Method and System for Evaluating the Operational Risks of Autonomous Taxis

Abstract

The present invention discloses a method and system for evaluating the operation risk of autonomous taxis, relating to the fields of intelligent transportation and autonomous driving. Aiming at the problems of insufficient adaptability to the dynamic traffic environment, rigid spatial decision-making, and lack of compatibility with vehicle operation rules in the prior art, this evaluation method dynamically divides the urban road network into hexagonal honeycomb cells, adjusts the honeycomb radius according to real-time travel demands, and selects key nodes; real-time accesses traffic flow density, the number of conflict events, communication delay, and environmental parameters to construct a risk index for key nodes; uses a hidden Markov chain model to simulate the behavioral responses of autonomous taxis at key nodes, defines the S1 safe state and the S2 risk state, and combines the Viterbi algorithm to infer the hidden state probability; finally, comprehensively determines the risk level according to the risk index and the state probability. This method realizes the dynamic evaluation and control of the operation risk of autonomous taxis through multi-source data fusion and scenario-based behavior modeling.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent transportation and autonomous driving, and more specifically, to a method and system for evaluating the operation risks of autonomous taxis. Background Art

[0002] As a core component of the intelligent transportation system, the large-scale deployment of autonomous taxis poses new challenges to the efficiency and safety of urban traffic networks. Existing technologies have significant limitations in coping with dynamic traffic environments, human-vehicle-road interaction modeling, elastic spatial decision-making, and vehicle operation law compatibility. The specific problems are as follows:

[0003] 1. Insufficient adaptability to dynamic traffic environments

[0004] Traditional risk assessment models rely on static road network data and historical statistical laws, and it is difficult to adapt to the time-varying characteristics of urban traffic flow. For example, traffic flow fluctuations during morning and evening rush hours, sudden road network state mutations caused by accidents or extreme weather may all lead to the disconnection between the deployment strategies of autonomous taxis and actual traffic demands. Existing methods lack real-time perception and dynamic response capabilities, and cannot accurately identify the risk evolution laws in traffic bottleneck areas (such as around schools and lane-changing nodes in commercial areas), resulting in a lack of scientific basis for setting vehicle access thresholds and path planning schemes, and prone to causing regional traffic efficiency decline or systemic safety risk accumulation.

[0005] 2. Limitations in human-vehicle-road interaction scenario modeling

[0006] Existing technologies lack the ability to model the interaction behaviors between autonomous taxis and heterogeneous traffic participants (pedestrians, non-motor vehicles, and manually driven vehicles) in a fine-grained manner. For example, at high-density intersections or sections without traffic lights, the perception delay and decision-making uncertainty of autonomous vehicles may trigger a chain reaction (such as continuous deceleration of rear vehicles caused by braking delay), but existing models cannot quantify the impact of such behaviors on the stability of the overall traffic flow. In addition, the risk boundaries of vehicles with different technical configurations (such as L3 / L4 level systems) vary significantly in complex scenarios, but traditional methods do not establish the mapping relationship between scenario features and technical capability parameters, resulting in strong subjectivity and poor universality of risk assessment results.

[0007] 3. Contradiction between rigid spatial decision-making and dynamic demands

[0008] Regional deployment strategies (such as restricting commercial areas and residential areas) are usually based on static road network characteristics (such as road grades and functional areas), and it is difficult to support rapid risk assessment under dynamic spatial range adjustment. Existing models lack spatial adaptability and cannot achieve dynamic segmentation of the road network and multi-scale traffic pattern recognition, resulting in vehicle access rules (such as time-limited flow control and area bans) being unable to flexibly respond to real-time traffic states, reducing the effectiveness of policy implementation and the utilization efficiency of road network resources.

[0009] 4. Lack of compatibility in vehicle operation patterns

[0010] The operation characteristics of autonomous taxis (such as high-frequency services, passenger boarding and alighting behaviors, and energy replenishment needs) are essentially different from those of traditional traffic participants. However, existing evaluation methods do not consider the long-term impact of their unique operation patterns on the traffic system. For example, shared mobility pick-up and drop-off points may cause congestion on surrounding roads due to vehicle queue overflow, and unreasonable charging station layouts may lead to unbalanced road network loads. Traditional models only focus on short-term traffic flow disturbances and ignore the correlation mechanism between vehicle scheduling strategies, energy replenishment behaviors, and traffic risk evolution, making it difficult to support full-life-cycle risk management and control.

[0011] The root cause of the above problems lies in the contradiction between the "static evaluation paradigm" of existing technologies and the "dynamic complexity" of urban traffic systems. Specifically: Difficulty in data fusion and real-time processing: The heterogeneity and spatio-temporal discreteness of multi-source traffic data (such as floating car trajectories, sensor data, and weather information) increase the complexity of dynamic modeling, and traditional algorithms are difficult to achieve real-time risk assessment with limited computing resources. Nonlinear characteristics of interaction behaviors: The interaction behaviors between autonomous vehicles and other traffic participants have strong coupling and uncertainty. Traditional microscopic simulation models are difficult to be widely applied in engineering practice due to difficulties in parameter calibration or high computational complexity. Technical bottleneck in dynamically dividing the spatial scope: Existing road network segmentation methods (such as fixed grid division) cannot adapt to the spatio-temporal evolution characteristics of traffic flow, resulting in the disconnection between risk assessment results and actual scenarios. Technical gap in quantifying long-term impacts: The spatio-temporal distribution patterns of the full-cycle behaviors of vehicle operation (such as charging and dwelling) are difficult to accurately predict through short-term data observations or simplified models and rely on large-scale historical data and complex simulation technologies.

[0012] In summary, existing technologies have systematic defects in aspects such as dynamic adaptability, scenario modeling accuracy, spatial decision-making flexibility, and long-term impact quantification. There is an urgent need for a new method that can integrate multi-source data, dynamically evaluate risks, and support flexible decision-making to support the safe and efficient deployment of autonomous taxis. Summary of the Invention

[0013] An object of the present invention is to solve the insufficient adaptability to dynamic traffic environments: Existing methods cannot perceive traffic flow changes in real time, resulting in the disconnection between the deployment strategies of autonomous taxis and actual needs, leading to a decline in regional traffic efficiency or an accumulation of safety risks.

[0014] Solve spatial division and path planning optimization: Traditional road network division methods cannot dynamically adjust cell boundaries, and path planning does not cover high-risk nodes, affecting the accuracy of risk assessment and the efficiency of vehicle scheduling.

[0015] Solving the parameter optimization of the hidden Markov chain model: The parameter settings of the state transition probability matrix and the emission probability matrix lack data support, resulting in insufficient accuracy of risk assessment.

[0016] Solving multi-source data fusion and real-time processing: The types of traffic data are single and the update frequency is low, which cannot comprehensively reflect complex traffic scenarios and affects the reliability of risk index calculation and model input.

[0017] Solving dynamic routing and path replanning: The existing path planning does not consider the real-time traffic state and long-term operation rules, resulting in insufficient balance between vehicle operation efficiency and safety risks.

[0018] Solving closed-loop verification and parameter iteration: There is a lack of a dynamic verification mechanism for path strategies, and the risk assessment model and scheduling strategy cannot be continuously optimized.

[0019] Solving Monte Carlo simulation and facility layout optimization: Traditional methods do not quantify the long-term impact of passenger boarding / alighting and energy replenishment behaviors on traffic flow, resulting in unreasonable layouts of charging stations and transfer points.

[0020] Solving the construction of digital twin environment and disturbance analysis: The existing simulation environment cannot truly reproduce traffic flow disturbances, affecting the accuracy of path strategy verification.

[0021] Solving human-machine collaborative safety response: The association between the safety officer response mechanism and traffic flow density is insufficient, resulting in lagging risk control measures.

[0022] Solving key node risk grading and dynamic monitoring: The existing methods do not achieve refined classification and differential monitoring of key nodes, reducing the pertinence of risk assessment.

[0023] To this end, the present invention provides a method for evaluating the operation risk of an autonomous driving taxi, including:

[0024] Dynamically dividing the urban road network into hexagonal honeycomb cells, with the dynamic radius R of each honeycomb cell t Calculated according to the real-time travel demand by the formula Where R

[0025] Is the reference radius, Q base Is the travel demand density within the honeycomb at the current time period, Q t Is the historical average demand density, Q avg And Q max And Q min Are the extreme values of the demand density, that is, the maximum and minimum values, α is the adjustment coefficient, and key nodes are selected in each hexagonal honeycomb cell;

[0026] Real-time access to multi-source traffic data, including traffic flow density, number of conflict events, communication delay data, and environmental parameters, to construct the key node risk index NRIj , and is calculated by the following formula: NRI j = ω1 × C j + ω2 × D j + ω3 × A j , where C j is the conflict rate, D j is the delay index, A j is the accident density, and ω1, ω2, ω3 are preset weights;

[0027] Simulate the behavioral responses of autonomous taxis at key nodes through a hidden Markov chain model, including: defining hidden states as S1 safe state and S2 risk state; collecting taxi target detection confidence, taxi obstacle avoidance response time, and taxi V2V communication delay data, and discretizing them into observation states;

[0028] Among them, according to the hidden states, construct a state transition probability matrix P; according to the observation states, construct an emission probability matrix E; use the Viterbi algorithm to infer the probabilities of each hidden state in the S1 safe state and S2 risk state at the current moment in combination with the state transition probability matrix P and the emission probability matrix E;

[0029] According to the key node risk index NRI j value and the inferred hidden state probabilities, comprehensively determine the risk level, output low, medium, and high risk warnings, and dynamically adjust the path planning strategy according to the risk level.

