Automobile driving test evaluation method fusing scene complexity

Abstract

The invention discloses a scene complexity-fused automobile driving test evaluation method, which comprises the following steps of: constructing a scene database, acquiring a plurality of driving scenes and storing the driving scenes in the scene database; a plurality of driving scenes are selected to be combined into a complex scene, the scene complexity of the complex scene is evaluated, and road requirements are generated according to the complex scene; constructing a digital twinborn model of the to-be-tested automobile, carrying out a virtual driving test in a complex scene based on the digital twinborn model, and obtaining virtual test data; setting a test road based on road requirements, performing actual driving test on the to-be-tested automobile on the test road, and obtaining actual test data; perception performance parameters, decision performance parameters and control performance parameters are respectively calculated according to the virtual test data and the actual test data, test evaluation is carried out in combination with scene complexity, and the performance of an automobile can be evaluated from multiple dimensions by contrasting and analyzing the virtual test data and the actual test data in a complex scene, so that optimization is facilitated.

Description

Technical Field

[0001] This invention relates to the field of automotive driving test technology, and in particular to an automotive driving test evaluation method that integrates scenario complexity. Background Technology

[0002] With the continuous improvement of living standards and intelligent technologies, the number of cars on the road is constantly increasing, and the requirements for driving performance are also getting higher and higher. In particular, the emerging autonomous driving requires mature and complete control technologies to ensure safety. Therefore, in order to verify the driving safety and driving system performance of vehicles, driving tests are necessary. The essence of driving tests is to systematically evaluate the perception, decision-making and control capabilities of vehicles by simulating various scenarios in real driving environments. As the automotive industry upgrades towards intelligence and connectivity, driving tests have evolved from traditional manual operation evaluation to a comprehensive testing system that includes multi-sensor fusion and autonomous driving algorithm verification. In order to adapt to complex road conditions, various scenarios are simulated during driving tests, exploring a variety of elements. The combination of different elements forms diverse test scenarios ranging from simple to complex. Whether it is the handling performance test of traditional fuel vehicles or the verification of autonomous driving functions of intelligent connected vehicles, the rationality and coverage of the scenario directly determine the validity of the test results. The quantification and integration of scenario complexity has become a key direction for improving the comprehensiveness of the test.

[0003] Currently, vehicle driving tests are mainly conducted using a combination of real-vehicle road tests and virtual simulation tests. Real-vehicle road tests obtain real driving data of vehicles by setting up specific scenarios on actual roads, while virtual simulation tests use driving simulators to build virtual testing environments to reduce testing costs and risks. To increase the challenge of testing, existing technologies often attempt to superimpose simple elements on a single scenario, such as combining rainy weather with obstacles in front to create a complex scenario, or adding non-motorized vehicle interference elements to urban road scenarios, thereby achieving vehicle driving tests in complex scenarios.

[0004] However, while such methods can increase the complexity of scenarios to some extent, they still have significant limitations. First, the scenario combinations lack dynamic adaptation and benchmarks. The superposition of simple elements does not establish a correlation between virtual and real scenarios, making it difficult to effectively compare virtual test data with real vehicle test data. This makes it difficult to quantify and analyze the performance deviation of the driving system in real complex scenarios, affecting the objectivity and credibility of the test results. Second, the current evaluation system focuses on a single test result and does not combine scenario complexity to build a hierarchical evaluation dimension. It cannot accurately identify the performance shortcomings of the system in scenarios with different levels of complexity. For example, a decision-making algorithm that performs well in low-complexity scenarios may experience response delays or misjudgments in high-complexity scenarios such as heavy rain, water accumulation, and sudden obstacles. Summary of the Invention

[0005] In view of this, the present invention proposes a vehicle driving test evaluation method that integrates scene complexity, which can conduct virtual and real comparative tests in combined complex scenarios to evaluate the actual driving performance of the vehicle under test.

[0006] The technical solution of this invention is implemented as follows: A vehicle driving test evaluation method that integrates scenario complexity includes the following steps: Step S1: Build a scenario database by acquiring and storing several driving scenarios in the scenario database; Step S2: Select multiple driving scenarios to form a complex scenario, evaluate the scenario complexity of the complex scenario, and generate road requirements based on the complex scenario; Step S3: Construct a digital twin model of the vehicle under test, conduct virtual driving tests in complex scenarios based on the digital twin model, and obtain virtual test data; Step S4: Set up the test road based on road requirements, conduct actual driving tests on the test road, and obtain actual test data; Step S5: Calculate the perception performance parameters, decision performance parameters, and control performance parameters based on the virtual test data and the actual test data, and evaluate them in conjunction with the scenario complexity.

