicle-road-cloud collaborative scheduling. The continuous improvement of BeiDou high-precision positioning, 5G vehicle-to-everything (V2X) technology, multi-source perception, and onboard computing capabilities makes it possible for IGVs to predict lane-changing risks, make autonomous decisions, and execute proactive safety controls based on real-time operating status and surrounding traffic conditions. This is crucial for ensuring the safety of horizontal transportation in port areas and improving operational efficiency and traffic resilience. Currently, port IGV lane-changing risk assessment and decision-making methods are mainly divided into rule-based models and data-driven approaches, both of which have significant limitations and are difficult to adapt to the high-density, highly interactive, and dynamically changing mixed-traffic operation scenarios in ports. Traditional risk assessment based on rules and safety distances relies heavily on preset safety thresholds, stopping distances, and simple car-following models to calculate risks. This only reflects the static safety boundaries of a single vehicle and fails to quantify the dynamic interactive impact between multiple vehicles, insufficiently characterizing the collaborative risks of vehicles in front and behind the target lane, and adjacent vehicles in the original lane. The port area experiences high vehicle density, significant speed fluctuations, and frequent lane changes, often involving manned trucks and port-operated equipment. Traditional methods are prone to delayed, missed, or misjudged risks, failing to support real-time and accurate lane-changing decisions. While risk estimation based on trajectory or state data improves real-time performance, it is generally limited to single-vehicle risk or local adjacent vehicle risk, lacking a unified model of the spatiotemporal interactions of multiple vehicles and failing to characterize the coupled impact of vehicle-to-vehicle and vehicle-to-road interactions during lane changes. These methods often assess risk based on single-point or short-term states, failing to utilize temporal characteristics for forward-looking predictions. If single-vehicle risk avoidance is not timely, it can easily trigger cascading collisions, congestion, or even multi-vehicle accidents, highlighting the contradiction with the port area's 24 / 7, highly reliable, and efficient operational requirements.

[0003] Therefore, how to uniformly model the multi-vehicle interaction relationships involved in IGV lane changing in a connected environment, quantify and predict the comprehensive risks of the entire lane changing process in real time and accurately, and trigger safety control based on the risk level and terminate lane changing when necessary, has become the core issue for improving the safety and decision-making robustness of IGV autonomous driving in port areas. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting lane change risks of IGVs in port areas based on multi-vehicle interaction graph modeling, which solves the problems of insufficient accuracy of existing IGV lane change risk assessment, lack of multi-vehicle interaction modeling, delayed risk prediction, and unreliable safety decisions.

[0005] A method for predicting lane-changing risks of IGVs in port areas based on multi-vehicle interaction graph modeling includes the following steps: (1) Multi-vehicle interaction diagram modeling; (1.1) Define lane-changing car groups: In a connected environment, the IGV master vehicle obtains the speed, acceleration, and location information of surrounding vehicles through the vehicle communication system. The communication distance of the IGV vehicle is defined as 150 m. The set of all other vehicles that directly interact with the IGV master vehicle within a 150 m radius of the lane-changing IGV master vehicle is defined as the lane-changing vehicle group. (1.2) Using graph theory to model the multi-vehicle interaction relationships in lane-changing vehicle groups: Representing multi-vehicle interaction relationships using graph theory: in, To represent the graph relationships between vehicles and other physical entities, then... Described as the first The graph relationships at different times; each vehicle or individual physical entity is considered as a graph node, where... It is the first A collection of timeline nodes. , It is the first The number of vehicles at any given time; It is the first The combination of time edges represents the first time. The interaction relationship between any two physical entities at any given moment; , It is the first The number of edges at any given time; It is the first Time map The adjacency matrix represents the connections between physical entities. Edges that do not directly generate interactive activities are denoted as 0, while those with direct interactive relationships are denoted as 1. Thus, for any given time... The interaction relationship of lane-changing vehicle groups is realized using a multi-vehicle interaction diagram. express; (1.3) Define the spatiotemporal multi-vehicle interaction diagram: Define a spatiotemporal multi-vehicle interaction graph, ( Characterize the spatiotemporal changes of multi-vehicle interaction relationships; introduce feature matrices and use feature matrices. Quantify the proprietary attribute information of internal nodes of the lane-changing vehicle group. ;in For the number of nodes, The length of the feature time series, The feature matrix for each node includes individual features, interaction features, and other features. Individual features include the speed, acceleration, and lateral and longitudinal coordinate information of each vehicle, carried by the vehicle node. Interaction features include information related to the interaction between the IGV master vehicle and surrounding vehicles, including differences in speed, acceleration, and lateral and longitudinal coordinates. Other features are information carried by roadside nodes and other physical entities, including lane speed and density. Therefore, changing lanes carries risks. This can be represented as a functional relationship between a graph that dynamically changes over time and its characteristic matrix. in, For time The risk at this moment, and the use of the diagram and its features arrive Real-time data, to achieve... arrive Feature prediction at time point The risk of time; (2) Lane change risk calculation: Adopting the lane change risk index Describe the comprehensive interaction risk between the IGV (In-Gear Vehicle) driver and surrounding vehicles during lane changes; and calculate the safe stopping distance (SSD) and the stopping distance index (SDI). (3) Lane change risk prediction based on multi-vehicle interaction graph modeling: The spatial features of the multi-vehicle interaction graph are extracted using a graph attention neural network (GAT), and the temporal features of the temporal multi-vehicle interaction graph are extracted using a gated recurrent neural network (GRN), thereby enabling lane change risk prediction. (4) Application of IGV driver lane change termination based on lane change risk prediction: Lane change risk index For lanes ∈ [0,1], the higher the risk index, the greater the risk of lane changing. When the risk index exceeds the threshold... This indicates that a collision would occur if the lane change continued. Therefore, when the IGV driver was performing a lane change, the risk index calculated in real time by the onboard system exceeded the threshold. The IGV autonomous driving system immediately terminates the lane change and maintains a straight-line driving posture, as shown in the following formula: in, This indicates the conditions under which the IGV (Incoming Traffic Vehicle) terminates lane changing.

