Driving decision-making method and system based on surrounding vehicle driving behavior tendency

Abstract

The invention relates to the technical field of automatic driving decision control, and discloses a driving decision method and system based on surrounding vehicle driving behavior tendentiousness, and the method comprises the steps: constructing a driving behavior causal graph structure, and separating the internal intention of surrounding vehicle drivers from the environmental constraint influence through anti-factual reasoning; establishing a hierarchical driving intention graph structure; constructing an intention competition dynamical model, and describing the mutual influence and dynamic evolution process among a plurality of driving intentions; learning parameters of the intention competition dynamical model by using a recurrent neural differential equation; constructing an intention evolution critical point detection algorithm, predicting a surrounding vehicle intention transition point, and generating a main vehicle defensive driving decision; according to the invention, the behavior change of surrounding vehicles can be predicted in advance, and the safety and stability of an automatic driving system in a complex traffic environment are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of automatic driving decision-making and control technology, and more specifically, to a driving decision-making method and system based on the driving behavior tendency of a surrounding vehicle. Background Art

[0002] With the rapid development of autonomous driving technology, accurately predicting the behavior of surrounding vehicles (surrounding vehicles) has become a key challenge to ensure the safety and efficiency of autonomous driving. In complex and changing traffic environments, the driving behavior of surrounding vehicles is often influenced by both the driver's internal intentions and external environmental constraints, showing diversity and uncertainty.

[0003] Existing vehicle behavior prediction technologies primarily rely on simple extrapolation based on observed historical trajectories and current states, using machine learning or rule-based models. These methods typically treat the surrounding vehicle as a physical entity following a deterministic motion pattern, ignoring the variability and diversity of the driver's inherent intentions. For example, when a vehicle slows down, traditional methods struggle to distinguish whether the driver is actively attempting to stop or being forced to do so by traffic ahead.

[0004] The existing technology has three major technical defects: it is difficult to effectively distinguish the true inner intentions of the driver of the surrounding vehicle from the performance behavior under environmental constraints, resulting in a significant decrease in prediction accuracy when environmental conditions change; it is unable to effectively handle complex situations where the driver has multiple conflicting driving intentions at the same time, such as the situation where the driver wants to overtake but is worried about safety risks and is hesitant; it lacks the ability to accurately predict the dynamic evolution process of multiple intentions in time series, and is unable to capture the evolution laws of intentions in long time series, resulting in a lag in the prediction of the turning point of the surrounding vehicle's behavior, which seriously affects the safety of the autonomous driving system. Summary of the Invention

[0005] The present invention provides a driving decision-making method and system based on the driving behavior tendency of the surrounding vehicle, which solves the technical problems in related technologies such as the difficulty in distinguishing the true internal intention of the surrounding vehicle driver from the influence of environmental constraints, the inability to handle the coexistence of multiple intentions, and the lack of the ability to predict the dynamic evolution of intentions.

[0006] The present invention provides a driving decision-making method based on the driving behavior tendency of the surrounding vehicle, comprising the following steps:

[0007] Obtain vehicle trajectory data and surrounding environment information through on-board sensors, including position, speed, acceleration and other time series data;

[0008] Based on the acquired data, the driving behavior causal graph structure is used to implement counterfactual reasoning, effectively distinguishing the influence of the driver's internal intention and external environmental constraints.

[0009] Based on the identified intrinsic intentions, a hierarchical driving intention graph is formed to represent the hierarchical relationship between different intentions. At the same time, an intention conflict measurement function is applied to quantitatively evaluate the degree of conflict between intentions.

[0010] Using the generated hierarchical intention graph structure and conflict measurement results, an intention competition dynamics model is established based on differential equations to characterize the mutual inhibition and promotion relationship between multiple driving intentions and their dynamic evolution over time.

[0011] The established intention competition dynamics model is combined with recursive neural differential equations to predict the changes in intention intensity and identify critical points. Based on the changes in intention intensity, the behavior changes of surrounding vehicles are predicted and the defensive driving decisions of the main vehicle are generated.

[0012] In a preferred embodiment, the step of implementing counterfactual reasoning using the driving behavior causal graph structure includes:

[0013] Construct a directed acyclic graph containing environment state nodes, surrounding vehicle observation state nodes, surrounding vehicle driving intention nodes, and surrounding vehicle behavior nodes to represent the causal structure of driving behavior;

[0014] Implementing structured causal models to infer how behavior changes when environmental conditions change but driving intentions remain constant;

[0015] Train an intent extraction network to extract intrinsic intent representations that are independent of environmental conditions.

[0016] In a preferred embodiment, the intent extraction network is trained by minimizing a combined loss function comprising behavior reconstruction loss and environment invariance loss, wherein the environment invariance loss is implemented through adversarial training to ensure that the extracted intent representation remains stable under different environmental conditions.

[0017] In a preferred embodiment, the hierarchical driving intention graph structure is a directed graph, wherein nodes represent driving intentions and edges represent hierarchical inclusion relationships between intentions;

[0018] The intention conflict measurement function quantifies the degree of conflict between two intentions based on the overlap of behavior distribution under intention conditions.

[0019] In a preferred embodiment, the intention competition dynamics model describes the change of each intention intensity value over time through a differential equation:

[0020] Each intention has an intrinsic growth rate parameter, which represents the strengthening tendency of the intention itself;

[0021] The upper capacity parameter indicates the maximum intensity that the intention can reach without external interference;

[0022] The inter-intention competition coefficient matrix represents the mutual inhibition relationship between different intentions;

[0023] and a scene-dependent external motivation function that maps environmental changes to changes in intent strength.

[0024] In a preferred embodiment, the recursive neural differential equation consists of the following parts:

[0025] Recursive structure to capture the evolving patterns of intent over long time series;

[0026] Parameterized neural networks for approximating differential equations of intention competition dynamics;

[0027] A mapping network that maps the current hidden state to the dynamics model parameters;

[0028] The differential equation is solved by numerical integration method to predict the intention intensity value at future time points.

[0029] In a preferred embodiment, the step of predicting the change in the intensity of the intention and identifying the critical point includes:

[0030] Calculate the second derivative of intention intensity;

[0031] When the absolute value of the second-order derivative exceeds the preset threshold and the sign of the first-order derivative changes, it is determined to be a potential critical point of intention;

[0032] Based on the identified critical points, the dominant intention after the transition is calculated, the possible behavior distribution of surrounding vehicles is predicted, and the defensive driving decision of the host vehicle is generated.

[0033] In a preferred embodiment, the defensive driving decision is generated by maximizing the safety margin, that is, among all possible actions of the host vehicle, the action that maximizes the safety margin under the most dangerous behavior of the surrounding vehicle is selected;

[0034] The safety function measures the safety of the main vehicle's actions when the surrounding vehicles perform specific behaviors.

[0035] In a preferred embodiment, the driving decision-making method based on the surrounding vehicle driving behavior tendency further includes real-time updating and adaptive adjustment steps:

[0036] As new observation data are continuously acquired, the parameters of the intention competition dynamics model are continuously updated;

[0037] Dynamically adjust the sensitivity threshold of critical point detection based on the accuracy of historical predictions;

[0038] Through the adaptive adjustment mechanism, the proportion of false warnings can be controlled while maintaining a high detection rate, thereby improving the reliability and effectiveness of defensive driving decisions.

[0039] In a preferred embodiment, a driving decision system based on the driving behavior tendency of a surrounding vehicle is used to execute a driving decision method based on the driving behavior tendency of a surrounding vehicle, including:

[0040] The sensor module is used to obtain vehicle trajectory data and surrounding environment information through on-board sensors, including position, speed, acceleration and other time series data;

[0041] The intention recognition module is used to implement counterfactual reasoning using the driving behavior causal graph structure, distinguish internal intentions from the influence of external environmental constraints, and form a hierarchical driving intention graph structure to represent the hierarchical relationship between different intentions;

[0042] The dynamics modeling module is used to establish an intention competition dynamics model based on differential equations, characterizing the mutual inhibition and promotion relationship between multiple driving intentions and their dynamic evolution over time;

[0043] A prediction module, which combines recursive neural differential equations to predict changes in intent intensity and identify critical points;

[0044] The decision-making module is used to generate defensive driving decisions for the host vehicle based on the intention prediction results.

[0045] The beneficial effects of the present invention are:

[0046] By separating the driver's intrinsic intention from the influence of environmental constraints through counterfactual reasoning, the accuracy and robustness of the prediction of the surrounding vehicle's driving intention are improved.

[0047] By using a hierarchical intention graph structure and an intention conflict measurement function, the system effectively handles complex situations where multiple intentions coexist. The system is able to simultaneously represent and process multiple potential driver intentions, and understand the hierarchical and conflicting relationships between them.

[0048] Based on a competitive dynamics model and recursive neural differential equations, the system accurately predicts the evolution and transition points of intent. It can capture the evolution of intent over long time series and predict the critical points of intent transition in advance.

