A method for quantifying heterogeneous risks of intelligent vehicle-human interaction conflicts and trajectory prediction

Abstract

This invention discloses a method for risk quantification and trajectory prediction considering the heterogeneity of interactions between intelligent vehicles and humans, relating to the fields of intelligent transportation systems, autonomous driving, and vehicle-road cooperative technologies. It generates a dynamic risk heatmap and constructs a model representing the strength of pedestrian and vehicle decision-making indices. The initial risk features extracted from the dynamic risk heatmap are concatenated with the model representing the strength of pedestrian and vehicle decision-making indices to form a risk feature vector F. risk ; Depth trajectory feature vector F traj With risk feature vector F risk After feature fusion, each future T is predicted. pred The system predicts the trajectory at each time step and calculates the estimated risk value for each prediction point. By probabilistically transforming the potential field and using a dynamic index of the strength relationship between pedestrians and vehicles, it describes behavioral uncertainty and analyzes the risk game from a heterogeneous perspective, thereby improving the accuracy and safety of trajectory prediction.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation systems, autonomous driving and vehicle-road cooperation technologies, specifically a method for quantifying and predicting the trajectory of heterogeneous risks in intelligent vehicle-human interaction conflicts. Background Technology

[0002] Unsignaled intersections are crucial nodes in urban and rural road networks. Due to the lack of clear right-of-way signals, the interactions between pedestrians, non-motorized vehicles, and motorized vehicles in these spaces exhibit high dynamism, randomness, and complexity. Traffic participants must rely on real-time inference and strategic maneuvering of each other's intentions and behaviors to ensure safe passage. This uncertainty makes unsignaled intersections a high-risk area for traffic accidents, especially those involving vulnerable road users. With the development of autonomous driving technology and the advancement of vehicle-to-infrastructure (V2I) systems, endowing machines with the same or even higher scene understanding and prediction capabilities as humans is a key prerequisite for achieving safe passage through such scenarios with high-level autonomous driving (Level 3 and above) and improving the overall safety efficiency of intelligent transportation systems. Therefore, accurately and dynamically quantifying the risk of pedestrian-vehicle conflicts at unsignaled intersections and achieving highly reliable trajectory prediction based on this quantification has significant theoretical and practical value.

[0003] In assessing pedestrian-vehicle conflict risks at unsignalized intersections, existing methods largely rely on static or discrete metrics. For example, they calculate potential conflicts by setting fixed distance or time thresholds or based on a single motion assumption (such as uniform speed). While computationally simple, these methods suffer from a core flaw: reducing a dynamic, continuous risk process to a static or fragmented judgment. They also implicitly assume that pedestrians and vehicles have homogeneous responses to risk, ignoring the structural differences in their physical attributes, decision-making mechanisms, and behavioral preferences. In reality, the intentions, acceleration changes, and interaction states of traffic participants are constantly evolving. Static metrics struggle to capture the spatiotemporal gradual changes and real-time urgency of risk, and cannot distinguish between the higher risk aversion sensitivity exhibited by pedestrians due to their vulnerability and the differentiated game strategies that vehicles might adopt due to their stronger control capabilities in the same risk scenario. This results in assessments that often lag behind actual scenario changes, failing to provide precise and timely input for prediction and decision-making.

[0004] In trajectory prediction, mainstream deep learning-based models, including the current advanced Transformer architecture, typically focus on learning patterns from historical trajectory sequences themselves. These models lack mechanisms for encoding and fusing explicit, quantitative risk characteristics; essentially, they remain "trajectory pattern recognizers." In normal driving conditions with low conflict risk, these models can accurately predict future vehicle and pedestrian trajectories. However, in complex traffic environments characterized by close-range interaction between vehicles and pedestrians, rapidly escalating risks, and rapidly evolving dynamics, the models fail to perceive the intensity of dynamic risks or distinguish the differences in the subjects influencing those risks. Consequently, the reliability and safety of their trajectory predictions significantly decrease.

