Gap game perception takeover decision method for highway cut-in scene

Abstract

The present application relates to a kind of gap game perception takeover decision method for highway cut-in scene, the method first identifies target lane three car body (Leader-TV-Follower) and constructs forward constraint, backward constraint and gap constraint, then data is processed by causal multi-head self-attention module, cross-flow attention module of causal constraint, obtain the gap sequence representation of fusion forward, backward risk, the importance of each time step data is evaluated, normalized, obtain time importance weight;Weighted aggregation is carried out, and global gap game representation is obtained, to build three car body risk game intensity index;Comprehensive safety margin evaluation is built and matched with preset threshold rule, and adaptive output grading takeover decision result, including no takeover, prompt takeover and forced takeover etc., to improve the accuracy and reliability of takeover decision in the high risk scene of highway cut-in.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicles and autonomous driving safety control technology, and relates to takeover decision and risk assessment under human-machine co-driving conditions in high-risk scenarios of highway cut-in. Specifically, it relates to a gap game perception takeover decision method for highway cut-in scenarios. Background Technology

[0002] With the development of intelligent connected vehicles and autonomous driving technology, L2 / L3 level systems have been applied in highway scenarios. However, under functional boundary conditions or sudden high-risk situations, it is still necessary to transfer control to the driver through takeover requests. Especially in highway cut-in scenarios, when a lane-changing vehicle forcibly inserts itself between the vehicle in front and the vehicle behind in a short period of time, the forward safety distance and rear-end collision margin are compressed simultaneously. The vehicle in front (Leader), the target vehicle (TV), and the vehicle behind (Follower) in the same lane form a strong coupling interaction around the limited available gap, significantly increasing the collision risk. Existing takeover triggering methods mostly rely on single or a few thresholds such as vehicle distance, relative speed, and time to collision (TTC), simplifying the scenario to a two-vehicle relationship of "the vehicle in front – the vehicle behind". They lack the characterization of the game relationship of the three vehicle bodies (Leader, TV, and Follower) under limited gap conditions, making it difficult to reflect the joint evolution process of forward risk, rear risk, and lane available gap. This can easily lead to problems such as insensitivity to high-risk cut-in situations or unreasonable takeover triggering timing. Therefore, it is necessary to propose a gap game perception takeover decision-making method for highway lane-cutting scenarios, which structurally models the risks and gap changes from the perspective of three vehicles to improve the accuracy and safety of takeover decisions. Summary of the Invention

[0003] Given that existing autonomous driving takeover decision-making methods mostly rely on single or limited thresholds such as vehicle distance, relative speed, and time-to-collision (TTC), they struggle to characterize the strong coupling and interaction relationships formed by the leading vehicle, target vehicle (TV), and following vehicle in the same lane within a limited available gap in highway cut-in scenarios. This leads to problems such as insensitivity to high-risk cut-in situations, unreasonable takeover trigger timing, and unstable decision output. Therefore, this invention proposes a gap-game perception-based takeover decision-making method for highway cut-in scenarios. This method first identifies the three vehicle bodies in the target lane (…). The system constructs a Leader-TV-Follower (TLD) framework and establishes risk-gap characteristics for forward constraints, backward constraints, and available gaps. Furthermore, it establishes a three-vehicle gap game state representation based on variables such as relative distance, relative speed, TTC, and gap change rate, thereby jointly reflecting the evolution trend of forward risk, backward risk, and gap resources over time. Based on this, a comprehensive safety margin assessment is constructed and matched with preset threshold rules to adaptively output graded takeover decision results, including no takeover required, prompted takeover, and forced takeover, to improve the accuracy and reliability of takeover decisions in high-risk cut-in scenarios on highways.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A gap-judgment game-theoretic takeover decision-making method for highway lane-cutting scenarios, comprising the following steps:

[0006] Step 1. Obtain the longitudinal distance, longitudinal relative speed, collision time, and first-order difference of TTC between the target vehicle and the vehicle in front in the same lane; the longitudinal distance, longitudinal relative speed, collision time, and first-order difference of TTC between the target vehicle and the vehicle behind in the same lane; and the longitudinal distance, relative speed, collision time, and rate of change of clearance between the vehicle in front in the same lane and the vehicle behind in the same lane.

