An automatic driving acceleration test scene extraction method and evaluation method

Abstract

The application is suitable for the field of automatic driving technology, and provides an automatic driving acceleration test scene extraction method and an evaluation method.The extraction method comprises: acquiring a plurality of driving data, obtaining a feature matrix of key driving interaction features based on a time sequence of the driving interaction data; solving an optimal segmentation scheme based on a total likelihood function estimation criterion, and obtaining segmented optimal segmentation subsegments based on the optimal segmentation number.The application has the advantages that by introducing a greedy Gaussian segmentation algorithm, segmentation of time sequence data is modeled based on a multivariate Gaussian distribution, effectively solving the problems of subjectivity, insufficient coverage of traditional rule-based methods, poor interpretability and high computational complexity of machine learning methods, reducing the dependence of the algorithm on data quality and manual annotation, providing efficient support for constructing an automatic driving test scene library that covers more comprehensively and is more challenging, and helping to improve the test efficiency and the reliability of system safety evaluation.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and in particular relates to a method for extracting and evaluating autonomous driving acceleration test scenarios. Background Technology

[0002] Autonomous driving testing faces a core dilemma: the volume of natural driving data is enormous—for example, a single test vehicle can generate terabytes of data per day. However, the vast majority of this data consists of routine scenarios, offering minimal value for verifying the system's limits or uncovering potential defects. Directly testing with the entire dataset would result in extremely low efficiency, potentially requiring a massive amount of time to cover critical risk points. Scenario extraction serves as a data filter and value extractor, filtering out redundant routine scenarios and retaining only the key scenarios that accelerate testing.

[0003] However, given the massive amount of natural driving data, how to efficiently extract representative and challenging test scenarios to accelerate the testing and verification of autonomous driving systems remains a critical issue that urgently needs to be addressed.

[0004] Among existing autonomous driving scene extraction technologies, one type is the driving scene extraction method based on preset rules. This method models driving behavior patterns and traffic scene features in natural driving data through expert experience, sets a series of rules and thresholds to filter the required scenes. The specific steps are to extract basic parameters such as vehicle trajectory, speed, and acceleration from the natural driving dataset, classify these parameters according to the rules defined by experts, and mark the scene fragments that meet the conditions as test scenes and store them. The other type is the driving scene extraction method based on machine learning. It uses machine learning algorithms to automatically analyze multi-dimensional features in natural driving data and generates a driving scene representation model through learning the features.

[0005] However, current methods all have many shortcomings: driving scene extraction methods based on preset rules have technical problems such as lack of universality, insufficient scene coverage, and strong subjectivity; driving scene extraction methods based on machine learning have technical problems such as low utilization of natural driving data and difficulty in efficiently capturing high-risk edge scenarios. Summary of the Invention

[0006] The purpose of this application is to provide a method for extracting autonomous driving acceleration test scenarios, which aims to solve the technical problems of driving scenario extraction methods based on preset rules, such as lack of universality, insufficient scenario coverage, and strong subjectivity; while driving scenario extraction methods based on machine learning have the problems of low utilization of natural driving data and difficulty in efficiently capturing high-risk edge scenarios.

[0007] This application provides a method for extracting autonomous driving acceleration test scenarios, the extraction method comprising:

[0008] Acquire a number of driving data points, and perform preliminary screening of the driving data based on preset typical driving scenarios to obtain a time series of driving interaction data that conforms to the preset scenarios.

[0009] Based on the time series of the driving interaction data, a feature matrix of key driving interaction features is obtained; the key driving interaction features include at least speed gradient features, acceleration gradient features, time headway features, and collision time features.

[0010] The feature matrix is ​​input into a greedy Gaussian segmentation model for segmentation, and the optimal segmentation scheme is solved based on the total likelihood function estimation criterion.

[0011] A preset segmentation number value is set for the greedy Gaussian segmentation model, and the segmentation number value is iterated to obtain the log-likelihood value during the iteration process. The optimal segmentation number is obtained corresponding to the abrupt inflection point of the log-likelihood value change curve. Based on the optimal segmentation number, the position of the segmentation point and the optimal segmentation sub-segment are obtained.

[0012] Preferably, before performing initial screening of the driving data, the method further includes preprocessing the driving data, wherein the preprocessing method is as follows:

[0013] Outliers in the driving data are removed to obtain the data to be denoised;

[0014] The data to be denoised is then subjected to denoising and smoothing processing based on the following formula:

[0015] ;

[0016] in, For vehicles The original data at sampling time k For the smoothed data, for Total number of sample points Represents the time span of the smoothing operation. Z represents the time interval between each frame.