[0007] Preferably, step S1 includes the following steps: Step S11: Extract the original scene from real road survey data, traffic accident data and standards and regulations; Step S12: Deconstruct the original scene into road elements, static obstacle elements, traffic participant elements, and environmental elements; Step S13: Classify and store road elements, static obstacle elements, traffic participant elements, and environmental elements into a structured scene database.

[0008] Preferably, step S2 includes the following specific steps: Step S21: When there is a subjective testing requirement, select the corresponding driving scenarios from the scenario database according to the subjective testing requirement to form a complex scenario; Step S22: When there is no subjective testing requirement, determine representative road elements based on the geographical location of the vehicle driving test; Step S23: Randomly select at least one element from static obstacle elements, traffic participant elements, and environmental elements to combine with road elements to form a complex scene; Step S24: Perform conflict detection on the combined complex scene. If a conflict exists, re-combine the complex scene. Step S25: Evaluate the complexity of the complex scene, extract the quantitative parameters of each element in the complex scene, and generate road requirements based on the quantitative parameters.

[0009] Preferably, step S3 includes the following specific steps: Step S31: Extract the physical parameters, decision rules, and control algorithms of the vehicle under test, and construct a digital twin model; Step S32: Import the complex scene into the simulation environment and place the digital twin model in the simulation environment for virtual driving test; Step S33: Extract virtual test data during the virtual driving test process through the simulation environment.

[0010] Preferably, step S4 includes the following specific steps: Step S41: Set up the driving scenario on the test road according to the road requirements; Step S42: Drive the vehicle to be tested onto the test road and conduct an actual driving test; Step S43: Collect actual test data during the actual driving test using onboard sensors and data recording equipment; Step S44: Perform spatiotemporal alignment between the virtual test data and the actual test data.

[0011] Preferably, both the virtual test data and the actual test data include perception data, decision data, and control data. The perception data includes the types and number of obstacles identified, the decision data includes the path to avoid obstacles and the decision selection of key nodes, and the control data includes acceleration and energy consumption.

[0012] Preferably, the specific steps for calculating the sensing performance parameters in step S5 include: Step S51: Compare the number of obstacles identified in the virtual test data and the actual test data, and calculate the false negative rate based on the difference in the number of obstacles; Step S52: Determine whether the types of obstacles identified at the same time point are consistent, and calculate the recognition accuracy based on the judgment results of whether the types of obstacles are consistent throughout the entire test process.

[0013] Preferably, the specific steps for calculating the decision performance parameters in step S5 include: Step S53: Extract obstacle avoidance paths from virtual test data and actual test data respectively, calculate the obstacle avoidance paths point by point, and obtain the longitudinal deviation and lateral deviation of the paths. Step S54: Assign weights to the longitudinal deviation and lateral deviation of the path respectively, and sum them to obtain the comprehensive path deviation; Step S55: Perform semantic recognition on the complex scene and extract key nodes. Obtain the decision-making choices of the virtual test data and the actual test data at the key nodes, and make a consistency judgment. Then, binarize the consistency judgment result.

[0014] Preferably, the specific steps for calculating the control performance parameters in step S5 include: Step S56: Extract acceleration values ​​from the virtual test data and the actual test data respectively, and calculate the acceleration value difference rate; Step S57: Extract the consumed energy from the virtual test data and the actual test data respectively, and calculate the energy difference rate.

[0015] Preferably, the specific steps of step S5, which combines scene complexity for testing and evaluation, include: Step S58: Construct an evaluation matrix. Standardize the perception performance parameters, decision performance parameters, and control performance parameters, and then fill them into the evaluation matrix along with the scenario complexity. Step S59: Calculate the eigenvalues ​​of the covariance matrix of the evaluation matrix, and conduct a test evaluation based on the difference between the largest eigenvalue and other eigenvalues.