[0007] Step 2. Construct forward constraints, backward constraints, and gap constraints based on the data obtained in Step 1;

[0008] Step 3. First, perform feature transformation on the forward constraint, backward constraint, and gap constraint, and superimpose sine-cosine positional encoding to obtain the initial hidden state. The initial hidden state is then processed by an L-layer causal multi-head self-attention module to output forward constraint features, backward constraint features, and gap constraint features. Then, a cross-current attention module based on forward causal constraints processes the forward constraint features and gap constraint features to obtain a gap sequence representation that integrates forward risk. Similarly, a cross-current attention module based on backward causal constraints processes the gap sequence representation that integrates forward risk and backward constraint features to obtain a gap sequence representation that integrates forward and backward risks. The data at each time step is evaluated for importance and normalized to obtain time importance weights. Weighted aggregation is then performed to obtain the global gap game representation. This is used to construct a risk game intensity index for three vehicle bodies. ;

[0009] Step 4. Based on the three-vehicle risk game intensity index, construct a historical comprehensive safety margin index for forward and rearward collision times. Based on the historical comprehensive safety margin index, predict the future comprehensive safety margin and compare it with the safety threshold to output the takeover decision.

[0010] As a preferred embodiment of the present invention, forward constraint The expression is:

[0011] ;

[0012] ;

[0013] Backward constraints The expression is:

[0014] ;

[0015] ;

[0016] Gap constraints The expression is:

[0017] ;

[0018] ;

[0019] In the formula, Let represent the longitudinal distance, longitudinal relative speed, collision time, and first-order difference of TTC between the target vehicle and the vehicle in front and the vehicle behind in the same lane at time t, respectively. This represents the longitudinal distance, relative speed, and collision time of the available clearance between vehicles in front and behind in the same lane at time t. This represents the rate of change in the clearance distance between vehicles in front and behind in the same lane at time t. , , They represent Forward constraints, backward constraints, and gap constraints at any given time. , For historical time steps.

[0020] As a preferred embodiment of the present invention, step 3 uses a single-layer linear mapping network to perform feature transformation on the forward constraint, backward constraint and gap constraint.

[0021] As a preferred embodiment of the present invention, in step 3, four layers of causal multi-head self-attention modules are superimposed in the time dimension. The l-th layer of causal multi-head self-attention module calculates the query, key, and value matrices of forward constraints, backward constraints, and gap constraints according to the input data, and introduces a causal mask matrix to calculate the masked attention weights. Finally, the attention output is obtained based on the masked attention weights. Then, the attention output and input are fused with the layer normalization through residual connections and fed into the feedforward network to achieve nonlinear transformation. After a second fusion through residual connections and layer normalization, forward constraint features, backward constraint features, and gap constraint features are obtained.

[0022] As a preferred embodiment of the present invention, in step 3, the cross-current attention module for forward causal constraints processes the forward constraint features and gap constraint features by using the current gap constraint feature as the Query and the entire forward constraint feature segment as the Key and Value to construct the cross-current attention output, thereby obtaining the risk-sensitive key at time t. Then, a one-dimensional gating mechanism is introduced to adaptively adjust the influence of the forward risk at the current time on the current gap state, thus obtaining the gap representation with fused forward risk at the current time t. The expression is:

[0023] ;

[0024] ;

[0025] In the formula, The currently available gap, i.e., the current gap constraint feature. For the Sigmoid function, This represents element-wise multiplication. The gating coefficient, This represents the high-dimensional representation of the gap after incorporating forward risks at time step t.

[0026] Preferably, in step 3, the cross-current attention module for backward causality constraints has the same structure as the cross-current attention module for forward causality constraints. The cross-current attention module for backward causality constraints is used to process the gap sequence representation fused with forward risk and the backward constraint features, thereby fusing the high-dimensional representation of the gap with forward risk. As a query, the entire backward constraint feature is used as the key and value.

