Method and system for real-time risk assessment of merging and diverging areas in human-machine mixed driving environment

Abstract

The present application relates to the field of intelligent transportation systems and road safety technology, in particular to a risk real-time evaluation method and system for merging area and diverging area in human-machine mixed driving environment, comprising the following steps: S1, real-time collection of multi-traffic data of merging area and diverging area, S2, real-time calculation and output of risk feature set based on multi-traffic data; S3, based on the risk feature set, through a dynamic weighted risk fusion model, real-time calculation and output of a comprehensive risk value representing the current regional safety situation; the dynamic weighted risk fusion model dynamically adjusts the fusion weight of each risk feature according to the real-time traffic state parameter; S4, according to the risk level of the comprehensive risk value, generating and issuing a customized intervention instruction matched with the driving demand of the merging area or diverging area; meeting the real-time requirement of highway risk response, realizing accurate real-time risk evaluation in human-machine mixed driving scene.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation systems and road safety technology, specifically to a method and system for real-time risk assessment of merging and separating zones in a human-machine mixed driving environment. Background Technology

[0002] With the rapid development and gradual commercialization of autonomous driving technology, the road traffic environment will remain in a complex state of coexistence between human-driven and autonomous vehicles for the foreseeable future. Merging zones (where vehicles merge onto the main road from ramps) and diverging zones (where vehicles exit onto ramps) on highways are key bottleneck areas prone to traffic conflicts and accidents. In these areas, frequent vehicle weaving and complex driving intentions pose unprecedented challenges to safety analysis and real-time intervention in human-machine hybrid driving environments.

[0003] Existing traffic safety risk assessment technologies suffer from the following limitations: First, traditional micro-level safety evaluation models, such as Time-to-Collision (TTC) and Post-Encroachment Time (PET), are largely based on assumptions of a purely human-driven environment, assuming that all traffic participants exhibit consistent and predictable behavioral patterns. These models struggle to accurately characterize the strong predictability, rapid response, and potential conservatism exhibited by autonomous vehicles due to safety regulations, and are even less capable of effectively depicting the complex interactions and game-theories arising from differences in behavioral patterns in human-machine hybrid driving environments. Second, existing risk assessment methods are mostly static and macro-level analyses, heavily reliant on historical accident statistics, and lack the ability to finely perceive and respond to real-time dynamic factors (such as changes in day and night illumination, rain and snow, and road surface slipperiness) and traffic flow conditions (such as density and speed distribution). More importantly, current research generally lacks a systematic framework capable of effectively quantifying the penetration rate of autonomous vehicles, their unique behavioral patterns, and key dynamic environmental variables, and coupling them into a unified, real-time risk assessment model.

[0004] Therefore, there is an urgent need for an innovative safety analysis technology that can adapt to the characteristics of human-machine mixed driving and comprehensively consider various dynamic and static risk factors, so as to achieve real-time risk assessment of safety risks in merging and diverging areas of highways. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a method and system for real-time risk assessment of merging and separating zones in a human-machine mixed driving environment.

[0006] To achieve the above objectives, the specific solution of the present invention is as follows:

[0007] This invention provides a method for real-time risk assessment of merging and separating zones in a human-machine hybrid driving environment, comprising the following steps:

[0008] S1. Real-time collection of multi-dimensional traffic data in merging and diverging areas, including vehicle trajectory data, environmental status data, and autonomous vehicle prediction signals obtained through vehicle-road cooperative communication;

[0009] S2. Based on the aforementioned multi-source traffic data, calculate and output a risk feature set in real time; the risk feature set includes micro-interaction conflict indicators, region-specific risk features corresponding to merging or diverging scenarios, environmental correction coefficients, and human-machine hybrid driving structural features.

[0010] S3. Based on the risk feature set, a comprehensive risk value representing the current regional security situation is calculated and output in real time through a dynamically weighted risk fusion model; the dynamically weighted risk fusion model dynamically adjusts the fusion weights of each risk feature according to real-time traffic state parameters.

[0011] S4. Based on the risk level of the comprehensive risk value, generate and issue customized intervention instructions that match the driving needs of the merging or diverging areas.

[0012] Furthermore, the region-specific risk characteristics for the merging zone include: merging gap sufficiency calculated based on the gap between the main road traffic flow and the minimum gap required for the merging vehicle to safely merge, and acceleration coordination coefficient calculated based on the synchronicity of the acceleration curves of the merging vehicle and the vehicles in the adjacent lanes of the main road.

[0013] Furthermore, the region-specific risk features for the diversion zone include: an exit interference coefficient calculated based on the frequency of lane changes by exiting vehicles and the resulting fluctuations in the speed of traffic flow on the main road, and a lane departure warning degree calculated based on the real-time offset of exiting vehicles relative to the lane centerline.

[0014] Furthermore, the human-machine hybrid driving structure features of the present invention are obtained through the following steps:

[0015] Within the set dynamic calculation window, the number of autonomous vehicles and manually driven vehicles is counted, and the ratio of human-machine mixed driving is calculated.

[0016] Extract the instantaneous speed sequence of all vehicles within the window, and calculate the traffic flow speed variance to characterize the speed dispersion.

[0017] The human-machine hybrid driving ratio and the normalized speed variance are weighted and fused to obtain the structural characteristics of the human-machine hybrid driving.

[0018] Furthermore, in step S3, the dynamically weighted risk fusion model adopts a hybrid computing architecture, specifically including:

[0019] S31: Store the mapping relationship between the high-frequency feature combinations obtained from pre-training and the risk values ​​as a lookup table, and perform matching calculations first;

[0020] S32: When the lookup table is not found, a lightweight neural network is started for calculation; wherein, the lightweight neural network is a three-layer fully connected structure, and its network weights are pre-trained and fixed through historical scene data.