[0030] Preferably, in the method for evaluating the operation risk of the autonomous taxi, it includes:

[0031] When dynamically dividing the urban road network into hexagonal honeycomb cells, use the Voronoi diagram algorithm to dynamically adjust the honeycomb boundary to ensure that adjacent honeycombs do not overlap and cover the entire area;

[0032] The reference radius R base ranges from 1 to 3 kilometers, and the adjustment coefficient α is determined by fitting historical traffic data to be 0.5 - 1.2; the weight coefficients ω1, ω2, ω3 of the key node risk index NRI j respectively correspond to the conflict rate C j , the delay index D j , and the accident density A j , and their value ranges are: ω1 is 0.4 - 0.6, ω2 is 0.3 - 0.5, and ω3 is 0.1 - 0.3;

[0033] Among them, R base is determined by fitting the urban road network density, and α is obtained through historical traffic flow regression analysis.

[0034] Preferably, in the method for evaluating the operation risk of the autonomous taxi, it includes:

[0035] In the transition probability matrix P of the hidden states S1 (safe state) and S2 (risk state) of the hidden Markov chain model, the probability of S1 transitioning to S1 is 0.85 - 0.95, the probability of S1 transitioning to S2 is 0.05 - 0.15, the probability of S2 transitioning to S1 is 0.40 - 0.60, and the probability of S2 transitioning to S2 is 0.40 - 0.60;

[0036] The emission probability matrix E is constructed based on the test data of multi - brand vehicles in a closed test field. Among them, when the confidence level of taxi target detection ≥ 0.9, the emission probability is 0.90; when the obstacle avoidance response time of the taxi ≤ 0.5 seconds, the emission probability is 0.85; when the V2V communication delay of the taxi ≤ 50 milliseconds, the emission probability is 0.92.

[0037] Preferably, in the method for evaluating the operation risk of the autonomous taxi, the real - time access to multi - source traffic data further includes:

[0038] Average vehicle speed, proportion of autonomous vehicles, number of emergency braking events, cooperative lane - changing failure rate, visibility, road surface adhesion coefficient, and signal light phase;

[0039] And update the multi - source traffic data at a set frequency. The multi - source traffic data is used to calculate the key node risk index NRI j , and combined with the taxi target detection confidence level, taxi obstacle avoidance response time, and taxi V2V communication delay data, discretize them jointly to obtain the observation states of the hidden Markov chain model.

[0040] Preferably, in the method for evaluating the operation risk of the autonomous taxi, it further includes:

[0041] For the test autonomous taxi, based on the dynamic radius R of the cellular unit t Determine the driving path planning range;

[0042] Within the planning range, generate an initial path set, satisfying that the single - run mileage is 0.5 - 4 times the cellular radius and covering all key nodes with NRI j ≥ 0.4;

[0043] With the goal of minimizing the comprehensive path risk value, that is, the weighted sum of efficiency risk and safety risk, and maximizing the service response speed, construct an optimized driving model,

[0044] In the optimized driving model, dynamically adjust the cruising paths of idle vehicles according to the traffic flow density, and preferentially keep idle vehicles at NRI jThe area around high-risk nodes with a value ≥ 0.7, and based on Monte Carlo simulation, predict the impact of the behavior of passengers getting on and off the vehicle on the traffic flow density within the next 30 minutes;

[0045] Execute the safety response mechanism. When the traffic flow density within the honeycomb is detected to reach 80 vehicles per kilometer, automatically extend the abnormal event handling duration to 1.2 times the original handling duration, and at the same time trigger a path replanning instruction.

[0046] Preferably, in the method for evaluating the operation risk of the autonomous taxi, it further includes:

[0047] Establish a closed-loop verification system. The closed-loop verification system simulates the disturbance impact of path adjustment on the actual traffic flow in the digital twin environment, updates the parameters of the risk assessment model according to the simulation results, and applies the parameters to the honeycomb radius R in the dynamic path planning framework in real time t Calculation;

[0048] The dynamic path planning framework and the real-time scheduling strategy are coordinated through the travel demand density Q within the honeycomb at the current time t Data to achieve coordination. The output data of the Monte Carlo simulation acts on both the optimization of the charging station location and the adjustment of the transfer point layout at the same time. The execution result of the path replanning instruction is fed back to the closed-loop verification system in real time to form an iterative optimization loop.

[0049] Preferably, in the method for evaluating the operation risk of the autonomous taxi, the method realizes the collaborative optimization of Monte Carlo simulation and dynamic routing strategy through the following steps:

[0050] Input parameters: Passenger getting on and off behavior data, where the mean of the Poisson distribution is 3 minutes and the standard deviation is 1 minute; Energy replenishment behavior data, where the charging time is 0.5 - 2 hours;

[0051] Set the queuing overflow probability matrix of the transfer points to quantify the vehicle retention risk of each transfer point in different time periods;

[0052] Set the load imbalance index of the road network around the charging station to reflect the traffic flow density fluctuation within the service radius of the charging station;

[0053] According to the queuing overflow probability matrix, dynamically adjust the location and capacity of the transfer points, and relocate the high-risk transfer points to the low-load areas;

[0054] Based on the load imbalance index, optimize the location of the charging stations, and preferentially deploy them in areas where the road network load balance degree ≥ 0.8;

[0055] Among them, when the traffic flow density within the honeycomb ≥ 80 vehicles per kilometer, automatically start path replanning;

[0056] Prioritize sections that have not reached the load threshold, i.e., traffic flow density < 60 vehicles / km; dynamically adjust the path weight coefficient, increase the efficiency risk weight by 30%, and synchronize the path re-planning results to the digital twin environment.

[0057] Preferably, in the method for evaluating the operation risk of the autonomous taxi, the established closed-loop verification system is specifically as follows:

[0058] The closed-loop verification system is constructed based on digital twin technology, and this system realizes data interaction with the traffic data collection module that collects multi-source traffic data in real time, the dynamic path planning framework, and the risk assessment model;

[0059] The traffic data collection module transmits the collected data to the closed-loop verification system; in the digital twin environment, a virtual scene consistent with the actual traffic scene is constructed based on the real-time traffic data, and the scene covers road networks, traffic lights, other traffic participants including pedestrians, and ordinary vehicle elements;

[0060] At the same time, the initial path set and the adjusted path plan generated by the dynamic path planning framework are input into the digital twin environment;

[0061] According to the adjusted path plan, simulate the driving process of the autonomous taxi in the virtual scene, and analyze the disturbance impact of the path adjustment on the surrounding traffic flow. The specific quantification indicators include:

[0062] Local traffic flow change: Calculate the change rate of traffic flow density and the change value of average vehicle speed on the surrounding sections before and after the path adjustment;

[0063] Evaluate the impact of the path adjustment on the average travel time and vehicle delay time within the entire cell;

[0064] After simulating the path adjustment, evaluate the change in the probability and frequency of conflicts between the autonomous taxi and other traffic participants through the risk assessment model.

[0065] Preferably, in the method for evaluating the operation risk of the autonomous taxi, the hidden state also includes a human-machine collaboration safety strategy, including the following steps:

[0066] It is clear that the average response delay of the safety officer is 30 seconds, and the average abnormal event handling duration is 90 seconds;

[0067] Establish the correlation between the handling duration and the traffic flow density: When the traffic flow density within the cell is greater than or equal to 80 vehicles / km, the handling duration increases by 20% on the original basis;

[0068] The safety risk will increase due to the changes in the response delay and handling duration, and the increased amplitude is positively correlated with the adjustment of the handling duration;

[0069] When the increase rate of safety risk reaches or exceeds the threshold of 30%, or the increase rate of efficiency risk reaches or exceeds another threshold of 20%, reduce the upper limit of the access density of autonomous taxis in this cell, adjust it to 70% of the original, and increase the weight of safety risk by 30%. At the same time, push warning information to the remote monitoring center.

[0070] Preferably, in the method for evaluating the operation risk of autonomous taxis, the key node risk index NRI j is constructed and the risk grading method is as follows:

[0071] Real-time access to multi-source data such as the movement trajectories, accident records, and weather information of test autonomous taxis, combined with the traffic flow density, number of conflict events, communication delay data, and environmental parameters described above;

[0072] Fuse and process the above data, and extract the conflict rate, delay index, and accident density of key nodes;

[0073] Calculate the key node risk index NRI according to the conflict rate, delay index, and accident density j , and according to the risk index NRI j Divide the key nodes into three categories:

[0074] Type I nodes, NRI j ≥0.7: Implement real-time monitoring and focus on analyzing its disturbing impact on the regional traffic flow;

[0075] Type II nodes, 0.4 ≤ NRI j <0.7: Conduct periodic evaluation and dynamically adjust the monitoring frequency;

[0076] Type III nodes, NRI j <0.4: Implement basic monitoring and regularly update the basic data.

[0077] The technical problems solved by the present invention also include:

[0078] 1. The problem that the existing risk assessment system for autonomous taxis is difficult to dynamically adapt to the complex changes of urban road networks. Traditional methods use fixed area division or simple grid models, which cannot dynamically adjust the coverage range and granularity of road network units according to real-time travel demands, resulting in insufficient risk assessment accuracy, especially difficult to accurately locate risk areas during peak hours or sudden traffic events.

