Automatic driving network connection cloud control site test method fusing vehicle road data

Abstract

The invention relates to the technical field of automatic driving test, and discloses an automatic driving network connection cloud control site test method fusing vehicle-road data, comprising the following steps: S1, constructing a dangerous scene set based on roadside sensing data; s2, selecting a test scene; constructing a continuous test task sequence, and solving an optimal solution of the scene risk degree and the coverage degree; executing target object process allocation and global trajectory planning based on the task sequence, and generating an automatic driving continuous test scheme; s3, establishing real-time communication connection between the controlled target object terminal and the cloud; receiving a planning track issued by the cloud at a controlled target object terminal, and constructing a system state model to realize tracking control; a global multi-target task state is monitored at a cloud end, a scene operation cost index is constructed, a multi-target trajectory planning scheme is dynamically adjusted, and a scene task execution state is researched and judged; and S4, constructing a digital twinborn model, and mapping and visualizing test data in real time. According to the invention, an optimization scheme is provided for the high-order intelligent driving function closed site test.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving testing technology, and in particular to a method for testing an autonomous driving networked cloud-controlled site by integrating vehicle-road data. Background Art

[0002] Currently, the development and validation of autonomous driving technology relies heavily on field testing. Traditional field testing approaches primarily focus on verifying single target conditions, such as discrete scenarios like accelerating into a neighboring vehicle, following a curve, and avoiding fixed obstacles. This technical implementation typically involves digitally constructing the static environment of the test site (e.g., road structure, traffic signs, and fixed obstacles) using dedicated scenario modeling tools. Dynamic scenario setup tools are then used to generate specific traffic participants (e.g., vehicles and pedestrians) and their pre-defined motion trajectories. Finally, these scenarios are replicated in the physical field using high-precision object control systems (e.g., robotic vehicles and moving dummies). Response data from the automated driving system (ADS) is collected and evaluated based on quantitative metrics such as detection accuracy, recall, and longitudinal and lateral control distance errors. This approach has played a crucial role in validating basic perception and control capabilities. However, with the rapid advancement of automotive intelligence, particularly the application of advanced artificial intelligence technologies such as end-to-end large models, and the continuous improvement of big data and algorithm capabilities, intelligent driving is undergoing rapid iteration and technological breakthroughs. Driven by these new data technologies, intensified competition is also pushing intelligent driving to accelerate market penetration, driving system functions from basic assisted driving (ADAS) to high-level intelligent driving with higher autonomous decision-making capabilities.

[0003] The core characteristic of advanced intelligent driving functions (such as urban navigation assistance, unprotected traffic at complex intersections, and long-distance point-to-point autonomous driving) lies in their need to handle complex scenarios with strong coupling and high interaction between the "driver, vehicle, road, and driving tasks" over long periods of time in a continuous, dynamically changing real-world traffic environment. These scenarios are no longer isolated, single-objective challenges with preset parameters, but rather systemic issues involving the intertwined influence of the intentions of multiple participants (motor vehicles, non-motor vehicles, and pedestrians), complex road topology (multiple intersections, irregular lanes), dynamic traffic regulations (signals, variable lanes, temporary traffic controls), and diverse driving tasks (navigation, merging, avoidance, parking, etc.). Current evaluation schemes based on typical scenarios in traditional field locations present several problems in this context. First, existing evaluation schemes are primarily designed for single-objective conditions and short-term scenarios. They are unable to adapt to the complex, long-term coupling of the "driver, vehicle, road, and driving tasks" in continuous traffic environments, and cannot replicate the complexity of element relationships and state evolution in the real world. Second, its evaluation indicators mainly focus on perception accuracy and basic control errors at specific moments. They are not suitable for the verification needs of mid- and high-level intelligent driving functions in complex scenarios, and it is difficult to effectively evaluate their functional characteristics.