[0021] Furthermore, the rules for dynamically adjusting the fusion weights include:

[0022] For merging zones, when the traffic density on the main road is higher than a first preset threshold, the weight of the merging gap sufficiency is increased; and / or, when the proportion of autonomous vehicles is higher than a second preset threshold, the weight of the acceleration coordination coefficient is increased.

[0023] For diversion zones, when the queue length of diverted vehicles exceeds a third preset threshold, the weight of the exit interference coefficient is increased; and / or, when the ambient light intensity is below a fourth preset threshold, the weight of the lane departure warning degree is increased.

[0024] Furthermore, the risk fusion step of this invention also includes an iterative optimization mechanism:

[0025] If no preset high-risk characteristics are detected, the comprehensive risk value is updated iteratively in the first cycle.

[0026] When a preset high-risk feature is detected, the system automatically switches to emergency iteration mode, shortens the iteration cycle to a second cycle, and assigns a higher calculation priority to the high-risk feature; wherein the second cycle is shorter than the first cycle.

[0027] Furthermore, in step S4, the generation and issuance of the customized intervention instruction includes:

[0028] When the comprehensive risk value indicates a medium risk level, a warning and auxiliary operation suggestions are sent to the vehicle terminal or driver of the relevant vehicle.

[0029] When the comprehensive risk value indicates a high risk level, a mandatory intervention plan is generated, which includes active vehicle control commands or roadside facility linkage commands, and is executed through vehicle-road cooperative communication.

[0030] Furthermore, the method of the present invention also includes a feedback optimization step:

[0031] Real-time monitoring of the execution effect of the customized intervention commands;

[0032] If risk mitigation fails to meet expectations within multiple consecutive assessment periods, the intervention intensity will be automatically escalated or the intervention strategy will be adjusted.

[0033] Another aspect of the present invention provides a real-time risk assessment system for merging and separating merging zones in a human-machine hybrid driving environment, used to implement the above-described method, the system comprising:

[0034] The data sensing module is configured to perform the data sensing steps and collect the multi-source traffic data in real time.

[0035] A feature calculation module, connected to the data perception module, is configured to execute the feature calculation step and output the risk feature set;

[0036] A risk fusion module, connected to the feature calculation module, is configured to execute the risk fusion step and output the comprehensive risk value;

[0037] The application output module is connected to the risk fusion module and is configured to execute the application output steps to generate and issue the customized intervention instructions.

[0038] This invention addresses the pain points of safety risk assessment in highway merging and diverging zones under human-machine hybrid driving environments. Through multi-dimensional technological innovation, it constructs a comprehensive "perception-feature-fusion-intervention-feedback" solution, offering the following significant advantages compared to existing technologies:

[0039] This invention addresses two key issues. First, it introduces "human-machine hybrid driving structural features" (integrating human-machine ratio, speed dispersion, and interaction uncertainty), combined with the acquisition of autonomous vehicle prediction signals, to accurately characterize the behavioral differences and interactive game processes of the two types of vehicles. This solves the problem of insufficient adaptability of traditional models to human-machine hybrid driving scenarios, improving the core risk identification accuracy by more than 40% compared to traditional models. Second, through the acquisition of multi-data such as "vehicle trajectory + environmental state + autonomous driving prediction" and the design of specific risk features for merging / diverging zones (such as sufficiency of merging gaps and exit interference coefficients), it achieves risk... The system features a detailed and scenario-based characterization, coupled with a mechanism of "dynamic calculation window (500-meter range, 1-second update) + millisecond-level synchronization + regular / emergency iteration". It also incorporates a hybrid computing architecture of "lookup table method + lightweight neural network" (single latency ≤ 500 milliseconds) and dynamic weighted fusion (adjusting weights in real time based on traffic density, autonomous driving ratio, etc.). This ensures that core risk factors are given priority consideration in key scenarios (improving assessment accuracy by 30% compared to fixed-weight models) and meets the real-time requirements of highway risk response, enabling accurate real-time risk assessment in human-machine hybrid driving scenarios. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the real-time risk assessment method for merging and separating merging zones in a human-machine hybrid driving environment, as described in this invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but this is not to limit the scope of the invention to this.

[0042] like Figure 1 As shown in this embodiment, a real-time risk assessment method for merging and separating merging zones in a human-machine hybrid driving environment includes the following steps:

[0043] Step S1: Collect multi-dimensional traffic data in the merging and diverging areas in real time. The multi-dimensional traffic data includes vehicle trajectory data, environmental status data, and prediction signals of autonomous vehicles obtained through vehicle-road cooperative communication.

[0044] Specifically, this step focuses on real-time performance and collects unique, multi-dimensional data tailored to the differentiated risk scenarios of the two types of areas. For merging areas, roadside millimeter-wave radar, high-definition cameras, and geomagnetic sensors at ramp entrances, acceleration lanes, and main road merging points are deployed to capture the following data in real time: the continuous speed and rate of change of acceleration of vehicles on the ramp from the entrance to the end of the acceleration lane; the remaining available length of the acceleration lane; the continuous time interval of the traffic flow in the target merging lane on the main road; whether vehicles on the main road exhibit obvious yielding behavior (such as deceleration or yaw); and the precise location of the intersection point between the merging vehicle and the main road lane line.

[0045] For the diversion zone, sensors (such as radar and cameras) and traffic counters deployed downstream of the main road and at the exit ramps collect data in real time: the continuous deceleration rate of diverted vehicles from the start of deceleration on the main road to the end of the deceleration lane, the distance from the deceleration start point to the exit ramp, the rate of change of following distance of vehicles in the subsequent lanes on the main road, whether vehicles on the main road show an intention to overtake, and the lateral distance between diverted vehicles and adjacent lanes at the turning point where the deceleration lane separates from the main road.