[0005] Existing research attempting to combine risk perception and prediction often simply incorporates risk indicators as static features into the model. This approach not only fails to adaptively adjust the contribution weights of risk features and original trajectory features based on real-time risk levels, but more importantly, the risk quantification sources it relies on typically lack the ability to characterize the heterogeneity of people and vehicles, resulting in structural biases inherent in the risk features input to the model. Given these inherent deficiencies in risk quantification sources, simple feature fusion cannot fundamentally improve predictive performance. Therefore, there is an urgent need for a solution that can describe behavioral uncertainty using probabilistic methods, analyze risk games from a heterogeneous perspective, and drive the predictive model's dynamic focusing and differentiated responses. This would systematically fill the multiple gaps in current technology regarding dynamic risk quantification, heterogeneous modeling, and adaptive fusion. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention aims to provide a method for quantifying and predicting trajectory risks that consider the heterogeneity of interactions between intelligent vehicles and people. By using a probabilistic potential field and a dynamic index representing the strength of the interaction between pedestrians and vehicles, the method describes behavioral uncertainty and analyzes risk games from a heterogeneous perspective. Furthermore, through a gating adaptive fusion mechanism, the prediction model can dynamically adjust its focus on risk characteristics based on real-time risk levels, thereby significantly improving the accuracy and safety of trajectory prediction at critical moments when conflicts are frequent.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for quantifying and predicting the trajectory of heterogeneous risks in interactions between intelligent vehicles and humans includes: (1) Grid the intersection, obtain the trajectory of each traffic individual at the intersection, and use the probabilistic artificial potential field to calculate the total risk potential field value of each pixel in the style of each traffic individual, and generate a dynamic risk heat map. (2) Based on evolutionary game theory, the conflict loss between pedestrians and vehicles is dynamically mapped in a differentiated manner, and a model is constructed to represent the decision strength index of pedestrians and vehicles. (3) Analyze the individual traffic trajectories, assemble the analyzed motion data into a state sequence, and then use the Transformer encoder to extract the depth trajectory feature vector F. traj The initial risk features of the dynamic risk heatmap are extracted using CNN, and the initial risk features are concatenated with the risk feature vector F formed by the model representing the decision strength index of pedestrians and vehicles. risk ; (4) Adaptively allocate depth trajectory feature vector F based on the real-time risk urgency of the current scenario. traj With risk feature vector F risk Feature fusion is performed using weights; (5) Input the fused features into the Transformer decoder to predict each future T pred Trajectory prediction at each time step involves performing a secondary probabilistic artificial potential field risk calculation on the target vehicle based on the future location coordinates of the traffic individual, thereby obtaining the estimated risk value corresponding to each prediction point.

[0008] In this invention, for any traffic participant, the total risk potential field U is defined as follows: For any location coordinate q to be evaluated at an intersection. total (q) is the superposition of the gravitational potential and the repulsive potential generated by all obstacles, and the calculation formula is as follows: ; in, Represents gravitational potential. This represents the repulsive potential generated by the i-th obstacle; Gravitational potential Calculation formula: ; q goal Assign values ​​to the target location coordinates of individual traffic vehicles based on their trajectories; K att These are probabilistic gravity gain coefficients; ; ; ; in, m att , s att They represent the probabilistic gravity gain coefficients, respectively. The mean and standard deviation, k 1 and b 1θ is a calibration constant greater than zero, representing the angle between the current orientation of the traffic individual and the target direction; v represents the current speed of the traffic individual; r traffic Indicates the density of vehicles in the surrounding area; This represents random noise that follows a standard normal distribution. No. i An obstacle generates a repulsive potential. Calculation formula: ; in, d is the probabilistic repulsive force gain coefficient; i Let q be the distance from the location of the obstacle. Euclidean distance, ; This represents the probabilistic influence of obstacles on distance;

[0009] They represent the probabilistic repulsion gain coefficients, respectively. The mean and standard deviation of the values; TTC is the predicted collision time between the two; I right-of-way This is the right-of-way indicator function; , β are respectively , The weighting coefficient; σ TTC The uncertainty of TTC forecasts based on historical data or model estimates; For σ TTC Weighting coefficients; Basic behavioral fluctuation term; Let t represent the mean and standard deviation of the probabilistic obstacle influence distance, respectively; react Typical perception and reaction time; a dec For comfortable deceleration; d safe The minimum safe interval after coming to a standstill; Indicates the proportionality coefficient; The visibility factor is normalized based on weather and lighting conditions; This represents random noise that follows a standard normal distribution.

[0010] In this invention, the total risk potential value of each pixel in the style is assigned to the heatmap matrix. The corresponding elements in the matrix; finally, the matrix... Normalization is performed, and different colors are mapped according to the value to generate a dynamic risk heatmap.

[0011] In this invention, a model is constructed to characterize the decision-making strength index of pedestrians and vehicles. or ( t): We construct dynamic equations to describe the evolution of the probability x(t) of pedestrians choosing to cross and the probability y(t) of vehicles choosing to cross: ; Where J represents pedestrian yielding loss; M represents pedestrian waiting loss; R(t) represents pedestrian conflict loss at time t; K represents vehicle yielding loss; N represents vehicle waiting loss; and S(t) represents vehicle conflict loss at time t. ; R0 and S0 are the baseline collision losses for pedestrians and vehicles, respectively. and The heterogeneity coefficient characterizing the inherent risk sensitivity of pedestrians and vehicles; function and The continuous physical risk values ​​for pedestrians and vehicles are mapped to conflict losses in the game-theoretic decision-making model, thereby coupling physical risk with behavioral decision-making logic. The calculation formula is as follows: ; in, functions respectively and The magnitude parameter determines the maximum incremental loss that the risk can cause; Representing functions respectively and The slope parameter; Representing functions respectively and The risk perception threshold parameter controls the position of the central growth point of the control function; ; in, Based on the base delay time, and These represent the unit time costs for pedestrians and vehicles, respectively. and These are the time value sensitivity coefficients for pedestrians and vehicles, respectively. and These are fixed penalties for pedestrians and vehicles, respectively. and These are the amplification factors for the mutual loss of pedestrians and vehicles, respectively. and These are additional game-theoretic penalties for pedestrians and vehicles, respectively. Represents a constant.