[0079] 2. The problem of insufficient multi-source traffic data fusion and risk quantification capabilities. Existing systems lack multi-dimensional real-time fusion analysis of traffic flow density, conflict events, communication delays, and environmental parameters, resulting in one-sided calculation of risk indexes and unable to accurately reflect the comprehensive risk status of key nodes.

[0080] 3. Low reliability in predicting autonomous vehicle behavior responses. Existing models struggle to simulate real-time changes in vehicle behavior in dynamic road networks. In particular, they lack the ability to discretize key parameters such as target detection, obstacle avoidance response, and communication latency, leading to significant deviations in risk state inference.

[0081] 4. Poor dynamic coordination between risk warning and route planning. Traditional methods rely on static risk classification and are unable to dynamically adjust route strategies based on real-time hidden state probabilities, resulting in delayed warnings or a disconnect between planning strategies and actual road conditions.

[0082] To this end, the present invention also provides an assessment system for the operating risk of a self-driving taxi, comprising:

[0083] The road network dynamic division module is used to dynamically divide the urban road network into hexagonal honeycomb units according to real-time travel demand, where the dynamic radius R of each honeycomb unit is t By formula Calculation, R base is the reference radius, Q t is the travel demand density within the cell during the current period, Q avg is the historical average demand density, Q max With Q min is the extreme value of demand density, α is the adjustment coefficient, and key nodes are selected in each hexagonal honeycomb unit;

[0084] Multi-source data access module for real-time acquisition of traffic density, number of conflict events, communication delay data, and environmental parameters;

[0085] Key node risk assessment module, used to calculate the key node risk index NRI based on the multi-source data j , where NRI j NRI by formula j =ω1×C j +ω2×D j +ω3×A j Calculation, C j is the conflict rate, D j is the delay index, A j is the accident density, ω1, ω2, ω3 are preset weights;

[0086] The behavioral response simulation module includes a hidden Markov chain model, which is used to define the hidden states S1, the safe state, and S2, the risk state. It discretizes the target detection confidence, obstacle avoidance response time, and V2V communication delay data of the self-driving taxi into observation states, constructs the state transition probability matrix P and the emission probability matrix E, and uses the Viterbi algorithm to combine the matrices P and E to infer the probability of each hidden state at the current moment;

[0087] A comprehensive risk decision-making module for comprehensively determining low, medium, and high risk levels based on the key node risk index NRI j and the hidden state probability, outputting early warning signals and dynamically adjusting the path planning strategy of the autonomous taxi.

[0088] The present invention has at least the following beneficial effects:

[0089] 1. Dynamic risk assessment and path optimization: Through the dynamic division of cellular units and the fusion of multi-source data, the real-time quantification of traffic risks and the dynamic adjustment of path strategies are realized, improving the scientificity and safety of the deployment of autonomous taxis.

[0090] 2. Spatial division and parameter optimization: The Voronoi diagram algorithm ensures the dynamic adaptation of cellular boundaries, and the scientific setting of parameter ranges enhances the universality of the model for different urban road networks.

[0091] 3. Probability model parameter calibration: Based on the state transition and emission probability matrices of measured data, the prediction accuracy of the hidden Markov chain model for vehicle behavior is improved, and the risk misjudgment rate is reduced.

[0092] 4. Multi-dimensional data support: By expanding the data types and update frequencies, the traffic status is comprehensively reflected, providing a richer basis for risk index calculation and model input.

[0093] 5. Dynamic routing and safety response: The path planning covers high-risk nodes, combines Monte Carlo simulation to predict long-term impacts, balances efficiency and safety; the path replanning triggered by traffic flow density improves the emergency response ability.

[0094] 6. Closed-loop verification and continuous optimization: The digital twin environment verifies the actual effects of path strategies, forms a "data - decision - verification" closed loop, and ensures the continuous evolution of strategies.

[0095] 7. Quantification of long-term impacts and facility optimization: Through Monte Carlo simulation, the long-term impacts of passenger behavior and energy replenishment are quantified, the layouts of charging stations and transfer points are optimized, and the imbalance of road network load is reduced.

[0096] 8. Perturbation analysis and accurate verification: The digital twin environment reproduces traffic flow perturbations, quantifies the actual impacts of path adjustments, and provides data support for strategy optimization.

[0097] 9. Enhancement of human-machine collaborative safety: The response of safety officers is associated with traffic flow density, and the disposal duration and access rules are dynamically adjusted to improve the risk control ability in complex scenarios.

[0098] 10. Differential monitoring of key nodes: Through risk grading, nodes are dynamically classified, the allocation of monitoring resources is optimized, and the pertinence and efficiency of risk assessment are improved.

[0099] 11. Dynamic road network division enhances the adaptability of risk assessment: Through the dynamic radius adjustment mechanism of hexagonal honeycomb cells, the road network coverage can be adaptively scaled according to the real-time travel demand density, solving the problem of the lag in response of traditional fixed grid division to traffic flow mutation scenarios. By adjusting the size of honeycomb cells in combination with the ratio of historical to current demand density, the honeycomb radius is reduced during peak hours to improve risk positioning accuracy, and the radius is enlarged during low-demand hours to reduce computational load, achieving the optimal balance between resource allocation and risk assessment efficiency.

[0100] 12. Multi-source data fusion enhances the comprehensiveness of risk quantification: By integrating multi-dimensional real-time data such as traffic flow density, number of conflict events, communication delay, and environmental parameters, and constructing a key node risk index (NRI j ) based on preset weights, the problem of one-sidedness in traditional single-index evaluation is solved. In the formula, the weighted calculation of the conflict rate (C j ), delay index (D j ), and accident density (A j ) can comprehensively quantify risks from three dimensions of traffic conflict, traffic efficiency, and safety accidents, significantly improving the scientificity and reliability of key node risk determination.

[0101] 13. Hidden Markov model optimizes behavior response prediction: By defining hidden states (safe / risky) and discretizing observed states (object detection confidence, obstacle avoidance response time, communication delay), combined with the state transition matrix and emission probability matrix, the Viterbi algorithm is used to dynamically infer the hidden state probability. This model can effectively capture the behavioral uncertainty of autonomous vehicles in complex road networks, especially reducing noise interference through discretization processing, improving the real-time performance and accuracy of risk state prediction, and providing a reliable basis for dynamic path planning.

[0102] 14. Comprehensive risk decision-making enhances the coordination of path planning: Based on the dual decision-making mechanism of NRI j and hidden state probability, the static risk index is combined with dynamic behavior response prediction to achieve multi-level early warning of low, medium, and high risk levels. By dynamically adjusting the path planning strategy, potential risk areas can be automatically avoided at high-conflict rate nodes, and redundant communication links can be switched when the communication delay is high, ensuring the optimal operation safety and efficiency of autonomous taxis in complex urban road networks. Description of the Drawings

[0103] Figure 1 : Schematic diagram of the ant climbing pole theory;

[0104] Figure 2 : Schematic diagram of the three-level architecture of "area - honeycomb - key node";

[0105] Figure 3:Operation model diagram of autonomous driving taxis in a cellular network;

[0106] Figure 4 :Schematic diagram of the road risk model for autonomous driving taxis at key nodes;

[0107] Figure 5 :Flowchart for formulating deployment parameters of autonomous driving taxis. Detailed implementation manners

[0108] The following further elaborates the present invention in conjunction with the accompanying drawings, so that those skilled in the art can implement it with reference to the text of the specification.

[0109] As Figures 1 - 5 shown: Figure 1 It is a schematic diagram of the ant climbing pole theory, Figure 1 which explains the theoretical basis for the dynamic division of cellular units through the bionics principle of ants climbing poles. When a large number of ants randomly crawl on the pole and turn back after meeting, their paths are ultimately equivalent to dividing the pole into two independent regions, and the number of ants in each region tends to be balanced. This phenomenon is mapped to the deployment strategy of autonomous driving taxis, indicating that the randomness of vehicle operation paths within dynamically adjusted cellular units can be equivalent to risk assessment within a fixed region, thus achieving spatial decoupling. Based on the ant colony algorithm, cellular units are dynamically generated centered around transportation hubs, and the Voronoi diagram algorithm is used to ensure that there are no overlapping boundaries and the entire area is covered, supporting the elastic calculation formula t of the cellular radius R , providing a spatial division basis for subsequent key node identification and risk assessment.

[0110] Figure 2 It is a schematic diagram of the "region - cellular - key node" three - level architecture. Figure 2 It shows the hierarchical architecture of the overall risk assessment: the bottom layer is the key node layer, which realizes local risk early warning by real - time monitoring of the NRI j value (such as a certain type of node with NRI j ≥0.7); the middle layer is the cellular layer, which aggregates the peak risk of nodes and calculates the access capacity, dynamically adjusting the vehicle density (such as path planning covering nodes with NRI j ≥0.4); the top layer is the region layer, which quantifies the cascading effect between cells through the risk propagation model and generates time - sharing and zoning strategies (such as traffic flow restriction during the morning rush hour). This architecture strengthens the "barrel effect" theory, that is, the regional risk is determined by the peak risk of key nodes, supports the application of the hidden Markov chain model and the Viterbi algorithm, and realizes the dynamic determination of risk levels through the state transition probability matrix P and the emission probability matrix E.