[0046] In addition, the system receives real-time merging or exit prediction signals (including expected trajectories and times) from autonomous vehicles within the area via a V2X communication unit. Simultaneously, it integrates weather sensors, a high-precision clock module, and road condition sensors to collect real-time environmental data such as light intensity, rainfall, visibility, and road friction coefficient.

[0047] All data from different sources are aligned using millisecond-level time synchronization technology to ensure spatiotemporal correlation with points of convergence or divergence, providing accurate and consistent input for subsequent evaluation.

[0048] Step S2: Based on the multi-source traffic data, calculate and output a risk feature set in real time; the risk feature set includes micro-interaction conflict indicators, region-specific risk features corresponding to merging or diverging scenarios, environmental correction coefficients, and human-machine mixed driving structural features.

[0049] Specifically, this step, based on the real-time data collected in step S2 for the merging and diverging areas, uses a multi-dimensional algorithm model to dynamically extract and quantify key risk characteristics, including the following:

[0050] A1. Calculation of traditional micro-level interaction conflict indicators, specifically including:

[0051] Based on the acquired real-time vehicle location, instantaneous speed, and relative distance data, traditional safety indicators are calculated using a time conflict model:

[0052] Time-based collision warning (TTC): The formula TTC = (relative distance between the two vehicles) / (speed of the following vehicle - speed of the preceding vehicle) is used, where the difference between the speed of the following vehicle and the speed of the preceding vehicle is taken as a positive value, so as to output the TTC value every 0.5 seconds in real time, reflecting the immediate collision risk between vehicles.

[0053] Collision Time Interval (PET): The arrival time difference between the two vehicles at the point of collision is calculated using a trajectory prediction algorithm. When PET < 2 seconds, it is marked as a high-risk interaction event.

[0054] A2. Dedicated calculation of core risk characteristics in merging areas: A three-level quantitative model is designed to address the core risk of "merging-main road coordination" in merging areas.

[0055] A21. Merging gap sufficiency calculation: Extract the time interval sequence of vehicles continuously passing through the merging zone on the main road; combine the expected merging speed of the merging vehicles to calculate the minimum gap threshold that satisfies safe merging, the formula is: minimum gap = main road vehicle speed × reaction time + merging vehicle acceleration distance, the reaction time is taken as the industry standard value of 1.2 seconds; use the ratio of actual gap / minimum threshold as the sufficiency quantification value, the range is [0,1], when the sufficiency quantification value is less than 0.6, it is judged as insufficient gap.

[0056] A22. Acceleration Coordination Coefficient Calculation: Real-time acquisition of acceleration curves of vehicles in adjacent lanes and merging vehicles on the main road; calculation of the synchronicity of acceleration changes between the two using the Pearson correlation coefficient, ranging from [-1,1]. Higher synchronicity indicates stronger coordination; weighted correction of the correlation coefficient based on whether vehicles on the main road actively decelerate, and finally outputting the coordination coefficient.

[0057] A23. Construction of Convergence Environment Correction Coefficient: Based on the environmental parameters collected by roadside equipment, the correction coefficient is output through a fuzzy comprehensive evaluation model: Set the safety threshold for each environmental factor; Use the analytic hierarchy process to allocate the weights of environmental factors, such as rainfall 0.4, visibility 0.3, and illumination 0.3; Calculate the comprehensive impact value and map it to the correction coefficient.

[0058] All characteristic values ​​of the merging zone are updated in real time through a 10-second sliding window to ensure accurate capture of instantaneous risk changes in the merging zone and to ensure that the output results are synchronized with changes in traffic flow status, providing highly timely quantitative input for subsequent risk assessment layers.

[0059] A3. Dedicated calculation of core risk characteristics in diversion zones: A three-level quantitative model is designed for the core risk of "exit-main road interference" in diversion zones.

[0060] A31. Calculation of Exit Interference Coefficient: Extract the real-time trajectory of vehicles exiting within a 1km range of the diversion zone and identify lane-changing behavior characteristics: Statistically count the number of consecutive lane changes by exiting vehicles per unit time and the lateral distance between them and vehicles in adjacent lanes during lane changes; collect vehicle speed fluctuation data of the affected lanes on the main road and calculate the speed standard deviation caused by the interference; construct the interference coefficient quantification formula: Exit Interference Coefficient = (Lane Change Interference Frequency × 0.6 + Speed ​​Fluctuation Intensity × 0.4) × Queue Correction Factor; where the queue correction factor is dynamically adjusted according to the queue length, with a coefficient range of [0,1], and a value higher than 0.7 is considered high-risk interference.

[0061] A32. Lane Departure Warning Calculation: Based on lane line recognition data from a high-definition camera, the lane centerline deviation of the departing vehicle is calculated in real time; combined with the vehicle's current speed, a dynamic deviation threshold is set: threshold = 0.1 × v / 3.6; the ratio of actual deviation to dynamic threshold is used as the basic value of the warning, and driver operation characteristics are added for correction, and the final warning level is output.

[0062] A33. Construction of Diversion Environment Correction Coefficient: Based on environmental parameters collected by roadside sensors, a multi-factor weighted model is used to calculate: the influence weight of each environmental factor is determined, such as light intensity 0.4, road surface humidity 0.3, and sign clarity 0.3, and a classification threshold is set; the influence degree of single factors is quantitatively scored; the weighted total score is calculated and mapped to a correction coefficient, ranging from [0.9, 1.3]. The higher the environmental risk, the larger the coefficient, which is used to amplify the characteristic values ​​of exit interference and lane departure.

[0063] All characteristic values ​​of the diversion zone are updated in real time through a 10-second sliding window to ensure accurate capture of instantaneous risk changes in the diversion zone and provide dynamic quantitative basis for risk assessment.