[0012] In this invention, the depth trajectory feature vector F is adaptively allocated. traj With risk feature vector F risk The feature fusion process based on the weights is as follows: F risk A scalar gating value α is generated through a multilayer perceptron (MLP). The magnitude of α directly reflects the system's assessment of the current risk level: α approaching 1 indicates a high-risk state, and the model needs to be highly vigilant; α approaching 0 indicates a low-risk normal state. ; The two types of features are dynamically weighted and fused using a gating value α to form the final contextual feature F. context : ; in, The activation function is sigmoid; W r and W t The depth trajectory feature vectors F are respectively traj With risk feature vector F risk Learnable linear transformation matrix.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a probabilistic artificial potential field, normalizing key parameters to enable the risk field to dynamically reflect behavioral uncertainty. Simultaneously, through a gating adaptive fusion mechanism, the prediction model can dynamically adjust its focus on risk characteristics based on real-time risk levels. This combination fundamentally upgrades the prediction model from a passive "pattern recognizer" to an active "risk-aware inference engine," thereby significantly improving the accuracy and safety of trajectory prediction during critical moments of high conflict incidence.

[0014] 2. This invention defines the gravitational and repulsive force gains and the influence distance as parameters that follow a normal distribution, enabling the model to characterize the general tendency of behavior and individual random fluctuations through the distribution mean and variance, respectively. This design makes the generated risk field a "fuzzy" probability field rather than a "hard" boundary field, which is more in line with the uncertainty and fuzzy judgment characteristics of human drivers and pedestrians in actual decision-making.

[0015] 3. This invention utilizes a risk-loss mapping function based on evolutionary game theory to map the same physical risk value into conflict loss parameters that differ between pedestrians and vehicles, thereby constructing a dynamic index reflecting the relative strength of the game between the two parties. This index enables the model not only to perceive "how great the risk is" but also to identify "who is more disadvantaged by the risk," thus more accurately determining which party will avoid the conflict and which party is likely to persist in passing during conflict evolution. This significantly improves the anthropomorphism and decision-making rationality of trajectory prediction in highly interactive and high-risk scenarios.

[0016] 4. The gating network introduced in this invention can automatically adjust the contribution weights of trajectory features and risk features in prediction based on the comprehensive risk value calculated in real time. Its advantages are twofold: in low-risk, normal scenarios, the model can primarily rely on efficient historical trajectory patterns for prediction, saving computational resources; in high-risk, conflict-critical scenarios, the model can immediately shift its focus to risk features, prioritizing the generation of risk avoidance predictions, thus intelligently balancing prediction efficiency and safety redundancy.

[0017] 5. This invention is not a "black box" end-to-end model. Its technical solution modules are clearly defined: a probabilistic potential field provides an interpretable risk map, dual encoders separate feature extraction, and a gating mechanism provides transparent logic for fusion decision-making. This design not only facilitates fault diagnosis and performance optimization, but more importantly, each frame of the predicted trajectory output is accompanied by a corresponding dynamic risk field and game strength index. This provides autonomous driving decision-making systems (such as planning and control modules) with an intuitive basis for "why the prediction is made this way" (i.e., where the high-risk areas are and what the game situation is), thereby supporting the generation of safer and more human-like driving strategies and making it easy to integrate and apply in real autonomous driving systems. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 This is a schematic diagram of the Transformer model structure of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] This invention discloses a method for quantifying the heterogeneous risk of interaction conflicts between intelligent vehicles and humans and for trajectory prediction, including: (i) Total risk potential field and dynamic risk density diagram of continuous space at intersection A probabilistic risk field dynamic construction module is used to quantify the collision risk of human-vehicle interaction in complex traffic environments. This method calculates the risk value of human-vehicle collision through an improved artificial potential field algorithm.

[0022] To overcome the fundamental shortcomings of traditional artificial potential field models, such as fixed parameters and inability to reflect the uncertainty of traffic behavior, this invention introduces probability distributions to describe key parameters and establishes a dynamic adaptive mechanism to construct a continuous risk field that closely reflects the stochastic characteristics of real-world human-vehicle interactions. The specific implementation of its technical solution is as follows: 1. Probabilistic definition and setting basis of core parameters The traditional artificial potential field method is a mathematical model that simulates the motion of an object in a virtual force field, with a total potential field U. total Typically determined by the gravitational potential U att and repulsive potential U rep The basic formula for superposition can be expressed as: ; in, This represents the agent's position in the current scene. The gravitational potential is proportional to the distance from the agent to the target point, while the repulsive potential increases sharply when the agent approaches an obstacle. The specific expression is generally written as: ; In the formula, d g This represents the distance from the agent to the target point. d o The distance from the agent to the obstacle. k att This is the gravitational gain coefficient. k rep The repulsive force gain coefficient, r 0 Distance is affected by obstacles. K att , k rep , r 0 These three parameters are preset fixed constants in traditional models, which cannot simulate the behavioral uncertainties caused by individual differences, driving styles, environmental interference, etc. in real traffic. Therefore, this invention replaces the above fixed constants with random parameters that follow a normal distribution.

