A dynamic path planning method fusing real-time road conditions and user driving habits

Abstract

This invention relates to the field of intelligent transportation technology, specifically to a dynamic path planning method that integrates real-time traffic conditions and user driving habits. The method acquires high-precision lane-level positioning information for vehicles through a ground augmentation system. This high-precision lane-level positioning information includes lane number, heading, speed, and positioning confidence. V2X data is acquired through V2X communication, and real-time traffic data is acquired through at least one of traffic cloud, floating cars, and roadside equipment. The V2X data and the real-time traffic data include signal phase. This invention advances path planning from the segment level to the lane and intersection turning levels, and maps user driving habits into a personalized dynamic risk potential field. This allows navigation paths to automatically adapt to user habits, reducing manual settings, improving the matching degree between the path and the user's actual driving behavior, and enhancing the intersection traffic experience and driving comfort.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically, to a dynamic path planning method that integrates real-time traffic conditions and user driving habits. Background Technology

[0002] Existing dynamic path planning methods typically acquire real-time traffic conditions through traffic clouds, floating cars, and roadside equipment. They use algorithms such as A and DLite to search for paths, at the cost of travel time, distance, and congestion levels. Some methods learn simple preferences based on users' historical trajectories, such as a preference for highways, avoiding toll stations, and minimizing congested sections, and add fixed or semi-fixed weights to the path cost. With the development of ground augmentation systems, RTK / PPP-RTK, V2X, and high-precision maps, vehicles have the ability to obtain lane-level positioning and lane-level traffic conditions. However, existing path planning methods still mostly remain at the road segment level and do not fully utilize information such as lane-level positioning, signal phase, queue length, and intersection turning probability.

[0003] Existing technologies mainly suffer from two problems. First, path planning lacks granularity and personalization. Traditional navigation uses road segments as the basic unit, merging different lanes and turns within a segment, failing to distinguish the differences in traffic flow between straight, left, and right turns, and between different lanes within the same road segment. User habit modeling relies heavily on historical frequency or manual weighting, making it difficult to express deep-seated habits such as starting response, lane change aversion, concerns about unprotected left turns, and caution at complex intersections. Real-time traffic conditions and user habits are often simply superimposed with independent weights, failing to form a unified cost model, resulting in significant differences in the passing strategies of different drivers at the same intersection. High-precision positioning... Secondly, location-level and lane-level maps and V2X data are not fully integrated into the core of route search. Furthermore, there is insufficient handling of road condition prediction errors, a lack of closed-loop simulation and personalized risk control. Existing methods mostly plan based on expected optimality or deterministic predictions, without considering the distribution of prediction errors. Once actual road conditions deviate from the prediction, the route quality drops sharply. Simultaneously, the lack of closed-loop simulation of actual user driving behavior makes it difficult to predict personalized travel time, energy consumption, risk, fatigue, and comfort. Risk handling often uses fixed thresholds or soft penalties, making it impossible to set personalized risk budgets and failing to cater to the different needs of conservative and aggressive users.

[0004] Based on this, the present invention discloses a dynamic path planning method that integrates real-time traffic conditions and user driving habits. Summary of the Invention

[0005] To address the problems mentioned in the background art, this invention provides a dynamic path planning method that integrates real-time traffic conditions and user driving habits, comprising: S1. High-precision lane-level positioning information of the vehicle is acquired through a ground-based augmentation system. This high-precision lane-level positioning information includes lane number, heading, speed, and positioning confidence level. V2X data is acquired through V2X communication. Real-time traffic data is also acquired through at least one of traffic cloud, floating car, and roadside equipment. The V2X data and the real-time traffic data include signal phase, queue length, and lane-level clearance probability. and the probability of turning at intersections, where ; S2 extracts user driving habit parameters from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data, and constructs a personalized dynamic risk potential field based on the user driving habit parameters. ,in , For lanes, For turning, For time, the personalized dynamic risk potential field is a weighted sum of the road potential field, obstacle potential field, congestion potential field, risk potential field, comfort potential field and V2X passage probability potential field. Each weight is adjusted online according to the user's driving habit parameters and the sum of the weights is 1. Injecting the lane-level passage probability into the personalized dynamic risk potential field in the form of a negative logarithm yields: in, The lower, The higher the value, the more likely the path search will avoid the corresponding lane or turn; S3, construct lane-level road network and turning map, and define the search state as: The dynamic costs of lane and steering are updated based on the injected personalized dynamic risk potential field, and the edge costs satisfy: in, , , , , For travel time, For energy consumption, For risk, For personalized dynamic risk potential fields, , , , These are the corresponding reference values. , , , As cost weight, and =1; The potential-field-guided A algorithm is used in the lane-level road network and turning map to perform path search with the updated dynamic cost, and outputs lane-level driving paths, intersection turning sequences, and lane change timing suggestions; the evaluation function of the potential-field-guided A algorithm is: ,in From the starting point to the current state The price has already been paid. For heuristic functions, , The distance from the current state to the target point is the road network distance. The heuristic function, which is the historical maximum speed of the current road segment, satisfies the acceptability requirement; the lane change timing suggestion includes the suggested lane, suggested lane change time, or suggested lane change interval. S4. Real-time monitoring of dynamic replanning trigger conditions. When the dynamic replanning trigger conditions are met, the existing search tree is reused for dynamic replanning, and the updated lane-level driving path, intersection turning sequence, and lane change timing suggestions are output. The dynamic replanning trigger conditions include at least one of the following: travel time change exceeding 10%, lane-level pass probability decrease exceeding 20%, vehicle deviation from the recommended path exceeding a preset distance, accident, construction, and signal phase change.

[0006] As a further improvement to this technical solution, the ground augmentation system includes a ground-based augmentation reference station, an RTK / PPP-RTK differential service broadcasting terminal, and an onboard high-precision positioning module. The positioning accuracy of the high-precision lane-level positioning information is from centimeter to decimeter level. The V2X data is obtained by communicating with the roadside unit (RSU) through the onboard unit (OBU) to obtain SPAT and MAP messages. The lane-level pass probability... Indicates time The probability of passing through an intersection in a designated lane with a designated turn. When the positioning confidence is lower than a preset confidence threshold, inertial measurement unit data and vehicle wheel speed data are used to compensate for the lane-level positioning information. The preset confidence threshold is 0.8 to 0.95.