[0064] A4. Human-machine hybrid driving structural feature calculation: To address the unpredictable behavioral risks in human-machine hybrid driving scenarios, a quantitative model for interaction uncertainty indicators and hybrid driving structure indicators is designed, and a dynamic window algorithm is used to ensure the timeliness of features. The specific process is as follows:

[0065] A41. Calculation of Interaction Uncertainty Index: Based on the identified vehicle type, match the behavioral feature databases of two types of vehicles (acceleration / deceleration response delay of autonomous vehicles and reaction time distribution of human drivers); collect in real time the relative distance, speed difference, and behavioral sequence of any two vehicles within a 500-meter window (including three combinations: autonomous driving-human driving, human driving-human driving, and autonomous driving-autonomous driving); calculate the behavior prediction error for different combinations: using the trajectory prediction model of autonomous vehicles as a benchmark, quantify the deviation between human driving behavior and the predicted trajectory, while considering the probability of sudden decision changes of autonomous vehicles in complex scenarios; use the entropy method to integrate multi-dimensional errors into an interaction uncertainty index, with a range of [0,1]. The higher the value, the more unpredictable the behavior. When the interaction ratio of autonomous driving-human driving is >30%, the index weight is increased by 20%.

[0066] A42. Calculation of Mixed Driving Structure Index: Count the number of autonomous vehicles and manually driven vehicles within a 500-meter dynamic window, calculate the human-machine mixed driving ratio, and calculate the human-machine mixed driving ratio = number of autonomous vehicles / total number of vehicles, accurate to 0.01; extract the instantaneous speed of all vehicles within the window, with a sampling frequency of 5Hz, calculate the speed variance to reflect the dispersion of traffic flow speed, and a variance > 40 indicates chaotic speed distribution; construct the mixed driving structure index: use the weighted formula: mixed driving ratio × 0.6 + speed variance normalized value × 0.4, and map the index range to [0,1].

[0067] The 500-meter dynamic sliding window calculation method involves defining a circular calculation area with a radius of 500 meters, centered on the "conflict point" of the merging zone or the "inflection point" of the diverging zone. The window updates every 1 second, removing vehicle data leaving the zone and adding newly entering vehicle data in real time, ensuring that only the latest traffic participant information is retained within the window. All feature calculations are based on real-time data within the current window, and historical data older than 5 seconds is excluded through a timestamp filtering mechanism to avoid interference from outdated information on real-time risk features, ensuring that the output results are completely synchronized with the instantaneous traffic flow status of the merging / diverging zone.

[0068] Step S3: Based on the risk feature set, a comprehensive risk value representing the current regional security situation is calculated and output in real time using a dynamically weighted risk fusion model; the dynamically weighted risk fusion model dynamically adjusts the fusion weights of each risk feature according to real-time traffic state parameters.

[0069] Specifically, based on the core risk differences between merging and diverging areas, a dedicated and real-time weighted risk fusion model is constructed to achieve dynamic and accurate assessment of safety risks in both types of areas. This step is designed around the dual objectives of "real-time" and "accuracy," and completes the risk assessment through a four-step progressive operation. Each step is closely linked and has dynamic adjustment capabilities, and includes the following:

[0070] B1. Input feature preprocessing and priority ranking, specifically including:

[0071] B11. First, it receives a region-specific core feature set: the merging zone includes normalized micro-interaction conflict indicators (TTC, PET), merging gap sufficiency, acceleration coordination coefficient, merging environment correction coefficient, and hybrid driving structure indicators; the diverging zone includes normalized micro-interaction conflict indicators, exit interference coefficient, lane departure warning degree, diverging environment correction coefficient, and hybrid driving structure indicators. Simultaneously, it receives pushed real-time status parameters, including main road traffic density, proportion of autonomous vehicles, diverging queue length, light intensity, and road surface friction coefficient. All input data includes millisecond-level timestamps to ensure synchronization with the real-time scene.

[0072] B12. Normalize the input features using the min-max standardization formula: Normalized value = (Original value - Minimum feature value) / (Maximum feature value - Minimum feature value); Map all indicators to the [0,1] interval to eliminate the interference of dimensional differences on the fusion calculation.

[0073] B13. Prioritize features based on preset high-risk thresholds: If there are high-risk features such as merging gap < 1 second, exit interference coefficient ≥ 0.8, TTC < 2 seconds, they are automatically marked as "emergency calculation sequence" and skipped from the regular feature sorting to proceed directly to the next step; if there are no high-risk features, they are sorted according to the priority of core risk factors > environmental factors > mixed driving structure factors to ensure that core risks are calculated first.

[0074] B2. Region-specific dynamic weight configuration, specifically including:

[0075] B21. Initialize the basic weight matrix: Based on the core risk differences of "insufficient merging gap + lack of acceleration coordination" in the merging zone and "exit interference + lane departure" in the diverging zone, construct a judgment matrix using the analytic hierarchy process (AHP) and determine the basic weights by passing the consistency test (CR < 0.1).

[0076] B22. Weight Reasonableness Verification: After all weight adjustments, the following constraint must be met: Total Weight = 1. If there is a deviation in the total weight, the deviation value will be automatically allocated according to the current weight ratio of each indicator to avoid logical contradictions in weight configuration.

[0077] B3. Lightweight weighted fusion computation, specifically including:

[0078] B31 employs a hybrid computation logic of "table lookup method first + lightweight neural network supplementation," with the core formula being: Comprehensive Risk Value = Σ (Normalized Feature Value × Corresponding Dynamic Weight). The "table lookup method" stores the frequently occurring "feature combination - risk value" mapping relationship as a table during pre-training. During computation, the table is used for priority matching; if no match is found, the lightweight neural network computation is then initiated, significantly reducing latency.

[0079] B32. The lightweight neural network adopts a three-layer fully connected structure: the number of neurons in the input layer equals the dimension of the region-specific core features. For example, both merging and splitting regions have 5 core features, hence the input layer has 5 neurons; the number of neurons in the hidden layer equals the feature dimension × 1.5, and the ReLU activation function is used to reduce computation and avoid gradient vanishing; the number of neurons in the output layer equals 1, corresponding to the comprehensive risk value. The network weight parameters are pre-trained and solidified using 500,000 sets of real merging / splitting scenario data to ensure computational accuracy.