[0023] Probabilized Gravitational Gain Coefficient K att This parameter characterizes the strength of a vehicle's or pedestrian's tendency towards a predetermined target (such as a desired lane or waypoint). Its probability distribution is determined by the mean. m att and standard deviation s att Decide.

[0024] Among them, the mean m att The calculation is directly related to driving intentions and efficiency requirements. At intersections, it can be based on the angle θ (range [0,π]) between the current orientation and the target direction, and the current speed of the individual traffic vehicle. v Configure: ; in, k 1 and b 1 The calibration constant is greater than zero. This design ensures the strongest gravitational reference for straight-ahead (θ=0) intentions and weakens the gravitational reference when turning (θ increases), consistent with the general behavioral characteristics of drivers who are typically more cautious and have a relatively lower target orientation when turning. Current speed of the individual in traffic. v The introduction of this reflects the potential need for path tracking efficiency at higher speeds.

[0025] Standard deviation s att The calculation is used to characterize the deterministic differences in behavior exhibited by different individuals when approaching the same goal, and is modeled as a correlation with traffic flow conditions (surrounding vehicle density). r traffic Related functions: ; in, k 2 , b 2 This is a calibration constant greater than zero. This formula implies that at intersections with heavy traffic and complex interactions (…), r traffic Larger differences in driver behavior patterns emerge. s att (Increase), for example, some people may stick to their original driving intentions, while others are more easily distracted and change their driving intentions.

[0026] Probabilized repulsion gain coefficient This parameter characterizes the strength of a traffic participant's willingness to avoid conflict risk. Its probability distribution is determined by the mean. and standard deviation Decide.

[0027] mean The calculation is related to the instantaneous conflict risk and traffic rule weights. For an interaction between a vehicle and a pedestrian or other vehicle, it is defined as: ; Where TTC is the estimated collision time between the two. −1 The larger the value, the more urgent the risk. right-of-way This is a right-of-way indicator function. If the vehicle has the right-of-way (e.g., proceeding straight on a green light), this value can be set to a constant c1; if it does not have the right-of-way (e.g., yielding when turning), it is set to a constant c2, where c2>c1, to reflect the mandatory yielding obligation required by regulations. ,β are respectively , The weighting coefficient is used to reconcile the contribution ratio of immediate risk and rule constraints.

[0028] Standard deviation The calculation: used to characterize the differences in risk tolerance among different participants, and is positively correlated with risk uncertainty: ; Where, σ TTC The TTC prediction uncertainty is estimated based on historical data or models. γ is the sensitivity conversion coefficient, a positive-zero proportionality coefficient that measures the proportion of risk uncertainty that the system converts into behavioral uncertainty. δ is the underlying behavioral volatility term, a positive-zero constant that characterizes the inherent differences in individual decision-making styles. The more difficult it is to accurately predict risk, the greater the differences in hedging strategies exhibited among individuals.

[0029] Probabilistic influence distance This parameter defines the spatial threshold at which other traffic individuals begin to have a significant risk impact on traffic individual i. Its probability distribution is determined by the mean. and standard deviation Decide.

[0030] mean The calculation focuses on the safety buffer distance, which covers the distance required for the perception-reaction-braking process and is related to speed. ; in, v Let t be the current speed of traffic individual i. react This represents typical perception and reaction time (e.g., 1.5 seconds). dec For comfortable deceleration. d safe This represents the minimum safe distance after the object has come to a standstill. This formula provides a physical benchmark for distance-affecting deterministic calculations based on kinematics.

[0031] Standard deviation The calculation is used to characterize fluctuations in perceived safety boundaries caused by distraction, environmental disturbances (such as weather and lighting), or individual differences in vigilance. The model is as follows: ; in, or η is a proportionality coefficient. According to research in traffic engineering and driving behavior, drivers' judgments of safe following distance typically vary by about 10% to 25% when following or avoiding obstacles. Therefore, η can be set within this range, for example, 0.15. Visibility_Factor is a visibility factor normalized based on weather and lighting conditions (1 represents best, 0 represents worst). The lower the visibility, the greater the difference in safe distance judgments among different drivers.

[0032] 2. Probabilistic Potential Field Calculation and Risk Map Generation Process Based on the aforementioned probabilistic definition of the core parameters, this invention calculates the dynamic risk field and generates a risk map and risk value through the following steps: (2.1) Probabilistic parameter sampling For the i-th obstacle in the scene The sampling formula is: ; in, It is random noise sampled from a standard normal distribution.

[0033] (2.2) Calculate the total risk potential field of each traffic parameter point by point. For any traffic participant, at any position q to be evaluated within the intersection, its total risk potential field U total (q) is the superposition of the gravitational potential and the repulsive potential generated by all obstacles, and the calculation formula is as follows: ; in, Represents gravitational potential. This represents the repulsive potential generated by the i-th obstacle; Gravitational potential Calculation: Propelling individual traffic vehicles toward their target coordinates q goal .