[0017] Preferably, the feature matrix is ​​input into a greedy Gaussian segmentation model for segmentation based on the following formula:

[0018] ;

[0019] ;

[0020] in:

[0021] This is the characteristic matrix, which has a total of There are n timestamps, each with a dimension of n;

[0022] The segmentation points of the feature matrix are: It is divided into 100 parts. The segment, the dividing point satisfies:

[0023] ;

[0024] This represents the mean. Describing covariance, Represents timestamp, mean and covariance The value is only at the split point Changes only occur at the branching points:

[0025] ;

[0026] superscript in the above formula Indicates the first Each segment This represents the mean value corresponding to each segment. Let be the covariance of each segment.

[0027] Preferably, the total likelihood function is:

[0028] ;

[0029] ;

[0030] in, Let be the objective function. For a sub-item in the objective function, All of these are the optimal segmentation point positions.

[0031] Preferably, the extraction method further includes:

[0032] During the iterative process of calculating the number of segments, the covariance is constrained from overfitting based on the following formula:

[0033] ;

[0034] in, For regularization terms, This is the optimized objective function.

[0035] Preferably, the extraction method further includes:

[0036] Calculate the mean and variance of multiple autonomous driving interaction sub-time series elements to obtain the mean information and variance information of multiple autonomous driving interaction sub-time series elements.

[0037] Based on the mean information and variance information of the multiple autonomous driving interaction sub-time series elements, multiple distribution intervals of autonomous driving interaction sub-time series elements are generated; each distribution interval of autonomous driving interaction sub-time series elements includes the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval; the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval correspond one-to-one with the autonomous driving interaction sub-time series information.

[0038] When the minimum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is greater than the left endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, and the maximum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is less than the right endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, then the autonomous driving interaction sub-time sequence information is used as autonomous driving scene extraction information.

[0039] Another objective of this application is to provide a method for evaluating autonomous driving acceleration test scenarios, the evaluation method comprising:

[0040] A real-time driving risk assessment method based on artificial potential field theory is constructed as a hazard model, wherein the hazard model is as follows:

[0041] ;

[0042] in, For the risk level model, and These represent the risks associated with interactive vehicles and the risks associated with speed, respectively. These are the weighting coefficients, The decay parameter for the exponential term; in the interactive risk term middle, The introduced virtual energy characterizes the contribution of vehicle size and speed to risk. and These are the vehicle's length and width, respectively. The angle between the directions of travel of the two vehicles; Let be the relative distance vector between the two vehicles. Its Euclidean distance; in the speed risk item middle, The target speed of the main vehicle, The actual speed of the main vehicle;

[0043] Based on the aforementioned hazard model, the evaluation results of the autonomous driving acceleration test scenario extraction method described above are obtained.

[0044] Another objective of this application is to provide an autonomous driving acceleration test scenario extraction device, the device comprising:

[0045] The driving data acquisition module is used to acquire a number of driving data, and to perform preliminary screening of the driving data based on a preset typical driving scenario to obtain a time series of driving interaction data that conforms to the preset scenario.

[0046] The feature matrix acquisition module is used to obtain a feature matrix of key driving interaction features based on the time series of the driving interaction data; the key driving interaction features include at least speed gradient features, acceleration gradient features, time headway features, and collision time features.

[0047] The optimal segmentation scheme acquisition module is used to input the feature matrix into the greedy Gaussian segmentation model for segmentation and solve the optimal segmentation scheme based on the total likelihood function estimation criterion.

[0048] The optimal segment acquisition module is used to preset the number of segments for the greedy Gaussian segmentation model, iterate the number of segments, obtain the log-likelihood value during the iteration process, obtain the optimal number of segments corresponding to the abrupt inflection point of the log-likelihood value change curve, and obtain the position of the segmentation point and the optimal segment based on the optimal number of segments.

[0049] Another objective of this application is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the autonomous driving acceleration test scenario extraction method as described above.

[0050] Another objective of this application is to provide an autonomous driving acceleration test scene extraction system, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the autonomous driving acceleration test scene extraction method as described above.

[0051] This application provides a method for extracting autonomous driving acceleration test scenarios. Its key advantage lies in its introduction of a greedy Gaussian segmentation algorithm. By modeling the segmentation process of time-series data based on a multivariate Gaussian distribution, it effectively solves the problems of subjectivity and insufficient coverage inherent in traditional rule-based methods, as well as the poor interpretability and high computational complexity of machine learning methods. This method significantly improves its adaptability to different datasets by dynamically optimizing segmentation breakpoints and regularizing covariance estimation. It can accurately capture the spatiotemporal features of vehicle interactions in complex traffic scenarios, and is particularly efficient in extracting high-risk edge scenarios missed by traditional rules. Furthermore, by using the log-likelihood function curve to determine the optimal segmentation point, it reduces the algorithm's dependence on data quality and manual annotation, providing efficient support for building a more comprehensive and challenging autonomous driving test scenario library, and helping to improve testing efficiency and the reliability of system safety assessments. Attached Figure Description

[0052] Figure 1 An application environment diagram for an autonomous driving acceleration test scenario extraction method provided in this application embodiment;

[0053] Figure 2 A flowchart illustrating an autonomous driving acceleration test scenario extraction method provided in this application embodiment;

[0054] Figure 3 This application provides a schematic diagram of an overtaking scenario.