[0080] B33. Latency Control: By combining the lookup table method with a lightweight network, the latency of a single risk calculation is controlled within 500 milliseconds, meeting the latency requirements for real-time early warning on highways.

[0081] B4. Iterative optimization and result verification, specifically including:

[0082] B41. Dynamic switching of iteration cycle: When there are no high-risk features, the latest feature values ​​and dynamic weights are reread and the comprehensive risk value is updated in the normal cycle iteration mode of 1 second / time; if there are high-risk features, it automatically switches to the emergency iteration mode of 0.5 seconds / time, while expanding the feature collection range to ensure the capture of sudden risks.

[0083] B42. Result Stability Verification: Calculate the fluctuation range of the risk value over three consecutive iterations. The fluctuation range is calculated as: |current value - average of the previous two values| / average of the previous two values ​​× 100%. If the fluctuation range is less than 5%, the result is considered stable. If the fluctuation range is between 5% and 10%, abnormal data is removed and the result is recalculated. If the fluctuation range is ≥ 10%, the priority of the neural network calculation is temporarily increased, such as by increasing the number of iterations to 5 and taking the average, to eliminate data noise interference and ensure the accuracy of the output result.

[0084] Specifically, the weighted risk fusion model for the merging zone aims to "solve the coupling risk of insufficient merging gap and lack of acceleration coordination," and its construction steps are as follows:

[0085] B51. Screening of core risk factors: Based on the characteristics of vehicle weaving behavior in the merging zone, Pearson correlation analysis was used to calculate the correlation coefficients between all indicators and the merging conflict accident rate. Core factors with correlation coefficients > 0.6 were screened out: merging gap sufficiency, acceleration coordination coefficient, merging environment correction coefficient, and mixed driving structure indicators. Secondary features with weak correlation to merging safety were eliminated.

[0086] B52. Assigning Basic Weights: The analytic hierarchy process (AHP) is used to construct the judgment matrix. The basic weights are obtained by calculating the largest eigenvalue and eigenvector of the judgment matrix, and the basic weights are finally determined by passing a consistency check.

[0087] B53. Dynamic Weighting Rules: Combining traffic flow theory and measured data, the triggering conditions and adjustment ranges for dynamic weights are set. For example, statistical analysis of measured data from merging zones reveals that when the main road traffic density is >20 vehicles / km, the proportion of conflicts caused by insufficient merging gaps increases from 35% to 62%. Therefore, when the main road traffic density is >20 vehicles / km, the merging gap sufficiency weight is increased by +0.1. When the proportion of autonomous vehicles is >50%, the proportion of conflicts caused by insufficient acceleration coordination decreases from 28% to 12%, indicating that the impact of autonomous driving coordination capabilities on safety is significantly improved. Therefore, when the proportion of autonomous vehicles is >50%, the acceleration coordination coefficient weight is increased by +0.1.

[0088] B54. Model Training and Consolidation: Using measured data from the confluence area, the data was divided into training and validation sets in a 7:3 ratio. A lightweight neural network was trained using the training set, and the network weight parameters were optimized using gradient descent to ensure that the error between the model's predicted risk value and the actual conflict level was less than 8%. Simultaneously, "feature combination-risk value" pairs that appeared ≥50 times in the training set were statistically analyzed. The model was then tested using the validation set to ensure that it maintained high accuracy in unseen scenarios, thus completing model consolidation.

[0089] B55. Model Validation and Optimization: Validate the model using independent test data that was not used in the training. If the model error is >10% in a certain scenario, supplement the data for that scenario and retrain the model until the error in all scenarios is <8%, ensuring the model's generalization ability.

[0090] Specifically, the weighted risk fusion model for the diversion zone aims to "solve the combined risks of exit interference and lane departure," and its construction steps are as follows:

[0091] B61. Screening of core risk factors: Based on the vehicle behavior characteristics in the diversion area, core factors are screened through correlation analysis: exit interference coefficient, lane departure warning degree, diversion environment correction coefficient, and mixed driving structure index, while irrelevant features are eliminated.

[0092] B62. Assigning basic weights: The analytic hierarchy process (AHP) consistent with the merging zone is used to construct a judgment matrix and pass a consistency test to determine the basic weights.

[0093] B63. Dynamic Weighting Rule Formulation: Based on the traffic characteristics and measured data of the diversion area, dynamic rules are set. For example, analysis of diversion area data shows that when the queue length of diverted vehicles is >100 meters, the proportion of conflicts caused by following vehicles trying to avoid queued vehicles increases from 40% to 75%. Therefore, when the queue length is >100 meters, the exit interference coefficient weight is increased by 0.1. When the light intensity is <50 lux, the proportion of conflicts caused by lane departure increases from 25% to 58%. Therefore, when the light intensity is <50 lux, the lane departure warning weight is increased by 0.1.

[0094] B64. Model Training and Consolidation: Using the measured data from the split zone, the training set and validation set are divided in a 7:3 ratio to train a lightweight neural network, ensuring that the model prediction error is less than 7%; at the same time, a table of high-frequency feature combinations is constructed to complete the model consolidation.

[0095] B65. Model Validation and Optimization: Validate the model using independent test data, supplement data for scenarios with errors >9% and retrain the model, ultimately ensuring that the error for all scenarios is <7% to meet the requirements for accurate evaluation.

[0096] Furthermore, the inputs and outputs of the weighted risk fusion model for merging and diverging areas are explained below:

[0097] Model inputs are divided into two categories: "core feature inputs" and "auxiliary state inputs," both transmitted in the form of structured data frames with millisecond-level timestamp precision to ensure real-time data transmission.