[0006] Furthermore, the specific steps of step (2) are as follows: (2.1) Calculate the parking sight distance: The SSD calculation formula is: in, and These represent the safe stopping distances of the vehicle in front and behind the IGV, respectively. and These are the speeds of the vehicle in front and the vehicle behind before they braked; The IGV braking reaction time is set to 0.3s; and This is the acceleration of the car in front and the car behind, taken as 3.4 m / s². 2 ; (2.2) Calculate the parking distance index: SDI is used to measure whether a following vehicle has collided with the preceding vehicle after the preceding vehicle has braked. The calculation formula is as follows: The length of the front vehicle is The distance between the two vehicles before braking ;when , , The sum is greater than If the conditions are met, no collision will occur; otherwise, a collision will occur. (2.3) Calculate the risk exposure level and severity level: Lane change risk exposure level (REL) represents the ratio of the number of times the SDI equals 1 during the lane change time to the total assessment time. The higher the value, the greater the probability of a collision during the assessment time. Lane change risk severity level (RSL) represents the ratio of the maximum SDI that can be reached during the lane change time to the theoretical extreme value of SDI, reflecting the intensity of the collision. The calculation formula is as follows: in, It's time to change lanes. Take 0.1s; ; yes The value in extreme cases, i.e., the result of the vehicle in front braking to a stop without any stopping distance. When SDI equals 1 during the evaluation process The maximum value; (2.4) Calculate the interaction risk between any two vehicles: The interaction risk between the two vehicles is obtained by multiplying REL and RSL, that is: in, For any two vehicles, the interaction risk is... These represent the interaction risks between the IGV driver and the vehicle in front in the original lane, the vehicle behind in the target lane, and the vehicle in front in the target lane, respectively. (2.5) Calculate the lane-changing risk index: The fault tree FTA evaluation method is used to calculate the comprehensive interaction risk between the master vehicle and all vehicles throughout the entire evaluation period, namely: in, The number of vehicles that directly interact with the IGV master vehicle is 3, which represents the vehicle in front of the IGV, the vehicle behind the IGV, and the vehicle in front of the intended lane.