[0049] By generating defensive driving decisions based on intention prediction, the system enhances the safety of autonomous driving systems in complex traffic situations. The system can proactively respond to potential behavioral changes of surrounding vehicles, reducing the risk of collisions.

[0050] Strong generalization capability: Because the present invention is based on causal reasoning and dynamic models rather than simple statistical data association, it can still demonstrate good generalization capabilities in new and unseen scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the driving decision-making method based on the driving behavior tendency of the surrounding vehicle of the present invention;

[0052] Figure 2 This is a bar chart comparing the accuracy of intention prediction between the method of the present invention and the traditional method in different scenarios;

[0053] Figure 3 is a line graph showing the evolution of the intensity of the intention of the present invention over time;

[0054] Figure 4 is a scatter plot of the relationship between the intended prediction lead time and the collision risk reduction rate of the present invention;

[0055] Figure 5 It is a radar chart comparing the multi-dimensional performance of the present invention and the traditional method;

[0056] Figure 6 It is an area diagram of the evolution of multiple intention intensities in the highway ramp merging scenario of the present invention. DETAILED DESCRIPTION

[0057] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0058] At least one embodiment of the present invention discloses a driving decision-making method based on the driving behavior tendency of the surrounding vehicle, such as Figure 1 As shown, the following steps are included:

[0059] Step 1: Obtain vehicle trajectory data and surrounding environment information through on-board sensors, including position, speed, acceleration and other time series data;

[0060] The specific steps include:

[0061] Step 1.1, collect vehicle trajectory data and environmental information;

[0062] Collect historical trajectory data of surrounding vehicles, including time series information such as position, speed, acceleration, and direction angle, and obtain corresponding traffic environment information such as road layout, traffic rules, and surrounding vehicle status.

[0063] The system obtains these data through vehicle sensors (such as radar, camera) and vehicle network, forming a structured driving scene data set D = {(X t ,E t ,A t )|t=1,2,...,T}, where Xt represents the state vector of the perimeter vehicle at time t, E t represents the environment state vector, A t represents the observed behavior vector of the weekly vehicle; T represents the total number of time steps.

[0064] Step 1.2, construct the driving behavior causal graph structure;

[0065] In one embodiment of the present application, a directed acyclic graph G=(V, ε) is constructed as a causal structure representation of driving behavior.

[0066] The node set V in the figure includes: environment state node E, surrounding vehicle observation state node X, surrounding vehicle driving intention node I and surrounding vehicle behavior node A.

[0067] The edge set ε represents the causal relationship between nodes, including: the edge from environment to intention E→I (the influence of environment on intention), the edge from intention to behavior I→A (intention leads to behavior), the edge from environment to behavior E→A (the environment directly constrains behavior), and the edge from state to intention X→I (the current state affects the formation of intention).

[0068] This causal graph structure clearly distinguishes the two different causal paths of "environment→behavior" and "intention→behavior", laying the foundation for the subsequent separation of the influence of intention and environment on behavior.

[0069] Step 1.3, implement the counterfactual reasoning model;

[0070] Based on the constructed causal graph, a structured causal model is implemented to infer “how the behavior will change if the environmental conditions change but the driving intention remains unchanged”.

[0071] Specifically, a structured causal model can be implemented:

[0072] P(A|do(E=e),I=i)=∫P(A|E=e,I=i,X=x)P(X|E=e)dx;

[0073] Here, do(E=e) represents an intervention in setting the environmental variable E to value e, artificially fixing the environmental condition to a specific state e; I=i represents driving intention i, assuming the driver's intrinsic intention remains in state i; A represents the surrounding vehicle's behavioral variables, such as acceleration and steering; X represents the surrounding vehicle's state variables, such as position and velocity; P(A|E=e, I=i) represents the conditional probability distribution of behavior A given the environment e and intention i; P(A|E=e, I=i, X=x) represents the conditional probability distribution of behavior A given the environment e, intention i, and state x; and P(X|E=e) represents the conditional probability distribution of state X given the environment e; ∫ represents the integral symbol, used to calculate the weighted average of all possible states X; and dx represents the infinitesimal volume element in the state space, used to implement the integration operation over the continuous state space. This model can calculate the behavior distribution given the given environment and intention, thereby separating the influence of environmental constraints and intrinsic intention on behavior.

[0074] In an optional embodiment of the present application, the counterfactual reasoning model may be represented by a structural equation model.

[0075] Specifically, the following structural equations can be implemented:

[0076] X=f X (E,U X )I=f I (E,X,U I )A=f A (I,E,X,U A );

[0077] Among them, f X 、f I and f A are the generating functions of state, intention and behavior respectively. X represents the state variables of the surrounding vehicle, such as position and speed, E represents the environment state vector, which contains information such as road layout, traffic rules, and surrounding vehicle status, I represents the driving intention of the surrounding vehicle, such as changing lanes, overtaking, and deceleration, and A represents the behavioral variables of the surrounding vehicle, such as acceleration, steering, and other actions. U X 、U I and U A These are exogenous variables representing the vehicle state variables, driving intention, and behavior variables, representing unobservable random factors or noise that influence the generation of state, intention, and behavior, respectively. According to some embodiments of the present application, these functions can be parameterized using a neural network and trained using observed data.

[0078] The generating function f can be realized by multi-layer perceptron X 、f I and f A , for example, for fI A three-layer structure can be used, with the input layer receiving the concatenated vector of E and X, the hidden layer using the ReLU activation function, and the output layer using the Softmax function to generate a probability distribution of intent. During the training phase, stochastic gradient descent and the cross-entropy loss function can be used to optimize network parameters. The learning rate can be set to 0.001, the batch size to 64, and the number of training iterations to 10,000.

[0079] In another optional embodiment, the counterfactual reasoning model can be implemented using a variational autoencoder structure. Specifically, the encoder network encodes the observed (X, E, A) into a latent variable representation of the intent I, and the decoder network reconstructs the behavior A based on the intent I and the environment E. By minimizing the reconstruction error and KL divergence, the model learns to decompose the behavior into components jointly determined by the intent and the environment.

[0080] The optimization objective function can be expressed as:

[0081]

[0082] Among them, L VAE is the loss function of the variational autoencoder, which is used to optimize the model parameters; Denotes the encoder distribution q φ Expected value under (I|X, E, A); q φ (I|X, E, A) is an encoder network with parameter φ, which maps the observed state X, environment E and behavior A to the posterior distribution of intention I; logp θ (A|I, E) is the reconstruction term, which represents the log-likelihood of behavior A given intention I and environment E, and is calculated by the decoder network with parameters θ; D KL (q φ (I|X, E, A)||p(I)) is the KL divergence term, which measures the difference between the posterior distribution of the encoder output and the prior distribution p(I); p(I) is the prior distribution of intent I, which is set to a standard normal distribution

[0083] The advantage of this method is that it can explicitly express the uncertainty of intention and is more robust to noise in the data.

[0084] In another optional embodiment, the counterfactual reasoning model can be implemented using a diffusion model. The diffusion model performs well in processing complex distributions by gradually adding noise to the data and then learning the inverse process to generate data.

[0085] A diffusion process that generates behavior conditioned by intention and environment can be defined as:

[0086]

[0087] Among them, q(A t |A t-1 ) represents the behavior state A from the previous time step t-1 t-1 Transfer to the behavior state A at the current time step t t The conditional probability distribution of ; t represents the time step number of the noise addition; A t A represents the behavior after adding noise in step t; t-1 represents the behavior of the previous time step; β t is the noise scheduling parameter, which controls the intensity of the noise added at each step; I represents the identity matrix, which is used to define the covariance structure; represents a normal distribution; is the mean, β t I is the covariance matrix.

[0088] The inverse process is performed by the neural network ∈ θ Parameterization, where θ represents the network parameters, ∈ represents the predicted noise, and the training objective is:

[0089]

[0090] Among them, L diff is the loss function of the diffusion model, represents the expectation of time step t, original behavior A0 and noise ∈; ∈ is the standard Gaussian noise added to the original behavior; ∈ θ is a neural network with parameters θ, used to predict the added noise; A t is the noisy behavior at step t in the diffusion process; t represents the time step of the diffusion process; I represents the driving intention; E represents the environmental state; ||·|| 2 Represents the square of the Euclidean distance.

[0091] This approach can generate more diverse and realistically distributed behavior predictions and is particularly suitable for complex multimodal behavior modeling.

[0092] Step 1.4, train the intent extraction network;

[0093] According to one embodiment of the present application, a neural network model f is constructed θ (X, E, A), receives the state of the surrounding vehicle, the state of the environment and the observed behavior as input, and outputs the potential intention representation I of the surrounding vehicle.