[0007] As a further improvement to this technical solution, the user driving habit parameters in S2 include start-up reaction time, following distance, lane change aversion, unprotected left turn aversion, curvature sensitivity, congestion tolerance, caution at complex intersections, energy consumption sensitivity, lateness risk tolerance, and nighttime risk preference. These user driving habit parameters are updated online through inverse reinforcement learning. The inverse reinforcement learning takes state-action pairs from historical trajectories as input and driving habit parameters as output, solving for the likelihood of the expert trajectory and the learned policy trajectory by maximizing the likelihood. The inverse reinforcement learning uses a neural network as a policy function approximator. The neural network's input... The input layer dimension is the number of state features, the output layer dimension is the number of driving habit parameters, the hidden layer consists of two fully connected layers, the activation function is the ReLU function, the loss function is the negative log-likelihood of the expert trajectory and the learning policy trajectory, the optimizer is the Adam optimizer, the learning rate is 0.0013, the batch size is 52, and the number of training epochs is 118. The online adjustments include: increasing the lane change penalty coefficient in the comfort potential field when lane change aversion increases; increasing the risk potential field corresponding to the left turn when unprotected left turn aversion increases; increasing the congestion potential field weight when congestion tolerance decreases; and increasing the energy consumption weight in the edge cost when energy consumption sensitivity increases.

[0008] As a further improvement to this technical solution, the road potential field in S2 is obtained by weighting the normalized values ​​of lane curvature, slope, number of lanes, and intersection complexity; the obstacle potential field is obtained by weighting the normalized values ​​of accidents, construction, large vehicles, and vulnerable road users; the congestion potential field is obtained by weighting the normalized values ​​of queue length, signal waiting time, and number of stops; the risk potential field is obtained by weighting the normalized values ​​of collision risk, violation risk, and fatigue risk; and the comfort potential field is obtained by weighting the normalized values ​​of lane change frequency, acceleration / deceleration frequency, and road surface bumpiness. The weighted values ​​are obtained by weighting the normalized values. Each normalized value is the ratio of the corresponding physical quantity to its preset maximum value, and the values ​​range from 0 to 1. Each weighting coefficient is determined by offline calibration or online least squares fitting. The offline calibration involves traversing the value combinations of each weighting coefficient in the simulation environment and selecting the optimal weighting coefficient with the goal of minimizing the comprehensive cost of path planning. The online least squares fitting involves using real-time collected driving operation data as the observation value and the normalized value of each sub-potential field as the independent variable, and solving for the weighting coefficient by recursive least squares method.

[0009] As a further improvement to this technical solution, the construction of the personalized dynamic risk potential field, the injection of the lane-level passage probability, and the path search of the A* algorithm guided by the potential field are deployed on a mobile edge computing node. The mobile edge computing node is connected to the ground augmentation system and the V2X communication module through a low-latency link, with an end-to-end latency of less than 10 milliseconds. The mobile edge computing node outputs lane-level driving paths, intersection turning sequences, and lane change timing suggestions to the in-vehicle navigation terminal. The mobile edge computing node works in collaboration with the cloud server. The mobile edge computing node performs the real-time potential field construction and path search, while the cloud server performs digital twin model training and global road network optimization. Data synchronization between the mobile edge computing node and the cloud server adopts an incremental synchronization method, synchronizing only the changed road network data and user driving habit parameters.

[0010] As a further improvement to this technical solution, the preset distance in S4 is 50 meters to 200 meters. The dynamic replanning triggering condition also includes the user driving habit drift detection value exceeding a threshold. The user driving habit drift detection value is obtained by comparing the mean deviation and variance deviation of the user driving habit parameters in the current time window and the historical time window. When the mean deviation or the variance deviation exceeds the corresponding threshold, dynamic replanning is triggered. The signal phase change includes the deviation between the current signal phase and the V2X predicted signal phase exceeding a preset phase deviation threshold. The preset phase deviation threshold is... Instant .

[0011] As a further improvement to this technical solution, T1 acquires high-precision lane-level positioning information of the vehicle through a ground augmentation system. The high-precision lane-level positioning information includes lane number, heading, speed, and positioning confidence. User driving habit parameters are extracted from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data. A user driving digital twin model is constructed based on the user driving habit parameters. The user driving digital twin model includes reaction time, following distance, lane change preference, acceleration and braking style, energy consumption characteristics, fatigue accumulation, risk sensitivity, congestion tolerance, toll sensitivity, and comfort preference. The user driving digital twin model is trained through behavioral cloning or inverse reinforcement learning, and the training data is the state-action sequence in the historical trajectory. T2. Construct a set of uncertain distributions for real-time traffic conditions. This set is constructed using at least one of the following: a scene set, a Wasserstein sphere, an interval distribution, and a Gaussian mixture distribution. When using a Wasserstein sphere, the sphere radius is determined based on the confidence level of the real-time traffic prediction error. For candidate paths... In scenarios with uncertain road conditions The comprehensive cost is calculated by weighting the normalized values ​​of time, energy consumption, comfort, risk, fatigue, and charges, with the sum of each weight being 1. T3 uses a split-bar and conditional value at risk (CVaR) optimization model: in, For a distribution in an uncertain distribution set, In distribution The expectations below For distribution Lower confidence level Conditional Value at Risk: ; in , For user risk budget parameters, and Based on the user driving habit parameters, the CVaR optimization model is solved by linear programming or stochastic gradient descent. T4. Shadow simulation is performed on multiple candidate paths in the user driving digital twin model. The shadow simulation refers to the parallel simulation of the vehicle driving along the candidate path in the user driving digital twin model. GPU parallel computing is used, the simulation time step is 0.1 seconds, and the number of scene samples is 1000. The personalized travel time, energy consumption, comfort, risk, fatigue and toll pre-simulation results of each candidate path are obtained. Based on the pre-simulation results, the NSGA-II algorithm is used to generate a Pareto optimal path set. When the user does not interact, the multi-objective weights are determined according to the user driving habit parameters. The path with the minimum weighted comprehensive cost is selected from the Pareto optimal path set. T5 feeds the actual driving results back to the user driving digital twin model and the user driving habit parameters for closed-loop updates. When real-time road conditions change or the actual driving deviates from the recommended path, dynamic replanning is triggered, and existing shadow simulation results are reused. Among them, the candidate path The lane-level driving path generated by any one of the methods described in claims 1 to 6 constitutes the driving path, or the lane-level driving path generated by any one of the methods described in claims 1 to 6 constitutes the driving path together with historically frequently used paths.