[0004] To summarize, existing field evaluation solutions need to expand scenarios and update technologies for advanced intelligent driving functions, and require a method for testing connected, cloud-controlled field systems for autonomous driving that integrates vehicle-road data. Summary of the Invention

[0005] The present invention aims to provide a method for testing connected cloud-controlled sites for autonomous driving that integrates vehicle-road data. The method is suitable for closed-site testing scenarios of high-level intelligent driving functions, can achieve high testing efficiency and testing stability, and help accelerate the testing and verification process of high-level autonomous driving technology, and improve the reliability of performance testing of high-level autonomous driving technology.

[0006] The basic solution provided by the present invention is: a method for testing an autonomous driving networked cloud-controlled field that integrates vehicle-road data, comprising the following steps:

[0007] S1, constructs a set of dangerous scenarios based on roadside perception data;

[0008] S2, based on the closed site test objectives and resource constraints, selects test scenarios from the set of dangerous scenarios generated in S1;

[0009] Construct a continuous test task sequence containing multiple test scenarios and use optimization algorithms to find the optimal solution for scenario risk and coverage;

[0010] Execute target object process allocation and global trajectory planning based on the task sequence, and generate an autonomous driving continuous test plan;

[0011] S3, establish a real-time communication connection between the controlled target terminal and the cloud;

[0012] At the controlled target terminal, the planned trajectory sent from the cloud is received, and a system state model is constructed to implement tracking control.

[0013] In the cloud: monitor the global multi-target mission status, build a scenario operation cost index, dynamically adjust the multi-target trajectory planning scheme, and analyze the scenario mission execution status;

[0014] S4, builds a digital twin model that matches the actual test site, mapping and visualizing test data in real time.

[0015] Furthermore, in S1, the following sub-steps are included:

[0016] S1.1, collect structured data of traffic participants as roadside perception data through roadside perception equipment, including target type and running space-time trajectory information;

[0017] S1.2, construct a spatiotemporal joint hazard probability density function based on roadside sensing data to screen potential hazard candidate targets;

[0018] S1.3, extract risk features from the screened potential dangerous candidate targets, generate dangerous scene slices, and then obtain scene slice data;

[0019] S1.4, extract core subordinate target objects through principal component analysis to form a standardized set of dangerous scenarios;

[0020] S1.5, convert the scene slice data into the OpenSCENARIO standard scene description format.

[0021] Furthermore, the spatiotemporal joint hazard probability density function is set to f(t, s, v); wherein t is a time parameter, s is a space parameter, and v is a speed parameter;

[0022] Set the risk threshold η; define the set of potential dangerous candidate targets as follows:

[0023] C={c|P c (f(t,s,v)>η)≥0.95};

[0024] Among them, P c is the spatiotemporal joint danger probability of target c; C is the set of potential dangerous candidate targets.

[0025] Furthermore, when extracting risk features and generating dangerous scene slices, the following sub-steps are included:

[0026] From the perspective of the main vehicle, calculate the subordinate participants T around the target c i ,Risk degree of road boundary map;

[0027] Define the risk factor R i :

[0028]

[0029] Among them, TTC i is the collision time risk between the master vehicle and the slave participant target, R type (T i ) is the subordinate participant type risk, R road (Map) is the conflict risk between the road map topology and driving rules; α1, α2, and α3 are the impact weights of the corresponding risk types;

[0030] Through the relative position relationship and risk coefficient R i Construct a dynamic risk potential field for the scenario and quantify the risk distribution U of the main vehicle scenario at time t risk (x,y), defined as follows:

[0031]

[0032] Among them, (x, y) is the spatial position of the main vehicle, (x i ,y i ) is the spatial position of the subordinate target, and k is the risk coefficient, which is used to dynamically adjust the threat impact range;

[0033] Set risk thresholds based on statistical data, complete dangerous scene data slicing, and obtain scene slicing data.

[0034] Furthermore, in S2, when constructing a continuous test task sequence containing multiple test scenarios, a scene transition model is first established based on the Markov decision process, and the scene space S and the scene transition probability matrix P(S t+1 |S t ); then generate the scene sequence S through dynamic iterative optimization strategy π:S→A seq , and S seq Satisfy the resource consumption constraint C max Under this condition, the cumulative reward function R(S i )maximize.