[0034] ; q goal To determine the destination of each individual traffic vehicle, a target coordinate is assigned based on its historical trajectory. For example, for vehicles, the vehicle's entrance lane and direction of travel are analyzed based on its historical trajectory, and its exit lane is predicted, thus assigning a target coordinate. For pedestrians, their crossing direction is analyzed based on their historical trajectory, and a target coordinate is assigned to them.

[0035] Repulsive potential calculation : Generated by the i-th obstacle, protecting the area around it.

[0036] ; in, Let q be the distance from the location of the obstacle. The Euclidean distance.

[0037] (2.3) Generation of risk heat map The intersection is meshed according to practical application requirements. Each pixel (x, y) in the mesh is traversed, and its corresponding physical location is q(x, y). The total potential field value at that location for each individual traffic is calculated using the formula described above. This value is then assigned to the heatmap matrix. The corresponding element in the matrix. Mathematically, this can be represented as: Finally, for the matrix The data is normalized and mapped to different colors based on its value (e.g., a continuous color gradient of "green-yellow-red" can be used to represent the transition from low risk to high risk). The resulting image is a dynamic risk heatmap, where darker colors (such as red) indicate higher instantaneous conflict risks, providing the most intuitive global risk spatial distribution for autonomous driving systems.

[0038] (2.4) Structured output of risk value To provide downstream algorithm modules such as game strength index and trajectory prediction with accurate quantitative analysis, the probabilistic artificial potential field model synchronously outputs structured risk data corresponding to the heatmap. Specifically, for each traffic participant, it outputs the individual risk value corresponding to the current location of pedestrians and vehicles in real time, denoted as Risk. p (t) and Risk v (t). These values ​​are obtained by mapping the real-time coordinates of pedestrians and vehicles to the risk matrix. These values, obtained through interpolation, directly and quantitatively reflect the level of conflict risk currently faced by all parties and serve as the core input for the subsequent heterogeneous game analysis module involving people and vehicles.

[0039] pedestrian With vehicles At any moment The corresponding individual risk value is determined by the value of the total risk potential field function at its respective real-time location, and its mathematical expression is as follows: To address the fundamental shortcomings of heterogeneity in logic and behavior, this invention, based on a probabilistic risk field, further introduces a collision risk heterogeneity modeling module. This module, based on evolutionary game theory, performs differentiated dynamic mapping of conflict losses between vehicles and pedestrians, calculates key indices representing the strength of decisions, and ultimately achieves precise utilization of heterogeneous risk characteristics through an improved adaptive gating mechanism.

[0040] 1. Probabilistic Risk Field and Heterogeneous Loss Mapping Unlike traditional methods, this invention argues that the same physical risk value implies different "decision losses" for pedestrians and vehicles. Pedestrians are more vulnerable, and their conflict loss R is more sensitive to changes in risk; vehicles, due to their stronger control and protection, exhibit different patterns of change in their conflict loss S. Therefore, an independent risk-loss mapping function is designed: ; Where R(t) and S(t) represent the conflict loss between pedestrians and vehicles at time t, respectively, and R0 and S0 represent the baseline conflict loss between pedestrians and vehicles, respectively. Based on the concept of quantifying gains and losses in game theory, the baseline loss can be associated with the typical behavioral costs (such as comfort deceleration) caused by traffic conflicts. Under different traffic scenario data, the average deceleration intensity of pedestrians and vehicles in conflict interactions can be statistically obtained. After normalization, these values ​​are used as the initial estimates of R0 and S0, respectively. The initial value of R0 is greater than that of S0 to reflect the higher basic risk sensitivity of pedestrians. and Heterogeneity coefficient (usually) characterizing the inherent risk sensitivity of pedestrians and vehicles ).

[0041] function and The continuous physical risk values ​​for pedestrians and vehicles are mapped to conflict losses in the game-theoretic decision-making model, thereby coupling physical risk with behavioral decision-making logic. The calculation formula is as follows: ; in, functions respectively and The magnitude parameter determines the maximum incremental loss that the risk can cause. Representing functions respectively and The slope parameter; Representing functions respectively and The risk perception threshold parameter controls the location of the central growth point of the control function; the smaller the C value, the more significant the increase in loss is perceived at a lower risk level. These parameters can be initialized based on prior knowledge from traffic engineering, accident statistics, and psychological research, and used as trainable parameters. During the training of the trajectory prediction model, they are jointly optimized to convergence along with the risk perception and trajectory prediction network weights.

[0042] 2. Evolutionary Dynamics and Quantification of Strength-Weakness Relationships in Human-Vehicle Conflicts Substituting the aforementioned dynamic conflict losses R(t) and S(t), along with pedestrian waiting losses M, vehicle waiting losses N, pedestrian yielding losses J, and vehicle yielding losses K, into the evolutionary game dynamics model, the evolutionary process of the probability x(t) of pedestrians choosing to pass and the probability y(t) of vehicles choosing to pass is described by replicating the dynamic equations: ; Based on this, the present invention defines a strength index for the relationship between pedestrians and vehicles. or ( t ): ; Where ϵ is a minimal constant to prevent division by zero. η(t)∈[−1,1], the closer its value is to 1, the more dominant the pedestrian is in the game (the stronger the tendency of "cars to yield to pedestrians"); the closer it is to -1, the more dominant the vehicle is (the stronger the tendency of "pedestrians to yield to vehicles"); and the closer it is to 0, the more it indicates a stalemate or mutual yielding.