[0007] Furthermore, the specific steps of step (3) are as follows: (3.1) Graph attention neural network extracts multi-vehicle interaction graph space features: set up For the node state of the IGV master vehicle, the attention mechanism updates the target vehicle's state based on the degree of influence of surrounding vehicles, with the following update rules: in, and These represent the state of the node at the current time and the previous time, respectively; To pay attention to the number of heads, For nodes The neighborhood group, for neighboring nodes The state; For the head Attention score, This is the weight matrix for the multi-head self-attention mechanism; The spatial features of the multi-vehicle interaction graph at any given time are represented by the final state of the target vehicle after two iterations, calculated as follows: (3.2) Gated recurrent neural network extracts temporal features of multi-vehicle interaction graph: After the spatial features of the continuous time multi-vehicle interaction graph are extracted by GAT, a spatial vector of time series is formed. The correlation of the multi-vehicle interaction graph across time series is extracted by gated recurrent unit (GRU), and an attention mechanism is used to assign weights to the aggregated features at different times. (3.2.1) Initialize weights and biases: Initialize the weights and biases of the GRU, including resetting the weight matrix of the gates. and bias matrix Update the weight matrix of the gate. and bias matrix The weight matrix of the current input and the hidden state at the previous time step. and bias ; (3.2.2) Calculate the reset gate: Reset door The calculation determines how the hidden state of the previous time step affects the candidate hidden state of the current time step, as follows: Here, sigmoid is the activation function, which compresses the input values ​​to between 0 and 1, thus determining the degree to which information passes through; Represents the hidden state at the previous moment. and the input at the current moment splicing; (3.2.3) Calculate the update gate: The update gate determines how much information from the previous hidden state needs to be retained in the current time step. It uses the sigmoid activation function to obtain a value between 0 and 1, determining the degree to which the previous hidden state is retained. The calculation is as follows: (3.2.4) Calculate the candidate hidden states: Candidate hidden state The calculation is based on the current input and the hidden state from the previous moment after the reset gate adjustment, as follows: in, represents element-level multiplication, i.e., the reset gate is multiplied by each element of the hidden state at the previous time step; tanh is the activation function used to compress the input value to between -1 and 1; (3.2.5) Update hidden status: Update the hidden state at the current moment based on the value of the update gate and the candidate hidden states. The calculation is as follows: (3.2.6) Attention mechanism weighting: Use attention mechanism for arbitrary time states of GRU Calculate weights ,in and The first thing to pay attention to is the weights and biases of the hidden layers. and The second point to note is the weights and biases of the hidden layer, calculated using the following formula: (3.3) Spatiotemporal attention graph neural network predicts lane change risk: A spatiotemporal attention graph neural network model is constructed by combining GAT, GRU, and attention mechanisms to predict lane-change risk; firstly, The multi-vehicle interaction graph at each time step is input into the GAT model, and after the attention coefficient is adjusted to balance the importance of information, the result is obtained. A hidden spatial state , i.e., spatiotemporal features { , }; Secondly, by As input, it is obtained through GRU time vectors { Then, the time vector is input into the attention model to assign weights to each time state. The weights are calculated by SOFTMAX and the fully connected network: { The attention mechanism uses a weighted summation method to calculate the spatiotemporal interaction information of the multi-vehicle interaction graph. Ultimately, the lane-changing risk indicator is output by the fully connected dense array. The calculation is as follows: .

[0008] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention utilizes real-time vehicle operation status data acquired in a connected environment to uniformly model the spatiotemporal interaction relationships of multiple vehicles during lane changing using graph theory. It accurately quantifies and predicts the comprehensive safety risks throughout the entire IGV lane changing process and, based on the predicted risks, achieves safety control for IGVs to autonomously terminate lane changes. This effectively solves the problems of traditional port area vehicle risk assessment methods, which rely heavily on static safety rules or single-vehicle status calculations, making it difficult to characterize the dynamic interaction impact of multiple vehicles. These methods suffer from delayed risk prediction, insufficient accuracy, and inability to support real-time safety decision-making. This invention provides significant theoretical support and engineering application value for IGV lane changing decision-making, multi-vehicle collaborative safety early warning, and port area horizontal transportation risk management in the context of vehicle-road cooperation and autonomous driving within ports. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the lane-changing vehicle group of the present invention; Figure 3 This is a multi-vehicle interaction diagram of the present invention; Figure 4 This is a time-multiple-vehicle interaction diagram of the present invention; Figure 5 This is a diagram illustrating the calculation of the parking distance index in this invention. Figure 6 This is a schematic diagram of the spatiotemporal attention neural network of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0011] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0012] It should be noted that the labels and letters in the following figures represent similar items, therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0013] A method for predicting lane-changing risks of IGVs in port areas based on multi-vehicle interaction graph modeling includes the following steps: Step 1: Multi-vehicle interaction diagram modeling This invention proposes a modeling method based on multi-vehicle interaction graphs, which uses graph theory to represent the relationship between the intelligent guided vehicle (IGV master vehicle, i.e., the IGV vehicle that wants to change lanes) and its surrounding vehicles.