[0016] Compared with the prior art, the beneficial effects of the present invention are: ① After storing several driving scenarios in the scenario database, several driving scenarios can be extracted from the scenario database to form a complex scenario. Based on the complex scenario, virtual car driving tests and actual car driving tests can be carried out separately. Through the combination of various different driving scenarios, the car under test can be subjected to comprehensive driving tests, simulating various different actual road conditions, thereby obtaining driving data of the car under test under different road conditions, evaluating the performance of the car under test in order to optimize the performance. ② During testing, the process is divided into two groups. First, a digital twin model of the vehicle under test is constructed. Then, the digital twin model can undergo virtual driving tests in a complex scenario virtual simulation environment. Simultaneously, the complex scenario can be reproduced on actual test roads. The vehicle under test can then undergo actual driving tests in the reproduced complex scenario, including autonomous driving or manual driving. Data is collected in both virtual and actual driving tests. The collected virtual test data and actual test data are compared and calculated to obtain perception performance parameters, decision performance parameters, and control performance parameters. Finally, by combining the perception performance parameters, decision performance parameters, and control performance parameters, the driving test of the vehicle under test can be evaluated. This includes evaluating the vehicle's ability to perceive obstacles, its decision-making ability after recognizing obstacles, and its control ability to control changes in the vehicle's driving state. This achieves a comprehensive and multi-dimensional evaluation of the vehicle's driving performance, allowing for the assessment of its driving performance. Based on the test data, the vehicle can be optimized to improve the driving experience. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a vehicle driving test evaluation method that integrates scene complexity according to the present invention; Figure 2 This is a flowchart of step S1 of a vehicle driving test evaluation method that integrates scene complexity according to the present invention; Figure 3 This is a flowchart of step S2 of a vehicle driving test evaluation method that integrates scene complexity according to the present invention; Figure 4 This is a flowchart of step S3 of a vehicle driving test evaluation method that integrates scene complexity according to the present invention; Figure 5 This is a flowchart of step S4 of a vehicle driving test evaluation method that integrates scene complexity according to the present invention. Figure 6 This is a flowchart of step S5 of a vehicle driving test evaluation method that integrates scene complexity according to the present invention. Detailed Implementation

[0019] To better understand the technical content of this invention, a specific embodiment is provided below, and the invention will be further described in conjunction with the accompanying drawings.

[0020] See Figures 1 to 3 The present invention provides a vehicle driving test evaluation method that integrates scenario complexity, comprising the following steps: Step S1: Build a scenario database by acquiring and storing several driving scenarios in the scenario database; Step S2: Select multiple driving scenarios to form a complex scenario, evaluate the scenario complexity of the complex scenario, and generate road requirements based on the complex scenario; Step S3: Construct a digital twin model of the vehicle under test, conduct virtual driving tests in complex scenarios based on the digital twin model, and obtain virtual test data; Step S4: Set up the test road based on road requirements, conduct actual driving tests on the test road, and obtain actual test data; Step S5: Calculate the perception performance parameters, decision performance parameters, and control performance parameters based on the virtual test data and the actual test data, and evaluate them in conjunction with the scenario complexity.

[0021] This invention provides a vehicle driving test evaluation method that integrates scenario complexity. It is applicable not only to driving tests of autonomous vehicles but also to driving tests of manually driven vehicles. It transforms various situations encountered in actual driving into driving scenarios, such as rain-soaked roads, large obstacles on the road, and forced lane changes from adjacent lanes. These different driving scenarios are stored in a scenario database, which testers can access at any time for single-scenario simulated driving tests. Current autonomous driving technology has relatively mature single-scenario simulated driving tests. For example, when encountering rain-soaked roads, it can automatically identify the location and size of the water, thereby changing the vehicle's direction to avoid the flooded area as much as possible. If it cannot avoid the flooded area in time, it can reduce speed to avoid splashing water. The surrounding area and vehicle chassis are considered. However, in actual driving, there are often various driving scenarios on the road. For example, when there is standing water in the rain and there are vehicles driving in the adjacent lane, it is necessary to determine whether avoiding the standing water will affect the normal driving of vehicles in the adjacent lane. However, current driving methods rarely conduct car driving tests in multiple scenarios, making it impossible to determine whether the vehicle's recognition and decision-making capabilities meet the requirements. This invention constructs a scenario database, from which several driving scenarios can be selected and combined to obtain complex scenarios. At the same time, the overall scenario complexity is calculated based on the complexity of different driving scenarios. Then, virtual driving tests and actual driving tests can be conducted separately based on the complex scenarios. Finally, the evaluation of car driving tests is achieved by comparing the virtual driving tests and actual driving tests.