[0012] As a further improvement to this technical solution, the high-precision positioning and V2X data acquisition module is configured to acquire high-precision lane-level positioning information of vehicles through a ground augmentation system, acquire V2X data through V2X communication, and acquire real-time traffic data through at least one of traffic cloud, floating cars, and roadside equipment. The high-precision lane-level positioning information includes lane number, heading, speed, and positioning confidence level; the V2X data and the real-time traffic data include signal phase, queue length, and lane-level passage probability. and the probability of turning at intersections, where ; The driving habit and potential field construction module is configured to extract user driving habit parameters from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data, and construct a personalized dynamic risk potential field based on the user driving habit parameters. and the lane-level pass probability is... Inject the personalized dynamic risk potential field; The lane-level potential field search and replanning module is configured to construct lane-level road networks and turning maps, defining the search state as follows: Based on the injected personalized dynamic risk potential field, the dynamic costs of lane and steering are updated, and the edge costs satisfy... ,in , , , The potential field-guided A* algorithm is used for path search, outputting lane-level driving paths, intersection turning sequences and lane change timing suggestions. When the dynamic replanning trigger condition is met, the existing search tree is reused for dynamic replanning. The digital twin and robust optimization module is configured to construct a user driving digital twin model, construct a set of real-time road condition uncertainty distributions, perform shadow simulations on candidate paths, solve the paths using a split-loop and CVaR optimization model, and update the user driving digital twin model and the user driving habit parameters in a closed loop based on the actual driving results. The high-precision positioning and V2X data acquisition module outputs the high-precision lane-level positioning information, the V2X data, and the real-time traffic data to the driving habit and potential field construction module. The driving habit and potential field construction module outputs the personalized dynamic risk potential field to the lane-level potential field search and replanning module. The lane-level potential field search and replanning module outputs the lane-level driving path, the intersection turning sequence, and the lane change timing suggestion. The digital twin and robust optimization module receives the lane-level driving path and performs shadow simulation and robust optimization to output a personalized lane-level navigation path. The system is used to execute the method of any one of claims 1 to 7.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention advances route planning from the segment level to the lane level and intersection turning level, and maps user driving habits into a personalized dynamic risk potential field, enabling navigation routes to automatically adapt to user habits, reducing manual settings, improving the matching degree between the route and the user's actual driving behavior, and improving the intersection passage experience and driving comfort.

[0014] 2. This invention optimizes the road conditions by using a split-loop bar and conditional value of risk (CVaR) and combines user driving digital twins and shadow simulations to form a closed-loop pre-simulation. This can effectively handle real-time road condition prediction errors, provide personalized risk control for different users, enhance the robustness of the route under uncertain road conditions, reduce risk exposure, and balance traffic efficiency and safety.

[0015] 3. This invention adopts mobile edge computing and cloud-based collaborative deployment to meet the latency requirements of real-time path planning, and has the ability to degrade operation in abnormal scenarios such as V2X communication interruption, ensuring continuous system availability, while reducing energy consumption and emissions, and has good prospects for industrial application. Attached Figure Description

[0016] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart of the overall process of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In the following embodiments, unless otherwise specified, all hardware used are commercially available vehicle-mounted or roadside-mounted products. The ground augmentation system adopts commercially available RTK / PPP-RTK differential services. The positioning accuracy of the vehicle-mounted high-precision positioning module is from centimeter to decimeter level. The V2X communication adopts OBU and RSU that comply with the 3GP PRELEASE 14 / 15 standard. The vehicle-mounted computing platform adopts NVIDIA Jetson AGX Orin or edge computing nodes with equivalent computing power. The simulation verification platform adopts the SUMO, CARLA and VISSIM joint simulation environment.

[0019] The performance testing methods involved are as follows: the comprehensive cost of path planning is calculated according to the edge cost formula of this invention, and a weighted sum is taken after normalization; the number of lane changes is counted as the number of times the steering wheel angle exceeds a preset threshold and crosses lane lines during the process from the start to the end; the number of complex intersection exposures is counted as the number of times the vehicle passes through intersections with more than or equal to 4 lanes and more than or equal to 6 turning conflict points; the number of unprotected left turns is counted as the number of times the vehicle completes a left turn without a dedicated left turn signal phase; the travel time is the actual travel time of the vehicle from the start to the end; energy consumption is obtained by reading the motor or engine power through the CAN bus and integrating it over time; and risk exposure is determined by collision exposure. The number of events with a time-to-completion (TTC) of less than 2 seconds was counted. User acceptance was assessed using a 5-level Likert scale, with drivers rating route comfort, safety, and conformity to their habits. Replanning delay was defined as the time from triggering the replanning condition to outputting the new route. Positioning accuracy was obtained by comparing the fixed solution of RTK with the true value. The probability accuracy was calculated by comparing the predicted lane-level pass probability with the actual pass result and calculating the AUC value. The CVaR improvement was compared to the tail risk reduction of the traditional expected optimal path and the CVaR optimized path of this invention. In the following embodiments, each set of performance data is the mean of 10 independent experiments, and the standard deviation is in parentheses.

[0020] Example 1 This embodiment is a lane-level dynamic path planning method based on V2X lane-level passage probability and personalized dynamic risk potential field.

[0021] The ground augmentation system includes a ground-based augmentation reference station, an RTK / PPP-RTK differential service broadcasting terminal, and a vehicle-mounted high-precision positioning module. The reference station broadcasts differential correction data at a frequency of 100Hz. The vehicle-mounted module outputs the vehicle's lane number, heading, speed, and positioning confidence level. The positioning accuracy is from centimeter to decimeter level. When the positioning confidence level is lower than 0.85, inertial measurement unit data and vehicle wheel speed data are used to compensate for the lane-level positioning information.

[0022] The V2X communication module communicates with the roadside RSU via the OBU to obtain SPaT and MAP messages. The message fields include intersection number, signal phase, phase remaining time, queue length, and lane-level pass probability. And the probability of turning at an intersection, where the probability of passing at the lane level is... The time value is calculated by the roadside edge computing nodes based on historical statistics and real-time detection. The probability of passing through an intersection in a specified lane with a specified turn, ranging from [value range missing]. The driving habit extraction module extracts user driving habit parameters from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data. These parameters include start reaction time, following distance, lane change aversion, unprotected left turn aversion, curvature sensitivity, congestion tolerance, caution at complex intersections, energy consumption sensitivity, lateness risk tolerance, and nighttime risk preference. These parameters are updated online through inverse reinforcement learning. Inverse reinforcement learning uses a neural network as the policy function approximator. The input layer has a dimension of 8, and the state features include vehicle speed, vehicle acceleration, distance to the vehicle in front, lane departure, steering wheel angle, intersection type, signal phase, and queue length. The output layer has a dimension of 10, corresponding to 10 driving habit parameters. The hidden layers are two fully connected layers with 64 neurons each. The activation function is the ReLU function, the loss function is the negative log-likelihood of the expert trajectory and the learned policy trajectory, the optimizer is the Adam optimizer, the learning rate is 0.0013, the batch size is 52, and the number of training epochs is 118.