[0035] Furthermore, in S2, the objective function of global trajectory planning is:

[0036]

[0037] Among them, ω is the weight coefficient, s t is the tth scene, s t+1 is the t+1th scene;

[0038] Risk(s t →s t+1 ,T) are s t Switch to s t+1 , execution time, energy consumption and risk value of trajectory T.

[0039] Furthermore, in S3, at the controlled target terminal, the system state model constructed includes:

[0040] Control panel state model:

[0041] Where x∈R n is the state vector, u∈R m is the control input, A and B are system matrices, and ω(t) is the process noise;

[0042] When performing tracking control, the LQR control algorithm is used to perform trajectory control to solve the optimization problem, and the system cost function is constructed as follows:

[0043]

[0044] Among them, Q is the state error weight matrix, R is the control input weight matrix;

[0045] Solve the Riccati equation to obtain the optimized feedback control matrix K and obtain the optimal control:

[0046] u(t)=-Kx(t).

[0047] Furthermore, in S3, in the cloud, the scenario running cost index is:

[0048]

[0049] Among them, R collision is the collision probability, E efficiency is the delay efficiency, W offset is the lateral trajectory deviation tolerance, D offset is the longitudinal trajectory deviation tolerance; β1 is R collision The influence weight of E efficiency The influence weight of offset and D offset The influence weight of .

[0050] Furthermore, in S4, it also includes: the cloud determines the scenario execution status by integrating the cluster target scenario operation cost;

[0051] When the running cost of the comprehensive cluster target scenario determines that the scenario cannot meet the regulatory standards or there is a huge safety risk, the safety protection measures are automatically triggered.

[0052] Furthermore, the scenario execution state is defined as follows:

[0053]

[0054] Among them, θ safe The preset safety status threshold is: when state(t) = 0, the safety protection measure is triggered - the cloud automatically issues an emergency intervention instruction and forces the current scenario to be restarted.

[0055] The working principle and advantages of the present invention are:

[0056] The present invention provides a cloud-controlled field testing method for autonomous driving that integrates vehicle-road data. This method is suitable for closed-field testing of advanced intelligent driving functions and can achieve high testing efficiency and stability. This method helps accelerate the testing and verification of advanced autonomous driving technologies and improves the reliability of performance testing of advanced autonomous driving technologies. Key points:

[0057] First, this solution reuses basic roadside data for complex scenario mining, identifying potentially dangerous scenarios within traffic flows and expanding the complex scenario library for autonomous driving. It then screens high-conflict candidate targets based on a spatiotemporal joint probability model and, combined with a dynamic risk field, quantifies the interaction risk between the primary vehicle and subordinate targets, fully exploring scenario characteristics. This solution transcends the limitations of traditional manual construction of single operating conditions and is able to extract complex risk characteristics associated with the strong coupling of "driver, vehicle, and road" in continuous traffic environments, forming a standardized library of dangerous scenarios. Furthermore, compared to existing technologies that rely on static modeling, automated mining based on roadside data helps improve the efficiency of scenario library construction, reduce the required mileage for real-world road testing, and provide a more realistic testing foundation for advanced intelligent driving functions.

[0058] Second, this solution specifically develops end-cloud dual closed-loop robust control technology. During the test execution phase, a high-precision trajectory tracking algorithm is used on the end side to ensure the stable operation of a single controlled tablet under various test conditions and non-ideal communication conditions (such as communication delays or interference conditions). This allows the device to safely and accurately execute cloud control commands, support the field restoration of dynamic and complex test scenarios, and provide a consistent test environment for the performance evaluation of high-level autonomous driving systems. The cloud monitors the global status in real time, comprehensively evaluates key indicators such as collision probability, delay efficiency, and trajectory deviation, and dynamically adjusts multi-target motion planning. At the same time, based on the Markov decision process, the scenario sequence arrangement is optimized to maximize risk coverage and test continuity under the constraints of site resources. This can improve the utilization rate of the driving tablet, reduce the waiting interval and execution time of test tasks, and upgrade the traditional discrete single-scenario test to a continuous dynamic scenario chain test, thereby significantly improving the testing efficiency of complex dynamic scenarios of autonomous driving and accelerating the test and verification process of high-level autonomous driving technology.