[0022] In virtual driving tests, a digital twin model is first constructed based on the relevant parameters of the vehicle under test, ensuring consistency in all aspects. This digital twin model simulates the entire driving process of the vehicle, including obstacle recognition, decision-making upon obstacle detection, and changes in various control parameters during driving. After constructing the digital twin model, complex scenarios can be virtually simulated. The digital twin model of the vehicle under test can then undergo virtual driving tests in these complex virtual environments. In these virtual tests, the vehicle will identify, perceive, and avoid obstacles in a standard manner, obtaining corresponding virtual test data. Because the virtual driving test is a 1:1 recreation of the vehicle under test within a complex scenario, the results can serve as benchmark data to determine whether any issues arise during actual driving tests. These variations necessitate the generation of road requirements based on complex scenarios before conducting actual driving tests. These requirements are then used to set up real-world test roads, reproducing the complex scenarios. Finally, the test vehicle can be controlled to perform actual driving tests in these reproduced complex scenarios, collecting real-world test data. By comparing virtual and real-world test data, the differences can be used to calculate perception performance parameters, decision-making performance parameters, and control performance parameters. Based on these differences, it can be determined whether there are deficiencies in obstacle recognition, decision-making after obstacle recognition, and vehicle control direction during the actual test. This process evaluates the driving performance of the test vehicle, allowing for subsequent optimization of its performance and related decision-making rules, ultimately improving the driving experience and safety performance of the test vehicle.

[0023] Preferably, step S1 includes the following steps: Step S11: Extract the original scene from real road survey data, traffic accident data and standards and regulations; Step S12: Deconstruct the original scene into road elements, static obstacle elements, traffic participant elements, and environmental elements; Step S13: Classify and store road elements, static obstacle elements, traffic participant elements, and environmental elements into a structured scene database.

[0024] The scenario database is a structured scenario database, which can store data in a structured manner according to different driving scenarios. In actual driving, data on the road and its surroundings can be acquired through the vehicle's own data collection equipment. Therefore, some original scenarios can be extracted from the vehicle's real road data. In addition, traffic accidents often contain a large amount of scenario data, so original scenarios can also be extracted from traffic accident data. Finally, to enrich the amount of original scenario data, classic original scenarios can be extracted from standard regulations. The original scenarios contain different elements, which can be used as a single driving scenario or combined to form complex scenarios. The elements are mainly divided into road elements, static obstacle elements, traffic participant elements, and environmental elements. Road elements mainly include road materials, lane lines, etc. Static obstacles include traffic cones, warning signs, stones, and tree branches on the road. Traffic participants include motor vehicles and non-motor vehicles in other lanes, as well as pedestrians on zebra crossings or sidewalks. Environmental elements include lighting, rain, or snow. After decomposing the original scenario into the above-mentioned multiple elements, the different elements can be stored in the structured scenario database according to their categories for later retrieval.

[0025] Preferably, step S2 includes the following specific steps: Step S21: When there is a subjective testing requirement, select the corresponding driving scenarios from the scenario database according to the subjective testing requirement to form a complex scenario; Step S22: When there is no subjective testing requirement, determine representative road elements based on the geographical location of the vehicle driving test; Step S23: Randomly select at least one element from static obstacle elements, traffic participant elements, and environmental elements to combine with road elements to form a complex scene; Step S24: Perform conflict detection on the combined complex scene. If a conflict exists, re-combine the complex scene. Step S25: Evaluate the complexity of the complex scene, extract the quantitative parameters of each element in the complex scene, and generate road requirements based on the quantitative parameters.