[0027] As a preferred embodiment of the present invention, the historical comprehensive safety margin index at time t in step 4 is... for:

[0028] ;

[0029] ;

[0030] In the formula, Set them to 0.2, 0.2, 0.3, and 0.3 respectively. The modulation factor is set to 0.4; These are the normalized risk game intensity indicators, and the forward and backward collision times, respectively. This is a risk interaction item.

[0031] As a preferred embodiment of the present invention, step 4 inputs the obtained historical comprehensive safety margin index into the long short-term memory network, and obtains the features through time-series modeling. Then, the MLP is used to predict the future comprehensive safety margin. The MLP consists of two fully connected layers and a ReLU activation function.

[0032] As a preferred embodiment of the present invention, the safety threshold in step 4 is set to The overall safety margin at time t in the future When the system is deemed to be in a stable and safe state, no intervention is required; continue automatic driving control. When the situation is deemed a risk warning, a takeover alert is issued, requesting the driver to prepare for takeover; when At that time, it was determined to be a high-risk situation and was forcibly taken over.

[0033] As a preferred embodiment of the present invention, the attention weights with masks are:

[0034] ;

[0035] ;

[0036] ;

[0037] In the formula, These are the attention weights with masks for the forward constraint, backward constraint, and gap constraint of the l-th layer. These are the query matrices for the forward constraint, backward constraint, and gap constraint at the l-th layer, respectively. These are the forward constraint, backward constraint, and gap constraint key matrices for the l-th layer, respectively. Indicates the embedding dimension. Represents the transpose sign, mask matrix satisfy:

[0038] ;

[0039] in, Represents the mask matrix The element in the i-th row and j-th column, i.e., the i-th time step, can only focus on the current moment and its historical moments. , For historical time steps.

[0040] Advantages and beneficial effects of the present invention:

[0041] (1) This invention starts from the perspective of three-vehicle structure and models the competitive relationship between the vehicle in front (Leader), the target vehicle (TV) and the vehicle behind (Follower) in the same lane around the limited available gap as a "gap game structure". It breaks through the traditional takeover decision method based only on all vehicles in the surrounding environment and realizes the systematic characterization of the joint evolution of forward risk, backward risk and gap resources in the cut-in scenario. It improves the integrity and physical consistency of takeover decision from the structural level.

[0042] (2) This invention constructs three types of temporal features: forward constraint flow, backward constraint flow and gap flow, and uses a causal consistent multi-layer self-attention mechanism for historical encoding, so that the model can explicitly learn the trend information of risk evolution over time, avoid the problem of ignoring risk accumulation and accelerated deterioration in traditional instantaneous threshold judgment, thereby enhancing the forward-looking identification ability of high-speed cut-in sudden risks.

[0043] (3) This invention introduces a "risk injection-gated fusion" cross-flow modeling mechanism. Taking the current gap state as the core, forward and backward risks are injected into the gap representation through a cross-flow attention structure, and the influence intensity of the risks is adaptively adjusted through a gating function to achieve causal driving modeling of risks on gap resources. This structure can explicitly characterize the coupling effect of "risk compressing gaps and gaps amplifying risks", giving the takeover decision a clear game theory physics explanation.

[0044] (4) This invention proposes a safety margin index, which integrates the risk game intensity with the forward and backward time-to-collision through risk normalization, nonlinear mapping and risk-gap coupling enhancement term, and satisfies monotonicity constraints, so that the safety margin simultaneously reflects the competition intensity of the three vehicles, the forward collision risk, the rear-end collision risk and the risk evolution trend, and utilizes historical safety margins. Predicting future overall safety margin This metric provides higher sensitivity and interpretability compared to the simple TTC threshold method, enabling safety alerts in cut-in scenarios.

[0045] (5) The present invention constructs a hierarchical takeover state machine model, which maps continuous risk indicators to stable and safe states, risk warning states and high-risk mandatory takeover states, effectively avoiding frequent oscillations of the takeover level near the threshold, and improving the stability of system decision-making and the comfort of driver interaction.