[0059] Third, this solution innovatively integrates digital twin technology to build a proactive safety protection system. By establishing a real-time virtual map of the test site and continuously calculating the comprehensive cost index of multi-target cluster operations, it can accurately identify abnormal conditions such as trajectory deviation and collision risk based on the cloud. Safety measures are automatically triggered when abnormal conditions occur, transforming traditional passive protection into predictive intervention. This significantly reduces the risk of testing accidents and provides a reliable verification foundation for the rapid iteration of advanced autonomous driving technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A schematic diagram of a method flow of an embodiment of a method for testing an autonomous driving networked cloud-controlled field that integrates vehicle-road data according to the present invention;

[0061] Figure 2 for Figure 1 A partial schematic diagram of the steps for constructing an autonomous driving risk test scenario corresponding to S1;

[0062] Figure 3 for Figure 1 A partial schematic diagram of the steps for constructing the closed-field autonomous driving continuous test task corresponding to S2;

[0063] Figure 4 for Figure 1 A partial diagram of the multi-target traffic participant cluster control steps corresponding to S3’s end-cloud collaboration;

[0064] Figure 5 for Figure 1 A partial schematic diagram of the steps of the digital twin real-time supervision and safety assurance mechanism for the test process corresponding to S4;

[0065] Figure 6 This is a schematic diagram of the end-cloud dual closed-loop multi-objective cluster control model structure of an embodiment of the present invention's autonomous driving networked cloud-controlled field testing method that integrates vehicle-road data. DETAILED DESCRIPTION

[0066] The following is a further detailed description through specific implementation methods:

[0067] The embodiment is basically as shown in the attached Figure 1 A method for testing an autonomous driving networked cloud-controlled field that integrates vehicle-road data is shown, comprising the following steps:

[0068] S1, such as Figure 2 As shown in the figure, a set of dangerous scenarios is constructed based on roadside perception data.

[0069] It includes the following sub-steps:

[0070] S1.1. Structured data of traffic participants is collected as roadside sensing data through roadside sensing equipment, including target type and time-space trajectory information. The target type specifically refers to the type of traffic participant, such as vehicles, pedestrians, and non-motorized vehicles; the time-space trajectory information specifically refers to the target's time-space trajectory information, such as its location coordinates, heading angle, speed, size, and stop status.

[0071] S1.2, construct a spatiotemporal joint hazard probability density function based on roadside perception data to screen potential hazard candidate targets.

[0072] The spatiotemporal joint hazard probability density function is set to f(t, s, v); wherein t is a time parameter, s is a space parameter, and v is a speed parameter.

[0073] Set the risk threshold η; define the set of potential dangerous candidate targets as follows:

[0074] C={c|P c (f(t,s,v)>η)≥0.95};

[0075] Among them, Pc is the spatiotemporal joint danger probability of target c; C is the set of potential dangerous candidate targets.

[0076] S1.3, extract risk features from the screened potential dangerous candidate targets, generate dangerous scene slices, and then obtain scene slice data.

[0077] It includes the following sub-steps:

[0078] From the perspective of the main vehicle, calculate the subordinate participants T around the target c i ,Risk degree of road boundary map;

[0079] Define the risk factor R i :

[0080]

[0081] Among them, TTC i is the collision time risk between the master vehicle and the slave participant target, R type (T i ) is the subordinate participant type risk, R road (Map) is the conflict risk between the road map topology and driving rules; α1, α2, and α3 are the impact weights of the corresponding risk types.

[0082] Through the relative position relationship and risk coefficient R i Construct a dynamic risk potential field for the scenario and quantify the risk distribution U of the main vehicle scenario at time t risk (x,y), defined as follows:

[0083]

[0084] Among them, (x, y) is the spatial position of the main vehicle, (x i ,y i ) is the spatial position of the subordinate target, and k is the risk coefficient, which is used to dynamically adjust the threat impact range;

[0085] Set risk thresholds based on statistical data, complete dangerous scene data slicing, and obtain scene slicing data.