[0026] When conducting driving tests, complex scenarios need to be generated. The generation of complex scenarios requires first determining whether the testers have subjective testing needs. For example, if the testers want to test whether a vehicle can identify a flooded area on a road with pedestrians on the adjacent sidewalk, and how it makes decisions to avoid splashing pedestrians, then a complex scenario can be created by directly selecting corresponding driving scenarios from the scenario database based on subjective testing needs. If there are no subjective testing needs, complex scenarios can be randomly generated based on the geographical location of the test site. Based on the geographical location, road elements can be determined first; for example, in cities with undulating terrain, sloping road sections can be selected. Then, at least one element can be randomly selected from static obstacle elements, traffic participant elements, and environmental elements for combination. After obtaining the complex scenario, conflict detection is required. For example, high temperatures and snowfall cannot occur simultaneously. If a conflict exists, the complex scenario needs to be regenerated. Finally, the scenario complexity is determined through expert evaluation, and quantitative parameters of each element are extracted from the complex scenario, such as rainfall, size and depth of the flooded area, etc. Converting these quantitative parameters into road requirements facilitates the subsequent reproduction of the complex scenario on the test road.

[0027] Preferably, step S3 includes the following specific steps: Step S31: Extract the physical parameters, decision rules, and control algorithms of the vehicle under test, and construct a digital twin model; Step S32: Import the complex scene into the simulation environment and place the digital twin model in the simulation environment for virtual driving test; Step S33: Extract virtual test data during the virtual driving test process through the simulation environment.

[0028] Before conducting virtual driving tests, a digital twin model needs to be constructed. After extracting the physical parameters, decision rules, and control algorithms of the vehicle under test, it can be used to construct the digital twin model. The physical parameters include the length, width, height, body design, speed, turning radius, etc. of the vehicle under test. The decision rules include the obstacle recognition algorithm used and the specific handling rules after the obstacle is identified. The control algorithm includes the control commands taken to the vehicle's components based on the decision. Then, the complex scenario is imported into the simulation environment. In the simulation environment, virtual driving tests can be conducted through the digital twin model. The virtual test data during the entire driving test process can be directly extracted from the simulation environment for subsequent comparison.

[0029] Preferably, step S4 includes the following specific steps: Step S41: Set up the driving scenario on the test road according to the road requirements; Step S42: Drive the vehicle to be tested onto the test road and conduct an actual driving test; Step S43: Collect actual test data during the actual driving test using onboard sensors and data recording equipment; Step S44: Perform spatiotemporal alignment between the virtual test data and the actual test data.

[0030] According to road requirements, complex scenarios can be reproduced on actual roads. Then, the vehicle under test can be driven on the test road. During the driving test, data is collected through sensors and data recording devices. The collected data is the actual test data. Since the actual test data needs to be compared with the virtual test data, it is necessary to ensure the consistency of the spatiotemporal dimensions of the two. That is, the virtual test data and the actual test data need to be spatiotemporally aligned.

[0031] Preferably, both the virtual test data and the actual test data include perception data, decision data, and control data. The perception data includes the types and number of obstacles identified, the decision data includes the path to avoid obstacles and the decision selection of key nodes, and the control data includes acceleration and energy consumption.

[0032] Virtual test data and actual test data need to be compared, so the data in both are consistent and include perception data, decision data, and control data. Perception data includes the identification of the types and calculation of the number of obstacles on the road when the test vehicle is driving on the road. Decision data includes planning the vehicle's path to avoid obstacles after they are identified, as well as decision-making choices at some key nodes. For example, when a vehicle in an adjacent lane signals to change lanes, should the vehicle yield to leave space in front or accelerate away to leave space behind? Control data includes the vehicle's acceleration and total energy consumption. If there is a large difference in the vehicle's acceleration at a certain point in time, it indicates that there may be differences in decision-making or identification. If there is a large difference in total energy consumption, it may indicate that the vehicle is not being controlled properly, leading to an increase in fuel or electricity consumption.

[0033] Preferably, the specific steps for calculating the sensing performance parameters in step S5 include: Step S51: Compare the number of obstacles identified in the virtual test data and the actual test data, and calculate the false negative rate based on the difference in the number of obstacles; Step S52: Determine whether the types of obstacles identified at the same time point are consistent, and calculate the recognition accuracy based on the judgment results of whether the types of obstacles are consistent throughout the entire test process.

[0034] To determine the false negative rate, the perception performance parameters need to be calculated by comparing the number of obstacles identified in the two sets of test data. If the number of obstacles identified in the actual test data is less than the number identified in the virtual test data, it indicates a false negative. The false negative rate can be calculated based on the difference in the number of obstacles. Since spatiotemporal alignment is performed, the time points when obstacles are encountered are basically consistent. Judging whether the types of obstacles identified at the same time point are consistent can assess the obstacle recognition ability of the vehicle under test. The accuracy rate can be calculated based on the difference in the types of obstacles identified. The false negative rate and the accuracy rate are the corresponding perception performance parameters.