[0111] Figure 3 It is the operation model diagram of autonomous driving taxis in a cellular network. Figure 3Describes the operating rules of autonomous taxis within a cellular unit: The vehicle starts from transportation hubs (such as subway stations, charging stations), and the path covers a range of 0.5 - 4 times the cellular radius, ensuring service coverage of high-risk nodes (NRI j ≥0.4). The boarding and alighting stay time of passengers follows a Poisson distribution (mean 2 - 5 minutes). When the battery level is below the threshold, the vehicle automatically returns to the charging station (charging time is 0.5 - 2 hours). The empty-running dispatching is limited within the cellular unit based on reinforcement learning. This model is closely related to Monte Carlo simulation. By predicting the impact of passenger behavior and energy replenishment requirements on traffic flow density, it optimizes the location of charging stations and the layout of transfer points, balances the long-term operation rules and short-term traffic conditions, and achieves the goal of minimizing the comprehensive risk value of the path.

[0112] Figure 4 Schematic diagram of the road risk model for autonomous taxis at key nodes. Figure 4 Constructed a mathematical model for the risk evolution of key nodes. The inputs include the initial risk X f of the node, the static features θ, and the number of autonomous vehicles C max , and the output is the comprehensive risk Y f after deployment (coupling of efficiency and safety risks). This model is verified through microscopic simulation experiments. Combining the NRI j grading method (real-time monitoring for class I nodes, periodic assessment for class II nodes), it quantifies the impact of technical configurations (such as L3 / L4 level systems) on the risk boundary. For example, when the vehicle detects a pedestrian intrusion, the successful braking probability is 92% (parameter of the emission probability matrix E). False judgment or delayed response will trigger the intervention of the safety officer. Through the human-machine collaborative safety strategy, the handling duration and access density are adjusted to achieve dynamic risk control.

[0113] Figure 5 Flowchart for formulating deployment parameters of autonomous taxis. Figure 5 Shows the closed-loop process of "evaluation - decision - verification": First, divide the area based on GIS and generate a cellular structure, and configure spatio-temporal strategies (such as operating time periods, human-machine ratio); verify the disturbance impact of the path strategy through a digital twin environment (closed-loop verification system), and quantify the local traffic flow changes (change rate of traffic flow density, change value of average vehicle speed) and overall efficiency indicators (average travel time, delay time); finally, iteratively optimize the parameters according to the simulation results, such as adjusting the cellular radius R t in the calculation of the α coefficient (value range 0.5 - 1.2) or the path weight coefficient (efficiency risk weight increased by 30%). This process supports the collaborative optimization of Monte Carlo simulation and dynamic routing, ensures scientific decision-making on the layout of charging stations and the capacity of transfer points, and finally forms a vehicle deployment plan that meets the regional risk control.

[0114] According to an embodiment of the present invention, when dynamically dividing an urban road network into hexagonal honeycomb cells, the reference radius R base can be determined by fitting according to the urban road network density, and the value range is 1 - 3 kilometers (for example, 1 kilometer in high-density urban areas and 3 kilometers in suburban areas). The adjustment coefficient α is determined by historical traffic flow regression analysis, and the value range is 0.5 - 1.2 (such as α = 1.0 during the morning peak period and α = 0.6 during the flat peak period). This process can use a GIS geographic information system (such as ArcGIS) to integrate high-precision map data, and combine with a Voronoi diagram algorithm library (such as the scipy library in Python) to dynamically adjust the honeycomb boundary to ensure that adjacent honeycombs do not overlap and cover the entire area. The road network data is stored in a cloud server and transmitted to in-vehicle terminals in real time through edge computing nodes. The honeycomb division result is displayed through the in-vehicle navigation system, and the driver can view the distribution of key nodes within the current honeycomb through the human-machine interaction interface.

[0115] When real-time accessing multi-source traffic data, floating car trajectory data can be collected through a GPS module (such as Trimble BD992) installed on test vehicles, the number of conflict events is identified by an in-vehicle camera (such as Mobileye EyeQ5) combined with computer vision algorithms, and V2V communication delay data is obtained through a DSRC module (such as Cohda Wireless MK5). The data update frequency is set to 1 minute, and real-time data stream transmission is achieved through the Apache Kafka message queue, and distributed computing is performed by SparkStreaming. The weight coefficients of the key node risk index NRI j are: ω1 is 0.4 - 0.6 (corresponding to the conflict rate), ω2 is 0.3 - 0.5 (corresponding to the delay index), ω3 is 0.1 - 0.3 (corresponding to the accident density). After the calculation result is completed on the edge server, it is pushed to the control center through the 5G network.

[0116] In the state transition probability matrix P of the hidden Markov chain model, the probability of S1 transitioning to S1 is 0.85 - 0.95 (such as 0.90), the probability of S1 transitioning to S2 is 0.05 - 0.15 (such as 0.10), the probability of S2 transitioning to S1 is 0.40 - 0.60 (such as 0.50), and the probability of S2 transitioning to S2 is 0.40 - 0.60 (such as 0.50). The emission probability matrix E is constructed based on the test data of multi-brand vehicles in a closed test field. When the target detection confidence level ≥ 0.9, the emission probability is 0.90; when the obstacle avoidance response time ≤ 0.5 seconds, it is 0.85; when the V2V communication delay ≤ 50 milliseconds, it is 0.92. Data storage uses a MySQL database, and the transition probability matrix P performs matrix operations through the NumPy library in Python. When determining the risk level, the low-risk threshold is NRI j <0.4 and the probability of S1 ≥ 0.85, and the medium risk is 0.4 ≤ NRI j<0.7 or S2 probability ≥ 0.20, high risk is NRI j ≥ 0.7 or S2 probability ≥ 0.50. When the risk level is high, the system automatically reduces the vehicle access density in the honeycomb to 70% of the original upper limit and pushes a new route through the in-vehicle terminal (such as avoiding NRI j ≥ 0.7 nodes). Through actual measurement and verification, this method increases the average vehicle speed in the area by 15% and reduces conflict events by 22%, effectively balancing traffic efficiency and safety risks.

[0117] According to another embodiment of the present invention, when dynamically dividing the urban road network into hexagonal honeycomb cells, the reference radius R base is determined according to the density of the urban road network, and the specific range is 1.5 kilometers to 2.5 kilometers. The adjustment coefficient α is obtained through historical traffic flow regression analysis, and the value range is 0.7 to 1.0. The real-time travel demand density Q t is collected by multi-source traffic sensors deployed at road intersections and main roads. For example, Siemens Sicore traffic flow detectors are installed on the top of street lamp poles or traffic signal poles, and the outer shell is made of UV-resistant ABS plastic and fixed by bolts. The computing unit selects the NVIDIA Jetson AGX Xavier embedded system and is installed in a waterproof metal box, and the box body is fixed in the roadside control cabinet through a bracket. During the working process, the sensor uploads vehicle trajectory data to the computing unit every 5 minutes, and the computing unit dynamically adjusts the honeycomb radius R t according to the formula and updates the honeycomb boundary through the Voronoi diagram algorithm. For example, when Q t increases by 20% compared with the historical average value, R t expands from 2 kilometers to 2.8 kilometers to cover the newly added congested area.

[0118] The calculation of the key node risk index NRI j relies on the fusion of multi-source traffic data. The conflict rate C j is collected by video surveillance equipment deployed at intersections. For example, Hikvision DS-2CD2346G2-IU cameras are installed at a height of 6 meters, fixed to the surveillance pole through a universal bracket, and the outer shell is made of aluminum alloy. The delay index D j is calculated based on the difference between the real-time vehicle speed and the historical average vehicle speed. The vehicle speed data is provided by GPS floating cars, and the threshold is set to trigger an alarm when the vehicle speed drops by more than 30%. The accident density A jObtained from the traffic management department's database, with a statistical period of 30 days. Data processing uses Dell's PowerEdge R750 servers, deployed in a cloud data center, and the cabinet is equipped with redundant power supplies and cooling systems. During operation, cameras capture traffic flow videos in real time, and edge computing nodes extract the number of conflict events per hour. If the number of conflict events at a certain node is ≥ 3 times per hour, then C j = 0.8; combined with a 40% decrease in real-time vehicle speed (D j = 0.6) and an accident density of 0.3, the NRI j is calculated to be 0.73, and it is determined as a high-risk node.

[0119] The hidden states of the Hidden Markov Model include S1 (safe) and S2 (risk), and the observation states are generated by discretizing in-vehicle sensor data. The target detection confidence threshold is set to ≥ 0.85, corresponding to an emission probability of 0.9; the obstacle avoidance response time threshold is ≤ 0.6 seconds, corresponding to an emission probability of 0.8; the V2V communication delay threshold is ≤ 60 milliseconds, corresponding to an emission probability of 0.9. The in-vehicle computing unit uses Aurora Innovation's autonomous driving domain controller, integrated into the vehicle's central control system, with a magnesium alloy housing, and is fixed in the trunk by bolts; the V2X communication module selects Huawei's MH5000-31 module, supporting 5G and DSRC protocols, and is installed in the vehicle's pre-installed T-Box. During operation, the vehicle collects the target detection confidence (such as 0.92), the obstacle avoidance response time (such as 0.55 seconds), and the V2V delay (such as 45 milliseconds) in real time. After inputting into the model, the probability of being in the S1 state is calculated to be 0.88 through the Viterbi algorithm. Combining with NRI j = 0.73, the system comprehensively determines it as a high-risk level, triggers a route replanning instruction, and guides the vehicle to a section with a traffic flow density lower than 60 vehicles / km.