[0094] The network can be trained by minimizing the following loss function:

[0095]

[0096] Among them, L(θ) is the total loss function, θ represents the network parameters; is the behavior reconstruction loss, which is used to measure the behavior of the model reconstruction The difference between the extracted intent and the actual observed behavior A ensures that the observed behavior can be reconstructed from the extracted intent; L inv (I, E) is the environment invariance loss, which is used to measure the independence of the extracted intent representation I from the environment variable E, ensuring that the extracted intent representation remains stable under different environmental conditions; λ is the weight parameter that balances the two losses and controls the importance of the environment invariance loss in the total loss. A larger λ value will make the model focus more on learning environment-invariant intent representations.

[0097] In some embodiments of the present application, the intent extraction network adopts an encoder-decoder structure. The encoder part can be composed of three components: state encoder Enc X , Environment Encoder Enc E and behavioral encoder Enc A , which respectively map the state, environment, and behavior into latent space.

[0098] Intent representation I is obtained by:

[0099] h X =Enc x (X);

[0100] h E =Enc E (E)

[0101] h A =Enc A (A);

[0102] I=f merge ([h X ,h E ,h A ]);

[0103] Among them, h X The encoding vector representing the state feature is generated by the state encoder Enc X Encode the input state X; h E The encoding vector representing the environmental features is generated by the environment encoder Enc E Encode the input environment E to obtain; h A The encoding vector representing the behavior feature is generated by the behavior encoder Enc A The input behavior A is encoded; I is the intention representation, which is obtained by the fusion function f merge The three encoding vectors h X 、h E and h A Combination generation; f mergeIt can be a multi-layer perceptron that fuses these three encoding vectors and extracts a pure intent representation that is independent of the context.

[0104] The decoder part can contain a behavior decoder Dec A , used to reconstruct behavior from intention and environment:

[0105]

[0106] in, Represents the reconstructed behavior prediction result, Dec A represents the behavior decoder function, I represents the extracted driving intention representation, and E represents the environment state. This formula describes how the decoder maps the extracted intention representation and the environment state into a predicted driving behavior, making the reconstructed behavior as close as possible to the observed actual behavior A.

[0107] Environmental invariance loss L inv This can be achieved through adversarial training. For example, the system can train an environment predictor to try to predict the environment E from the intent representation I, while the intent extraction network tries to generate a representation that makes the environment predictor fail. This adversarial mechanism can force the network to learn a pure intent representation that is unrelated to the environment.

[0108] Step 2: Based on the acquired data, the driving behavior causal graph structure is used to implement counterfactual reasoning, effectively distinguishing the influence of the driver's internal intention and external environmental constraints;

[0109] The specific steps include:

[0110] Step 2.1, define the basic driving intention set;

[0111] First, we define the basic intention set I in the driving scenario B ={i1, i2, ..., i n}, where i1, i2, i n "" represents the first, second, and nth basic driving intentions, respectively; n represents the total number of basic driving intentions. These basic intentions are the fundamental units for the system to identify and predict. In the embodiment of this application, based on historical driving data and expert knowledge, the system pre-defines 30 basic driving intentions, covering common driving scenarios.

[0112] Step 2.2, construct a hierarchical intent graph structure;

[0113] In one embodiment of the present application, a directed graph G is constructed. I =(V I , E I ) represents the hierarchical relationship between driving intentions. IIndicates all basic intentions and compound intentions formed by combining basic intentions; E I Indicates the hierarchical inclusion relationship between intentions. For example, the intention of "overtaking" may contain two sub-intentions: "accelerating" and "changing lanes".

[0114] In the formal representation, if the intention i p It is intention q sub-intention, there can be a directed edge (i q ,i p )∈E I , where i p Indicates sub-intention, i q Indicates the father's intention, E I Represents the set of edges in the intent graph.

[0115] For each composite intent i c , define its sub-intent set:

[0116] Sub(i c )={i|(i c , i)∈E I};

[0117] Among them, i c Indicates composite intent, Sub(i c ) means i c The set of all sub-intentions of , each element i in the set is c Through the directed edge (i c ,i) Connected sub-intentions,E I Represents the set of edges in the intent graph.

[0118] In addition, this hierarchical structure enables the system to understand driving intentions at different levels of abstraction simultaneously, improving the flexibility and expressiveness of intention representation.

[0119] In some embodiments of the present application, the hierarchical intent graph structure can be represented by a hybrid of a tree structure and a directed acyclic graph. For example, the system can organize intents into three levels:

[0120] Top-level intention: represents the high-level driving goal, such as "reaching the destination", "avoiding danger", "obeying traffic rules", etc.

[0121] Mid-level intentions: These represent strategies for achieving top-level goals, such as “overtaking,” “changing lanes,” “slowing down,” and “following a vehicle.”

[0122] Low-level intent: Indicates the specific operations for executing mid-level strategies, such as "accelerate," "steering," and "braking."

[0123] Step 2.3, construct the intention conflict measurement function;

[0124] In order to evaluate the degree of conflict between different intentions, an intention conflict measurement function S is constructed. conf , which can quantify the degree of conflict between two intentions based on the overlap of behavior distribution under intention conditions:

[0125]

[0126] Among them, S conf (i1, i2) represents the conflict measurement function between intentions i1 and i2; i1 and i2 represent two different driving intentions, such as "lane change" and "accelerate", respectively; P(A|I=i1) represents the behavior distribution set under the condition of intention i1, that is, the set of behaviors that the driver may take when he has intention i1; P(A|I=i2) represents the behavior distribution set under the condition of intention i2; ∩ represents the intersection operation of sets, which calculates the common part of two behavior distribution sets; ∪ represents the union operation of sets, which calculates all elements of two behavior distribution sets; |·| represents the cardinality (number of elements) of a set, which is used to calculate the number of elements in a set.

[0127] The range of this metric function is [0, 1]. Smaller values indicate more severe conflict between the two intents (fewer shared behaviors), while a value of 1 indicates complete compatibility between the two intents (complete overlap in behavior distribution). When the behaviors corresponding to the two intents are completely different, the function value is 0, indicating the most severe conflict.

[0128] In some embodiments of the present application, the intention conflict metric function is calculated by the following steps:

[0129] Behavior space discretization: discretize the continuous behavior space (acceleration, steering angle, etc.) into a finite set of grid points A dis For example, the acceleration range is [-5m / s 2 , 5m / s 2 ] is divided into 20 equally spaced points.

[0130] Conditional probability estimation: For each intention i and discrete behavior point a∈A dis , the system can estimate the conditional probability P(a|I=i) based on historical data.

[0131] The system can adopt the kernel density estimation method, using the Gaussian kernel function:

[0132]

[0133] Where P(a|I=i) represents the conditional probability of executing behavior a under intention i, N i is the number of historical behavior samples associated with intention i, a jis the behavior value observed in the jth historical sample, K is the Gaussian kernel function, h is the bandwidth parameter, and it is a hyperparameter that controls the smoothness of the probability distribution. The larger the value, the smoother the distribution. For all N i This formula uses the kernel density estimation method to calculate the probability distribution of a specific behavior under a given intention based on historical observation data.

[0134] Overlap calculation: Based on the estimated conditional probability distribution, the overlap of the behavior distributions of the two intentions is calculated to obtain the final conflict measurement value.

[0135] In addition to the overlap of behavior distribution, the system can also consider the semantic compatibility of the intent target. For example, a semantic compatibility matrix C∈[0,1]n×n can be defined, where C ij Indicates the degree of compatibility between intention i and intention j at the semantic level and incorporates it into the calculation of the conflict measure:

[0136]

[0137] in, represents the enhanced conflict metric function value considering semantic compatibility, which is used to evaluate the comprehensive conflict degree between intention i1 and intention i2; S conf (i1, i2) represents the original conflict metric function value calculated based on the overlap of behavior distribution; It represents the compatibility between intention i1 and intention i2 at the semantic level, and its value range is [0, 1]. A larger value indicates a higher semantic compatibility. α∈[0, 1] is a weight parameter that balances the conflict measurement at the behavioral level and the compatibility at the semantic level. When α is close to 1, the system pays more attention to the conflict at the behavioral level; when α is close to 0, the system pays more attention to the compatibility at the semantic level.

[0138] This enhanced conflict measurement function can simultaneously consider conflicts at the behavioral level and the target semantic level, providing a more comprehensive intention conflict assessment.

[0139] Step 2.4, calculate the probability of intention realization;

[0140] For each identified intent, the system needs to calculate its probability of realization under the current environmental conditions.

[0141] Use the counterfactual intervention model to estimate the probability of intention realization:

[0142] P(Realize(i)|E=e=∫P(Realize(i)|A=a,E=e)P(A=a|do(I=i),E=e)da;

[0143] Among them, P(Realize(i)|E=e) represents the probability that intention i is realized under the environmental condition E=e; Realize(i) represents the event that intention i is realized; P(Realize(i)|A=a, E=e) represents the conditional probability that intention i is realized when behavior A=a is performed under the environmental condition E=e; P(A=a|do(I=i), E=e) represents the probability of choosing behavior a when the intervention sets intention to i under the environmental condition E=e; do(I=i) represents a causal intervention operation that forces the intention to be set to i; ∫(·)da represents the integration of all possible behaviors a, that is, considering all possible behavior choices.