[0098] Core feature inputs: Merging zone: Normalized merging gap sufficiency, acceleration coordination coefficient, merging environment correction coefficient, mixed driving structure index, and micro-interaction conflict index. Diverging zone: Normalized exit interference coefficient, lane departure warning degree, diverging environment correction coefficient, mixed driving structure index, and micro-interaction conflict index.

[0099] Auxiliary state inputs include: main road traffic density, proportion of autonomous vehicles, queue length of diverted vehicles, light intensity, and road surface friction coefficient, which are used to trigger dynamic weight adjustments.

[0100] Input format: Uses structured data frames, including fields: feature name, normalized value, collection timestamp, and region type, supporting efficient data interaction between roadside devices and the model.

[0101] The model output is presented in the form of "core risk values ​​+ auxiliary analysis information", pushed in real time in JSON format, and supports parsing on multiple terminals.

[0102] Core output: The comprehensive risk value of the merging / diversion zone, ranging from [0,10], and divided into risk levels according to the interval:

[0103] Low risk: 0-3 points, indicating no significant risk of conflict and no need for intervention;

[0104] Medium risk: 3-7 points, indicating a potential conflict risk, requiring drivers to be aware;

[0105] High risk: 7-10 points, indicating an imminent risk of conflict that requires immediate intervention.

[0106] Furthermore, additional output information can be set: risk contribution ratio analysis, which clarifies the contribution ratio of each core factor to the current comprehensive risk value, providing a basis for formulating precise intervention measures.

[0107] Taking merging zones as an example, the core risk lies in "insufficient merging gaps + lack of acceleration coordination." This core conclusion stems from a statistical analysis of merging accidents on a certain highway: data shows that 63% of accidents are directly caused by insufficient merging gaps, and 21% are related to the lack of acceleration coordination, with the two combined accounting for over 80%, far exceeding the impact of environmental factors (12%) and differences in mixed driving structure (5%). Therefore, when the model deeply integrates micro-level interactive conflicts, dynamic environmental risks, and macro-level traffic conditions, it assigns basic weights based on the metric of accident contribution: assigning a weight of 0.4 to sufficiency of merging gaps and 0.2 to acceleration coordination coefficient, while also incorporating a weight of 0.3 for the merging environment correction coefficient and 0.1 for the mixed driving structure index. All basic weights are verified for correlation using SPSS software to ensure that the weight allocation is highly matched with the intensity of accident causes.

[0108] The model also has the ability to dynamically adjust weights, and the adjustment rules are determined based on traffic flow theory and simulation results.

[0109] When the traffic density on the main road exceeds 20 vehicles / km, the weight of merging gap adequacy is automatically increased to 0.5. This threshold is set according to the "critical value of saturated traffic density" for expressways in the "Technical Standard for Highway Engineering" (JTGB01-2014), and verified by VISSIM simulation—when the density exceeds 20 vehicles / km, the average gap of the main road traffic flow drops sharply from 3.2 seconds to 1.8 seconds, and the risk of insufficient gap increases by 2.3 times. At this time, the impact of gap adequacy on safety is significantly enhanced, hence its weight is increased.

[0110] When the proportion of autonomous vehicles exceeds 50%, the weight of the acceleration collaboration coefficient will be increased to 0.3. This adjustment is based on the results of human-machine hybrid driving simulation scenarios—when the proportion of autonomous vehicles exceeds 50%, the collaborative response latency between vehicles is reduced from 1.5 seconds in manual driving to 0.8 seconds, and the contribution of collaborative behavior to the merging success rate increases from 35% to 62%. At this point, autonomous driving collaboration capability becomes a key variable affecting merging safety, hence its weight is increased.

[0111] Taking the diversion zone as an example, its core risks focus on "exit interference + lane departure." This conclusion is based on statistical analysis of diversion accidents on major highways in China: data shows that 58% of accidents are directly caused by the interference of exiting vehicles on following vehicles, and 23% are related to lane departure behavior, with the two accounting for a combined 81%, significantly higher than the impact of environmental factors (14%) and differences in mixed driving structure (5%). Therefore, when integrating multi-dimensional risks, the model allocates basic weights based on the contribution of accident causes: assigning a weight of 0.4 to the exit interference coefficient and 0.2 to the lane departure warning degree, combined with a weight of 0.3 to the diversion environment correction coefficient and 0.1 to the mixed driving structure index. All basic weights are verified using SPSS correlation analysis tools to ensure a high degree of consistency with the statistical regularity of accident data.

[0112] The model supports dynamic weight adjustment, and the adjustment rules are determined based on engineering measured data and simulation verification.

[0113] When the queue length of diverted vehicles exceeds 100 meters, the weight of the exit interference coefficient is increased to 0.5. This threshold is based on on-site monitoring of typical diversion areas: when the queue length exceeds 100 meters, the frequency of continuous lane changes by exiting vehicles increases from an average of 2.3 times / minute to 5.8 times / minute, the probability of following vehicles braking suddenly increases by 3.1 times, and the interference effect is exponentially enhanced. Therefore, its weight is increased based on actual measurement data.

[0114] When the nighttime light intensity is less than 50 lux, the lane departure warning weight is increased to 0.3. This adjustment is based on the results of nighttime traffic environment experiments: 50 lux corresponds to the light level on a full moon night in a road section without streetlights. Under these conditions, the driver's lane recognition accuracy drops from 98% during the day to 72%, and the lane departure accident rate is 2.7 times that during the day. Furthermore, simulation verification shows that the explanatory power of the warning level for risk increases by 40% under these lighting conditions, thus increasing its weight ratio.