[0023] The personalized dynamic risk potential field construction module maps driving habit parameters to potential field parameters, constructing a total potential field. This total potential field is obtained by weighting and summing the road potential field, obstacle potential field, congestion potential field, risk potential field, comfort potential field, and V2X potential field using probabilistic potential fields. Each weight is adjusted online based on the user's driving habit parameters, and the sum of the weights is 1. The road potential field is obtained by weighting the normalized values ​​of lane curvature, gradient, number of lanes, and intersection complexity. The obstacle potential field is obtained by weighting the normalized values ​​of accidents, construction, large vehicles, and vulnerable road users. The congestion potential field is obtained by weighting the normalized values ​​of queue length, signal waiting time, and number of stops. The risk potential field is obtained by weighting the normalized values ​​of collision risk and traffic violations. The normalized values ​​of risk and fatigue risk are weighted together; the comfort potential field is obtained by weighting together the normalized values ​​of lane change frequency, acceleration / deceleration frequency, and road bumpiness. Each normalized value is the ratio of the corresponding physical quantity to its preset maximum value, and the values ​​range from 0 to 1. Each weighting coefficient is determined by offline calibration or online least squares fitting. Offline calibration involves iterating through the combinations of weighting coefficient values ​​in the SUMO simulation environment and selecting the optimal weighting coefficient with the goal of minimizing the overall cost of path planning. Online least squares fitting uses real-time collected driving operation data as observations and the normalized values ​​of each sub-potential field as independent variables, and solves for the weighting coefficients by recursive least squares method.

[0024] The mapping relationship between driving habit parameters and potential field weights is as follows: when lane change aversion increases, the lane change penalty coefficient in the comfort potential field is increased; when unprotected left turn aversion increases, the risk potential field corresponding to the left turn is increased; when congestion tolerance decreases, the weight of the congestion potential field is increased; when energy consumption sensitivity increases, the energy consumption weight in the edge cost is increased.

[0025] The lane-level passage probability is injected into a personalized dynamic risk potential field in the form of a negative logarithm, resulting in: , The lower, The higher the value, the more the path search tends to avoid the corresponding lane or turn. The lane-level road network and turn map construction module constructs a lane-level road network that includes lane, turn, and intersection connection relationships. The search state is defined as follows: Edge costs are calculated using a normalized weighted sum: in, , , , , For travel time, For energy consumption, For risk, For personalized dynamic risk potential fields, , , , These are the corresponding reference values. , , , As cost weight, and =1, in this embodiment =60 seconds, =0.5 kilowatts, =1, =1, =0.38、 =0.22、 =0.21、 =0.19.

[0026] The potential field path search module uses the potential field-guided A* algorithm, with the following evaluation function: ,in From the starting point to the current state The price has already been paid. For heuristic functions, , The distance from the current state to the target point is the road network distance. Given the current road segment's historical maximum speed, this heuristic function satisfies the acceptability requirement, i.e. ,in To achieve the minimum residual cost, the guiding role of the personalized dynamic risk potential field is reflected in the potential field term in the edge cost. That is, the higher the potential field of a lane or turn, the greater its edge cost, and the more the path search tends to avoid it. During dynamic replanning, only the subtrees affected by the triggering conditions in the existing search tree are updated, while unaffected nodes are retained.

[0027] The timing of lane changes is determined by a combination of factors, including the time difference between the current lane and the target lane, the difference in V2X pass probability, and the user's lane change aversion. When the time to pass in the target lane is more than 8 seconds less than that in the current lane, the pass probability of the target lane is more than 0.15 higher than that in the current lane, and the user's lane change aversion is less than 0.6, a lane change suggestion is output. The lane change suggestion includes the suggested lane, the suggested time, or the suggested interval.

[0028] Dynamic replanning is triggered by at least one of the following conditions: a change in travel time of more than 10%, a decrease in lane-level passability of more than 20%, a vehicle deviating from the recommended path of more than 100 meters, an accident, construction, or a signal phase change. A signal phase change includes a deviation between the current signal phase and the V2X predicted signal phase of more than ±4 seconds.

[0029] In this embodiment, the construction of personalized dynamic risk potential fields, the injection of lane-level pass probabilities, and the path search of the potential field-guided A* algorithm are deployed on mobile edge computing nodes. The mobile edge computing nodes are connected to the ground augmentation system and the V2X communication module through 5G URLLC low-latency links, with an end-to-end latency of less than 10 milliseconds. The mobile edge computing nodes will send the output lane-level driving paths, intersection turning sequences, and lane change timing suggestions to the vehicle navigation terminal. The mobile edge computing nodes and the cloud server work together. The mobile edge computing nodes perform the real-time potential field construction and path search, while the cloud server performs digital twin model training and global road network optimization. Data synchronization between the two adopts an incremental synchronization method, specifically a version vector-based incremental synchronization, which only transmits data items with inconsistent version numbers.

[0030] The simulation was validated in a SUMO and CARLA co-simulation environment. The scenario included 12 signalized intersections and a total length of 8 kilometers of urban roads. The simulation time periods were set as follows: morning peak from 7:30 to 9:00, evening peak from 17:30 to 19:00, and off-peak from 10:00 to 16:00. The road types included urban arterial roads, urban secondary roads, and expressway ramps. During the morning and evening peak congestion scenarios, the vehicle speed ranged from 5 to 20 kilometers per hour, and the queue length ranged from 20 to 50 vehicles. During the off-peak smooth flow scenarios, the vehicle speed ranged from 30 to 60 kilometers per hour, and the queue length ranged from 0 to 5 vehicles. In the rainy weather scenario, the road surface friction coefficient was reduced to 0.4 to 0.6.