[0144] Step 3: Based on the identified intrinsic intentions, a hierarchical driving intention graph is formed to represent the hierarchical relationship between different intentions. At the same time, an intention conflict measurement function is applied to quantitatively evaluate the degree of conflict between intentions.

[0145] The specific steps include:

[0146] Step 3.1, define the intention competition dynamics model;

[0147] Based on the classic theory of population competition dynamics, a differential equation model is constructed to describe the mutual influence and dynamic evolution process among multiple driving intentions.

[0148] For each intention i, its strength value I i The variation of (t) with time t can be described by the following differential equation:

[0149]

[0150] Among them, I i (t), I j (t) represents the intensity value of intention i and intention j at time t, with a value range of [0, 1], indicating the activeness of the driving intention at the current moment; α i is the intrinsic growth rate of intention i, indicating the strengthening trend of the intention itself and reflecting the driver's natural tendency towards the intention when there is no external influence; K i is the upper limit of the capacity of intention i, indicating the maximum intensity that the intention can reach without external interference; β ij is the competition coefficient of intention j to intention i, indicating the mutual inhibition relationship between the two intentions, β ij The larger the value, the stronger the inhibition of intention j on intention i, reflecting the degree of mutual exclusion between different driving intentions; i (S scene (t)) is a scene-related external incentive function, which indicates the influence of environmental factors on intention i; S scene (t) represents the scene state at time t, including environmental conditions and vehicle status; It represents the rate of change of the intensity of intention i over time, describing how the intensity of the intention evolves dynamically over time.

[0151] Step 3.2, construct the competition coefficient matrix;

[0152] In one embodiment of the present application, a competition coefficient matrix B=[β ij ], where n1 is the number of intentions considered, β ij Indicates the degree of competitive impact of intention j on intention i. This matrix can have the following characteristics:

[0153] Diagonal elements β ii =0, indicating that the intention is not to compete with itself;

[0154] Off-diagonal elements β ij ≥0, the larger the value, the more intense the competition;

[0155] The matrix is not necessarily symmetric, i.e. ij May not be equal to β ji ,The competitive relationship between the representation intentions may be asymmetric;

[0156] The initial competition coefficient matrix can be determined based on the intention conflict metric function calculated in step 2:

[0157] β ij =1-S conf (i i ,i j )

[0158] Among them, β ij Indicates the competition coefficient of intention j to intention i. The larger the value, the more intense the competition. conf (i i ,i j ) express intention i and intention j The conflict measurement function value between them ranges from [0, 1]; 1-S conf (i i ,i j ) represents the conversion of conflict degree into competition coefficient.

[0159] When two intentions are highly conflicting (S conf close to 0), the competition coefficient β ij Close to 1, indicating strong competitive relationship;

[0160] When two intentions hardly conflict (S conf close to 1), the competition coefficient β ij Close to 0, indicating weak competitive relationship;

[0161] This means that the higher the conflict degree (S conf The smaller the , the greater the competition coefficient.

[0162] Use machine learning methods to learn the competitive coefficient matrix from historical driving data instead of manually setting it based on prior knowledge. For example, you can use gradient descent to minimize the prediction error and optimize the competitive coefficient:

[0163]

[0164] Among them, B * is the optimized competition coefficient matrix, argmin B Represents finding the matrix B that minimizes the objective function, represents the sum of all moments in the time series from 1 to T; T represents the total number of time steps; ||·|| 2 represents the square of the Euclidean norm, I t+1 is the intention intensity vector at time t+1, I t is the intention intensity vector at time t, Δt is the time step, S scene (t) is the scene state at time t, F is the intent evolution function, and contains the competition coefficient matrix B as a parameter. This formula represents the optimization of the competition coefficient matrix by minimizing the squared error between the predicted intent strength and the actual intent strength.

[0165] Step 3.3, construct the external activation function;

[0166] According to one embodiment of the present application, a scene-dependent external excitation function γ is constructed. i (S scene (t)), which is used to capture the direct impact of environmental state changes on intention strength.

[0167] This function can take the following forms:

[0168]

[0169] Among them, γ i (S scene (t)) represents the external incentive function of the scene to the intention i, S scene (t) represents the scene state at time t, f k (S scene (t)) is the state of the scene S scene The kth feature extracted from (t), w ik is the weight of the feature's influence on intention i, indicating the importance of feature k to intention i. m is the total number of features, and ∑ represents the weighted sum of all m features. This formula describes how to integrate multiple scenario features into the incentive strength for a specific driving intention.

[0170] Step 3.4, apply intention hierarchy constraints;

[0171] Based on the hierarchical intent graph structure constructed in step 2, hierarchical constraints are imposed on the dynamics model:

[0172] If the intention p It is intention q 's sub-intention, then I p (t)≤I q (t), that is, the intensity of the child intention cannot exceed that of the parent intention;

[0173] The strength calculation of a composite intent can take into account the strength of its sub-intents:

[0174] I c (t) = g({I s (t)|s∈Sub(c)});

[0175] Among them, I c (t) represents the intensity value of the composite intention c at time t, reflecting the activity level of the composite intention; I s (t) represents the intensity value of sub-intention s at time t, indicating the activity level of each sub-intention constituting the composite intention; s∈Sub(c) indicates that s is a sub-intention of composite intention c, and Sub(c) represents the set of all sub-intentions of intention c; g is an aggregation function used to merge the intensity values of multiple sub-intentions into the intensity value of the composite intention, which can take different forms such as minimum value, maximum value, and weighted average.

[0176] The system can use a probabilistic graphical model to represent the intent hierarchy and convert hierarchical constraints into conditional probability relationships:

[0177] P(I c =a ctive|{I s =active},s∈Sub(c));

[0178] Among them, P(I c =active|{I s = a ctive}, s∈Sub(c)) represents the sub-intention set {I s} is in the activated state, the composite intention I c Conditional probability of being in the activated state; I c Indicates a composite intention, i.e., the parent intention; I s Represents a sub-intention; active indicates that the intention is in an activated state; s∈Sub(c) indicates that s is a sub-intention of c, and Sub(c) represents the set of all sub-intentions of intention c.

[0179] Step 4: Using the generated hierarchical intention graph structure and conflict measurement results, an intention competition dynamics model is established based on differential equations to characterize the mutual inhibition and promotion relationships between multiple driving intentions and their dynamic evolution over time.

[0180] The specific steps include:

[0181] Step 4.1, construct a recursive neural differential equation model;

[0182] Combining recurrent neural networks and differential equation modeling techniques, a unified framework is constructed to learn the parameters of the intention competition dynamics model. The model is expressed as:

[0183]

[0184] Where I(t) is the intensity vector of all intentions, indicating the activation degree of each intention at time t; S scene (t) is the scene state, which represents the traffic environment information at time t, including vehicle position, speed, road structure and other related features; f θ is a parameterized intent evolution function that describes the dynamics of how intent strength changes over time; θ is a parameter to be learned, which contains the weight coefficients that control the competition and cooperation relationships between intents; It represents the rate of change of the intention intensity vector over time, that is, the time derivative of the intention intensity.

[0185] f θ It is composed of a dynamic model based on prior knowledge and a data-driven neural network:

[0186] f θ (I(t),S scene (t))=f prior (I(t),θ prior )+NN θ (I(t),S scene (t));

[0187] Among them, f θ is a parameterized intention evolution function that describes the rate of change of intention intensity over time; I(t) represents the intention intensity vector at time t, which contains the intensity values of all intentions; S scene (t) represents the scene state vector at time t; f prior represents the prior dynamic model; θ prior is the parameter set of the prior dynamic model, such as the competition coefficient matrix between intentions, inhibition coefficient, etc.; NN θ For the neural network part.

[0188] Step 4.2, data collection and preprocessing;

[0189] To learn the model parameters, it is necessary to collect driving data from multiple drivers in various scenarios, including time-series data such as vehicle status, environmental conditions, and driver behavior. This data can come from real-world road tests, driving simulators, or existing driving datasets.

[0190] For the collected raw data, the intention extraction network in step 1 is used to extract the intensity value sequence of potential intentions {I(t1), I(t2), ..., I(t T )}, among which, I(t1), I(t2), I(t T ) represent the time t1, t2, t T The intent intensity vector is constructed to construct a dataset for training:

[0191]

[0192] Where D represents the constructed training dataset; I(t) represents the intention intensity vector at time t, which contains the intensity values of various driving intentions; S scene (t) represents the scene state vector at time t, which contains information such as environment and traffic conditions; It represents the rate of change of intention intensity I(t) with time t; T represents the total number of time steps.