[0035] Preferably, the specific steps for calculating the decision performance parameters in step S5 include: Step S53: Extract obstacle avoidance paths from virtual test data and actual test data respectively, calculate the obstacle avoidance paths point by point, and obtain the longitudinal deviation and lateral deviation of the paths. Step S54: Assign weights to the longitudinal deviation and lateral deviation of the path respectively, and sum them to obtain the comprehensive path deviation; Step S55: Perform semantic recognition on the complex scene and extract key nodes. Obtain the decision-making choices of the virtual test data and the actual test data at the key nodes, and make a consistency judgment. Then, binarize the consistency judgment result.

[0036] When calculating decision performance parameters, the main components include obstacle avoidance path decisions and key node decisions. The driving paths, including those before and after encountering obstacles, are extracted from both virtual and actual test data. Then, point-by-point calculations are performed on these driving paths to obtain longitudinal and lateral path deviations. Weighted summation of these deviations yields the comprehensive path deviation. A smaller comprehensive path deviation indicates more similar obstacle avoidance paths and closer decision performance. Furthermore, different driving scenarios involve different key nodes. For example, in a driving scenario including a zebra crossing, reaching the zebra crossing is a key node. Therefore, semantic recognition of complex scenarios can be performed to obtain several key nodes. Then, the decision choices at these key nodes are extracted from both virtual and actual test data, and their consistency is assessed. If consistent, a value of 1 is assigned; otherwise, a value of 0 is assigned. Binarization facilitates subsequent quantification for testing and evaluation.

[0037] Preferably, the specific steps for calculating the control performance parameters in step S5 include: Step S56: Extract acceleration values ​​from the virtual test data and the actual test data respectively, and calculate the acceleration value difference rate; Step S57: Extract the consumed energy from the virtual test data and the actual test data respectively, and calculate the energy difference rate.

[0038] When calculating control performance parameters, the difference rate of acceleration value and the difference rate of energy consumption can be directly calculated based on the differences in acceleration values ​​and energy consumption extracted from virtual test data and actual test data.

[0039] Preferably, the specific steps of step S5, which combines scene complexity for testing and evaluation, include: Step S58: Construct an evaluation matrix. Standardize the perception performance parameters, decision performance parameters, and control performance parameters, and then fill them into the evaluation matrix along with the scenario complexity. Step S59: Calculate the eigenvalues ​​of the covariance matrix of the evaluation matrix, and conduct a test evaluation based on the difference between the largest eigenvalue and other eigenvalues.

[0040] When evaluating performance in conjunction with scenario complexity, an evaluation matrix is ​​first constructed. Perception performance parameters, decision performance parameters, and control performance parameters are standardized and their dimensions are unified. Then, the corresponding false negative rate, accuracy rate, overall path deviation, binarized consistency judgment results, acceleration value difference rate, energy difference rate, and scenario complexity are filled into the evaluation matrix. Next, the covariance matrix of the evaluation matrix is ​​calculated, and its eigenvalues ​​are calculated. Since there may be multiple eigenvalues, the largest eigenvalue is selected. Its principal component direction corresponds to the direction of the most drastic data change. The largest eigenvalue is compared with other eigenvalues. If the largest eigenvalue is significantly larger than other eigenvalues ​​(e.g., accounting for more than 80% of the overall variance), it indicates that a common trend dominates the overall system performance in both virtual and real-world testing. If the largest eigenvalue is not significantly different from other eigenvalues, it indicates a large gap between virtual and real-world driving. In this case, testing in other complex scenarios or optimizing the performance of the vehicle under test or related rules can be performed.

[0041] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating a car driving test by fusing scenario complexity, characterized in that, The method comprises the following steps: Step S1, constructing a scene database, obtaining a plurality of driving scenes and storing them into the scene database; Step S2, selecting a plurality of driving scenes to combine into a complex scene, evaluating the scene complexity of the complex scene, and generating road requirements according to the complex scene; Step S3, constructing a digital twin model of the to-be-tested vehicle, performing virtual driving test in the complex scene based on the digital twin model, and obtaining virtual test data; Step S4, setting the test road based on the road requirements, and performing actual driving test on the test road by the to-be-tested vehicle, and obtaining actual test data; Step S5, calculating the perception performance parameters, decision performance parameters and control performance parameters according to the virtual test data and the actual test data respectively, and combining the scene complexity to evaluate the test.