[0014] 1. Define lane-changing vehicle groups Multi-vehicle interaction graph modeling requires defining an interaction system within the network's scope, namely, a lane-changing vehicle group, such as... Figure 2 As shown, the physical set of vehicles related to lane-changing behavior, consisting of the IGV master vehicle and its surrounding vehicles, is called the lane-changing vehicle group. Specifically, in a connected environment, the IGV master vehicle can obtain information such as the speed, acceleration, and position of surrounding vehicles through its onboard communication system. The communication distance of the IGV vehicles is defined as 150 m. That is, the set of all other vehicles that directly interact with the IGV master vehicle within a 150 m radius of the lane-changing IGV master vehicle is defined as the lane-changing vehicle group.

[0015] 2. Use graph theory to model the multi-vehicle interaction relationships in lane-changing vehicle groups. Based on the properties of graph theory in describing the topological relationships of multiple physical entities within a spatial range, this invention extends this theory to the representation of multi-vehicle interaction relationships. For example... Figure 3 As shown, To represent the graph relationships between vehicles and other physical entities, then... It can be described as the first The graph relationships at any given moment. Each vehicle or individual physical entity is considered a graph node, where... It is the first A collection of timeline nodes. , It is the first The number of vehicles at any given time. It is the first The combination of time edges represents the first time. The interaction relationship between any two physical entities at any given time. , It is the first The number of edges at any given time. It is the first Time map The adjacency matrix represents the connections between physical entities. Edges between entities that do not directly interact are denoted as 0; edges between entities that have direct interaction are denoted as 1. For any given time... The interaction relationships of lane-changing vehicle groups can all be represented by multi-vehicle interaction diagrams. express.

[0016] 3. Define the time-based multi-vehicle interaction diagram Lane-changing behavior is not instantaneous, but rather gradually established through dynamic interactions between the IGV (Integrated Vehicle Vehicle) and surrounding physical traffic entities within a certain timeframe. While multi-vehicle interaction diagrams depict the interaction states in arbitrary spatial segments of traffic entities such as vehicle-to-vehicle and vehicle-to-road interactions, they cannot depict the changes in these interaction states over time. Therefore, a temporal multi-vehicle interaction diagram is proposed to represent the spatiotemporal changes in multi-vehicle interaction relationships. For example... Figure 4As shown, the spatiotemporal relationships between multiple vehicles at consecutive times can be represented by a multi-vehicle interaction diagram with a time series, i.e., a spatiotemporal multi-vehicle interaction diagram: ( Furthermore, while adjacency matrices can characterize the interaction relationships between lane-changing vehicle groups, these relationships, in addition to the topological relationships between vehicles and physical entities, should also possess certain attributes to quantify the magnitude of the interaction impact between vehicles and between vehicles and other physical entities. Therefore, a feature matrix is ​​introduced, and this feature matrix is ​​used... Quantify the proprietary attribute information of internal nodes of the lane-changing vehicle group. .in For the number of nodes, The length of the feature time series, The number of features for each node is specified. If the number of features for nodes differs, a masking mechanism is used to fill in missing features for mathematical operations between the feature matrix and the adjacency matrix. The feature matrix includes individual features, interaction features, and other features. Specifically, as shown in Table 1, individual features include information such as the speed, acceleration, and lateral and longitudinal coordinates of each vehicle, carried by the vehicle node; interaction features include information related to the interaction between the IGV master vehicle and surrounding vehicles, including differences in speed, acceleration, and lateral and longitudinal coordinates; other features include information carried by roadside nodes or other physical entities, such as upstream and downstream traffic conditions that cannot be directly observed by the vehicle, such as lane speed and density.

[0017] Table 1 Elements of the Multi-Vehicle Interaction Graph Feature Matrix Therefore, lane change risk It can be represented as a functional relationship between a graph that changes dynamically over time and its characteristic matrix.

[0018] In the formula, For time The risk at this moment, and the diagram and its features are used arrive The data at any given moment is for the purpose of using... arrive Feature prediction at time point The purpose of this study is to predict lane change risks by establishing a mathematical model if we have sufficient historical data on IGV and the form of surrounding vehicles.

[0019] Step 2: Lane Change Risk Calculation 1. Calculate parking sight distance This invention uses a lane-changing risk index This describes the comprehensive interaction risk between the IGV (Integrated Vehicle Vehicle) and surrounding vehicles during lane changes. Calculation First, you need to calculate the Safe Stopping Distance (SSD) and the Stopping Distance Index (SDI). SSD is the distance a vehicle travels from its current driving speed to a complete stop during emergency braking.