[0193] The system uses smoothing techniques to deal with noise in the raw data, for example, using a sliding average:

[0194]

[0195] in, is the smoothed intent intensity value, I(t+j·Δt) represents the original intent intensity value at time t+j·Δt, w is the half-width of the window, which determines the size of the smoothing range, j is the index variable within the window, ranging from -w to w, Δt is the sampling interval, and 2w+1 represents the width of the entire window, which is used to normalize the summation result. This smoothing process can reduce fluctuations introduced by sensor noise or estimation errors and improve training stability.

[0196] Step 4.3, parameter learning algorithm;

[0197] Design a parameter learning algorithm based on gradient descent, and the optimization goal is to minimize the prediction error:

[0198]

[0199] Where L(θ) is the loss function used to evaluate the performance of the model parameters θ, θ is the network parameter; I is the driving intention intensity vector; S scene is the driving scene state; express That is, the rate of change of intention intensity; f θ (I, S scene ) is a parameterized neural network model that predicts the rate of change of intention given an intention and scenario; is the square error between the predicted rate of change and the true rate of change; D is the training dataset, which contains a triplet of intent strength, scene state, and intent change rate; R(θ) is the regularization term used to prevent model overfitting; λ is the regularization strength, which controls the influence of the regularization term.

[0200] Step 4.4, solving the differential equation;

[0201] Use numerical methods to solve differential equations and predict the evolution trajectory of intention intensity:

[0202]

[0203] Where I(t+Δt) represents the intention intensity vector at time point t+Δt; I(τ) represents the intention intensity vector at time point t; ∫ t t+Δt represents the definite integral from time t to time t+Δt; f θ represents the parameterized intention evolution function, which describes the dynamic rules of how the intention intensity changes with time; I(τ) represents the intention intensity vector at the time of the integral variable τ; S scene (τ) represents the scene state vector at the time of the integral variable τ, which contains information such as the environment and traffic conditions; θ1 represents the learning parameters of the model, which controls the dynamic characteristics of the evolution of the intention; dτ represents the differential of the integral variable.

[0204] Step 5: Combine the established intention competition dynamics model with recursive neural differential equations to predict changes in intention intensity and identify critical points. Based on the changes in intention intensity, the behavior changes of surrounding vehicles are predicted and defensive driving decisions are generated for the main vehicle.

[0205] The specific steps include:

[0206] Step 5.1, define the critical point of intention evolution;

[0207] The critical point of intention evolution refers to the moment when the intensity or dominance of the driving intention of the surrounding vehicle changes significantly, usually indicating a shift in future behavior patterns. The critical point can be defined as the moment when any of the following conditions are met:

[0208] Intention alternation critical point: When two dominant intentions I i and I j When the intensity relationship of I i (t)>I j (t) becomes I i(t+Δt) j (t+Δt), where I i (t) represents the intensity value of intention i at time t; I j (t) represents the intensity value of intention j at time t; t represents the current time; Δt represents the time increment, that is, the predicted future time window; I i (t+Δt), I j (t+Δt) represents the predicted strength values of intention i and intention j at the future time t+Δt respectively.

[0209] Intention Rising Critical Point: When an intention with low intensity I k The intensity increases significantly in a short period of time, exceeding the preset threshold δ rise , that is I k (t+Δt)-I k (t)>δ rise .

[0210] Intention fading critical point: When an intention that was originally strong l The intensity decreases significantly in a short period of time, exceeding the preset threshold δ fall , that is I l (t)-I l (t+Δt)>δ fall , where I l represents the intensity value of the lth intention, t+Δt represents the future time, Δt is the time window size, δ fall is the threshold parameter for the decrease in intention intensity.

[0211] In actual application scenarios, appropriate time window Δt and threshold parameter δ can be set as needed rise , δ fall to adjust the detection sensitivity.

[0212] Step 5.2, construct a critical point detection algorithm;

[0213] Based on the recursive neural differential equation model trained in step 4, a forward-looking critical point detection algorithm is constructed. The basic process of the algorithm includes:

[0214] Based on the currently observed intention state I(t) and scene state S scene (t), using a recursive neural differential equation model to predict the intention evolution trajectory in the future time window;

[0215] Detect whether there is a moment in the predicted trajectory that meets the definition of a critical point;

[0216] If a critical point is detected, its occurrence time, the type of intention involved, and the characteristics of the transition are calculated to provide a basis for subsequent decision-making; ​

[0217] Construct a probability-based critical point detection method, expressed as the probability of detecting a critical point in the future:

[0218]

[0219] in, represents the probability of the critical point event of alternating intentions i and j in the time interval [t1, t2]; Indicates the critical point event where intention i and intention j alternate within the time interval [t1, t2]; I i (t1) represents the intensity value of intention i at time t1; I j (t1) represents the intensity value of intention j at time t1; I i (t2) represents the intensity value of intention i at time t2; I j (t2) represents the intensity value of intention j at time t2; t1 represents the time when the observation starts; t2 represents the time when the observation ends; ∧ represents the logical "and" operation, indicating that two conditions are met at the same time.

[0220] Step 5.3, behavior prediction based on critical points;

[0221] When the critical point of intention evolution is detected, the system can predict the specific future behavior of the vehicle based on the change in intention. The mapping between intention and behavior can be expressed as:

[0222] P(A|I,E)=P(A|I dominant ,E);

[0223] Where P(A|I, E) represents the conditional probability of the predicted behavior A occurring under the given intention set I and environment state E; A is the predicted vehicle behavior; I is the intention set, which contains multiple possible driving intentions and their intensity values; I dominant represents the dominant intention; E represents the environmental state; P(A|I dominant , E) means only considering the dominant intention I dominant and the conditional probability of the predicted behavior under the environmental state E.

[0224] Step 5.4, generating defensive driving decisions;

[0225] Based on the detected critical points and the predicted surrounding vehicle behavior, the system can generate defensive driving decisions for the host vehicle to prevent potential dangerous situations. Defensive decisions can be based on the following factors:

[0226] Types and timing of tipping points;

[0227] Predicted behavior after Zhouche intention shift;

[0228] The current spatial relationship between the main vehicle and surrounding vehicles;

[0229] The overall condition of the traffic environment;

[0230] The system can generate a decision score function to evaluate different driving actions:

[0231] Score(a ego )=w safety Safety(a ego ,a surr )+w eff ·Efficiency(a ego )+w comf Comfort(a ego );

[0232] Among them, Score(a ego ) represents the comprehensive score of the candidate actions of the main vehicle, which is used to select the optimal driving decision; a ego Indicates the candidate actions of the main vehicle, including possible driving operations such as acceleration, deceleration, and lane change; a surr Indicates the predicted actions of surrounding vehicles (circumferential vehicles); Safety (a ego , a surr ) represents the security scoring function; Efficiency(a ego ) represents the efficiency scoring function, which evaluates the impact of the main vehicle's actions on driving efficiency; Comfort(a ego ) represents the comfort scoring function, which evaluates the impact of the main vehicle action on the passenger comfort; w safety 、w eff 、w comf They represent the weight coefficients of safety score, efficiency score, and comfort score respectively.

[0233] The core concept of defensive driving is to take preventive actions in advance while ensuring safety, rather than waiting until dangerous situations have emerged before reacting passively.

[0234] In some optional embodiments, a defensive decision-making system is constructed based on a risk assessment framework:

[0235]

[0236] Among them, Risk(a ego ) indicates that the host vehicle takes action a ego The risk value, a surr represents the possible actions of surrounding vehicles, P(a surr |CP) is the condition under which the peripheral vehicle takes action a given the detected critical point. surr Severity(a ego, a surr ) is when the host vehicle takes action a ego And the car takes action a surr The formula calculates the comprehensive risk value of a certain action of the main vehicle by taking a weighted sum of all possible surrounding vehicle behaviors.

[0237] The system chooses the action with the least risk:

[0238]

[0239] in, represents the optimal action of the main vehicle, that is, the action with the least risk; represents the main vehicle action parameter that makes the subsequent risk function obtain the minimum value; Risk(a ego ) indicates that the host vehicle takes action a ego The risk assessment function value when a ego Represents the candidate actions of the main vehicle and is a decision variable.

[0240] The system can learn the optimal defensive driving policy using reinforcement learning methods:

[0241]

[0242] Among them, π * (s) represents the optimal policy function; argmax a Indicates the selection of action a that maximizes the following function; Q * (s,a) represents the optimal action-value function, which evaluates the long-term value of taking action a in state s; Indicates that following the optimal strategy π * The mathematical expectation under the conditions of ; represents the sum of infinite time steps accumulated from the current moment; γ t represents the t-th power of the discount factor γ, which ranges from [0, 1] and is used to balance the importance of short-term and long-term rewards; r t+1 represents the immediate reward obtained at time t+1 after taking an action at time t; s 0=s, a0=a represents the condition where the initial state is s and the initial action is a; s represents the state vector; a represents the action that the vehicle can perform; r represents the reward function; γ represents the discount factor, which determines the importance of future rewards relative to immediate rewards. The closer the value is to 1, the more the system focuses on long-term benefits; Q * represents the optimal action-value function, which represents the maximum cumulative discounted reward that can be obtained after executing action a from state s under the optimal policy.