2. The method according to claim 1, wherein The specific steps of the step S1 include: Step S11, extracting original scenes from real road data, traffic accident data and standard regulations; Step S12, decomposing the original scenes into road elements, static obstacle elements, traffic participant elements and environment elements; Step S13, classifying and storing the road elements, static obstacle elements, traffic participant elements and environment elements into a structured scene database.

3. The method according to claim 2, wherein The specific steps of the step S2 include: Step S21, when there is a subjective test requirement, selecting corresponding driving scenes from the scene database according to the subjective test requirement to combine into a complex scene; Step S22, when there is no subjective test requirement, determining representative road elements according to the geographical location of the vehicle driving test; Step S23, randomly selecting at least one from the static obstacle elements, traffic participant elements and environment elements to combine with the road elements into a complex scene; Step S24, performing conflict detection on the combined complex scene, and recombining the complex scene if there is a conflict; Step S25, evaluating the scene complexity of the complex scene, extracting the quantization parameters of each element in the complex scene, and generating road requirements according to the quantization parameters.

4. The method of claim 1, wherein the method further comprises: The specific steps of the step S3 include: Step S31, extracting physical parameters, decision rules and control algorithms of the to-be-tested vehicle, and constructing a digital twin model; Step S32, importing the complex scene into a simulation environment, and placing the digital twin model in the simulation environment for virtual driving test; Step S33, extracting virtual test data in the virtual driving test process through the simulation environment.

5. The method of claim 1, wherein the method further comprises: The specific steps of the step S4 include: Step S41, arranging the driving scene on site in the test road according to the road requirements; Step S42, driving the to-be-tested vehicle into the test road and performing actual driving test; Step S43, collecting actual test data in the actual driving test process through vehicle-mounted sensors and data recording equipment; Step S44, time and space alignment of virtual test data and actual test data.

6. The method of claim 1, wherein the method further comprises: The virtual test data and the actual test data both include perception data, decision data and control data, the perception data includes identified obstacle types and quantities, the decision data includes path for avoiding obstacles and decision selection of key nodes, and the control data includes acceleration and energy consumption.

7. The method according to claim 6, wherein The specific steps of the step S5 of calculating the perception performance parameter include: A step S51, comparing the identified obstacle quantities in the virtual test data and the actual test data, and calculating a missed detection rate based on the difference of the obstacle quantities; A step S52, judging whether the identified obstacle types at the same time node are consistent, and calculating an identification accuracy based on the judgment result of whether the obstacle types are consistent in the whole test process.

8. The method of claim 6, wherein the method further comprises: The specific steps of the step S5 of calculating the decision performance parameter include: A step S53, respectively extracting the paths for avoiding obstacles from the virtual test data and the actual test data, point-by-point calculating the paths for avoiding obstacles, and obtaining path longitudinal deviation and path lateral deviation; A step S54, respectively weighting the path longitudinal deviation and the path lateral deviation, and obtaining a comprehensive path deviation after summation; A step S55, performing semantic recognition on the complex scene, extracting key nodes, respectively obtaining the decision selections of the key nodes in the virtual test data and the actual test data, and performing consistency judgment, and binarizing the consistency judgment result.

9. The method of claim 6, wherein the method further comprises: The specific steps of the step S5 of calculating the control performance parameter include: A step S56, respectively extracting acceleration values from the virtual test data and the actual test data, and calculating an acceleration value difference rate; A step S57, respectively extracting energy consumption from the virtual test data and the actual test data, and calculating an energy difference rate.

10. The method of claim 1, wherein the method further comprises: The specific steps of the step S5 of combining the scene complexity for test evaluation include: A step S58, constructing an evaluation matrix, and filling the standardized perception performance parameter, the standardized decision performance parameter and the standardized control performance parameter together with the scene complexity into the evaluation matrix; A step S59, calculating eigenvalues of a covariance matrix of the evaluation matrix, and performing test evaluation according to the size difference between the largest eigenvalue and other eigenvalues.

Images