[0031] The mean results of 10 independent experiments show that, compared with the traditional road segment-level real-time traffic conditions plus fixed preference weight scheme, this embodiment reduces the number of lane changes by 31.7% (standard deviation 2.4%), the number of exposures at complex intersections by 28.3% (standard deviation 2.1%), the number of unprotected left turns by 41.2% (standard deviation 3.2%), the travel time increases by no more than 3.1% (standard deviation 0.8%), the user acceptance score increases from 3.2 to 4.5 (standard deviation 0.2), the average replanning delay is 85 milliseconds (standard deviation 8 milliseconds), the positioning accuracy is 0.1 meters, and the pass probability accuracy AUC is 0.91.

[0032] In congested morning and evening rush hour scenarios, this embodiment reduces lane change frequency by 26.4% (standard deviation 2.8%) and exposure frequency at complex intersections by 23.8% (standard deviation 2.5%). In off-peak, smooth-flowing scenarios, lane change frequency decreases by 34.2% (standard deviation 2.2%) and exposure frequency at complex intersections by 31.5% (standard deviation 2.0%). In rainy weather scenarios, due to the decreased probability of V2X passage, this embodiment automatically increases the congestion potential field weight, increasing travel time by no more than 5.2% (standard deviation 1.1%) and reducing risk exposure by 37.6% (standard deviation 3.0%).

[0033] Example 2 This embodiment is a personalized path planning method based on digital twin and distributed CVaR, which can be used as a supplement to embodiment one or implemented independently.

[0034] High-precision lane-level positioning information of the vehicle is obtained through a ground augmentation system. User driving habit parameters are extracted from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data. A user driving digital twin model is constructed based on the user driving habit parameters. The model includes reaction time, following distance, lane change preference, acceleration and braking style, energy consumption characteristics, fatigue accumulation, risk sensitivity, congestion tolerance, toll sensitivity, and comfort preference. The digital twin model is trained through behavioral cloning or inverse reinforcement learning. The training data is the state-action sequence in the historical trajectory. In this embodiment, the digital twin model adopts a neural network with an input layer dimension of 6. The state features include speed, acceleration, distance to the vehicle in front, lane departure, intersection type, and signal phase. The hidden layer consists of two fully connected layers with 128 neurons each and the activation function is ReLU. The output layer dimension is 2, outputting acceleration and steering angle respectively. The loss function is mean squared error (MSE). The optimizer is Adam, the learning rate is 0.0013, the batch size is 36, and the number of training epochs is 180.

[0035] A set of uncertain distributions of real-time traffic conditions is constructed using a Wasserstein sphere. The radius of the sphere is determined based on the confidence level of the real-time traffic prediction error. The mathematical relationship between the confidence level 1-δ and the mathematical relationship between the confidence level 1-δ is: ,in For the sample size, take... =250, confidence level 0.95, calculated as follows In this embodiment, 0.15, for candidate paths In scenarios with uncertain road conditions The overall cost is calculated by weighting the normalized values ​​of time, energy consumption, comfort, risk, fatigue, and charges, with the sum of each weight being 1.

[0036] The optimization model employs split-bar and CVaR: ;in, For an uncertain distribution set Distribution in In distribution The expectations below For distribution Lower confidence level Conditional Value at Risk: ;in ,Pick =0.9, =0.5, for conservative users, Take 0.95, For aggressive users, a value of 0.8 is used. Take 0.8, The optimization model with a VaR of 0.3 is solved by linear programming. The number of linear programming variables is the number of scenarios plus 1, and the number of constraints is the number of scenarios. When the number of scenarios is 1000, the interior point method is used to solve the problem, and the computation time is approximately 50 milliseconds.

[0037] Shadow simulations were performed on multiple candidate routes within a user driving digital twin model. The shadow simulation employed GPU parallel computing with a simulation time step of 0.1 seconds and 1000 scene samples. Scenes were generated using Monte Carlo sampling, following a normal distribution with a mean of the real-time traffic prediction and a standard deviation 1.5 times the prediction error. Personalized travel time, energy consumption, comfort, risk, fatigue, and toll simulation results were obtained for each candidate route. Based on these simulation results, the NSGA-II algorithm was used to generate a Pareto optimal route set with a population size of 100 and an iteration count of [number missing]. With a crossover probability of 0.9 and a mutation probability of 0.1, when the user does not interact, the multi-objective weights are determined based on the user's driving habit parameters. The path with the minimum weighted comprehensive cost is selected from the Pareto optimal path set. The method for determining the multi-objective weights is as follows: for the six objectives of time, energy consumption, comfort, risk, fatigue, and toll, normalize the user's driving habit parameters for lateness risk tolerance, energy consumption sensitivity, comfort preference, risk sensitivity, fatigue sensitivity, and toll sensitivity respectively. Then, take the reciprocal of the normalized value as the weight, and finally normalize so that the sum of the weights is 1.

[0038] The actual driving results are fed back to the user's digital twin model and user driving habit parameters for closed-loop updates. The updates employ a gradient update method with a learning rate of 0.01. When real-time road conditions change or the actual driving deviates from the recommended path, dynamic replanning is triggered, reusing existing shadow simulation results and candidate paths. It consists of lane-level driving paths generated by Example 1, or lane-level driving paths generated by Example 1 combined with historically frequently used paths.

[0039] The results were validated in the CARLA simulation environment. The mean results of 10 independent experiments showed that, in the extreme scenario with a prediction error of 20%, this embodiment improved CVaR by 35.2% with a standard deviation of 2.8% compared to the traditional expected optimal path, while only increasing travel time by 5.3% with a standard deviation of 1.2%, reducing risk exposure by 42.1% with a standard deviation of 3.1%, and achieving a user acceptance score of 4.3 with a standard deviation of 0.2. The average shadow simulation time was 120 milliseconds with a standard deviation of 12 milliseconds, meeting the real-time requirements. In the scenario with a prediction error of 10%, the CVaR improved by 22.7% with a standard deviation of 2.3%, and in the scenario with a prediction error of 30%, the CVaR improved by 41.5% with a standard deviation of 3.5%. For conservative users, risk exposure was reduced by 48.3% with a standard deviation of 3.4%, while travel time increased by 7.1% with a standard deviation of 1.5%. For aggressive users, travel time was reduced by 3.2% with a standard deviation of 1.0%, while risk exposure was reduced by 28.6% with a standard deviation of 2.7%.

[0040] Example 3 This embodiment is a combination of Embodiment 1 and Embodiment 2, and adopts mobile edge computing and cloud collaborative deployment.