[0243] The system can learn this policy using deep Q-learning, policy gradient, or other reinforcement learning algorithms.

[0244] In some embodiments, defensive decision-making can be structured in a hierarchical structure, including strategic, tactical, and operational layers:

[0245] Strategic-level decisions deal with long-term planning, such as whether to change lanes, overtake, or stay behind the vehicle;

[0246] Tactical-level decisions deal with mid-term actions, such as accelerating, decelerating, or maintaining a constant speed when approaching surrounding vehicles;

[0247] Operational-level decisions handle short-term control, such as specific steering angle, accelerator, and brake pedal positions

[0248] This hierarchical structure enables the system to coordinate decisions at different time scales, taking into account long-term safety while ensuring flexibility in short-term responses.

[0249] Step 5.5, system integration and optimization;

[0250] In one embodiment of the present application, the intent evolution critical point detection algorithm can be seamlessly integrated with the vehicle control system to form a closed-loop autonomous driving architecture:

[0251] The perception module collects information about the surrounding environment and vehicle status;

[0252] The intention extraction module identifies and quantifies the driving intention of the vehicle based on the observed data;

[0253] The intention evolution prediction module uses recursive neural differential equations to predict the change of intention intensity;

[0254] The critical point detection module monitors potential critical points in the intention prediction results;

[0255] The behavior prediction module predicts the future behavior of the vehicle based on the critical point of intention;

[0256] The decision-making planning module generates corresponding defensive driving strategies;

[0257] The control execution module converts the decision into specific control commands;

[0258] Information flows smoothly between the various modules of the system, and each module has adaptive and online learning capabilities, which can continuously optimize performance as driving experience accumulates.

[0259] In some embodiments, the system can further optimize performance by:

[0260] Online parameter adjustment: Dynamically adjust critical point detection thresholds and prediction model parameters based on real-time detection prediction errors;

[0261] Scenario Adaptation: Use specially optimized models and parameters for different driving scenarios (such as highways, urban roads, parking lots, etc.);

[0262] Driving style matching: learning and adapting to driving behavior characteristics in different regions and cultural backgrounds;

[0263] Dynamic allocation of computing resources: Dynamically adjust the computing resources allocated to each module based on the complexity of the scenario and the potential risk level;

[0264] In practical applications, the performance indicators of the system may include:

[0265] Critical point detection time: On average, behavioral changes can be detected 1 to 3 seconds earlier than traditional methods;

[0266] Prediction accuracy: Achieve over 85% accuracy in predicting intent changes across various scenarios.

[0267] Missed detection rate and false detection rate: Maintain a missed detection rate and false detection rate below 5% in safety-critical scenarios;

[0268] Behavior prediction accuracy: lateral position prediction error is less than 0.3 meters, and longitudinal speed prediction error is less than 2 meters / second;

[0269] Timely decision-making: The system's end-to-end latency is less than 100 milliseconds, meeting real-time decision-making needs;

[0270] Through these optimization and integration measures, the system can accurately predict changes in surrounding vehicle behavior in various complex traffic scenarios, make defensive decisions in advance, and improve the safety, comfort and efficiency of autonomous driving.

[0271] Application examples of this implementation:

[0272] The driving decision-making method of this embodiment has been applied in the complex traffic environment of a highway. The test environment was a busy section of a six-lane highway (three lanes in each direction), containing multiple entrances and exits, as well as areas of dense traffic. The test period was chosen during the morning rush hour (7:00-9:00) and evening rush hour (17:00-19:00) on weekdays. These periods are characterized by high traffic volume and diverse driving behaviors, making them suitable for verifying the effectiveness of this method.

[0273] The test vehicle was equipped with the following sensing equipment:

[0274] 360-degree surround-view camera system for obtaining visual information of surrounding vehicles;

[0275] Millimeter-wave radar system for accurately measuring the distance, speed, and relative position of surrounding vehicles;

[0276] GPS positioning system, providing precise vehicle location and road information;

[0277] Vehicle status sensor, which records its own speed, acceleration, steering angle and other information.

[0278] During testing, the system must handle typical scenarios including:

[0279] Lane change scenarios in high-density traffic, where surrounding vehicles need to change lanes to avoid slower vehicles or head towards an exit;

[0280] In the ramp merging area, the surrounding vehicle enters the main lane from the ramp and needs to coordinate its speed and position with the vehicles in the main lane;

[0281] Sudden deceleration scenario: the vehicle ahead suddenly slows down due to road or traffic conditions, triggering a chain reaction;

[0282] In the hesitant decision-making scenario, the driver of the vehicle is hesitant between multiple potential driving intentions, showing mixed behavioral characteristics.

[0283] These scenarios are highly complex, involving multi-vehicle interactions and complex driving intention evolution, and are ideal environments for validating the advantages of our approach.

[0284] Counterfactual reasoning and intent extraction examples:

[0285] In this application example, the counterfactual reasoning and intent extraction steps are implemented in the following way:

[0286] First, the system collects historical trajectory data for the surrounding vehicle, ID-07, from onboard sensors. This includes 100 data points of continuous position, velocity, and acceleration over a 10-second period, sampled at a 10Hz frequency. It also acquires environmental information, including lane position, traffic density (average of 15 vehicles per kilometer per lane), and the status of the vehicle ahead.

[0287] The system constructs a driving behavior causal graph structure using four types of nodes: environmental state node E (including lane layout, traffic density, surrounding vehicle positions, etc.), surrounding vehicle observation state node X (including position, speed, acceleration, etc.), surrounding vehicle driving intention node I (including lane change intention, acceleration intention, following vehicle intention, etc.) and surrounding vehicle behavior node A (including lateral displacement, longitudinal acceleration, etc.).

[0288] On a road with heavy traffic, when vehicle ID-07 showed a slight deceleration (from 95 km / h to 88 km / h) and a slight deviation to the right (0.2 meters), traditional methods found it difficult to determine whether this was the driver's inherent intention to change lanes or the influence of the deceleration of the vehicle ahead.

[0289] The system performs the following process through counterfactual reasoning:

[0290] Original observation: The pericycle slowed down and deviated slightly to the right;

[0291] Construct a counterfactual scenario: "How would the driver of the vehicle have driven if no vehicle ahead had slowed down?"

[0292] Reasoning Result: The system calculates that in the counterfactual scenario, there is a 78% probability that Zhou Che will still exhibit similar behavior;

[0293] Based on the above reasoning, the system extracts the surrounding car's inherent intention to "prepare to change lanes right" (strength value 0.67), rather than simply reacting to the environment. To achieve this, the system trained an intention extraction network using a three-layer architecture:

[0294] Encoder network: encodes the environment state E, vehicle state X and observed behavior A respectively;

[0295] Fusion network: Integrates the three codes into intent representation I;

[0296] Reconstruction network: Reconstruct the observed behavior A based on the intention I and the environment E;

[0297] The network is trained by minimizing a combination of a behavior reconstruction loss and an environment invariance loss. Environment invariance is achieved through adversarial training, ensuring that the extracted intent representation remains stable under varying environmental conditions. In this example, the intent extraction network successfully extracts the driver's lane change intention from subtle behavioral cues, even though the intention is not yet expressed through overt physical actions (such as turning on the turn signal).

[0298] Actual tests show that in 500 recognition tasks involving subtle behavioral changes, the system's intention extraction accuracy reached 81.3%, 32.5 percentage points higher than the traditional rule-based method, especially when there is a conflict between the driver's intention and the environment's forced behavior.

[0299] Hierarchical intent graph structure and intent conflict measurement example:

[0300] In this application example, the system constructs a hierarchical intent graph with a three-layer structure for highway scenarios:

[0301] Top layer (strategic intent): includes three high-level intent nodes: "efficient driving", "safe driving" and "destination navigation".

[0302] Middle layer (tactical intention): includes middle-level intention nodes such as "lane change", "overtaking", "following", "avoidance", and "exiting the highway".

[0303] Bottom layer (operation intention): includes basic operation intention nodes such as "accelerate", "slow down", "maintain speed", "turn left", "turn right", and "keep straight".

[0304] The system uses directed edges to represent the hierarchical relationships between intents. For example, the directed edge from the "overtaking" mid-level intent to the "acceleration" and "lane change" bottom-level intents indicates that overtaking behavior typically consists of acceleration and lane changes. The directed edge from the "efficient driving" top-level intent to the "overtaking" mid-level intent indicates that drivers may adopt overtaking strategies to achieve efficient driving.