[0120] According to another embodiment of the present invention, when the dynamic road network division uses the Voronoi diagram algorithm, commercial GIS software (such as ArcGIS) can be used in cooperation with the real-time traffic data interface. The reference radius R base ranges from 1 km (applicable to high-density urban areas) to 3 km (applicable to suburban areas), and the adjustment coefficient α is fitted to be 0.5 (off-peak hours) to 1.2 (peak hours) according to historical traffic data. The selection of key nodes can combine floating car trajectory data, and use transportation hubs (such as subway stations, commercial areas) as initial seed points to identify high-conflict areas in the road network through the Dijkstra algorithm. In terms of equipment, lidar (such as Velodyne HDL-64E) and cameras (such as Basler acA2040-90um) deployed on road infrastructure can be used to collect traffic flow data in real time.

[0121] The reference radius R baseThe determination of [[ID=]] requires calling the urban road network density data. For example, the density value can be calculated by obtaining the road length and regional area through OpenStreetMap. The fitting process of the adjustment coefficient α can use SPSS software for multiple linear regression analysis. The independent variables include traffic flow, accident rate, and weather data in historical periods. The weight coefficient ω1 is taken as 0.5 (dominated by the conflict rate), ω2 is taken as 0.4 (the delay index follows), and ω3 is taken as 0.1 (assisted by the accident density). This combination is calculated through the Analytic Hierarchy Process (AHP). In terms of equipment, the Huawei Atlas 500 intelligent edge computing platform can be used for real-time data processing, with a memory configuration of 16GB DDR4.

[0122] Key Node Risk Index NRI j The calculation of [[ID=]] requires real-time access to the traffic management platform data, including the traffic flow density obtained through the V2X communication module (such as the Huawei 5G vehicle terminal MH5000) and the number of conflict events counted through the video analysis system (such as Hikvision iDS-9600). The data update frequency is set to 10Hz to ensure dynamicity. When calculating, the conflict rate C j is obtained by dividing the number of emergency braking events within a unit time by the total traffic volume. The delay index D j adopts the ratio of the actual travel time to the free flow time. The accident density A j is the historical accident count divided by the node area. In terms of equipment, NVIDIA JetsonAGX Orin can be used as an edge computing node, equipped with TensorRT for accelerated inference.

[0123] By dynamically adjusting the cellular boundary through the Voronoi algorithm, the road network segmentation accuracy can be improved by more than 30%, effectively reducing the boundary overlap problem caused by traditional fixed grid division. The scientific setting of the parameter value range improves the model adaptability. In tests of different city scales, the error rate of the reference radius R base is controlled within 15%. The optimized combination of weight coefficients makes the NRI j The correlation coefficient between the calculation result and the actual accident rate reaches 0.82, improving the prediction accuracy by 45% compared with the traditional single-index method. This implementation provides a reliable basis for spatial division and risk quantification for the dynamic deployment of autonomous driving taxis.

[0124] According to another embodiment of the present invention, the probability of the hidden state S1 transitioning to S1 is set to 0.90 (typical value), the probability of S1 transitioning to S2 is 0.10, the probability of S2 transitioning to S1 is 0.50, and the probability of S2 transitioning to S2 is 0.50. These parameters are based on the statistical data measured in a closed test field. For example, in 5000 urban road tests, the safety state retention rate is 88.7%. In terms of equipment, a CompactRIO system from National Instruments can be used for data acquisition, and LabVIEW software is used to record vehicle state transition events. This system can be installed in the trunk of an autonomous taxi to obtain real-time data from the vehicle control unit through the CAN bus interface.

[0125] When the object detection confidence level ≥ 0.9, the emission probability is set to 0.90, and this threshold can be obtained by training the YOLOv5s model on the Cityscapes dataset. The emission probability for an obstacle avoidance response time ≤ 0.5 seconds is 0.85, and the data is from the test of a fusion perception system of a lidar (such as Ouster OS1-64) and a millimeter-wave radar (such as Bosch MRR440). The emission probability for a V2V communication delay ≤ 50 ms is 0.92, and the test is carried out using the MK5 OBU device from Cohda Wireless under the 802.11p protocol. These devices are installed on the front bumper (lidar), side mirror (millimeter-wave radar), and roof (OBU antenna) of the vehicle respectively.

[0126] When initializing the hidden Markov chain model, the state probability vector is set to [0.9, 0.1]. Observation data is received every 500 ms and input into the model after discretization processing. The Viterbi algorithm runs on the NVIDIA Jetson AGX Orin platform, and the inference is accelerated through TensorRT. Parameter calibration uses the Baum-Welch algorithm and is iteratively optimized based on 100 hours of actual road test data. In terms of equipment, Renesas R-Car H3 can be used as the in-vehicle computing platform to run the state estimation program with a Linux operating system.

[0127] The probability matrix based on the measured data enables the state prediction accuracy to reach 92.3%, which is 18 percentage points higher than the traditional empirical value method. The precise threshold setting of the emission probability controls the false alarm rate below 6.8%, meeting the requirements of the ISO 26262 functional safety standard. In the mixed traffic scenario test, the response delay of the model to the risk state is shortened to 120 ms, effectively supporting the emergency decision-making of the autonomous driving system. This embodiment provides a reliable probability modeling basis for dynamic risk assessment and improves the behavior prediction ability of autonomous taxis in complex scenarios.

[0128] According to another embodiment of the present invention, the newly added data includes average vehicle speed (0 - 120 km / h), proportion of autonomous vehicles (10% - 80%), number of emergency braking events (0 - 5 times / minute), cooperative lane change failure rate (0.1% - 5%), visibility (50 m - 2000 m), road surface adhesion coefficient (0.1 - 1.0), signal light phase (cycle 60 - 180 seconds). Among them, the visibility threshold is set as: low (<200 m), medium (200 - 500 m), high (>500 m). In terms of equipment, a Bosch BMP280 barometric pressure sensor can be used to measure visibility, which is installed inside the vehicle's front windshield; a Honeywell MLX90640 infrared thermal imager is used to detect the road surface temperature and calculate the adhesion coefficient, which is installed at the front end of the chassis.

[0129] The multi-source data is transmitted to the edge computing node in real time through a 5G communication module (such as Huawei MH5000), and the update frequencies are set to 10 Hz (high-frequency data) and 1 Hz (low-frequency data). The average vehicle speed is obtained through a GPS module (such as u-blox NEO-M8N), which is installed on the vehicle roof; the proportion of autonomous vehicles is obtained by statistically analyzing the identity information of surrounding vehicles through V2X communication, and the OBU device is installed at the rear of the vehicle. The cooperative lane change failure rate is recorded by an in-vehicle camera (such as Sony IMX490) for lane change events and analyzes failure cases, and the camera is installed on the rearview mirror housing.

[0130] The multi-source data is fused through the Kalman filter algorithm to generate a data stream with a unified timestamp. The conflict rate C j is calculated as the number of emergency braking events divided by the total traffic volume, and the delay index D j uses the ratio of travel time to free flow time, and the accident density A j is the number of historical accidents divided by the node area. The observation state discretization uses K-means clustering to map continuous data into discrete values of 0 - 2 (such as visibility 0 = low, 1 = medium, 2 = high). In terms of equipment, Renesas R-Car H3 can be used as an in-vehicle computing platform, and a Linux system is installed to run the data processing program.

[0131] Expanding the data type increases the calculation dimension of the risk index by 40%, and the correlation with the actual accident rate is increased to 0.85. The high-frequency data update (10 Hz) shortens the model response delay to 80 ms, which is 3 times faster than the traditional method. The discretization process reduces the data dimension by 60% and improves the inference efficiency of the hidden Markov chain model by 25%. In the heavy rain weather test, the addition of visibility data increases the risk warning accuracy rate from 72% to 89%, effectively enhancing the risk identification ability in complex environments. This embodiment provides more comprehensive data support for dynamic risk assessment and improves the multi-dimensional perception ability of the model for the traffic system.

[0132] According to another embodiment of the present invention, in terms of determining the dynamic path planning range, the radius R of the cellular unit t adopts a reference value of 1 - 3 kilometers, and combines the real-time travel demand density Q t for dynamic adjustment. For example, during the morning rush hour, when Q t exceeds 120% of the historical average, R t can be shortened to 0.8 times the reference value. The path planning range is set to 0.5 - 4 times the cellular radius, and the specific multiple is adjusted in real time according to the traffic flow state. The Velodyne HDL-64E lidar is used for environmental perception, and the HERE high-precision map is combined to achieve dynamic division of the cellular boundary. The in-vehicle GPS module obtains the position information in real time, and calculates the current cellular coverage range through the Voronoi diagram algorithm to ensure that adjacent cells do not overlap and the whole area is covered.

[0133] The initial path set is generated based on the NVIDIA DRIVE PX2 computing platform, and the real-time road conditions are obtained by combining the TomTom traffic data API. The path planning needs to cover all key nodes with NRI j ≥0.4, and the single-run mileage is controlled within 1.5 - 3 times the cellular radius. For example, in a cellular unit with a cellular radius of 2 kilometers, the path length is set to 3 - 6 kilometers. The Dijkstra algorithm is used to generate candidate paths, and the path combination with a balance of efficiency risk and safety risk is selected through multi-objective optimization. The cruising path of idle vehicles is dynamically adjusted according to the traffic flow density, and is preferentially deployed within 500 meters around high-risk nodes with NRI j ≥0.7.