[0046] (6) This invention mainly relies on the conventional perception modules of existing autonomous driving systems (cameras, millimeter-wave radar, lidar, positioning system and high-precision map) to achieve the goal. There is no need to add expensive sensing equipment. It can be directly integrated into the existing autonomous driving control architecture. It is suitable for typical high-risk scenarios such as highway cut-in and has good engineering implementation value and industrial application prospects. Attached Figure Description

[0047] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0048] Figure 1 A flowchart of the gap game perception takeover decision-making method for highway lane-cutting scenarios provided by the present invention;

[0049] Figure 2 This is a flowchart for single-stream causal timing coding. Detailed Implementation

[0050] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0051] like Figure 1 As shown, this embodiment provides a gap-game perception-based takeover decision-making method for highway lane-cutting scenarios. The method includes the following steps:

[0052] Step 1. Scene Recognition and Traffic Participant Acquisition:

[0053] Step 1.1. Multi-source external environment data acquisition:

[0054] The system collects external environment perception data and road structure information of vehicles during highway driving, including target detection / tracking results (target position, speed, acceleration, heading angle, etc.) output by the vehicle-mounted perception system, and lane geometry information (lane centerline, lane boundary, lane topology, etc.) output by high-precision maps or lane line perception. The external environment perception data can be acquired collaboratively by cameras, millimeter-wave radar, lidar, or V2X; the road structure information can be obtained by high-precision maps and positioning systems, or by vision / laser-based lane line extraction modules.

[0055] Step 1.2. Data Extraction:

[0056] Obtain the longitudinal distance between the target vehicle (TV) and the vehicle ahead (Leader) in the same lane: Longitudinal relative velocity Collision Time Tolerance (TTC): The first difference of TTC: The longitudinal distance between the target vehicle (TV) and the following vehicle (autonomous vehicle) in the same lane: Longitudinal relative velocity Collision Time Tolerance (TTC): The first difference of TTC: The longitudinal distance of available clearance between the Leader and Follower vehicles in the same lane. }, relative velocity Collision Time Tolerance (TTC): }, rate of change of gap distance }, For historical time steps.

[0057] Step 2. Construct relevant features for gap game:

[0058] Step 2.1. Based on the Leader, the target vehicle (TV), and the Follower in the same lane identified in Step S1, calculate the forward constraints, backward constraints, and risk-gap features related to the target lane gap resource quantity within the observation time window to form a structured input for the three-vehicle gap game state. Specifically, this can be expressed as:

[0059] Forward constraints:

[0060]

[0061]

[0062] Backward constraints:

[0063]

[0064]

[0065] Gap constraint (gap flow):

[0066]

[0067]

[0068] In the formula, Let T represent the longitudinal distance, longitudinal relative velocity, collision time TTC, and first difference of TTC between the target vehicle TV and the vehicle in front and the vehicle behind in the same lane at time t, respectively. This represents the longitudinal distance, relative speed, and collision time of the available clearance between the Leader and Follower vehicles in the same lane at time t. This represents the rate of change in the clearance distance between the Leader and Follower vehicles in the same lane at time t, used to characterize whether the target lane clearance is narrowing or widening. , , They represent Forward constraints, backward constraints, and gap constraints at any given time. , For historical time steps.

[0069] Step 3. Three-vehicle risk game:

[0070] Step 3.1. Single-stream causal timing coding:

[0071] Forward constraints respectively Backward constraints With gap constraints The specific steps for performing causal consistency historical coding are as follows:

[0072] Step 3.1.1. Apply a single-layer linear mapping network to the forward constraints. Backward constraints With gap constraints Perform feature transformation and superimpose a sine-cosine positional code to obtain the initial hidden state. , can be represented as:

[0073]

[0074]

[0075]

[0076] In the formula, This represents a linear layer network (linear mapping network), and PE represents positional encoding. , , They represent lengths of The forward, backward, and gap initial hidden state sequences, , , Represent The initial hidden state of the forward constraint, the hidden state of the backward constraint, and the hidden state of the gap constraint at each time step.