[0305] In actual application, when the system observes vehicle ID-12 driving in the middle lane, there is a slow-moving truck (75 km / h) 150 meters ahead, and there is a certain amount of traffic in the left lane. The system uses the intent extraction network to simultaneously identify multiple potential intentions of the vehicle:

[0306] “Accelerate” intention (strength value 0.35);

[0307] “turn left” intention (strength value 0.28);

[0308] “following the car” intention (intensity value 0.42);

[0309] “overtaking” intention (intensity value 0.55);

[0310] Based on the hierarchical relationship, the system identifies "overtaking" as a mid-level intent, supported by the bottom-level intents of "accelerating" and "turning left," and conflicting with the "following" intent. The system calculates the conflict metric between the intents:

[0311] Degree of conflict between intentions of “overtaking” and “following”: S conf =0.23 (indicating high conflict);

[0312] Conflict between "acceleration" and "deceleration" intentions: S conf =0.12 (indicating extremely high conflict);

[0313] Degree of conflict between intentions of “turn left” and “turn right”: S conf =0.08 (indicating extremely high conflict);

[0314] Degree of conflict between the intentions of “overtaking” and “acceleration”: S conf =0.87 (low conflict, high compatibility);

[0315] The conflict metric is calculated based on the overlap of behavior distributions under the intent condition. The system first performs a kernel density estimate on the distributions of the two intents in the behavior space and then calculates the overlap between these two distributions. For example, for the "overtake" and "follow" intents, the system calculates the overlap of their distributions in the two-dimensional behavior space of acceleration and steering angle:

[0316]

[0317] In this example, the hierarchical intent structure enables the system to understand that the driver of the surrounding vehicle may have multiple conflicting intentions simultaneously. Specifically, the driver of the surrounding vehicle desires to overtake (strength 0.55), but the current environment is constrained (traffic in the left lane), resulting in a high strength of the following intention (0.42). This state of multiple coexisting and competing intentions is difficult to effectively handle using traditional models that assume a single intention.

[0318] Experimental data shows that in 2,000 multi-intention recognition tasks, the system was able to simultaneously identify the driver's primary and secondary intentions with an average accuracy of 75.8%, while traditional methods achieved only 41.2% accuracy in multi-intention recognition. The system performed particularly well in handling driver hesitation caused by conflicting intentions, predicting the driver's ultimate intention 1.7 seconds in advance, on average.

[0319] Intention competition dynamics model and recursive neural differential equation example:

[0320] In this application example, the system constructed a differential equation model of the evolution of intent intensity based on competitive dynamics theory. Taking vehicle ID-23 merging onto a ramp as an example, the system identified three primary intents: "accelerate to merge," "slow down and wait," and "maintain speed," with initial intensity values of 0.30, 0.45, and 0.25, respectively.

[0321] For the "accelerated import" intention, the system sets the parameters as follows: α 加速汇入 =0.22, K 加速汇入 =0.90. In terms of competition coefficient, the inhibitory coefficient β of the intention of "slowing down and waiting" on the intention of "speeding up and merging" is 加速汇入，减速等待 =0.58, indicating a strong inhibitory effect.

[0322] The system constructs a 3×3 competition coefficient matrix to represent the mutual inhibition relationship between the three intentions:

[0323]

[0324] The rows and columns correspond to the intentions of "accelerate to merge," "slow down and wait," and "maintain speed," respectively. Larger values in the matrix indicate stronger inhibition. For example, a value of 0.62 indicates a strong inhibitory effect of "accelerate to merge" on "slow down and wait."

[0325] The impact of environmental conditions on intention is captured through an external incentive function. The system detects a gap (approximately 35 meters) in the main lane traffic flow 120 meters ahead of the ramp, sufficient for a safe merge. As the surrounding vehicle approaches this gap, the external incentive value for the "accelerate merge" intention increases from 0.05 to 0.45, reflecting the positive impact of this environmental change on the intention's strength.

[0326] To accurately learn these parameters, the system uses a recursive neural differential equation model consisting of the following components:

[0327] Recursive network: uses an LSTM structure with a hidden state dimension of 64 to capture the temporal dependencies of intent evolution;

[0328] Parameter network: a three-layer MLP structure (input 64-hidden 128-hidden 64-output 15), mapping the LSTM hidden state to the dynamic model parameters;

[0329] Differential equation solver: Dormand-Prince method (DOPRI5) with adaptive step size;

[0330] The system trained the RNDE model using 20,000 driving data segments (5 to 15 seconds each) collected over three months. Each segment was divided into a training set (70%), a validation set (15%), and a test set (15%). The model was optimized using the Adam algorithm with an initial learning rate of 0.001 and a batch size of 64. Learning rate decay and early stopping strategies were used during training.

[0331] In actual application, the system predicted the evolution of the intention of the surrounding vehicle ID-23: as the vehicle approached an available gap, the intensity of the "accelerate to merge" intention was estimated to rise from 0.30 to 0.65 within 2.5 seconds, surpassing the "slow down and wait" intention (which dropped from 0.45 to 0.32) to become the dominant intention. This prediction enabled the host vehicle to recognize the surrounding vehicle's impending acceleration and merging maneuver three seconds earlier, approximately 1.8 seconds earlier than traditional methods based on observed behavior.

[0332] Model effectiveness evaluation showed that in 1,000 ramp-merge scenario tests, the system predicted intention changes with an average lead time of 2.7 seconds, a root mean square error (RMSE) of 0.092 (with an intention intensity range of 0-1), and an intention change prediction accuracy of 83.5%. Especially in situations where the driver is hesitant, the system can still accurately capture subtle changes in intention intensity through the competitive dynamics model and predict the final intention.

[0333] Examples of detecting critical points in intent evolution and making defensive decisions:

[0334] In this application example, the system implements forward-looking detection of critical points in intent evolution based on the prediction results of the intent competition dynamics model. The system defines three types of critical points:

[0335] Intention alternation critical point: when the intensity relationship between the two dominant intentions is reversed;

[0336] Intention Rising Tipping Point: When the strength of an initially weak intention increases significantly in a short period of time;

[0337] Intention fading critical point: when the intensity of a strong intention weakens significantly in a short period of time;

[0338] Taking a sudden deceleration scenario on a highway as an example, the system tracked the evolving intentions of vehicle ID-35. Initially, the vehicle had two primary intentions: "maintain speed" (intensity 0.72) and "slow down" (intensity 0.28). The system used a recursive neural differential equation to predict the evolution of these intentions over the next five seconds.

[0339] When the system detects a sudden braking event of a vehicle 400 meters ahead, the prediction model shows that the surrounding vehicle's "slow down" intention will exceed the "maintain speed" intention in 2.3 seconds, forming a critical point of alternating intentions. The specific predicted trajectory is:

[0340] At 0 seconds: maintain vehicle speed (0.72), decelerate (0.28);

[0341] At 1 second: maintain vehicle speed (0.68), decelerate (0.32);

[0342] At 2 seconds: maintain vehicle speed (0.54), decelerate (0.46);

[0343] 2.3 seconds: critical point - maintain vehicle speed (0.50), slow down (0.50);

[0344] At 3 seconds: maintain vehicle speed (0.42), decelerate (0.58);

[0345] The system accurately identifies the tipping point by calculating the first and second derivatives of the intent intensity. Near the tipping point, the first derivative of the "slow down" intent reaches a maximum of 0.17 / second, and the second derivative turns from positive to negative, indicating that the growth rate of the intent intensity first increases and then decreases, consistent with the characteristics of a tipping point.

[0346] Based on this prediction, the system recognized that the surrounding vehicle's deceleration intention would take over 2.3 seconds before the vehicle actually braked, much earlier than traditional systems, which typically require a lag of 0.5 to 1 second to detect overt deceleration. This provided ample reaction time for the host vehicle, enabling the system to generate defensive driving decisions:

[0347] Start gradual deceleration 3 seconds in advance, increasing the following distance from 35 meters to 50 meters;

[0348] Reduce cruising speed from 105km / h to 85km / h;

[0349] Prepare emergency avoidance space on the right side, ensuring a lateral safety distance of at least 4 meters;

[0350] Subsequent verification showed that the vehicle did indeed begin to decelerate significantly about 0.8 seconds after the critical point predicted by the system, and the final deceleration reached -2.3m / s. 2 The main vehicle successfully avoided the potential rear-end collision risk by taking defensive measures in advance, while maintaining a smooth deceleration process (the maximum deceleration was only -0.9m / s 2 ), improving ride comfort.

[0351] In 1,500 sudden deceleration scenario tests, the system detected the critical point of surrounding vehicles' deceleration intentions an average of 2.1 seconds in advance, with a false alarm rate of only 6.3%. Compared with traditional prediction methods based solely on vehicle dynamic characteristics, this system reduced emergency braking incidents by 76.2% and close-following incidents (less than 15 meters apart) by 64.5%.