[0055] Figure 4 This is a schematic diagram of a risk assessment result for an overtaking scenario provided in an embodiment of this application;

[0056] Figure 5 A schematic diagram of the log-likelihood value variation curve provided in an embodiment of this application;

[0057] Figure 6 This is a schematic diagram of an optimized risk assessment result for an overtaking scenario provided in an embodiment of this application;

[0058] Figure 7 A schematic diagram showing the comparison of risk assessment results before and after optimization, provided for an embodiment of this application;

[0059] Figure 8 This application provides a structural block diagram of an autonomous driving acceleration test scene extraction device according to an embodiment of the present application.

[0060] Figure 9 This is a block diagram of the internal structure of a computer device in one embodiment. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] It is understood that the terms "first," "second," etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish the first unit or module from another unit or module. For example, without departing from the scope of this application, the first script may be referred to as the second script, and similarly, the second script may be referred to as the first script.

[0063] Figure 1 This is an application environment diagram of the autonomous driving acceleration test scenario extraction method provided in the embodiments of this application, such as... Figure 1 As shown, the application environment includes a detection device 110 and a computer device 120.

[0064] The computer device 120 can be an independent or networked vehicle-mounted computer, an independent physical server or terminal, a server cluster consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud servers, cloud storage, and CDN.

[0065] The detection device 110 may be a lidar, visual radar, millimeter-wave radar, or a device that can transmit data of the vehicle's surrounding environment to the computer device 120, but it is not limited to these.

[0066] The detection device 110 and the computer equipment 120 can be connected via a network, which is not a limitation of this application.

[0067] like Figure 2 As shown, in one embodiment, a method for extracting autonomous driving acceleration test scenarios is proposed. This embodiment mainly applies this method to the above-mentioned scenarios. Figure 1 Taking computer device 120 as an example, a method for extracting an autonomous driving acceleration test scenario may specifically include the following steps:

[0068] Step S10: Acquire a number of driving data points, and perform preliminary screening of the driving data based on preset typical driving scenarios to obtain a time series of driving interaction data that conforms to the preset scenarios.

[0069] In this embodiment, the driving data includes at least lane data, the vehicle's own driving data, and driving data of other vehicles in the surrounding area.

[0070] In this embodiment, the vehicle itself is first selected as the target vehicle, and other vehicles are selected as background vehicles. A target recognition algorithm is used to obtain some driving data.

[0071] Among them, the lane data in the driving data can include at least lane width data, and the acquired vehicle driving data can include one or more of the following data: lateral and longitudinal speed, acceleration, headway, and vehicle length and width data; similarly, the acquired driving data of several other vehicles can also include one or more of the above data.

[0072] In this embodiment, due to the potentially large number of vehicles in the lanes and the resulting massive amount of data, initial screening of the data is necessary. Initial screening is used to filter out meaningless data after traversing all driving data, based on numerous typical scenarios such as following other vehicles, lane changing, and vehicles cutting in front. For example, in a cutting-in scenario, the system can extract the vehicle itself, the nearest vehicle within a threshold distance in front of the vehicle, and the cutting-in vehicle; in a lane changing scenario, it can extract the vehicle itself, the nearest vehicle in front of the vehicle, and the vehicle behind the vehicle within a threshold distance of the target lane.

[0073] In this embodiment, driving interaction data conforming to a preset scenario refers to, for example, in a following scenario, only the driving data of the vehicle itself and the nearest vehicle within a threshold distance (e.g., 200m) in front of it needs to be extracted, and this data is used as the driving interaction data for the preset scenario. Then, based on the time sequence of this data, the time sequence of ideal driving interaction data is obtained. The driving data conforming to the preset scenario can be defined as driving interaction data, which is the data of vehicles with interactive behavior.

[0074] Step S20: Based on the time series of the driving interaction data, obtain the feature matrix of key driving interaction features; the key driving interaction features include at least speed gradient features, acceleration gradient features, time headway features, and collision time features.

[0075] In this embodiment, a feature matrix is ​​constructed to capture the relative motion between vehicles in a more refined manner, avoiding the neglect of local dynamics.

[0076] In this embodiment, the traditional six-parameter overtaking kinematic model has been improved. Therefore, in addition to features such as speed gradient features and acceleration gradient features, the key driving interaction features can also include the following features: speed features, acceleration features, and vehicle spacing features.