[0077] Step 3.1.2. Stack L layers of causal multi-head self-attention modules along the time dimension. Each layer only allows explicit attention to itself and information from historical moments at the current moment, thus obtaining the attention output of the l-th layer. Specifically, it can be expressed as:

[0078] Let the input of the l-th layer be:

[0079]

[0080]

[0081]

[0082] Then its query, key, and value matrices are as follows:

[0083]

[0084]

[0085]

[0086] In the formula, Let be the learnable projection matrix of the l-th layer multi-head self-attention. These are the forward constraint query matrix, backward constraint query matrix, and gap constraint query matrix for the l-th layer, respectively. These represent the forward constraint key matrix, the backward constraint key matrix, and the gap constraint key matrix of the l-th layer, respectively. These are the forward constraint value matrix, backward constraint value matrix, and gap constraint value matrix of the l-th layer, respectively.

[0087] Building upon this, a causal masking matrix is ​​introduced, and the masked attention weights are defined as follows:

[0088]

[0089]

[0090]

[0091] In the formula, These are the attention weights with masks for the forward constraint, backward constraint, and gap constraint of the l-th layer. Represents the embedding dimension, mask matrix satisfy:

[0092]

[0093] in, Represents the mask matrix The element in the i-th row and j-th column, i.e., the i-th time step, can only focus on the current moment and its historical moments. This explicitly eliminates reliance on future information. Based on the above weights, the attention output of the l-th layer (L is usually set to 4 layers) can be obtained:

[0094] =

[0095] =

[0096] =

[0097] in, This is the attention output for the forward constraints, backward constraints, and gap constraints of the l-th layer.

[0098] Step 3.1.3. Output Attention With input , , The residual connections and layer normalization are fused together and then fed into a feed-forward network (FFN) to achieve nonlinear transformation. A second fusion is then performed using residual connections and layer normalization to obtain the final result. , Specifically, it is expressed as:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] in, The forward constraint features, backward constraint features, and gap constraint features are output from the l-th layer after the initial fusion. LayerNorm represents the layer normalization. This represents the forward constraint feature, backward constraint feature, and gap constraint feature output by the l-th layer causal multi-head self-attention module. In this embodiment, the forward constraint feature is finally output after passing through 4 layers of causal multi-head self-attention modules. Backward constraint features Gap constraint features Gap constraint features , represent The gap status at any given moment (available gap).

[0106] Step 3.2. Forward risk perception coding, the specific steps are as follows:

[0107] Step 3.2.1. At each time step t, adopt a cross-flow attention structure of "single-step Query – sequence Key / Value": use the current gap state (gap constraint feature) as the Query, and the entire forward stream encoding (forward constraint feature) as the Key and Value, construct the cross-flow attention output, and obtain the risk-sensitive key at time t. , can be represented as:

[0108]

[0109] In the formula, This represents a cross-stream attention module with causal constraints. This can be viewed as: under the condition of "the forward risk history observed up to the current time t", the current available gap This is a risk-sensitive re-representation. If the Leader has experienced multiple sharp decelerations and the TTC has remained consistently low over a period of time, then the corresponding geometric spacing... This will be encoded as a more "dangerous" state; conversely, under long-term stability and abundant TTC, It will be closer to a "relatively safe" gap representation.

[0110] Step 3.2.2. Introduce a one-dimensional gating mechanism to adaptively adjust the influence of the forward risk at the current time on the current gap state, and obtain the gap representation of the fused forward risk at the current time t. Specifically, it can be expressed as:

[0111]

[0112]

[0113] In the formula, For the Sigmoid function, This represents element-wise multiplication. This represents the high-dimensional representation of the gap after incorporating the forward risk at time step t. Gating coefficients. This reflects the model's subjective judgment on "the extent to which changes in the current available gaps are driven by forward risk": when the historical Leader-TV relationship exhibits a high-risk pattern such as continuous deceleration and a significant decrease in TTC, It will be learned to be closer to 1, at which point More reliant on risk-sensitive representation This amplifies the impact of forward risk on the current gap; conversely, when gap changes are more due to the Follower's own acceleration / deceleration or other external disturbances, It will approach 0, making Mainly retain the original geometric gap state This is to avoid overemphasizing forward risks.