[0352] Technical effect verification:

[0353] The core technical advantage of this implementation lies in its ability to accurately predict changes in surrounding vehicle driving intentions and proactively take defensive measures based on the predicted results. This is demonstrated below using two key technical performance indicators.

[0354] Lead time for intent change prediction:

[0355] The first key technical benefit of this method is that it can identify the transition point of the surrounding vehicle's driving intention in advance, providing sufficient reaction time for the main vehicle to make decisions. Table 1 shows the comparison of the lead time of intention change prediction between this method and traditional methods in different scenarios:

[0356] Table 1: Comparison of the lead time of intention change prediction between this method and traditional methods in different scenarios;

[0357]

[0358] Data shows that this method significantly predicts driving intention transition points in advance across all test scenarios, with an average lead time of 2.52 seconds, 1.54 seconds earlier than traditional methods, a relative improvement of 157.1%. In sudden deceleration scenarios, this method's lead time is particularly significant, reaching seven times that of traditional methods, providing a valuable window for response in emergencies.

[0359] Safety improvements in defensive driving decisions:

[0360] The second key technical effect of this method is that it generates defensive driving decisions based on intention prediction, significantly improving driving safety. Table 2 shows the comparison of safety indicators between this method and traditional methods in a 5,000-kilometer real-road test:

[0361] Table 2: Comparison of safety indicators between this method and traditional methods in a 5000 km real road test;

[0362] Safety indicators This method Traditional methods Improvement quantity Improvement ratio (%) Emergency braking events (times / 100km) 0.42 1.74 -1.32 -75.9％ Dangerous following incidents (times / 100km) 0.85 2.37 -1.52 -64.1％ Close range interaction events (times / 100km) 1.28 3.69 -2.41 -65.3％ Comfort violation events (times / 100km) 0.93 2.56 -1.63 -63.7％ <![CDATA[Standard deviation of acceleration (m / s 2 )]]> 0.24 0.51 -0.27 -52.9％ Average minimum time interval (TTC, seconds) 4.82 2.35 +2.47 +105.1％

[0363] Data shows that this approach achieved significant improvements across all safety metrics. Emergency braking events decreased by 75.9%, dangerous following incidents by 64.1%, and close-range interactions by 65.3%. Driving comfort was also improved, with the standard deviation of acceleration decreasing by 52.9%, indicating a smoother ride. Of particular note, the average time to collision (TTC) increased from 2.35 seconds to 4.82 seconds, a 105.1% increase, demonstrating that the system provides ample buffer time for potential conflicts.

[0364] Figures 2 to 6 They are the comparison of intention prediction accuracy between the method of the present invention and the traditional method in different scenarios; the curve of intention intensity evolution over time; the relationship between intention prediction lead time and collision risk reduction rate; the multi-dimensional performance comparison between the method of the present invention and the traditional method; and the evolution of multiple intention intensities in the highway ramp merging scenario.

[0365] Based on the above verification results, this implementation effectively solves the prediction lag problem existing in the existing technology by accurately predicting the changes in the driving intentions of surrounding vehicles and generating defensive decisions in advance, thereby significantly improving the safety and comfort of the autonomous driving system in complex traffic environments.

[0366] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A driving decision-making method based on the driving behavior tendency of a surrounding vehicle, characterized by: The following steps are involved: Obtain vehicle trajectory data and surrounding environment information through on-board sensors, including position, speed, acceleration and other time series data; Based on the acquired data, the driving behavior causal graph structure is used to implement counterfactual reasoning, effectively distinguishing the influence of the driver's internal intention and external environmental constraints. Based on the identified intrinsic intentions, a hierarchical driving intention graph is formed to represent the hierarchical relationship between different intentions. At the same time, an intention conflict measurement function is applied to quantitatively evaluate the degree of conflict between intentions. Using the generated hierarchical intention graph structure and conflict measurement results, an intention competition dynamics model is established based on differential equations to characterize the mutual inhibition and promotion relationship between multiple driving intentions and their dynamic evolution over time. The established intention competition dynamics model is combined with recursive neural differential equations to predict the changes in intention intensity and identify critical points. Based on the changes in intention intensity, the behavior changes of surrounding vehicles are predicted and the defensive driving decisions of the main vehicle are generated.

2. The driving decision-making method based on the driving behavior tendency of the surrounding vehicle according to claim 1 is characterized in that: The steps of implementing counterfactual reasoning using the driving behavior causal graph structure include: Construct a directed acyclic graph containing environment state nodes, surrounding vehicle observation state nodes, surrounding vehicle driving intention nodes, and surrounding vehicle behavior nodes to represent the causal structure of driving behavior; Implementing structured causal models to infer how behavior changes when environmental conditions change but driving intentions remain constant; Train an intent extraction network to extract intrinsic intent representations that are independent of environmental conditions.

3. The driving decision-making method based on the surrounding vehicle driving behavior tendency according to claim 2, characterized in that: The intent extraction network is trained by minimizing a combined loss function consisting of behavior reconstruction loss and environment invariance loss, where the environment invariance loss is implemented through adversarial training to ensure that the extracted intent representation remains stable under different environmental conditions.

4. The driving decision-making method based on the driving behavior tendency of the surrounding vehicle according to claim 1, characterized in that: The hierarchical driving intention graph structure is a directed graph, in which nodes represent driving intentions and edges represent hierarchical inclusion relationships between intentions; The intention conflict measurement function quantifies the degree of conflict between two intentions based on the overlap of behavior distribution under intention conditions.

5. The driving decision-making method based on surrounding vehicle driving behavior tendency according to claim 1, characterized in that: The intention competition dynamics model describes the change of each intention strength value over time through a differential equation: Each intention has an intrinsic growth rate parameter, which represents the strengthening tendency of the intention itself; The upper capacity parameter indicates the maximum intensity that the intention can reach without external interference; The inter-intention competition coefficient matrix represents the mutual inhibition relationship between different intentions; and a scene-dependent external motivation function that maps environmental changes to changes in intent strength.

6. The driving decision-making method based on the driving behavior tendency of the surrounding vehicle according to claim 1, characterized in that: The recursive neural differential equation consists of the following parts: Recursive structure to capture the evolving patterns of intent over long time series; Parameterized neural networks for approximating differential equations of intention competition dynamics; A mapping network that maps the current hidden state to the dynamics model parameters; The differential equation is solved by numerical integration method to predict the intention intensity value at future time points.

7. The driving decision-making method based on the driving behavior tendency of the surrounding vehicle according to claim 1, characterized in that: The steps of predicting the change in the intensity of the intention and identifying the critical point include: Calculate the second derivative of intention intensity; When the absolute value of the second-order derivative exceeds the preset threshold and the sign of the first-order derivative changes, it is determined to be a potential critical point of intention; Based on the identified critical points, the dominant intention after the transition is calculated, the possible behavior distribution of surrounding vehicles is predicted, and the defensive driving decision of the host vehicle is generated.

8. The driving decision-making method based on surrounding vehicle driving behavior tendency according to claim 7, characterized in that: Defensive driving decisions are generated by maximizing the safety margin, that is, choosing the action that maximizes safety under the most dangerous behavior of surrounding vehicles among all possible actions of the main vehicle; The safety function measures the safety of the main vehicle's actions when the surrounding vehicles perform specific behaviors.

9. The driving decision-making method based on surrounding vehicle driving behavior tendency according to claim 1, characterized in that: Also includes real-time update and adaptive adjustment steps: As new observation data are continuously acquired, the parameters of the intention competition dynamics model are continuously updated; Dynamically adjust the sensitivity threshold of critical point detection based on the accuracy of historical predictions; Through the adaptive adjustment mechanism, the proportion of false warnings can be controlled while maintaining a high detection rate, thereby improving the reliability and effectiveness of defensive driving decisions.

10. A driving decision system based on the driving behavior tendency of a surrounding vehicle, used to execute the driving decision method based on the driving behavior tendency of a surrounding vehicle according to any one of claims 1 to 9, characterized in that: include: The sensor module is used to obtain vehicle trajectory data and surrounding environment information through on-board sensors, including position, speed, acceleration and other time series data; The intention recognition module is used to implement counterfactual reasoning using the driving behavior causal graph structure, distinguish internal intentions from the influence of external environmental constraints, and form a hierarchical driving intention graph structure to represent the hierarchical relationship between different intentions; The dynamics modeling module is used to establish an intention competition dynamics model based on differential equations, characterizing the mutual inhibition and promotion relationship between multiple driving intentions and their dynamic evolution over time; A prediction module, which combines recursive neural differential equations to predict changes in intent intensity and identify critical points; The decision-making module is used to generate defensive driving decisions for the host vehicle based on the intention prediction results.

Images