[0041] The vehicle-mounted terminal uploads ground-based augmented positioning information, V2X data, and vehicle bus data to the mobile edge computing node. The mobile edge computing node runs the personalized dynamic risk potential field construction of Example 1, and generates an initial lane-level path through V2X probability injection and potential field-guided A* search. At the same time, the mobile edge computing node sends the initial path to the cloud server. The cloud server runs the digital twin shadow simulation and distributed bary CVaR optimization of Example 2 to perform robustness evaluation and Pareto optimization on the initial path. The optimized path is then sent to the mobile edge computing node. The mobile edge computing node updates the search tree based on the optimized path and outputs the final lane-level driving path, intersection turning sequence, and lane change timing suggestions.

[0042] Data synchronization between mobile edge computing nodes and cloud servers adopts an incremental synchronization method, specifically version vector-based incremental synchronization, which only synchronizes changed road network data and user driving habit parameters. The low-latency link adopts 5G URLLC, with end-to-end latency of less than 10 milliseconds. The digital twin model is trained in the cloud. After training, INT8 quantization compression is used to quantize the model parameters from 32-bit floating-point to 8-bit integers, compressing the model size to one-quarter of the original size and improving the inference speed by about 2 times. The model prediction accuracy loss after quantization is less than 2%, which meets the path planning requirements, and then it is sent to mobile edge computing nodes for inference.

[0043] Real-vehicle verification and the average results of 10 independent experiments show that, in mixed scenarios including urban expressways, arterial roads, and secondary roads, this embodiment reduces travel time by 7.2% (standard deviation 1.3%), risk exposure by 18.4% (standard deviation 2.2%), and user acceptance score by 4.6 (standard deviation 0.2) compared to using Embodiment 1 alone. Compared to using Embodiment 2 alone, replanning latency is reduced from 120 milliseconds to 75 milliseconds (standard deviation 7 milliseconds), and CO2 emissions are reduced by 6.1% (standard deviation 1.1%), verifying the effectiveness of the combined solution. In scenarios where V2X communication is interrupted, this embodiment automatically downgrades to using only ground-based augmented positioning and vehicle sensor data, with a travel time increase of no more than 8.5% (standard deviation 1.8%) and a risk exposure increase of no more than 12.3% (standard deviation 2.4%), still outperforming traditional road segment-level solutions.

[0044] It should be noted that Embodiment 1, Embodiment 2, and Embodiment 3 belong to the same inventive concept, namely, they are all based on the high-precision lane-level positioning of the ground augmentation system and user driving habit parameters to achieve the fusion of real-time road conditions and user driving habits, thereby automatically generating personalized navigation paths. Embodiment 1 focuses on the fusion of V2X lane-level pass probability and personalized dynamic risk potential field and lane-level potential field search, Embodiment 2 focuses on digital twin shadow simulation and sub-Browser CVaR optimization, and Embodiment 3 is a combination of the two. The three can be implemented independently or in combination, and all belong to the overall technical concept of the present invention of real-time road conditions plus user driving habits plus automatic navigation path planning.

[0045] Comparative Example 1 The method employs a traditional segment-level real-time traffic condition scheme with fixed preference weights. Real-time traffic conditions are obtained through a traffic cloud, and Dijkstra's algorithm is used for path planning. The edge cost is the weighted sum of travel time and fixed preference weights, which are manually set by the user. All other conditions are the same as in Example 1. After 10 independent experimental tests, the number of lane changes was 1.47 times that of Example 1, with a standard deviation of 0.08; the number of exposures at complex intersections was 1.39 times, with a standard deviation of 0.07; the number of unprotected left turns was 1.69 times, with a standard deviation of 0.09; the user acceptance score was 3.1, with a standard deviation of 0.3; and the risk exposure was 1.58 times that of Example 1, with a standard deviation of 0.11. Since lane-level pass probability and personalized potential fields were not used, the path had a low match with the user's actual driving habits. Therefore, this comparison does not involve CVaR optimization, and the CVaR improvement index is not applicable.

[0046] Comparative Example 2 Digital twins and expected optimal path planning were used, but without using the split bar and CVaR. All other conditions were the same as in Example 2. After 10 independent experimental tests, in a scenario with a prediction error of 20%, the CVaR improvement was only 8.3% (standard deviation 1.5%), risk exposure was 1.52 times that of Example 2 (standard deviation 0.10), travel time fluctuated significantly (standard deviation 2.1 times that of Example 2), lane change frequency was 1.28 times that of Example 2 (standard deviation 0.09), exposure frequency at complex intersections was 1.22 times that of Example 2 (standard deviation 0.08), and user acceptance was 3.4 (standard deviation 0.3). This indicates that the lack of CVaR resulted in insufficient tail risk control.

[0047] Comparative Example 3 Using the same proportions and processes as in Example 1, but with the potential field guidance coefficient set to 0.8 (exceeding the 0.1 to 0.5 range specified in this invention), after 10 independent experimental tests, the path search time increased by 42% with a standard deviation of 4.5%. Due to the unacceptable heuristic function, the generated path was not optimal. The travel time increased by 12.7% with a standard deviation of 2.1% compared to Example 1, the number of lane changes increased by 18.3% with a standard deviation of 2.6%, the number of complex intersection exposures increased by 14.5% with a standard deviation of 2.3%, the CVaR improved by 12.1% with a standard deviation of 1.8%, the user acceptance score was 3.6 with a standard deviation of 0.3, and the replanning delay was 121 milliseconds with a standard deviation of 11 milliseconds. This indicates that taking the potential field guidance coefficient outside the range of this invention leads to a decrease in search efficiency and a reduction in path quality.

[0048] The comparison results between each embodiment and the comparative example are as follows.

[0049] The reduction in lane change frequency was 31.7% in Example 1, -47% in Comparative Example 1, -28% in Comparative Example 2, and -18.3% in Comparative Example 3.

[0050] Exposure reduction index at complex intersections: Example 1: 28.3%, Comparative Example 1: -39%, Comparative Example 2: -22%, Comparative Example 3: -14.5%.

[0051] Unprotected left turn reduction index: Example 1: 41.2%, Comparative Example 1: -69%.

[0052] Travel time change indicators: Example 1 shows an increase of 3.1%, Example 2 shows an increase of 5.3%, Example 3 shows a decrease of 7.2%, and Comparative Example 3 shows an increase of 12.7%.

[0053] Risk exposure reduction indicators: 37.6% for Example 1, 42.1% for Example 2, 18.4% for Example 3, -58% for Comparative Example 1, and -52% for Comparative Example 2.