[0043] The specific values ​​of the above parameters are determined by the following formula: ; in, Based on the base delay time, and These represent the unit time costs for pedestrians and vehicles, respectively. and The time value sensitivity coefficients for pedestrians and vehicles are respectively ( This reflects that pedestrians are more sensitive to delays. and These are fixed penalties for pedestrians and vehicles, respectively. and These are the amplification factors for the mutual loss of pedestrians and vehicles, respectively. and These are additional game-theoretic penalties for pedestrians and vehicles, respectively.

[0044] and These parameters are initially calibrated based on traffic statistics and are used as trainable parameters. During the training process of the trajectory prediction model of this invention, they are jointly optimized to convergence along with the risk perception and trajectory prediction network weights.

[0045] (iii) Trajectory prediction for intelligent vehicles considering the heterogeneity of collision risk After completing the probabilistic dynamic risk quantification of the artificial potential field and the heterogeneity of human-vehicle collision risk, the core of this invention lies in how to deeply and adaptively integrate this dynamic risk information into the trajectory prediction model. To this end, this invention constructs a trajectory prediction model with a Transformer-based core architecture and an embedded dynamic risk quantification module. This model, through a unique dual-stream coding and gating fusion mechanism, achieves collaborative modeling and adaptive utilization of risk features and trajectory features. Its overall architecture and workflow are as follows.

[0046] 1. Technical architecture of trajectory prediction Trajectory prediction is performed using an "encode-fusion-decode" paradigm. The input consists of two paths: one is the historical trajectory sequence of the target pedestrian and surrounding vehicles; the other is a dynamic risk density map generated in real-time by the probabilistic artificial potential field model, corresponding to the current scene. These two data streams are processed by independent encoders for feature extraction. Subsequently, an adaptive gating fusion module dynamically fuses the two types of features based on the current overall risk level. Finally, a Transformer decoder generates a multi-time-future trajectory prediction distribution based on the fused contextual features rich in risk perception information. The entire processing flow achieves seamless integration from raw data to risk perception prediction. The overall network framework diagram is shown in the attached figure.

[0047] 2. Dual-stream feature extraction To fully extract features from heterogeneous input data, the model employs a parallel dual encoder: Track encoder: Employs a standard Transformer encoder. Based on the trajectories of each individual traffic participant, it extracts their motion data (including position, velocity, acceleration, etc.) and tracks all traffic participants (pedestrians P1, P2... vehicles V1, V2...) within a historical time window T. hist The state sequence within the sequence is input into the Transformer encoder, where the state at each time step includes position, velocity, acceleration, etc. The sequence is converted into a feature vector through an embedding layer, and spatiotemporal position encoding is added. The Transformer encoder's multi-layer self-attention mechanism effectively captures the complex relationships within the sequence (temporal dependencies of each participant) and between sequences (social interactions between participants), outputting a deep trajectory feature vector F. traj .

[0048] Risk Feature Encoder: Responsible for transforming dynamic risk density maps into high-level semantic risk features. This encoder is typically composed of a lightweight convolutional neural network. Each pixel value in the input risk density map represents the conflict potential at that spatial location. The CNN extracts spatial multi-scale patterns (such as the spatial distribution, shape, and intensity gradient of high-risk regions) of the risk map step by step through its convolutional and pooling layers.

[0049] The encoder first extracts initial risk features from a dynamic risk heatmap using a CNN network. These initial risk features are then concatenated with a game index and projected into a global context vector. Subsequently, through feature broadcasting or channel attention mechanisms, contextual information representing the global game situation and individual risk levels is deeply embedded into each location of the spatial feature map, achieving semantic and spatial alignment. Finally, these fused features are aggregated into a compact, fixed-dimensional risk feature vector F. risk .

[0050] 3. Feature fusion Feature fusion is not simply a matter of splicing features together; it aims to solve the problem of "how to effectively combine" risk features and trajectory features. Its core is to use a learnable gating network to adaptively assign importance weights to the two types of features in subsequent predictions based on the real-time risk urgency of the current scenario. The working steps are as follows: (3.1) Input depth trajectory feature vector F traj and risk feature vector F risk , (3.2) Generate the gating signal. First, set F... risk A scalar gating value α is generated using a multilayer perceptron (MLP) (its range is typically constrained to [0, 1] by a sigmoid function). The magnitude of α directly reflects the system's assessment of the current risk level: α approaching 1 indicates a high-risk state, requiring the model to be highly vigilant; α approaching 0 indicates a low-risk normal state.

[0051] ; in, α It is a gating signal. For the activation function sigmoid, W g It is a weight matrix. bg It is a bias term.