[0086] S1.4, extract the core subordinate target objects through principal component analysis to form a standardized dangerous scenario set S c .

[0087] Standardized hazard scenario set S c The main data elements include time series Time, environment map Map, main vehicle trajectory Path c and core subordinate target trajectories The expression is as follows:

[0088]

[0089] S1.5, convert the scene slice data into the OpenSCENARIO standard scene description format.

[0090] Specifically, after completing the element extraction of the standardized dangerous scene set in S1.4, a local coordinate system is established according to the main vehicle's perspective, and relative coordinate projection is completed from the global coordinates, and the scene slice data is further converted into the OpenSCENARIO standard scene description format.

[0091] S2, such as Figure 3 As shown, according to the closed site test objectives and resource constraints, the test scenarios are selected from the set of dangerous scenarios generated by S1;

[0092] Construct a continuous test task sequence containing multiple test scenarios and use optimization algorithms to find the optimal solution for scenario risk and coverage;

[0093] Target object process allocation and global trajectory planning are performed based on the task sequence, and an autonomous driving continuous test plan is generated.

[0094] Specifically, in this step, when constructing a continuous test task sequence containing multiple test scenarios, a scenario transition model is first established based on the Markov decision process, and the scenario space S = {s1,…,s n} and scene transition probability matrix P(S t+1 |S t ); where each scene s i Contains the basic classification attributes obtained through S1 calculation, including the scenario risk distribution U risk_i , risk type distribution Resource consumption C i ={C area ,C obj}, where C area 、C obj They represent site resource consumption and target object resource consumption respectively.

[0095] Then, through the dynamic iterative optimization strategy π:S→A, the scene sequence S is generated seq , and S seq Satisfy the resource consumption constraint C max Under this condition, the cumulative reward function R(S i )maximize;

[0096] Right now, and

[0097] Where k is the number of scenes, ΔT i =T i ∩T′ coverRepresents the contribution of scene type, T′ cover It indicates that the covered risk types occupy the complement of the whole set T, with weight w1+w2+w3=1, and the weight is dynamically allocated as the resource consumption accumulates.

[0098] Based on scene sequence S seq ={s t}Determine the target roles and starting and ending location information of each scene. The cloud completes the process allocation of the participating objects based on the improved Hungarian algorithm, and takes the optimal global path cost as the goal to obtain the global path planning T.

[0099] The global path cost objective function is defined as follows:

[0100]

[0101] Among them, ω is the weight coefficient, s t is the tth scene, s t+1 is the t+1th scene;

[0102] Risk(s t →s t+1 ,T) are s t Switch to s t+1 , execution time, energy consumption and risk value of trajectory T.

[0103] S3, such as Figure 4 As shown, a real-time communication connection is established between the controlled target terminal and the cloud to facilitate the execution of closed field testing;

[0104] The closed field test process includes three stages: "test scenario preparation - test scenario execution - test scenario termination". In this embodiment, the controlled target terminal and the cloud are combined to build a terminal-cloud dual closed-loop multi-target cluster control model to perform closed field tests, such as Figure 6 The cloud is responsible for low-frequency, global, multi-target scheduling closed-loop control, while the device side implements highly reliable and robust closed-loop control of a single target (the controlled target terminal), ensuring conflict-free and efficient operation of multiple controlled drive panels in continuous and complex scenarios.

[0105] Specifically, the controlled target terminal (including positioning equipment, drive motor, drive flatbed controller, and drive flatbed vehicle) receives the planned trajectory sent from the cloud, builds a system state model, and implements tracking control.

[0106] The constructed system state model includes:

[0107] Control panel state model:

[0108] Where x∈R nis the state vector, u∈R m is the control input, A and B are system matrices, and ω(t) is the process noise;

[0109] When performing tracking control, the LQR control algorithm is used to perform trajectory control to solve the optimization problem, and the system cost function is constructed as follows:

[0110]

[0111] Among them, Q is the state error weight matrix, R is the control input weight matrix;

[0112] Solve the Riccati equation to obtain the optimized feedback control matrix K and obtain the optimal control:

[0113] u(t)=-Kx(t).