[0054] CVaR improvement metrics: 35.2% for Example 2, 8.3% for Comparative Example 2, and 12.1% for Comparative Example 3.

[0055] User acceptance metrics: Example 1: 4.5, Example 2: 4.3, Example 3: 4.6, Comparative Example 1: 3.1, Comparative Example 2: 3.4, Comparative Example 3: 3.6.

[0056] Replanning delay metrics: 85 milliseconds for Example 1, 120 milliseconds for Example 2, 75 milliseconds for Example 3, and 121 milliseconds for Comparative Example 3.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic path planning method that integrates real-time traffic conditions and user driving habits, characterized in that: include S1. High-precision lane-level positioning information of the vehicle is acquired through a ground-based augmentation system. This high-precision lane-level positioning information includes lane number, heading, speed, and positioning confidence level. V2X data is acquired through V2X communication. Real-time traffic data is also acquired through at least one of traffic cloud, floating car, and roadside equipment. The V2X data and the real-time traffic data include signal phase, queue length, and lane-level clearance probability. and the probability of turning at intersections, where ; S2 extracts user driving habit parameters from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data, and constructs a personalized dynamic risk potential field based on the user driving habit parameters. ,in , For lanes, For turning, For time, the personalized dynamic risk potential field is a weighted sum of the road potential field, obstacle potential field, congestion potential field, risk potential field, comfort potential field and V2X passage probability potential field. Each weight is adjusted online according to the user's driving habit parameters and the sum of the weights is 1. Injecting the lane-level passage probability into the personalized dynamic risk potential field in the form of a negative logarithm yields: ;in, The lower, The higher the value, the more likely the path search will avoid the corresponding lane or turn; S3, construct lane-level road network and turning map, and define the search state as: The dynamic costs of lane and steering are updated based on the injected personalized dynamic risk potential field, and the edge costs satisfy: ;in, , , , , For travel time, For energy consumption, For risk, For personalized dynamic risk potential fields, , , , These are the corresponding reference values. , , , As cost weight, and =1; The potential-field-guided A algorithm is used in the lane-level road network and turning map to perform path search with the updated dynamic cost, and outputs lane-level driving paths, intersection turning sequences, and lane change timing suggestions; the evaluation function of the potential-field-guided A algorithm is: ,in From the starting point to the current state The price has already been paid. For heuristic functions, , The distance from the current state to the target point is the road network distance. The heuristic function, which is the historical maximum speed of the current road segment, satisfies the acceptability requirement; the lane change timing suggestion includes the suggested lane, suggested lane change time, or suggested lane change interval. S4. Real-time monitoring of dynamic replanning trigger conditions. When the dynamic replanning trigger conditions are met, the existing search tree is reused for dynamic replanning, and the updated lane-level driving path, intersection turning sequence, and lane change timing suggestions are output. The dynamic replanning trigger conditions include at least one of the following: travel time change exceeding 10%, lane-level pass probability decrease exceeding 20%, vehicle deviation from the recommended path exceeding a preset distance, accident, construction, and signal phase change.

2. The dynamic path planning method integrating real-time traffic conditions and user driving habits according to claim 1, characterized in that: The ground-based augmentation system includes a ground-based augmentation reference station, an RTK / PPP-RTK differential service broadcasting terminal, and a vehicle-mounted high-precision positioning module. The high-precision lane-level positioning information has a positioning accuracy of centimeters to decimeters. The V2X data is obtained by communicating with the roadside unit (RSU) via the on-board unit (OBU) to acquire SPAT and MAP messages. The lane-level pass probability... Indicates time The probability of passing through an intersection in a designated lane with a designated turn; when the positioning confidence is lower than a preset confidence threshold, inertial measurement unit data and vehicle wheel speed data are used to compensate for the lane-level positioning information, and the preset confidence threshold is 0.8 to 0.

95.

3. The dynamic path planning method integrating real-time traffic conditions and user driving habits according to claim 1, characterized in that: The user driving habit parameters mentioned in S2 include start reaction time, following distance, lane change aversion, unprotected left turn aversion, curvature sensitivity, congestion tolerance, caution at complex intersections, energy consumption sensitivity, lateness risk tolerance, and nighttime risk preference. The user driving habit parameters are updated online through inverse reinforcement learning. This inverse reinforcement learning takes state-action pairs from historical trajectories as input and driving habit parameters as output, solving for the likelihood of the expert trajectory and the learned policy trajectory. The inverse reinforcement learning uses a neural network as the policy function approximator. The input layer dimension of the neural network is the number of state features, the output layer dimension is the number of driving habit parameters, the hidden layer is a two-layer fully connected layer, the activation function is the ReLU function, the loss function is the negative log-likelihood of the expert trajectory and the learned policy trajectory, the optimizer is the Adam optimizer, the learning rate is 0.0013, the batch size is 52, and the number of training epochs is 118. The online adjustments include: increasing the lane change penalty coefficient in the comfort potential field when lane change aversion increases; increasing the risk potential field corresponding to the left turn when unprotected left turn aversion increases; increasing the weight of the congestion potential field when congestion tolerance decreases; and increasing the energy consumption weight in the edge cost when energy consumption sensitivity increases.

4. The dynamic path planning method integrating real-time traffic conditions and user driving habits according to claim 1, characterized in that: The road potential field described in S2 is obtained by weighting the normalized values ​​of lane curvature, slope, number of lanes, and intersection complexity. The obstacle potential field is obtained by weighting the normalized values ​​of accidents, construction, large vehicles, and vulnerable road users. The congestion potential field is obtained by weighting the normalized values ​​of queue length, signal waiting time, and number of stops. The risk potential field is obtained by weighting the normalized values ​​of collision risk, violation risk, and fatigue risk. The comfort potential field is obtained by weighting the normalized values ​​of lane change frequency, acceleration / deceleration frequency, and road surface bumpiness. Each normalized value is the ratio of the corresponding physical quantity to its preset maximum value, and the values ​​range from 0 to 1. Each weighting coefficient is determined by offline calibration or online least squares fitting. The offline calibration involves iterating through the combinations of weighting coefficient values ​​in the simulation environment and selecting the optimal weighting coefficient with the goal of minimizing the overall cost of path planning. The online least squares fitting uses real-time collected driving operation data as observations and the normalized values ​​of each sub-potential field as independent variables, and solves for the weighting coefficients using the recursive least squares method.