[0114] Step 3.2.3. Stacking the gap representations of all time steps sequentially yields the gap sequence representation of the fused forward risk. , can be represented as:

[0115]

[0116] Step 3.3. Backward Risk Perception Coding:

[0117] To further characterize the causal effect of the longitudinal behavior of the following vehicle on the available clearance, this invention... Based on this, we introduce crossflow modeling using the "backward constraint → gap flow" approach. The specific steps are as follows:

[0118] Step 3.3.1. At time step t, characterize the gaps that have been incorporated into the forward risk. As a query, it is encoded as a whole backward stream. Using these as Key and Value, a cross-stream attention module with causal constraints is constructed to obtain... , can be specifically expressed as:

[0119]

[0120] In the formula, and The structure is consistent, and it also uses a time-causal mask to ensure that at time t, only backward risk information from the historical interval [1, t] can be aggregated. Intuitively, This can be interpreted as: given the current "forward risk-geometric gap" state Under these conditions, the backward risk-driven effect is "traced back" from the acceleration and deceleration behavior of the Follower vehicle in the same lane and the TTC evolution of the TV-Follower.

[0121] Step 3.3.2. Introduce a one-dimensional gating mechanism to adaptively adjust the influence intensity of backward risk on the current gap state, and obtain the gap representation of fusing forward and backward risks at the current time t. Specifically, it can be expressed as:

[0122]

[0123]

[0124] In the formula, The sigmoid function has a gate coefficient. It describes the extent to which changes in currently available gaps are driven by backward risk. This is a high-dimensional representation feature of the gap that comprehensively considers both forward and backward risks.

[0125] Step 3.3.3. Convert the time steps... By stacking them sequentially, we can obtain the final gap sequence representation considering both forward and backward risk effects. .

[0126] Step 3.4. Game State Representation:

[0127] Step 3.4.1. Representation of the final gap sequence High-dimensional representation of each time step A nonlinear scoring function is constructed to evaluate its importance, resulting in... Specifically, it can be expressed as:

[0128]

[0129] In the formula, It is a nonlinear mapping function used to characterize the contribution of the state between each time step to the overall risk evolution.

[0130] Step 3.4.2. Obtain the time importance weights through softmax normalization. Specifically, it can be expressed as:

[0131]

[0132] In the formula, the weights It represents the relative influence of the time step t gap state in the overall risk game process.

[0133] Step 3.4.3. Based on the time weights in Step 3.4.2 By weighted aggregation of the gap sequences, a global gap game representation is obtained. , can be represented as;

[0134]

[0135] In the formula, the gap game represents Describe the overall evolution trend and stability characteristics of the available gap of the target lane under the combined effects of forward and backward risks at time t.

[0136] Step 3.5. Calculation of the intensity of the risk game among the three vehicles:

[0137] Based on global gap game representation Constructing a risk game intensity index for three vehicle groups This is used to quantify the level of interactive tension between the Leader (vehicle ahead), the target vehicle (TV), and the Follower (vehicles behind) in the same lane, surrounding a limited available gap. It can be expressed as:

[0138]

[0139] In the formula, Let be the mapping function used to project the high-dimensional gap game representation onto the risk metric space. When When the value is large, it indicates that the three vehicle bodies form a strong competitive relationship around the limited gap, and the cut-in risk increases significantly.

[0140] Step 4. Takeover Decision Generation:

[0141] After obtaining the game state representation of the three-vehicle gap and the comprehensive safety margin index, a takeover control strategy is generated based on a risk-driven hierarchical decision-making mechanism to achieve adaptive safety control in highway cut-in scenarios. The specific steps are as follows:

[0142] Step 4.1. Based on the risk game intensity index in Step 3.5 And combined with forward collision time Rearward collision time Construct a comprehensive safety margin index To obtain historical comprehensive safety margin The details are as follows:

[0143] First, the intensity of the risk game and the time of collision are normalized:

[0144]

[0145]

[0146]

[0147] In the formula, For observation window The maximum and minimum values ​​of the risk game intensity within the scope. The risk sensitivity coefficients are set to 0.4 and 0.6, and exp is an exponential function used to characterize the nonlinear relationship between TTC and collision risk.