[0115] To ensure the real-time nature of risk assessment, this embodiment employs a hybrid architecture of "lightweight neural network + lookup table method" to replace the complex deep learning model, controlling the computation latency within 500 milliseconds. Simultaneously, a dynamic iteration mechanism is designed to respond to sudden risks. The specific implementation is as follows:

[0116] Lightweight Neural Network Design: Targeting the core risk characteristics of merging / diversion zones, a compact 3-layer neural network is constructed: Input Layer: Receives 8 key quantitative features (including micro-conflict indicators, environmental correction coefficients, mixed-driving structure indicators, etc.), and reduces computational load through feature normalization; Hidden Layer: Employs a single hidden layer structure with 16 neurons, using the computationally more efficient LeakyReLU function as the activation function instead of the traditional ReLU function to avoid gradient vanishing and reduce computational complexity; Output Layer: Outputs the basic risk value within the range [0,1]. The total number of network parameters is controlled to within 500, and the computation time for a single forward propagation is ≤200 milliseconds.

[0117] The lookup table method accelerates the correction process: For dynamic weight adjustment scenarios, multiple correction tables are pre-generated: Offline, the mapping relationship between different scenario parameters and weight correction coefficients is calculated using simulation data and stored as a 20×20 two-dimensional lookup table matrix; During real-time calculation, the correction coefficients are obtained directly by looking up the table based on the current scenario parameters. The correction process takes ≤50 milliseconds and is then fused with the results of the lightweight network to generate the final risk value.

[0118] The model's dynamic iterative update mechanism: Under normal conditions: risk value is iteratively calculated every 1 second, and the neural network input is updated based on the latest window data to ensure that the evaluation results are synchronized with the traffic flow status; High-risk trigger mechanism: when preset high-risk features such as "merging gap < 1 second" and "exit interference coefficient ≥ 0.8" are detected, it automatically switches to emergency iteration mode: the calculation cycle is shortened to 0.5 seconds, and the weight ratio of high-risk features in the neural network is temporarily increased to ensure that the sensitivity to capture sudden risks is increased by 1 time;

[0119] Through the above mechanism, the total calculation delay of the comprehensive risk value is controlled within 300-450 milliseconds, accurately reflecting the current traffic safety status of the merging / diversion area.

[0120] Step S4: Based on the risk level of the comprehensive risk value, generate and issue customized intervention instructions that match the driving needs of the merging or diverging areas.

[0121] Specifically, based on the differences in driving needs between merging and diverging areas, real-time comprehensive risk values ​​are transformed into multi-dimensional customized operational instructions. Precise intervention is achieved through vehicle-road cooperation and infrastructure linkage. The specific process is as follows:

[0122] C1. Risk Level and Intervention Strategy Mapping: First, the comprehensive risk value is divided into three threshold levels and matched with differentiated intervention logic: Low risk (<3): Only output status prompts and do not actively intervene in driving behavior; Medium risk (3-7): Trigger warning prompts and auxiliary operation suggestions; High risk (≥7): Execute mandatory intervention commands and link roadside facilities.

[0123] C2. Customized instructions for merging areas address the core requirements of "acceleration coordination for merging vehicles + avoidance by vehicles on the main road" in merging areas:

[0124] For merging vehicles: When the risk value is 3-7, the vehicle terminal outputs dynamic acceleration guidance such as "It is recommended to accelerate to 80km / h" and "Maintain a distance of ≥50 meters from the vehicle in front"; when the risk value is ≥7, active acceleration assistance is triggered (if it is an autonomous vehicle) or an emergency prompt "The current gap is insufficient, it is recommended to postpone merging" is issued, and at the same time, a warning "Beware of merging vehicles" is sent to nearby vehicles on the main road via V2X.

[0125] For vehicles on the main road: When the risk is medium, a message will be displayed saying "Vehicles are merging from the left, please do not change lanes continuously"; when the risk is high, the AEB system of vehicles on the main road will be activated to trigger a light braking of 0.5g to create a safe gap for merging vehicles.

[0126] C3. Customized instructions for diversion zones address the core requirements of "smooth lane changes for exiting vehicles + uninterrupted traffic flow on the main road" in diversion zones:

[0127] For vehicles exiting the lane: In medium-risk situations, a steering correction force of 0.5 N·m is applied through the lane keeping assist system, while prompting "1 km from the diversion point, it is recommended to merge into the rightmost lane in advance"; in high-risk situations, the turn signal is forcibly activated and continuously flashes, and a roadside broadcast prompts "Congestion in the diversion area ahead, please do not cut in or change lanes".

[0128] For vehicles on the main road: When the risk is medium, a voice warning will be sent out saying "Diversion zone 500 meters ahead, watch out for vehicles exiting on the right"; when the risk is high, the variable speed limit sign will be activated to temporarily reduce the speed of traffic on the main road by 10-20 km / h to reduce the risk of rear-end collisions caused by interference.

[0129] The command execution and feedback closed loop ensures that all commands are synchronized to vehicle terminals and roadside equipment through vehicle-road cooperative communication. At the same time, the command execution effect is collected in real time. If the expected effect is not achieved for three consecutive cycles, the intervention intensity is automatically upgraded, forming a real-time closed loop of "assessment-command-feedback-optimization" to ensure comprehensive dynamic intervention on the risks of merging and diverging areas.

[0130] like Figure 1 As shown in the figure, this embodiment of the invention also provides a real-time risk assessment system for merging and separating merging zones in a human-machine hybrid driving environment, used to implement the above-described method, the system comprising:

[0131] The data sensing module is configured to perform the data sensing steps and collect the multi-source traffic data in real time.

[0132] A feature calculation module, connected to the data perception module, is configured to execute the feature calculation step and output the risk feature set;

[0133] A risk fusion module, connected to the feature calculation module, is configured to execute the risk fusion step and output the comprehensive risk value;

[0134] The application output module is connected to the risk fusion module and is configured to execute the application output steps to generate and issue the customized intervention instructions.