[0134] In the construction of the optimized driving model, the efficiency risk weight is initially set to 0.6, and the safety risk weight is 0.4. When the detected traffic flow density reaches 80 vehicles per kilometer, the path replanning mechanism is automatically triggered, and the section with a traffic flow density < 60 vehicles per kilometer is selected. The Mobileye EyeQ4 chip is used for driving decision-making, and V2X data interaction is realized through the 4G / 5G communication module. Monte Carlo simulation predicts the passenger boarding and alighting behavior within the next 30 minutes, and the input parameters include a Poisson distribution mean of 3 minutes and a standard deviation of 1 minute. When the traffic flow density exceeds the standard, the abnormal event handling duration is extended from 90 seconds to 108 seconds, and at the same time, the safety risk weight is increased to 0.52. The path replanning response time is controlled within 200 ms to ensure the real-time requirement.

[0135] Through dynamic route planning, coverage of high-risk nodes increased to 92%, and average service response time was reduced by 18%. Monte Carlo simulation prediction accuracy reached 85%, and the error in charging station site selection was controlled within 1.2 kilometers. A traffic density trigger mechanism reduced route replanning response time by 40%, and the accident rate decreased by 23% after extending the time required to handle abnormal events. This solution significantly improved safety performance in complex traffic scenarios while maintaining vehicle operating efficiency.

[0136] According to another embodiment of the present invention, the construction of a closed-loop verification system is based on digital twin technology. The system is capable of data exchange with the traffic data collection module, the dynamic path planning framework, and the risk assessment model. The traffic data collection module can use Huawei's Road-X perception device to collect data such as traffic density and the number of conflict events in real time at a frequency of 5Hz. In the digital twin environment, a virtual scene is constructed with the help of the Unity3D engine. The scene contains the road network provided by the HERE high-precision map and elements such as pedestrians and ordinary vehicles generated by the CARLA simulation platform. The initial path set and the adjusted path plan generated by the dynamic path planning framework will be synchronously input into the virtual scene to simulate the driving process of the self-driving taxi.

[0137] The parameter update mechanism operates according to a specific process. First, the digital twin environment calculates indicators such as the rate of change in traffic density (threshold set to ±15%) and the change in average vehicle speed (threshold set to ±8 km / h) on surrounding roads before and after the route adjustment. These indicators are then fed back to the risk assessment model, which adjusts the state transition probability matrix P of the hidden Markov chain accordingly. For example, the probability of transitioning from S1 to S2 can be dynamically adjusted by ±20% from the original 0.05-0.15. Simultaneously, the output data from the Monte Carlo simulation is used to optimize the location of charging stations, with a load balancing threshold set to 0.8. The capacity of the docking points is also adjusted, with a maximum capacity of 5 vehicles per docking point.

[0138] The collaborative optimization strategy is implemented in three aspects. First, the dynamic path planning framework will t Data adjustment cell radius R t , with an adjustment range of 0.8-1.2 times the baseline value. Second, when the traffic density reaches 80 vehicles per kilometer, the system automatically initiates route replanning, prioritizing sections with a traffic density of less than 60 vehicles per kilometer, while also increasing the efficiency risk weight from 0.6 to 0.78. Third, the results of route replanning are fed back to the closed-loop verification system in real time, forming an optimization loop with an iterative cycle of 15 minutes. Field tests have shown that this system can shorten the average travel time in the region by 12% and reduce the frequency of route strategy optimization by 40%.

[0139] As a result, the simulation accuracy of the digital twin environment reached 92%, with the error from actual traffic flow within ±10%. The error in charging station site selection was less than 1.5 kilometers, and queue overflow at docking points was reduced by 65%. The parameter update frequency of the risk assessment model was increased to once per minute, improving the system's response speed by 30%. Regional traffic flow stability indicators (such as speed variance) improved by 22%, significantly enhancing the overall performance of the system.

[0140] According to another embodiment of the present invention, the Monte Carlo simulation input parameters utilize a Poisson distribution model for passenger boarding and alighting behavior data, with a mean of 3 minutes and a standard deviation of 1 minute. For energy recharge behavior data, charging time is set to 0.5-2 hours, with the specific time dynamically adjusted based on the charging station power (e.g., 7kW slow charging or 120kW fast charging). Passenger boarding and alighting behavior samples can be obtained using TomTom's floating vehicle data API, and charging time statistics can be obtained from the State Grid charging station database. Data input is updated every minute to ensure simulation timeliness.

[0141] The optimization strategy for docking points and charging stations is implemented using two matrices. The docking point queue overflow probability matrix is divided into a 3×3 grid, with a maximum capacity of 5 vehicles per grid. A migration mechanism is triggered when the number of vehicles in the queue exceeds 80% of the capacity for 15 consecutive minutes. When calculating the load imbalance index for the road network surrounding the charging station, the service radius is set to 1.5 kilometers. When the traffic density of a road section fluctuates by more than ±25% of the historical average, it is marked as an imbalanced area. Huawei Atlas 500 Smart Station can be used for real-time analysis of traffic flow data, combined with Esri ArcGIS for geographic information processing.

[0142] The implementation of the path replanning strategy is divided into three steps. First, a traffic density threshold of 80 vehicles per kilometer is set. When the detection value reaches the threshold, the path replanning algorithm is automatically activated, giving priority to sections with a traffic density of less than 60 vehicles per kilometer. Secondly, the path weight coefficient is dynamically adjusted, increasing the efficiency risk weight from 0.6 to 0.78 and reducing the safety risk weight accordingly. Finally, the adjusted path plan is synchronized to the digital twin environment for verification, with a verification cycle set to 15 minutes. The algorithm can be deployed using Baidu Apollo's path planning module, and data synchronization is completed through the 5G communication module.

[0143] In this way, the queue overflow rate at the docking point was reduced by 65%, and the average detention time at a single node was shortened from 8 minutes to 3 minutes; the charging station site selection error was controlled within 1.2 kilometers, and the load imbalance index of the surrounding road network was improved by 38%; the path replanning response time was shortened from 45 seconds of the traditional method to 18 seconds; the Monte Carlo simulation prediction accuracy reached 85%, and the prediction error of passengers' boarding and disembarking behavior was less than ±1.5 minutes.

[0144] According to another embodiment of the present invention, in terms of the architecture design of the closed-loop verification system, Huawei Atlas 500 is used as an edge computing node and deployed at the intersection of transportation hubs and main roads to collect data such as traffic flow density and the number of conflict events in real time at a frequency of 5Hz. The digital twin environment is built based on the Unity 3D engine and integrates HERE high-precision map data, with the road network accuracy reaching the 0.1-meter level. The pedestrian behavior model in the virtual scene is trained using the CARLA open-source dataset, and the dynamic parameters of ordinary vehicles come from the TomTom floating car database. The path plan generated by the dynamic path planning framework is synchronized to the digital twin environment through the 5G communication module, and the data transmission delay is controlled within 20ms.

[0145] In terms of setting the quantitative indicators for perturbation analysis, the threshold of the local traffic flow change rate is set to ±15%, and the threshold of the average vehicle speed change is ±8km / h. The average travel time at the cellular unit level is calculated based on the weighted average of the travel times of all road segments in the road network, and the delay time is calculated using the Webster formula. The conflict probability calculation of the risk assessment model combines the state transition probability of the hidden Markov chain, and an early warning is triggered when the probability of the S2 state exceeds 0.6. The simulation time step is set to 0.1 second, and the single simulation duration covers the typical cycle of traffic flow changes (such as 15 minutes).

[0146] The implementation of the verification process is divided into three stages. First, the traffic data collection module transmits the real-time data to the cloud platform, and after preprocessing by Kalman filtering, it is input into the digital twin environment. Secondly, the initial path and the adjusted path generated by the dynamic path planning framework are respectively subjected to 100 Monte Carlo simulations to statistically obtain the confidence interval of the perturbation index. Finally, the parameters of the risk assessment model are adjusted according to the simulation results. For example, the weight coefficient ω1 of the key node risk index NRI j is dynamically calibrated from 0.5±0.1. The entire verification cycle is set to 15 minutes, and after the parameters are updated, they are synchronized to the autonomous vehicle terminal.

[0147] As a result, the simulation error of the digital twin environment is controlled within ±10%, and the matching degree with the actual traffic flow reaches 92%; the prediction accuracy of the impact of path adjustment on the surrounding traffic flow density is increased to 88%; the prediction error of the average travel time of the cellular unit is reduced from 25% of the traditional method to 12%; the parameter update frequency of the risk assessment model is increased to 1 time per minute, and the system response speed is increased by 30%.

[0148] According to another embodiment of the present invention, in terms of parameter setting of the human-machine collaborative security strategy, the average value of the safety officer's response delay is set to 30 seconds, which is determined by collecting historical data through an on-vehicle event recorder and performing normal distribution fitting. The initial value of the abnormal event handling duration is set to 90 seconds, and when the traffic flow density in the cell reaches 80 vehicles per kilometer, the handling duration is extended to 108 seconds. The Mobileye EyeQ5 chip can be used to monitor the traffic flow density in real time, and the data is transmitted to the remote monitoring center through a 4G communication module. Parameter calibration is completed through tests in a closed test field. The test vehicle is an L4-level autonomous driving taxi equipped with a Bosch ESP system for braking control.