[0148] To characterize the amplified competitive effect of the three vehicles under limited clearance resources, a risk interaction term is introduced. :

[0149]

[0150]

[0151] In the formula, Set them to 0.2, 0.2, 0.3, and 0.3 respectively. The risk characteristics at time t represent the historical overall safety margin. The modulation factor is set to 0.4; historical comprehensive safety margin. , Representing history Risk characteristics at any given moment.

[0152] Step 4.2. Based on the historical comprehensive safety margin obtained in Step 4.1 Features are obtained using LSTM time series modeling. Then, a simple MLP is used to predict the future comprehensive safety margin. :

[0153]

[0154]

[0155] in, For future time steps, the MLP consists of two fully connected layers and a ReLU activation function. For Long Short-Term Memory (LSTM) networks, Representing the future The overall safety margin at any given moment.

[0156] Step 4.3: Comprehensive security margin based on the future time t obtained in Step 4.2 A tiered takeover decision-making mechanism is constructed to map continuous risk quantification results into discrete, actionable takeover levels, thereby achieving a closed-loop safety control system of "stability—early warning—mandatory" states. Specifically, this can be described as follows:

[0157] The takeover level is defined as a set of three discrete states. :

[0158]

[0159] In the formula, To maintain a stable and safe state (no need for intervention, maintain automatic driving control); The status is a risk warning (indicating a takeover request and requesting the driver to prepare for takeover). The situation is considered high-risk (forced takeover or minimum-risk mobile control).

[0160] Safety determination: The safety threshold is set to... The takeover level is output based on the relationship between the obtained future comprehensive safety margin and the safety threshold: when At that time, it was determined to be ;when At that time, it was determined to be ;when At that time, it was determined to be .

[0161] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the gap game perception takeover decision method for highway lane-jumping scenarios described above.

[0162] The present invention also provides a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the gap game perception takeover decision-making method for highway lane-jumping scenarios described above.

[0163] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0164] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0165] The above-described specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A gap-game perception-based takeover decision-making method for highway lane-cutting scenarios, characterized in that, The method includes the following steps: Step 1. Obtain the longitudinal distance, longitudinal relative speed, collision time, and first-order difference of TTC between the target vehicle and the vehicle in front in the same lane; the longitudinal distance, longitudinal relative speed, collision time, and first-order difference of TTC between the target vehicle and the vehicle behind in the same lane; and the longitudinal distance, relative speed, collision time, and rate of change of clearance between the vehicle in front in the same lane and the vehicle behind in the same lane. Step 2. Construct forward constraints, backward constraints, and gap constraints based on the data obtained in Step 1; Step 3. First, perform feature transformation on the forward constraint, backward constraint, and gap constraint, and superimpose sine-cosine positional encoding to obtain the initial hidden state. The initial hidden state is then processed by an L-layer causal multi-head self-attention module to output forward constraint features, backward constraint features, and gap constraint features. Then, a cross-current attention module based on forward causal constraints processes the forward constraint features and gap constraint features to obtain a gap sequence representation that integrates forward risk. Similarly, a cross-current attention module based on backward causal constraints processes the gap sequence representation that integrates forward risk and backward constraint features to obtain a gap sequence representation that integrates forward and backward risks. The data at each time step is evaluated for importance and normalized to obtain time importance weights. Weighted aggregation is then performed to obtain the global gap game representation. This is used to construct a risk game intensity index for three vehicle bodies. ; In step 3, four layers of causal multi-head self-attention modules are superimposed in the time dimension. The l-th layer of causal multi-head self-attention module calculates the query, key, and value matrices of forward constraints, backward constraints, and gap constraints according to the input data, and introduces a causal mask matrix to calculate the masked attention weights. Finally, the attention output is obtained based on the masked attention weights. Then, the attention output and input are fused with the layer normalization through residual connections and fed into the feedforward network to achieve nonlinear transformation. After a second fusion through residual connections and layer normalization, the forward constraint features, backward constraint features, and gap constraint features are obtained. In step 3, the cross-current attention module for forward causal constraints processes the forward constraint features and gap constraint features by using the current gap constraint feature as the Query and the entire forward constraint feature segment as the Key and Value to construct the cross-current attention output, thus obtaining the risk-sensitive weight at time t. Then, a one-dimensional gating mechanism is introduced to adaptively adjust the influence of the forward risk at the current time on the current gap state, thus obtaining the gap representation with fused forward risk at the current time t. ; Step 4. Based on the three-vehicle risk game intensity index, construct a historical comprehensive safety margin index for forward and rearward collision times. Based on the historical comprehensive safety margin index, predict the future comprehensive safety margin and compare it with the safety threshold to output the takeover decision.

2. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 1, characterized in that, Forward constraints The expression is: ； ； Backward constraints The expression is: ； ； Gap constraints The expression is: ； ； In the formula, Let represent the longitudinal distance, longitudinal relative speed, collision time, and first-order difference of TTC between the target vehicle and the vehicle in front and the vehicle behind in the same lane at time t, respectively. This represents the longitudinal distance, relative speed, and collision time of the available clearance between vehicles in front and behind in the same lane at time t. This represents the rate of change in the clearance distance between vehicles in front and behind in the same lane at time t. , , They represent Forward constraints, backward constraints, and gap constraints at any given time. , This is a step in historical time.

3. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 2, characterized in that, In step 3, a linear mapping network is used to perform feature transformation on the forward constraint, backward constraint and gap constraint.

4. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 3, characterized in that, Gap representation of forward risk at current time t The expression is: ； ； In the formula, The currently available gap, i.e., the current gap constraint feature. For the Sigmoid function, This represents element-wise multiplication. The gating coefficient, This represents the high-dimensional representation of the gap after incorporating forward risks at time step t.

5. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 4, characterized in that, The cross-current attention module for backward causality constraints described in step 3 has the same structure as the cross-current attention module for forward causality constraints. The cross-current attention module for backward causality constraints is used to process the gap sequence representation fused with forward risk and the backward constraint features, and to transform the high-dimensional gap representation fused with forward risk into a higher-dimensional representation. As a query, the entire backward constraint feature is used as the key and value.

6. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 5, characterized in that, The historical comprehensive safety margin index at time t in step 4 for: ； ； In the formula, Set them to 0.2, 0.2, 0.3, and 0.3 respectively. The modulation factor is set to 0.4; These are the normalized risk game intensity indicators, and the forward and backward collision times, respectively. This is a risk interaction item.

7. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 6, characterized in that, Step 4 involves inputting the obtained historical comprehensive safety margin index into the Long Short-Term Memory network for time-series modeling to obtain features. Then, the MLP is used to predict the future comprehensive safety margin. The MLP consists of two fully connected layers and a ReLU activation function.

8. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 7, characterized in that, In step 4, the safety threshold is set to... The overall safety margin at time t in the future When the system is deemed to be in a stable and safe state, no intervention is required; continue automatic driving control. When the situation is deemed a risk warning, a takeover alert is issued, and the driver is requested to prepare to take over. when At that time, it was determined to be a high-risk situation and was forcibly taken over.

9. The gap-game perception-based takeover decision-making method for highway lane-cutting scenarios according to claim 8, characterized in that, The attention weights with masks are: ； ； ； In the formula, These are the attention weights with masks for the forward constraint, backward constraint, and gap constraint of the l-th layer. These are the query matrices for the forward constraint, backward constraint, and gap constraint at the l-th layer, respectively. These are the forward constraint, backward constraint, and gap constraint key matrices for the l-th layer, respectively. Indicates the embedding dimension. Represents the transpose sign, mask matrix satisfy: ； in, Represents the mask matrix The element in the i-th row and j-th column, i.e., the i-th time step, can only focus on the current moment and its historical moments. , This is a step in historical time.

Images