[0020] In the formula, the middle and These represent the safe stopping distances of the vehicle in front and behind the IGV, respectively. and These are the speeds of the vehicle in front and the vehicle behind before they braked. The IGV braking reaction time is taken as 0.3s. and This is the acceleration of the car in front and the car behind, taken as 3.4 m / s². 2 .

[0021] 2. Calculate the parking distance index SDI is a binary unit used to measure whether a following vehicle collided with the preceding vehicle after the preceding vehicle braked. For example... Figure 5 If the braking distance of the following vehicle exceeds the braking distance of the preceding vehicle, the length of the preceding vehicle ( ) and the distance between the two vehicles before braking ( If the two vehicles collide, then a collision will occur.

[0022] 3. Calculate the risk exposure level and severity level. Lane change risk exposure level (REL) refers to the ratio of the number of times the SDI equals 1 during the lane change time to the total assessment time. The higher the value, the greater the probability of a collision during the assessment time. Lane change risk severity level (RSL) refers to the ratio of the maximum SDI that can be reached during the lane change time to the theoretical extreme value of SDI. This value reflects the intensity of the collision.

[0023] In the formula, It's time to change lanes. Take 0.1s; ; yes The value in extreme cases, i.e., the result of the vehicle in front braking to a stop without any stopping distance. When SDI equals 1 during the evaluation process The maximum value.

[0024] 4. Calculate the interaction risk between any two vehicles. First, calculate the interaction risk between the IGV master vehicle and any adjacent vehicle. This risk is obtained by multiplying REL and RSL, which can be used to assess the interaction risk between any two vehicles during the assessment time. In the formula, For any two vehicles, the interaction risk is... These represent the interaction risks between the IGV driver and the vehicle in front in the original lane, the vehicle behind in the target lane, and the vehicle in front in the target lane, respectively.

[0025] 5. Calculate the lane-changing risk index Fault Tree Analysis (FTA) is used to calculate the comprehensive interaction risk between the primary vehicle and all other vehicles throughout the entire assessment period. When constructing the fault tree, the undesirable outcome (lane change failure) is first identified and placed at the top of the fault tree as the root event. Then, the various possible causes of this outcome (arbitrary two-vehicle collision) are analyzed layer by layer, and these causes are designated as fault events for the next level. In each branch of the fault tree, the upper-level fault event is the result of the lower-level fault event, while the lower-level event is the cause of the upper-level fault event. Through fault tree analysis, the lane change risk index is calculated.

[0026] In the formula, The number of vehicles that directly interact with the IGV master vehicle is 3, which represents the vehicle in front of the IGV, the vehicle behind the IGV, and the vehicle in front of the intended lane.

[0027] Step 3: Lane change risk prediction based on multi-vehicle interaction graph modeling Multi-vehicle interaction graphs are a data structure with graph and graph features. It is difficult for general deep learning methods to use this data structure as input. This invention uses graph attention neural networks to extract spatial features of multi-vehicle interaction graphs and gated recurrent neural networks to extract temporal features of temporal multi-vehicle interaction graphs, ultimately achieving the effect of predicting lane change risks.

[0028] 1. Graph Attention Neural Network Extracts Multi-Vehicle Interaction Graph Space Features The basic principle of a Graph Attention Neural Network (GAT) is to use hidden self-attention layers to aggregate information from nodes in a graph. It can not only recognize graph structures and extract spatial vectors from multi-vehicle interaction graphs, but also assign weights to features of surrounding vehicles based on their impact on the safety of the IGV (Integrated Vehicle Vehicle). Assuming... For the node state of the IGV master vehicle, the attention mechanism will update the target vehicle's state based on the degree of influence of surrounding vehicles, with the following update rules: in, and These represent the state of the node at the current moment and the previous moment, respectively. To pay attention to the number of heads, For nodes The neighborhood group, for neighboring nodes The state. For the head Attention score, This is the weight matrix for the multi-head self-attention mechanism.

[0029] The spatial features of the multi-vehicle interaction graph at any given time can be represented by the final state of the target vehicle after two iterations, calculated as follows: 2. Gated recurrent neural network for extracting temporal features of multi-vehicle interaction graphs After spatial feature extraction from the continuous-time multi-vehicle interaction graphs using GAT (Gate-Attribution Array), a spatial vector for the time series is formed. A Gated Recurrent Unit (GRU) is used to extract the correlations of the multi-vehicle interaction graphs across time series. Furthermore, to focus on important moments in the time series, an attention mechanism is used to weight the aggregated features at different moments.