5. The dynamic path planning method integrating real-time traffic conditions and user driving habits according to claim 1, characterized in that: The construction of the personalized dynamic risk potential field, the injection of the lane-level passage probability, and the path search of the A* algorithm guided by the potential field are deployed on mobile edge computing nodes. The mobile edge computing nodes are connected to the ground augmentation system and the V2X communication module through low-latency links, with an end-to-end latency of less than 10 milliseconds. The mobile edge computing nodes output lane-level driving paths, intersection turning sequences, and lane change timing suggestions to the in-vehicle navigation terminal. The mobile edge computing nodes work in collaboration with the cloud server. The mobile edge computing nodes perform the real-time potential field construction and path search, while the cloud server performs digital twin model training and global road network optimization. Data synchronization between the mobile edge computing nodes and the cloud server adopts an incremental synchronization method, synchronizing only the changed road network data and user driving habit parameters.

6. The dynamic path planning method integrating real-time traffic conditions and user driving habits according to claim 1, characterized in that: The preset distance in S4 is 50 meters to 200 meters. The dynamic replanning trigger condition also includes the user driving habit drift detection value exceeding a threshold. The user driving habit drift detection value is obtained by comparing the mean deviation and variance deviation of the user driving habit parameters in the current time window and the historical time window. When the mean deviation or the variance deviation exceeds the corresponding threshold, dynamic replanning is triggered. The signal phase change includes the deviation between the current signal phase and the V2X predicted signal phase exceeding a preset phase deviation threshold. The preset phase deviation threshold is... Instant .

7. A dynamic path planning method that integrates real-time traffic conditions and user driving habits, characterized in that, include: T1 acquires high-precision lane-level positioning information of the vehicle through a ground augmentation system. The high-precision lane-level positioning information includes lane number, heading, speed, and positioning confidence. User driving habit parameters are extracted from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data. A user driving digital twin model is constructed based on the user driving habit parameters. The user driving digital twin model includes reaction time, following distance, lane change preference, acceleration and braking style, energy consumption characteristics, fatigue accumulation, risk sensitivity, congestion tolerance, toll sensitivity, and comfort preference. The user driving digital twin model is trained through behavioral cloning or inverse reinforcement learning, and the training data is the state-action sequence in the historical trajectory. T2. Construct a set of uncertain distributions for real-time traffic conditions. This set is constructed using at least one of the following: a scene set, a Wasserstein sphere, an interval distribution, and a Gaussian mixture distribution. When using a Wasserstein sphere, the sphere radius is determined based on the confidence level of the real-time traffic prediction error. For candidate paths... In scenarios with uncertain road conditions The comprehensive cost is calculated by weighting the normalized values ​​of time, energy consumption, comfort, risk, fatigue, and charges, with the sum of each weight being 1. T3 uses a split-bar and conditional value at risk (CVaR) optimization model: ;in, For a distribution in an uncertain distribution set, In distribution The expectations below For distribution Lower confidence level Conditional Value at Risk: ; in , For user risk budget parameters, and Based on the user driving habit parameters, the CVaR optimization model is solved by linear programming or stochastic gradient descent. T4. Shadow simulation is performed on multiple candidate paths in the user driving digital twin model. The shadow simulation refers to the parallel simulation of the vehicle driving along the candidate path in the user driving digital twin model. GPU parallel computing is used, the simulation time step is 0.1 seconds, and the number of scene samples is 1000. The personalized travel time, energy consumption, comfort, risk, fatigue and toll pre-simulation results of each candidate path are obtained. Based on the pre-simulation results, the NSGA-II algorithm is used to generate a Pareto optimal path set. When the user does not interact, the multi-objective weights are determined according to the user driving habit parameters. The path with the minimum weighted comprehensive cost is selected from the Pareto optimal path set. T5 feeds the actual driving results back to the user driving digital twin model and the user driving habit parameters for closed-loop updates. When real-time road conditions change or the actual driving deviates from the recommended path, dynamic replanning is triggered, and existing shadow simulation results are reused. Among them, the candidate path The lane-level driving path generated by any one of the methods described in claims 1 to 6 constitutes the driving path, or the lane-level driving path generated by any one of the methods described in claims 1 to 6 constitutes the driving path together with historically frequently used paths.

8. A dynamic route planning system that integrates real-time traffic conditions and user driving habits, characterized in that: include: The high-precision positioning and V2X data acquisition module is configured to acquire high-precision lane-level positioning information of vehicles through a ground augmentation system, acquire V2X data through V2X communication, and acquire real-time traffic data through at least one of traffic cloud, floating cars, and roadside equipment. The high-precision lane-level positioning information includes lane number, heading, speed, and positioning confidence level. The V2X data and the real-time traffic data include signal phase, queue length, and lane-level passage probability. and the probability of turning at intersections, where ; The driving habit and potential field construction module is configured to extract user driving habit parameters from vehicle bus, inertial measurement unit, satellite navigation data, historical trajectory data, and driving operation data, and construct a personalized dynamic risk potential field based on the user driving habit parameters. and the lane-level pass probability is... Inject the personalized dynamic risk potential field; The lane-level potential field search and replanning module is configured to construct lane-level road networks and turning maps, defining the search state as follows: Based on the injected personalized dynamic risk potential field, the dynamic costs of lane and steering are updated, and the edge costs satisfy... ,in , , , The potential field-guided A* algorithm is used for path search, outputting lane-level driving paths, intersection turning sequences and lane change timing suggestions. When the dynamic replanning trigger condition is met, the existing search tree is reused for dynamic replanning. The digital twin and robust optimization module is configured to construct a user driving digital twin model, construct a set of real-time road condition uncertainty distributions, perform shadow simulations on candidate paths, solve the paths using a split-loop and CVaR optimization model, and update the user driving digital twin model and the user driving habit parameters in a closed loop based on the actual driving results. The high-precision positioning and V2X data acquisition module outputs the high-precision lane-level positioning information, the V2X data, and the real-time traffic data to the driving habit and potential field construction module. The driving habit and potential field construction module outputs the personalized dynamic risk potential field to the lane-level potential field search and replanning module. The lane-level potential field search and replanning module outputs the lane-level driving path, the intersection turning sequence, and the lane change timing suggestion. The digital twin and robust optimization module receives the lane-level driving path and performs shadow simulation and robust optimization to output a personalized lane-level navigation path. The system is used to execute the method of any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 7.

10. A computer device, comprising a processor and a computer program stored on a computer-readable storage medium and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.

Images