[0149] The implementation steps of the risk dynamic adjustment mechanism are as follows. First, when the traffic flow density exceeds the threshold, the system automatically triggers the mechanism for extending the handling duration, and the extension amplitude is 20% of the original duration. Second, the increase amplitude of the safety risk is calculated through a hidden Markov chain model. When it exceeds 30% of the threshold, the upper limit of the access density is reduced from 15 vehicles per kilometer to 10.5 vehicles per kilometer. At the same time, the safety risk weight coefficient is increased from 0.5 to 0.65, and the efficiency risk weight is correspondingly reduced. NVIDIA DRIVE Sim can be used for risk propagation simulation to verify the impact of weight adjustment on the system stability. The data update frequency is set to once per second to ensure real-time requirements.

[0150] The implementation of the warning and response process is divided into three links. First, the traffic flow density is monitored in real time through a Huawei RoadEye lidar. When the detected value exceeds the threshold for 5 consecutive minutes, a warning is triggered. Second, the system automatically generates an access density adjustment plan and pushes it to all autonomous driving vehicle terminals through a V2X communication module. Finally, the adjusted parameters are synchronized to the digital twin environment for verification, and the verification period is 30 minutes. TomTom Traffic Index API can be used to obtain regional traffic flow data and perform spatial analysis in combination with Esri ArcGIS. The measured data shows that this mechanism can reduce the intervention frequency of safety officers by 40% and shorten the emergency response time to within 15 seconds.

[0151] As a result, the misjudgment rate of safety risks is reduced by 28%, and the risk assessment accuracy rate is increased to 91%; the mechanism for extending the abnormal event handling duration reduces the accident rate by 23%; the dynamic adjustment of the access density improves the road network load balance by 35%; after the risk weight is optimized, the response speed of the system to sudden traffic events is increased by 40%.

[0152] According to another embodiment of the present invention, in terms of multi-source data access, a test self-driving taxi is equipped with a lidar (such as Velodyne HDL-32E) and a camera (such as Sony IMX490), and motion trajectory data is collected at a frequency of 10 Hz. Accident records are obtained through the in-vehicle EDR black box, and weather information comes from the meteorological bureau's API interface. Traffic flow data is collected in real time by geomagnetic sensors (such as 3M™ 7910) and microwave radars (such as Siemens SIRESPOT) deployed at intersections, with an update frequency of 5 Hz. The visibility in environmental parameters is monitored by a lidar visibility meter (such as Vaisala FD12P) installed on a street lamp pole, and the road surface adhesion coefficient is obtained through a friction coefficient sensor (such as Scantronix MC5) embedded in the road surface.

[0153] The data fusion processing flow is as follows. First, the Kalman filtering algorithm is used to denoise the vehicle trajectory data, and the positioning accuracy is controlled within 0.3 meters. Secondly, through the BP neural network, accident records, weather information, and real-time traffic flow data are fused to extract three core indicators: conflict rate (unit: times / hour), delay index (dimensionless), and accident density (times / km²). The key node risk index NRI j The calculation formula of is NRI j =0.5×C j +0.3×D j +0.2×A j , where C j Calculates the number of conflict events at intersections through video analysis, D j Is calculated based on the Webster delay formula, and A j Is interpolated from the historical accident database.

[0154] The steps for risk classification and monitoring strategy implementation are as follows. When NRI j ≥0.7, trigger the monitoring mechanism for type-I nodes, and perform real-time video analysis through edge computing nodes (such as Huawei Atlas 500) deployed at key locations to focus on monitoring high-risk behaviors such as pedestrians crossing the street and vehicles changing lanes. Type-II nodes use a rolling assessment with a period of 15 minutes and the support vector machine algorithm to predict the risk level in the future period. Type-III nodes update the basic information through crowdsourced data, and the data collection period is set to 24 hours. The monitoring resource allocation ratio is 50% for type-I nodes, 30% for type-II nodes, and 20% for type-III nodes, and the deployment density of cameras and sensors is optimized through the genetic algorithm.

[0155] Thus, the recognition accuracy of key nodes reaches 94%, and the false alarm rate is lower than 5%; the real-time monitoring of type-I nodes shortens the accident response time by 25%; the operation and maintenance cost is reduced by 30% after the optimized allocation of monitoring resources; the risk classification strategy improves the overall traffic efficiency of the road network by 18%.

[0156] The operation risk assessment system of the self-driving taxi of the present invention realizes accurate risk perception and dynamic path optimization through dynamic road network division, multi-dimensional data fusion and intelligent behavior prediction. The specific implementation process is as follows:

[0157] The first step: Dynamic road network division and key node identification:

[0158] The system obtains the travel demand density data of each region of the city in real time, including vehicle distribution heat map, order request volume and historical demand average value. Based on the dynamic adjustment algorithm, the urban road network is divided into hexagonal honeycomb cells, and the coverage range of each honeycomb automatically scales with the demand density: when the real-time demand is higher than the historical average level, the honeycomb radius shrinks to improve the risk positioning accuracy; when the demand is lower than the average value, the radius expands to reduce the calculation load. Within each honeycomb cell, by analyzing historical accident data and real-time traffic flow trajectories, areas with frequent vehicle lane changes, long signal waiting times or high accident rates are selected as key nodes, which are the core areas for subsequent risk assessment.

[0159] The second step: Multi-source data fusion and risk index calculation:

[0160] The system accesses and integrates multi-dimensional traffic data in real time: the traffic flow density is collected through roadside sensing devices and in-vehicle sensors, multi-modal fusion technology is used to detect vehicle emergency braking, lane change conflicts and pedestrian intrusion events, and the conflict frequency per unit time is counted; the communication delay data between the vehicle and the cloud is monitored synchronously, and the impact of the environment on traffic is evaluated in combination with meteorological parameters. Based on the above data, three core indicators are calculated for each key node:

[0161] 1. Conflict rate: The ratio of the number of conflict events to the total number of passing vehicles, reflecting the traffic conflict risk;

[0162] 2. Delay index: The average deviation between the actual travel time of the vehicle and the theoretical free flow time, representing the loss of traffic efficiency;

[0163] 3. Accident density: The spatial distribution statistics of the number of historical accidents around the node;

[0164] The comprehensive risk index is generated by weighted fusion of the above indicators. The weight allocation gives priority to the real-time conflict rate, followed by the delay index and the historical accident density. The system updates the risk level every five minutes, divides it into three threshold levels of high, medium and low, and triggers the corresponding early warning mechanism.

[0165] The third step: Behavior prediction and dynamic path decision-making:

[0166] Construct a vehicle behavior response prediction framework based on the Hidden Markov Model, discretize the object detection confidence, obstacle avoidance response speed, and communication delay data into three-level observation states, and define two hidden states: "safe" and "risk". Train the state transition probability and the distribution law of the observation state through historical data, and use the dynamic programming algorithm to infer the probability that the vehicle is in a risk state in real time. Combine the risk index of key nodes with the behavior prediction results to execute hierarchical decision-making:

[0167] When the node risk index reaches the high-risk threshold and the vehicle behavior risk probability continues to increase, immediately send a detour instruction to surrounding vehicles and update the no-go area map;

[0168] If either the node risk index or the behavior risk probability exceeds the standard alone, restrict the vehicle's driving speed and activate the redundant communication link;

[0169] For low-risk nodes, maintain regular monitoring and periodically refresh the evaluation data. The decision instructions are sent in real time through the vehicle-road collaborative network, synchronously optimizing the risk heat map of the cloud road network to achieve global path planning collaboration.

[0170] This system reduces the positioning error of traditional fixed grids through dynamic road network division, and the risk identification accuracy during peak hours is improved by more than 40%; the multi-dimensional risk assessment mechanism controls the false alarm rate within 7%, and the reliability is significantly enhanced compared with the single-index method; the behavior prediction model has an identification delay of less than 2 seconds for sudden risks, and the path dynamic adjustment enables the detour rate of vehicles in high-risk areas to exceed 95%, and the traffic efficiency is improved by about 20%.

[0171] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to the specific details and the illustrated and described examples here.

Claims

1. An evaluation method for the operation risks of autonomous taxis, characterized in that, Including: Dynamically divide the urban road network into hexagonal honeycomb cells, with the dynamic radius R of each honeycomb cell t According to real-time travel demands, calculate according to the formula calculate where R base is the reference radius, Q t is the travel demand density within the honeycomb during the current period, Q avg is the historical average demand density, Q max and Q min is the extreme value of the demand density, α is the adjustment coefficient, and key nodes are selected in each hexagonal honeycomb cell; Real-time access to multi-source traffic data, including traffic flow density, number of conflict events, communication delay data, and environmental parameters, to construct the key node risk index NRI j , and through the formula NRI j =ω1×C j +ω2×D j +ω3×A j Calculate, where C j is the conflict rate, D j is the delay index, A j is the accident density, and ω1, ω2, ω3 are preset weights; Simulating the behavioral responses of autonomous taxis at key nodes through a hidden Markov chain model, including: defining the hidden states as S1 (safe state) and S2 (risk state); collecting data on taxi target detection confidence, taxi obstacle avoidance response time, and taxi V2V communication delay, and discretizing them into observation states; Among them, according to the hidden states, a state transition probability matrix P is constructed. The construction method of the state transition probability matrix P is as follows: the probability of transitioning from S1 to S1 is 0.85 - 0.95, the probability of transitioning from S1 to S2 is 0.05 - 0.15, the probability of transitioning from S2 to S1 is 0.40 - 0.60, and the probability of transitioning from S2 to S2 is 0.40 - 0.60; according to the observation states, an emission probability matrix E is constructed. The emission probability matrix E is constructed based on the test data of multi-brand vehicles in a closed test field. Among them, when the taxi target detection confidence ≥ 0.9, the emission probability is 0.90; when the taxi obstacle avoidance response time ≤ 0.5 seconds, the emission probability is 0.85; when the taxi V2V communication delay ≤ 50 milliseconds, the emission probability is 0.92; using the Viterbi algorithm, combined with the state transition probability matrix P and the emission probability matrix E, infer the probabilities of each hidden state in the S1 safe state and S2 risk state at the current moment; According to the key node risk index NRI j Based on the value and the inferred hidden state probability, comprehensively determine the risk level, output low, medium, and high risk warnings, and dynamically adjust the path planning strategy according to the risk level.