[0030] The computation process of GRU with integrated attention mechanism is as follows: (1) Initialize weights and biases Before starting the calculation, the weights and biases of the GRU need to be initialized. This includes resetting the weight matrix of the gates ( ) and bias matrix ( ), update the gate weight matrix ( ) and bias matrix ( The weight matrix of the current input and the hidden state at the previous time step ( ), and bias ( ).

[0031] (2) Calculate the reset door Reset door ( This determines how the hidden state of the previous time step affects the candidate hidden state of the current time step, calculated as follows: The sigmoid function is an activation function that compresses input values ​​to between 0 and 1, thus determining the degree to which information passes through. Represents the hidden state at the previous moment. and the input at the current moment splicing.

[0032] (3) Calculate the update gate The update gate determines how much information from the hidden state of the previous time step needs to be retained in the current time step. Similar to the reset gate, the update gate also uses the sigmoid activation function to obtain a value between 0 and 1, which determines the degree to which the hidden state of the previous time step is retained. The calculation is as follows: (4) Calculate the candidate hidden state Candidate hidden state ( The calculation is based on the current input and the hidden state of the previous time step after the reset gate adjustment, and is as follows: In the formula, This represents element-wise multiplication, where the reset gate is multiplied by each element of the previous hidden state. `tanh` is the activation function, which compresses the input values ​​to between -1 and 1.

[0033] (5) Update hidden status Finally, update the hidden state at the current moment based on the value of the updated gate and the candidate hidden states. The calculation is as follows: (6) Attention mechanism empowerment Use attention mechanism for arbitrary time states of GRU Calculate weights ,in and The first thing to pay attention to is the weights and biases of the hidden layers. and The second thing to pay attention to is the weights and biases of the hidden layers.

[0034] 3. Spatiotemporal attention graph neural network for predicting lane change risk like Figure 6 This paper combines GAT, GRU, and attention mechanisms to construct a Spatial-temporal Attention Graph Neural Network (STAG). First, it combines... The multi-vehicle interaction graph at each time step is input into the GAT model, and after the attention coefficient is adjusted to balance the importance of information, the result is obtained. A hidden spatial state (Graph attention), i.e., spatiotemporal features { , Secondly, by As input, it is obtained through GRU time vectors { Then, the time vectors are input into the attention model to assign weights to each time state. Each The weights are calculated by SOFTMAX and the fully connected network: { The attention mechanism uses a weighted summation method to calculate the spatiotemporal interaction information of the multi-vehicle interaction graph. Ultimately, the lane-changing risk indicator is output by the fully connected dense array. The overall calculation for STAG is as follows: Step 4: Termination of lane-changing application by the IGV master vehicle based on lane-changing risk prediction Lane change risk index For lanes ∈ [0,1], the higher the risk index, the greater the risk of lane changing. When the risk index exceeds a certain threshold... This indicates that continuing to change lanes would result in a collision. Therefore, when the IGV driver performs a lane change, if the LCRI calculated in real time by the onboard system exceeds a certain threshold, the IGV autonomous driving system must terminate the lane change and maintain a straight-line posture. in, This indicates the conditions under which the IGV (Incoming Traffic Vehicle) terminates lane changing.

[0035] The above description constitutes an embodiment of the present invention. The foregoing descriptions are preferred embodiments of the present invention. Unless there is a clear contradiction or a prerequisite for a particular preferred embodiment, the preferred embodiments can be arbitrarily combined and used. The embodiments and specific parameters described are merely for clearly illustrating the verification process of the invention and are not intended to limit the scope of patent protection of the present invention. The scope of patent protection of the present invention is still determined by its claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention should also be included within the scope of protection of the present invention.