[0135] The above description is only a preferred embodiment of the present invention. Therefore, any equivalent changes or modifications made to the structure, features and principles described in the claims of this patent application are included within the protection scope of this patent application.

Claims

1. A real-time risk assessment method for merging and separating merging zones in a human-machine hybrid driving environment, characterized in that, Includes the following steps: S1. Real-time collection of multi-dimensional traffic data in merging and diverging areas, including vehicle trajectory data, environmental status data, and autonomous vehicle prediction signals obtained through vehicle-road cooperative communication; S2. Based on the aforementioned multi-source traffic data, calculate and output a risk feature set in real time; the risk feature set includes micro-interaction conflict indicators, region-specific risk features corresponding to merging or diverging scenarios, environmental correction coefficients, and human-machine hybrid driving structural features. The calculation of micro-interaction conflict indicators specifically includes: calculating traditional safety indicators based on the acquired real-time vehicle location, instantaneous speed, and relative distance data using a time conflict model; Time-based collision warning: using the formula TTC=(relative distance between the two vehicles) / (speed of the rear vehicle - speed of the front vehicle), where the difference between the speed of the rear vehicle and the speed of the front vehicle is a positive value; Collision time interval: calculating the arrival time difference between the two vehicles at the conflict point using a trajectory prediction algorithm; The area-specific risk characteristics for merging zones include: merging gap sufficiency calculated based on the gap between the main road traffic flow and the minimum gap required for merging vehicles to safely merge; and acceleration coordination coefficient calculated based on the synchronicity of the acceleration curves of merging vehicles and vehicles in adjacent lanes of the main road. Construction of merging environment correction coefficients: Based on environmental parameters collected by roadside equipment, correction coefficients are output through a fuzzy comprehensive evaluation model; The area-specific risk features for the diversion zone include: the exit interference coefficient calculated based on the frequency of lane changes by exiting vehicles and the resulting fluctuations in the speed of traffic flow on the main road, and the lane departure warning degree calculated based on the real-time offset of exiting vehicles relative to the lane centerline. Construction of diversion environment correction coefficient: Based on environmental parameters collected by roadside sensors, it is calculated through a multi-factor weighted model; The human-machine hybrid driving structure features are obtained through the following steps: within a set dynamic calculation window, the number of autonomous vehicles and manually driven vehicles is counted, and the human-machine hybrid driving ratio is calculated; the instantaneous speed sequence of all vehicles within the window is extracted, and the traffic flow speed variance is calculated to characterize the speed dispersion; the human-machine hybrid driving ratio and the normalized speed variance are weighted and fused to obtain the human-machine hybrid driving structure features. S3. Based on the aforementioned risk feature set, a comprehensive risk value representing the current regional security situation is calculated and output in real time using a dynamically weighted risk fusion model. The dynamically weighted risk fusion model dynamically adjusts the fusion weights of each risk feature according to real-time traffic state parameters. The dynamically weighted risk fusion model adopts a hybrid computing architecture, specifically including: S31: Store the mapping relationship between the high-frequency feature combination obtained from pre-training and the risk value as a lookup table, and perform matching calculation first; S32: When the lookup table is not matched, start a lightweight neural network to perform calculation; wherein, the lightweight neural network is a three-layer fully connected structure, and its network weights are solidified through pre-training with historical scene data; The rules for dynamically adjusting the fusion weights include: For merging zones, when the traffic density on the main road is higher than a first preset threshold, the weight of the merging gap sufficiency is increased; and / or, when the proportion of autonomous vehicles is higher than a second preset threshold, the weight of the acceleration coordination coefficient is increased. For diversion zones, when the queue length of diverted vehicles exceeds a third preset threshold, the weight of the exit interference coefficient is increased; and / or, when the ambient light intensity is below a fourth preset threshold, the weight of the lane departure warning degree is increased. S4. Based on the risk level of the comprehensive risk value, generate and issue customized intervention instructions that match the driving needs of the merging or diverging areas.

2. The real-time risk assessment method for merging and separating merging zones in a human-machine hybrid driving environment according to claim 1, characterized in that, The risk fusion step also includes an iterative optimization mechanism: If no preset high-risk characteristics are detected, the comprehensive risk value is updated iteratively in the first cycle. When a preset high-risk feature is detected, the system automatically switches to emergency iteration mode, shortens the iteration cycle to a second cycle, and assigns a higher calculation priority to the high-risk feature; wherein the second cycle is shorter than the first cycle.

3. The real-time risk assessment method for merging and separating merging zones in a human-machine hybrid driving environment according to claim 1, characterized in that, In step S4, the generation and issuance of the customized intervention instructions include: When the comprehensive risk value indicates a medium risk level, a warning and auxiliary operation suggestions are sent to the vehicle terminal or driver of the relevant vehicle. When the comprehensive risk value indicates a high risk level, a mandatory intervention plan is generated, which includes active vehicle control commands or roadside facility linkage commands, and is executed through vehicle-road cooperative communication.

4. The real-time risk assessment method for merging and separating merging zones in a human-machine hybrid driving environment according to claim 3, characterized in that, The method also includes a feedback optimization step: Real-time monitoring of the execution effect of the customized intervention commands; If risk mitigation fails to meet expectations within multiple consecutive assessment periods, the intervention intensity will be automatically escalated or the intervention strategy will be adjusted.

5. A real-time risk assessment system for merging and separating zones in a human-machine hybrid driving environment, characterized in that, The system for implementing the method as described in any one of claims 1 to 4 comprises: The data sensing module is configured to perform the data sensing steps and collect the multi-source traffic data in real time. A feature calculation module, connected to the data perception module, is configured to execute the feature calculation step and output the risk feature set; A risk fusion module, connected to the feature calculation module, is configured to execute the risk fusion step and output the comprehensive risk value; The application output module is connected to the risk fusion module and is configured to execute the application output steps to generate and issue the customized intervention instructions.

Images