[0052] (3.3) Dynamic weighting and fusion of features. The two types of features are dynamically weighted and fused using a gate value α to form the final context feature F. context : ; Among them, W r W t The risk feature vectors F are respectively risk Depth trajectory feature vector F traj A learnable linear transformation matrix is ​​used to map two classes of features to the same semantic space and perform scale adjustment. This formula enables dynamic allocation of model attention: under high conflict risk, the model primarily focuses on F. risk The focus is on risk avoidance in trajectory prediction, guiding the prediction towards avoiding high-risk areas; in safe cruising mode, the model mainly relies on F... traj The encoded social interactions and movement patterns can be used for efficient prediction.

[0053] 4. Risk Perception Trajectory Decoding and Output The fused contextual features F context It is input into the Transformer decoder. The decoder is connected to F... context Perform cross-attention calculations to decode the future T for each individual in the traffic flow. predThe trajectory prediction at each time step is performed. Based on the future location coordinates of individual traffic vehicles, a secondary probabilistic artificial potential field risk calculation is performed on the target vehicle (the vehicle that needs to make driving decision-making and planning). The estimated risk value corresponding to each prediction point is obtained, forming an integrated "trajectory-risk" prediction result, which provides accurate and safe information input for autonomous driving decision-making and planning.

[0054] 5. Training the trajectory prediction model Data is collected from different traffic individuals at the intersection, and the collected data is used to train the trajectory prediction model. All behavioral preference coefficients, whose initial values ​​are calibrated based on traffic statistics, are used as trainable parameters and, within the end-to-end training framework of this invention, are jointly optimized until convergence along with the risk perception and trajectory prediction network weights. The specific process is as follows: Backpropagation and gradient update This is the core of "training." The loss function calculates the predicted trajectory. With the actual trajectory The differences between them.

[0055] The specific expression for its overall loss function is as follows: ; in, and These are preset positive hyperparameters used to balance the weights of various losses. This is the trajectory prediction loss, the dominant loss term, used to measure the difference between the model-predicted multimodal trajectory distribution and the true trajectory. It is typically expressed as a negative log-likelihood loss. ; In the formula, This represents the total number of time steps for prediction. This represents the total number of trajectory modes predicted by the model. For the model number The confidence (probability) of each predicted mode assignment satisfies ; Let be the probability density function of a multivariate Gaussian distribution; For at any time The actual trajectory coordinates; The model predicts the first The modality at time... The mean and covariance matrix of the trajectory points.

[0056] This is the risk regression loss, an auxiliary loss used to constrain the model's prediction of individual risk values, ensuring consistency with the true risk value. ; In the formula This represents the total number of traffic participants (pedestrians and vehicles) in the current training sample. The first calculated by the model The risk value of each participant (i.e.) ); For the first Each participant's risk truth value label. This label can be applied during the preprocessing stage based on established rules (such as time-to-collision). reciprocal (Generated from data)

[0057] This is the regularization loss, an optional regularization term used to prevent overfitting of the game parameters and guide them towards a reasonable numerical range. Here, regularization is applied to the four game parameters. Regularization: ; In the formula, The four game parameters that need to be trained are: pedestrian waiting loss, vehicle waiting loss, and pedestrian yielding loss. Representing vectors Norm, which here is the sum of squares of the parameters.

[0058] loss The generated gradient is propagated along the backpropagation algorithm. The path is reversed.

[0059] Optimizers (such as Adam) based on the propagation to The gradient on the learning rate is used to update the values ​​of these four parameters. The optimization logic is: if the current... Value leads to game index Inaccuracy leads to larger trajectory prediction biases, so these inaccuracies will be adjusted in the hope of producing more accurate results in the next iteration. And more accurate predictions.

[0060] Training convergence and parameter solidification The above process is repeated hundreds or thousands of times on the training set. When the trajectory prediction loss... Training stops when the performance on the validation set no longer decreases significantly, or when the performance on the validation set reaches its optimum. At this point, The values ​​no longer change; they have achieved co-optimization with the parameters of other functional modules in the model.

[0061] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for quantifying the heterogeneity risk and predicting the trajectory of interactions between intelligent vehicles and humans, characterized in that, include: (1) Grid the intersection, obtain the trajectory of each traffic individual at the intersection, and use the probabilistic artificial potential field to calculate the total risk potential field value of each pixel in the style of each traffic individual, and generate a dynamic risk heat map. (2) Based on evolutionary game theory, the conflict loss between pedestrians and vehicles is dynamically mapped in a differentiated manner, and a model is constructed to represent the decision strength index of pedestrians and vehicles. (3) Analyze the individual traffic trajectories, assemble the parsed motion data into a state sequence, and then input it into the Transformer encoder to extract the depth trajectory feature vector F. traj The initial risk features of the dynamic risk heatmap are extracted using CNN, and the initial risk features are concatenated with the risk feature vector F formed by the model representing the decision strength index of pedestrians and vehicles. risk ; (4) Adaptively allocate depth trajectory feature vector F based on the real-time risk urgency of the current scenario. traj With risk feature vector F risk Feature fusion is performed using weights; (5) Input the fused features into the Transformer decoder to predict each future T pred Trajectory prediction at each time step involves performing a secondary probabilistic artificial potential field risk calculation on the target vehicle based on the future location coordinates of the traffic individual, thereby obtaining the estimated risk value corresponding to each prediction point.