Claims

1. A method for predicting lane-changing risks of IGVs in port areas based on multi-vehicle interaction graph modeling, characterized in that, Includes the following steps: (1) Multi-vehicle interaction diagram modeling; (1.1) Define lane-changing car groups: In a connected environment, the IGV master vehicle obtains the speed, acceleration, and location information of surrounding vehicles through the vehicle communication system. The communication distance of the IGV vehicle is defined as 150 m. The set of all other vehicles that directly interact with the IGV master vehicle within a 150 m radius of the lane-changing IGV master vehicle is defined as the lane-changing vehicle group. (1.2) Using graph theory to model the multi-vehicle interaction relationships in lane-changing vehicle groups: Representing multi-vehicle interaction relationships using graph theory: in, To represent the graph relationships between vehicles and other physical entities, then... Described as the first The graph relationships at different times; each vehicle or individual physical entity is considered as a graph node, where... It is the first A collection of timeline nodes. , It is the first The number of vehicles at any given time; It is the first The combination of time edges represents the first time. The interaction relationship between any two physical entities at any given moment; , It is the first The number of edges at any given time; It is the first Time map The adjacency matrix represents the connections between physical entities. Edges that do not directly generate interactive activities are denoted as 0, while those with direct interactive relationships are denoted as 1. Thus, for any given time... The interaction relationship of lane-changing vehicle groups is realized using a multi-vehicle interaction diagram. express; (1.3) Define the spatiotemporal multi-vehicle interaction diagram: Define a spatiotemporal multi-vehicle interaction graph, ( Characterize the spatiotemporal changes of multi-vehicle interaction relationships; introduce feature matrices and use feature matrices. Quantify the proprietary attribute information of internal nodes of the lane-changing vehicle group. ;in For the number of nodes, The length of the feature time series, The feature matrix for each node includes individual features, interaction features, and other features. Individual features include the speed, acceleration, and lateral and longitudinal coordinate information of each vehicle, carried by the vehicle node. Interaction features include information related to the interaction between the IGV master vehicle and surrounding vehicles, including differences in speed, acceleration, and lateral and longitudinal coordinates. Other features are information carried by roadside nodes and other physical entities, including lane speed and density. Therefore, changing lanes carries risks. This can be represented as a functional relationship between a graph that dynamically changes over time and its characteristic matrix. ； in, For time The risk at this moment, and the use of the diagram and its features arrive Real-time data, to achieve... arrive Feature prediction at time point The risk of time; (2) Lane change risk calculation: Adopting the lane change risk index Describe the comprehensive interaction risk between the IGV (In-Gear Vehicle) driver and surrounding vehicles during lane changes; and calculate the safe stopping distance (SSD) and the stopping distance index (SDI). (3) Lane change risk prediction based on multi-vehicle interaction graph modeling: The spatial features of the multi-vehicle interaction graph are extracted using a graph attention neural network (GAT), and the temporal features of the temporal multi-vehicle interaction graph are extracted using a gated recurrent neural network (GRN), thereby enabling lane change risk prediction. (4) Application of IGV driver lane change termination based on lane change risk prediction: Lane change risk index For lanes ∈ [0,1], the higher the risk index, the greater the risk of lane changing. When the risk index exceeds the threshold... This indicates that a collision would occur if the lane change continued. Therefore, when the IGV driver was performing a lane change, the risk index calculated in real time by the onboard system exceeded the threshold. The IGV autonomous driving system immediately terminates the lane change and maintains a straight-line driving posture, as shown in the following formula: ； in, This indicates the conditions under which the IGV (Incoming Traffic Vehicle) terminates lane changing.

2. The port area IGV lane-changing risk prediction method based on multi-vehicle interaction graph modeling according to claim 1, characterized in that, Step (2) consists of the following steps: (2.1) Calculate the parking sight distance: The SSD calculation formula is: ； ； in, and These represent the safe stopping distances of the vehicle in front and behind the IGV, respectively. and These are the speeds of the vehicle in front and the vehicle behind before they braked; The IGV braking reaction time is set to 0.3s; and This is the acceleration of the car in front and the car behind, taken as 3.4 m / s². 2 ; (2.2) Calculate the parking distance index: SDI is used to measure whether a following vehicle has collided with the preceding vehicle after the preceding vehicle has braked. The calculation formula is as follows: ； The length of the front vehicle is The distance between the two vehicles before braking ;when , , The sum is greater than If the conditions are met, no collision will occur; otherwise, a collision will occur. (2.3) Calculate the risk exposure level and severity level: Lane change risk exposure level (REL) represents the ratio of the number of times the SDI equals 1 during the lane change time to the total assessment time. The higher the value, the greater the probability of a collision during the assessment time. Lane change risk severity level (RSL) represents the ratio of the maximum SDI that can be reached during the lane change time to the theoretical extreme value of SDI, reflecting the intensity of the collision. The calculation formula is as follows: ； ； in, It's time to change lanes. Take 0.1s; ; yes The value in extreme cases, i.e., the result of the vehicle in front braking to a stop without any stopping distance. When SDI equals 1 during the evaluation process The maximum value; (2.4) Calculate the interaction risk between any two vehicles: The interaction risk between the two vehicles is obtained by multiplying REL and RSL, that is: ； in, For any two vehicles, the interaction risk is... These represent the interaction risks between the IGV driver and the vehicle in front in the original lane, the vehicle behind in the target lane, and the vehicle in front in the target lane, respectively. (2.5) Calculate the lane-changing risk index: The fault tree FTA evaluation method is used to calculate the comprehensive interaction risk between the master vehicle and all vehicles throughout the entire evaluation period, namely: ； in, The number of vehicles that directly interact with the IGV master vehicle is 3, which represents the vehicle in front of the IGV, the vehicle behind the IGV, and the vehicle in front of the intended lane.