[0114] In the cloud - monitor the global multi-target task status, build a scenario operation cost index, dynamically adjust the multi-target trajectory planning scheme, and analyze the scenario task execution status.

[0115] Specifically, the cloud receives real-time pose data (including position, velocity, and acceleration) reported by each target, calculates the planned trajectory deviation (lateral / longitudinal offset) and environmental map data, predicts the target trajectory through Kalman filtering, and calculates the following indicators based on the actual trajectory and predicted trajectory at the current time t:

[0116] Collision probability (R collision ): Calculate the path overlap probability and collision risk based on the predicted trajectory;

[0117] Delay efficiency (E efficiency ): Calculate the deviation between the average speed of all targets in the scene and the planned speed to predict the delay rate;

[0118] Lateral track deviation tolerance W offset , longitudinal track deviation tolerance D offset : Set the horizontal offset threshold and the vertical offset threshold, and calculate the target ratio that exceeds the target.

[0119] Then, based on the collision probability, delay efficiency, and trajectory deviation tolerance, the scenario operation cost index at the current time t is comprehensively calculated, and the threshold is adjusted according to the actual situation to trigger dynamic trajectory optimization.

[0120] Specifically, the scenario operation cost index is:

[0121]

[0122] Among them, R collision is the collision probability, E efficiency is the delay efficiency, W offsetis the lateral trajectory deviation tolerance, D offset is the longitudinal trajectory deviation tolerance; β1 is R collision The influence weight of E efficiency The influence weight of offset and D offset The influence weight of .

[0123] The cloud dynamically adjusts the global multi-target trajectory planning scheme based on the scenario operation cost index and analyzes the current task execution status.

[0124] S4, such as Figure 5 As shown, a digital twin model matching the actual test site is built to map and visualize test data in real time.

[0125] Furthermore, the cloud determines the scenario execution status by integrating the cluster target scenario operation cost;

[0126] When the running cost of the comprehensive cluster target scenario determines that the scenario cannot meet the regulatory standards or there is a huge safety risk, the safety protection measures are automatically triggered.

[0127] The scenario execution state is defined as follows:

[0128]

[0129] Among them, θ safe The preset safety status threshold is: when state(t) = 0, the safety protection measure is triggered - the cloud automatically issues an emergency intervention instruction and forces the current scenario to be restarted.

[0130] This embodiment provides a method for testing connected cloud-controlled autonomous driving sites that integrates vehicle-road data. The method is suitable for closed-field testing scenarios of advanced intelligent driving functions, can achieve high testing efficiency and stability, and helps accelerate the testing and verification process of advanced autonomous driving technologies, thereby improving the reliability of performance testing of advanced autonomous driving technologies.

[0131] The above is only an embodiment of the present invention. Common knowledge such as the specific structure and characteristics of the scheme is not described in detail here. Ordinary technicians in the relevant field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the guidance of this application. Some typical well-known structures or well-known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for testing an autonomous driving networked cloud-controlled field that integrates vehicle-road data, characterized in that: The following steps are involved: S1, constructs a set of dangerous scenarios based on roadside perception data; S2, based on the closed site test objectives and resource constraints, selects test scenarios from the set of dangerous scenarios generated in S1; Construct a continuous test task sequence containing multiple test scenarios and use optimization algorithms to find the optimal solution for scenario risk and coverage; Execute target object process allocation and global trajectory planning based on the task sequence, and generate an autonomous driving continuous test plan; S3, establish a real-time communication connection between the controlled target terminal and the cloud; At the controlled target terminal, the planned trajectory sent from the cloud is received, and a system state model is constructed to implement tracking control. In the cloud: monitor the global multi-target mission status, build a scenario operation cost index, dynamically adjust the multi-target trajectory planning scheme, and analyze the scenario mission execution status; S4, builds a digital twin model that matches the actual test site, mapping and visualizing test data in real time.

2. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 1, characterized in that: In S1, the following sub-steps are included: S1.1, collect structured data of traffic participants as roadside perception data through roadside perception equipment, including target type and running space-time trajectory information; S1.2, construct a spatiotemporal joint hazard probability density function based on roadside sensing data to screen potential hazard candidate targets; S1.3, extract risk features from the screened potential dangerous candidate targets, generate dangerous scene slices, and then obtain scene slice data; S1.4, extract core subordinate target objects through principal component analysis to form a standardized set of dangerous scenarios; S1.5, convert the scene slice data into the OpenSCENARIO standard scene description format.

3. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 2, characterized in that: The spatiotemporal joint hazard probability density function is set to f(t, s, v); wherein t is a time parameter, s is a space parameter, and v is a speed parameter; Set the risk threshold η; define the set of potential dangerous candidate targets as follows: C={c|P c (f(t,s,v)>η)≥0.95}; Among them, P c is the spatiotemporal joint danger probability of target c; C is the set of potential dangerous candidate targets.

4. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 2, characterized in that: When extracting risk features and generating dangerous scene slices, the following sub-steps are included: From the perspective of the main vehicle, calculate the subordinate participants T around the target c i ,Risk degree of road boundary map; Define the risk factor R i : Among them, TTC i is the collision time risk between the master vehicle and the slave participant target, R type (T i ) is the subordinate participant type risk, R road (Map) is the conflict risk between the road map topology and driving rules; α1, α2, and α3 are the impact weights of the corresponding risk types; Through the relative position relationship and risk coefficient R i Construct a dynamic risk potential field for the scenario and quantify the risk distribution U of the main vehicle scenario at time t risl (x,y), defined as follows: Among them, (x, y) is the spatial position of the main vehicle, (x i ,y i ) is the spatial position of the subordinate target, and k is the risk coefficient, which is used to dynamically adjust the threat impact range; Set risk thresholds based on statistical data, complete dangerous scene data slicing, and obtain scene slicing data.

5. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 1, characterized in that: In S2, when constructing a continuous test task sequence containing multiple test scenarios, a scene transition model is first established based on the Markov decision process, and the scene space S and the scene transition probability matrix P(S t+1 |S t ); then generate the scene sequence S through dynamic iterative optimization strategy π:S→A seq , and S seq Satisfy the resource consumption constraint C max Under this condition, the cumulative reward function R(S i )maximize.

6. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 1, characterized in that: In S2, the objective function of global trajectory planning is: Among them, ω is the weight coefficient, s t is the tth scene, s t+1 is the t+1th scene; s respectively t Switch to s t+1 , execution time, energy consumption and risk value of trajectory T.

7. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 1, characterized in that: In S3, at the controlled target terminal, the system state model constructed includes: Control panel state model: Where x∈R n is the state vector, u∈R m is the control input, A and B are system matrices, and ω(t) is the process noise; When performing tracking control, the LQR control algorithm is used to perform trajectory control to solve the optimization problem, and the system cost function is constructed as follows: Among them, Q is the state error weight matrix, R is the control input weight matrix; Solve the Riccati equation to obtain the optimized feedback control matrix K and obtain the optimal control: u(t)=-Kx(t).

8. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 1, characterized in that: In S3, in the cloud, the cost index for running the scenario is: Among them, R collision is the collision probability, E efficiency is the delay efficiency, W offset is the lateral trajectory deviation tolerance, D offset is the longitudinal trajectory deviation tolerance; β1 is R collision The influence weight of E efficiency The influence weight of offset and D offset The influence weight of .

9. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 1, characterized in that: In S4, it also includes: the cloud determines the scenario execution status by integrating the cluster target scenario running cost; When the running cost of the comprehensive cluster target scenario determines that the scenario cannot meet the regulatory standards or there is a huge safety risk, the safety protection measures are automatically triggered.

10. The method for testing an autonomous driving networked cloud-controlled field using vehicle-road data integration according to claim 9, characterized in that: The scenario execution state is defined as follows: Among them, θ safe The preset safety status threshold is: when state(t) = 0, the safety protection measure is triggered - the cloud automatically issues an emergency intervention instruction and forces the current scenario to be restarted.

Images