2. The method for quantifying and predicting the trajectory considering the heterogeneity of interaction conflicts between intelligent vehicles and humans according to claim 1, characterized in that, For any traffic participant, at any coordinate q of the location to be evaluated at the intersection, the total risk potential field U is... total (q) is the superposition of the gravitational potential and the repulsive potential generated by all obstacles, and the calculation formula is as follows: ； in, Represents gravitational potential. This represents the repulsive potential generated by the i-th obstacle; Gravitational potential Calculation formula: ； q goal Assign values ​​to the target location coordinates of individual traffic vehicles based on their trajectories; K att These are probabilistic gravity gain coefficients; ； ； ； in, μ att , σ att They represent the probabilistic gravity gain coefficients, respectively. The mean and standard deviation, k 1 and b 1 θ is a calibration constant greater than zero, representing the angle between the current orientation of the traffic individual and the target direction; v represents the current speed of the traffic individual; ρ traffic Indicates the density of vehicles in the surrounding area; Represents random noise that follows a standard normal distribution; No. i An obstacle generates a repulsive potential. Calculation formula: ； in, d is the probabilistic repulsive force gain coefficient; i Let q be the distance from the location of the obstacle. Euclidean distance, ; This represents the probabilistic influence of obstacles on distance; ； ； ； ； ； ； They represent the probabilistic repulsion gain coefficients, respectively. The mean and standard deviation of the values; TTC is the predicted collision time between the two; I right-of-way This is the right-of-way indicator function; , β are respectively , The weighting coefficient; σ TTC The uncertainty of TTC forecasts based on historical data or model estimates; For σ TTC Weighting coefficients; Basic behavioral fluctuation term; , Let t represent the mean and standard deviation of the probabilistic obstacle influence distance, respectively; react Typical perception and reaction time; a dec For comfortable deceleration; d safe The minimum safe interval after coming to a standstill; Indicates the proportionality coefficient; The visibility factor is normalized based on weather and lighting conditions; , This represents random noise that follows a standard normal distribution.

3. The method for quantifying and predicting the trajectory of heterogeneous interaction conflicts between intelligent vehicles and humans according to claim 1, characterized in that, Assign the total risk potential value of each pixel in the style to the heatmap matrix. The corresponding elements in the matrix; finally, the matrix... Normalization is performed, and different colors are mapped according to the value to generate a dynamic risk heatmap.

4. The method for quantifying and predicting the trajectory considering the heterogeneity of interaction conflicts between intelligent vehicles and humans according to claim 1, characterized in that, Construct a model representing the strength of pedestrian and vehicle decision-making indices. η ( t ): We construct dynamic equations to describe the evolution of the probability x(t) of pedestrians choosing to cross and the probability y(t) of vehicles choosing to cross: ； ； Where J represents pedestrian yielding loss; M represents pedestrian waiting loss; R(t) represents pedestrian conflict loss at time t; K represents vehicle yielding loss; N represents vehicle waiting loss; and S(t) represents vehicle conflict loss at time t. ； ； R0 and S0 are the baseline collision losses for pedestrians and vehicles, respectively. and The heterogeneity coefficient characterizing the inherent risk sensitivity of pedestrians and vehicles; function and The continuous physical risk values ​​for pedestrians and vehicles are mapped to conflict losses in the game-theoretic decision-making model, thereby coupling physical risk with behavioral decision-making logic. The calculation formula is as follows: ； ； in, , functions respectively and The magnitude parameter determines the maximum incremental loss that the risk can cause; , Representing functions respectively and The slope parameter; , Representing functions respectively and The risk perception threshold parameter controls the position of the central growth point of the control function; ； ； ； ； ； in, Based on the base delay time, and These represent the unit time costs for pedestrians and vehicles, respectively. and These are the time value sensitivity coefficients for pedestrians and vehicles, respectively. and These are fixed penalties for pedestrians and vehicles, respectively. and These are the amplification factors for the mutual loss of pedestrians and vehicles, respectively. and These are additional game-theoretic penalties for pedestrians and vehicles, respectively. Represents a constant.

5. The method for quantifying and predicting the trajectory of heterogeneous interaction conflicts between intelligent vehicles and humans according to claim 1, characterized in that, Adaptively assign depth trajectory feature vector F traj With risk feature vector F risk The feature fusion process based on the weights is as follows: F risk A scalar gating value α is generated through a multilayer perceptron (MLP). The magnitude of α directly reflects the system's assessment of the current risk level: α approaching 1 indicates a high-risk state, and the model needs to be highly vigilant; α approaching 0 indicates a low-risk normal state. ； The two types of features are dynamically weighted and fused using a gating value α to form the final contextual feature F. context : ； in, sigmoid; W r and W t The depth trajectory feature vectors F are respectively traj With risk feature vector F risk Learnable linear transformation matrix.

Images