3. The port area IGV lane-changing risk prediction method based on multi-vehicle interaction graph modeling as described in claim 1, characterized in that, Step (3) The specific steps are as follows: (3.1) Graph attention neural network extracts multi-vehicle interaction graph space features: set up For the node state of the IGV master vehicle, the attention mechanism updates the target vehicle's state based on the degree of influence of surrounding vehicles, with the following update rules: ； in, and These represent the state of the node at the current time and the previous time, respectively; To pay attention to the number of heads, For nodes The neighborhood group, for neighboring nodes The state; For the head Attention score, This is the weight matrix for the multi-head self-attention mechanism; The spatial features of the multi-vehicle interaction graph at any given time are represented by the final state of the target vehicle after two iterations, calculated as follows: ； (3.2) Gated recurrent neural network extracts temporal features of multi-vehicle interaction graph: After the spatial features of the continuous time multi-vehicle interaction graph are extracted by GAT, a spatial vector of time series is formed. The correlation of the multi-vehicle interaction graph across time series is extracted by gated recurrent unit (GRU), and an attention mechanism is used to assign weights to the aggregated features at different times. (3.2.1) Initialize weights and biases: Initialize the weights and biases of the GRU, including resetting the weight matrix of the gates. and bias matrix Update the weight matrix of the gate. and bias matrix The weight matrix of the current input and the hidden state at the previous time step. and bias ; (3.2.2) Calculate the reset gate: Reset door The calculation determines how the hidden state of the previous time step affects the candidate hidden state of the current time step, as follows: ； Here, sigmoid is the activation function, which compresses the input values ​​to between 0 and 1, thus determining the degree to which information passes through; Represents the hidden state at the previous moment. and the input at the current moment splicing; (3.2.3) Calculate the update gate: The update gate determines how much information from the previous hidden state needs to be retained in the current time step. It uses the sigmoid activation function to obtain a value between 0 and 1, determining the degree to which the previous hidden state is retained. The calculation is as follows: ； (3.2.4) Calculate the candidate hidden states: Candidate hidden state The calculation is based on the current input and the hidden state from the previous moment after the reset gate adjustment, as follows: ； in, represents element-level multiplication, i.e., the reset gate is multiplied by each element of the hidden state at the previous time step; tanh is the activation function used to compress the input value to between -1 and 1; (3.2.5) Update hidden status: Update the hidden state at the current moment based on the value of the update gate and the candidate hidden states. The calculation is as follows: ； (3.2.6) Attention mechanism weighting: Use attention mechanism for arbitrary time states of GRU Calculate weights ,in and The first thing to pay attention to is the weights and biases of the hidden layers. and The second point to note is the weights and biases of the hidden layer, calculated using the following formula: ； ； (3.3) Spatiotemporal attention graph neural network predicts lane change risk: A spatiotemporal attention graph neural network model is constructed by combining GAT, GRU, and attention mechanisms to predict lane-change risk; firstly, The multi-vehicle interaction graph at each time step is input into the GAT model, and after the attention coefficient is adjusted to balance the importance of information, the result is obtained. A hidden spatial state , i.e., spatiotemporal features { , }; Secondly, by As input, it is obtained through GRU time vectors { Then, the time vector is input into the attention model to assign weights to each time state. The weights are calculated by SOFTMAX and the fully connected network: { The attention mechanism uses a weighted summation method to calculate the spatiotemporal interaction information of the multi-vehicle interaction graph. Ultimately, the lane-changing risk indicator is output by the fully connected dense array. The calculation is as follows: ； ； ； ； 。

Images