2. The method for evaluating the operation risk of an autonomous taxi as described in claim 1, wherein, Including: When dynamically dividing the urban road network into hexagonal honeycomb cells, the Voronoi diagram algorithm is used to dynamically adjust the honeycomb boundary to ensure that adjacent honeycombs do not overlap and cover the entire area; The reference radius R base has a value range of 1 - 3 kilometers, and the adjustment coefficient α is determined by fitting historical traffic data to be 0.5 - 1.2; the key node risk index NRI j has weight coefficients ω1, ω2, and ω3 corresponding to the conflict rate C j , the delay index D j , and the accident density A j , and their value ranges are: ω1 is 0.4 - 0.6, ω2 is 0.3 - 0.5, and ω3 is 0.1 - 0.3; where R base is determined by fitting the urban road network density, and α is obtained through historical traffic flow regression analysis.

3. The method for evaluating the operation risk of an autonomous taxi according to claim 1, wherein The real-time access to multi-source traffic data further includes: Average vehicle speed, proportion of autonomous vehicles, number of emergency braking events, collaborative lane change failure rate, visibility, road surface adhesion coefficient, and signal light phase; And update the multi-source traffic data at a set frequency, where the multi-source traffic data is used to calculate the key node risk index NRI j , and combine the taxi target detection confidence, taxi obstacle avoidance response time, and taxi V2V communication delay data, and discretize them together to obtain the observation state of the hidden Markov chain model.

4. The method for evaluating the operation risk of an autonomous taxi according to claim 1, characterized in that, Also including: For the test self-driving taxi, determine the driving path planning range based on the dynamic radius R of the cellular unit t ​ Generate an initial path set within the planning scope, satisfying that the single-run mileage is 0.5 to 4 times the cellular radius and covering all NRIs within the cell j Key nodes with a value of ≥ 0.4; With the goal of minimizing the comprehensive path risk value, that is, the weighted sum of efficiency risk and safety risk, and maximizing the service response speed, an optimized driving model is constructed, In the optimized driving model, the cruising paths of idle vehicles are dynamically adjusted according to the traffic flow density, and idle vehicles are preferentially parked in the surrounding areas of high-risk nodes with j ≥0.7, and the impact of passenger boarding and alighting behavior on traffic flow density within the next 30 minutes is predicted based on Monte Carlo simulation;​ Execute the safety response mechanism. When the traffic flow density in the honeycomb is detected to reach 80 vehicles / km, automatically extend the abnormal event handling duration to 1.2 times the original handling duration, and at the same time trigger a path replanning instruction.

5. The method for evaluating the operation risk of an autonomous taxi according to claim 4, wherein Also including: Establish a closed-loop verification system that simulates the disturbance effect of path adjustment on the actual traffic flow in a digital twin environment, updates the parameters of the risk assessment model according to the simulation results, and applies the parameters to the cellular radius R in the dynamic path planning framework in real time t Calculation; The dynamic path planning framework and the real-time scheduling strategy are coordinated through the travel demand density Q within the cellular network during the current period. t The output data of the Monte Carlo simulation acts on both the optimization of the charging station location and the adjustment of the feeder point layout simultaneously. The execution result of the path replanning instruction is fed back to the closed-loop verification system in real time to form an iterative optimization loop.

6. The method for evaluating the operation risk of an autonomous taxi according to claim 5, wherein, The collaborative optimization of Monte Carlo simulation and dynamic routing strategy is achieved through the following steps: Input parameters: Passenger boarding and alighting behavior data, where the mean of the Poisson distribution is 3 minutes and the standard deviation is 1 minute; energy replenishment behavior data, where the charging time is 0.5 - 2 hours; Set the queuing overflow probability matrix for transfer points to quantify the vehicle retention risk at each transfer point during different time periods; Set the load imbalance index around the charging station to reflect the traffic flow density fluctuation within the service radius of the charging station; According to the queuing overflow probability matrix, dynamically adjust the location and capacity of the transfer points, and relocate the high-risk transfer points to low-load areas; Based on the load imbalance index, optimize the location of the charging stations and preferentially deploy them in areas where the road network load balance degree ≥ 0.8; Among them, when the traffic flow density in the honeycomb ≥ 80 vehicles / km, automatically start path replanning; Prioritize selecting sections that do not reach the load threshold, that is, the traffic flow density < 60 vehicles / km; dynamically adjust the path weight coefficient, increase the efficiency risk weight by 30%, and synchronize the path replanning result to the digital twin environment.

7. The method for evaluating the operation risk of an autonomous taxi according to claim 5, wherein The established closed-loop verification system is as follows: The closed-loop verification system is built based on digital twin technology, and this system realizes data interaction with a traffic data collection module that collects multi-source traffic data in real time, a dynamic path planning framework, and a risk assessment model; The traffic data collection module transmits the collected data to the closed-loop verification system; in the digital twin environment, a virtual scene consistent with the actual traffic scene is constructed based on real-time traffic data, and the scene covers road networks, traffic lights, and other traffic participants including pedestrians and ordinary vehicle elements; At the same time, the initial path set and the adjusted path plan generated by the dynamic path planning framework are input into the digital twin environment; According to the adjusted path plan, simulate the driving process of the autonomous driving taxi in the virtual scene, and analyze the disturbance impact of the path adjustment on the surrounding traffic flow. The specific quantification indicators include: Local traffic flow change: Calculate the change rate of the traffic flow density and the change value of the average vehicle speed on the surrounding road sections before and after the path adjustment; Evaluate the impact of the path adjustment on the average travel time and vehicle delay time within the entire cell; After simulating the path adjustment, evaluate the change in the probability and frequency of conflicts between the autonomous driving taxi and other traffic participants through the risk assessment model.

8. The evaluation method for the operation risk of an autonomous taxi according to claim 1, wherein The hidden state also includes a human-machine collaboration safety strategy, which includes the following steps: It is clear that the average response delay of the safety officer is 30 seconds, and the average duration for handling abnormal events is 90 seconds; Establish the correlation between the handling duration and the traffic flow density: When the traffic flow density within the cell is greater than or equal to 80 vehicles per kilometer, the handling duration increases by 20% on the original basis; The safety risk will increase due to the changes in the response delay and handling duration, and the increased amplitude is positively correlated with the adjustment of the handling duration; When the increased amplitude of the safety risk reaches or exceeds the threshold of 30%, or the increased amplitude of the efficiency risk reaches or exceeds another threshold of 20%, reduce the upper limit of the access density of autonomous driving taxis within the cell, adjust it to 70% of the original, and increase the weight of the safety risk by 30%; at the same time, push a warning message to the remote monitoring center.

9. An evaluation system for the operating risks of autonomous taxis, characterized in that, Including: The road network dynamic division module is used to dynamically divide the urban road network into hexagonal honeycomb units according to real-time travel demands. Among them, the dynamic radius R of each honeycomb unit t is calculated through the formula where R base is the reference radius, Q t is the travel demand density within the honeycomb during the current period, Q avg is the historical average demand density, Q max and Q min are the extreme values of the demand density, α is the adjustment coefficient, and key nodes are selected in each hexagonal honeycomb unit; A multi-source data access module for obtaining traffic flow density, the number of conflict events, communication delay data, and environmental parameters in real time; A key node risk assessment module for calculating a key node risk index NRI based on the multi-source data j , where NRI j is calculated through the formula NRI j = ω1 × C j + ω2 × D j + ω3 × A j . C j is the conflict rate, D j is the delay index, A j is the accident density, and ω1, ω2, and ω3 are preset weights; The behavior response simulation module includes a hidden Markov chain model, which is used to define the hidden states S1 (safe state) and S2 (risk state), discretize the target detection confidence, obstacle avoidance response time, and V2V communication delay data of the autonomous taxi into observation states, and construct the state transition probability matrix P and the emission probability matrix E. The construction method of the state transition probability matrix P is as follows: the probability of transitioning from S1 to S1 is 0.85 - 0.95, the probability of transitioning from S1 to S2 is 0.05 - 0.15, the probability of transitioning from S2 to S1 is 0.40 - 0.60, and the probability of transitioning from S2 to S2 is 0.40 - 0.

60. The emission probability matrix E is constructed based on the test data of multiple brand vehicles in a closed test field. Among them, when the target detection confidence of the taxi is ≥ 0.9, the emission probability is 0.90; when the obstacle avoidance response time of the taxi is ≤ 0.5 seconds, the emission probability is 0.85; when the V2V communication delay of the taxi is ≤ 50 milliseconds, the emission probability is 0.

92. The Viterbi algorithm is used to infer the probabilities of each hidden state at the current moment by combining matrices P and E; A comprehensive risk decision-making module for comprehensively determining low, medium, and high risk levels based on the key node risk index NRI j and the hidden state probability, outputting a warning signal and dynamically adjusting the path planning strategy of the autonomous driving taxi.

Images