[0077] In this embodiment, vehicle spacing refers to the distance between the vehicle and the vehicle in front, directly reflecting the spatial relationship between the two vehicles. Headway time is calculated as vehicle spacing / vehicle speed, indicating how long it takes to reach the position of the vehicle in front at the current speed; it reflects risk better than simple spacing. Time to collision (TTC) is calculated as vehicle spacing / relative speed, indicating how many seconds it will take for a collision to occur if the current relative speed remains constant; the smaller the TTC value, the higher the level of danger.

[0078] Step S30: Input the feature matrix into the greedy Gaussian segmentation model for segmentation, and solve the optimal segmentation scheme based on the total likelihood function estimation criterion.

[0079] Step S40: Preset the number of segments for the greedy Gaussian segmentation model, iterate the number of segments, obtain the log-likelihood value during the iteration process, obtain the optimal number of segments corresponding to the abrupt inflection point of the log-likelihood value change curve, and obtain the optimal segmentation sub-segment based on the optimal number of segments.

[0080] In this embodiment, preferably, the feature matrix can first be processed using principal component analysis (PCA) to reduce data dimensionality, retain the most important variation information in the data, and improve the efficiency and robustness of subsequent segmentation algorithms. Then, the processed multivariate time series is input into a greedy Gaussian segmentation model to segment it as time-series data.

[0081] In this embodiment, during the segmentation of time-series data into subsequences, the data corresponding to each subsequence segment is modeled as a multivariate Gaussian distribution with different means and covariances. This transforms the multivariate time series segmentation problem into a maximum likelihood problem, which is then used to solve a simplified combinatorial optimization problem of searching for possible breakpoints or boundaries. This allows the multivariate time series to be reasonably divided into multiple subsequence segments. The greedy Gaussian segmentation model used in this embodiment assumes that each segment follows a multivariate Gaussian distribution and finds the optimal breakpoint by maximizing the likelihood function.

[0082] In this embodiment, a greedy Gaussian segmentation model is used to determine the number of segmentation points. The process is iterated, and the log-likelihood value is observed as the number of split points increases. The optimal number of split points is determined by analyzing the changing curve of the log-likelihood function value curve. and the optimal number of split points The breakpoint values. Theoretically, the log-likelihood function value will increase with the number of breakpoints. The log-likelihood function value increases with the number of split points. The growth process resembles a logarithmic curve; after a certain point, the growth of the log-likelihood value tends to plateau, and the curve becomes nearly parallel to the x-axis. A straight line along the axis indicates that the log-likelihood value hardly changes with the number of dividing points. The number of split points increases with the increase of [the number of split points], and the turning point is the optimal number of split points. The final model outputs the optimal number of segmentation points and the position of each segmentation point. Based on the position of the optimal segmentation points, the data in the feature matrix is ​​segmented to obtain multiple optimal segmentation sub-segments.

[0083] In this embodiment, the optimal segmentation obtained through the above method is one of the core aspects of researching autonomous driving acceleration test scenarios. The feature matrix of each segment can be used for classification or clustering to identify scenario types, such as following, lane changing, and cutting in. Because the features within each segment are uniform, they are easier to match with preset typical driving scenarios. By extracting key scenario fragments, i.e., the optimal segmentation, the entire driving process does not need to be reproduced during testing; the focus can be placed on the critical or typical scenarios represented by these segments, thereby significantly accelerating testing efficiency.

[0084] The advantage of this application lies in its introduction of a greedy Gaussian segmentation algorithm. By modeling the segmentation process of time-series data based on a multivariate Gaussian distribution, it effectively solves the problems of subjectivity and insufficient coverage of traditional rule-based methods, as well as the poor interpretability and high computational complexity of machine learning methods. This method significantly improves its adaptability to different datasets by dynamically optimizing segmentation breakpoints and regularizing covariance estimation. It can accurately capture the spatiotemporal features of vehicle interactions in complex traffic scenarios, and is particularly efficient in extracting high-risk edge scenarios missed by traditional rules. Furthermore, by using the log-likelihood function curve to determine the optimal segmentation point, it reduces the algorithm's dependence on data quality and manual annotation, providing efficient support for building a more comprehensive and challenging autonomous driving test scenario library, and helping to improve testing efficiency and the reliability of system safety assessment.

[0085] In this application embodiment, driving interaction features refer to features collected in scenarios where actions occur between vehicles, such as... Figure 3 The diagram illustrates an overtaking scenario where a typical six-parameter overtaking kinematics model can be used. The features involved in this model are shown in the table below:

[0086]

[0087] There are many other driving interaction features. For example, the features involved in lane changing scenarios can be shown in the table below:

[0088]

[0089] The key driving interaction features used in this embodiment refer to features selected from the above-mentioned parameter features, including at least speed gradient features, acceleration gradient features, time headway features (thw), and collision time features (ttc), and may also include the six types of features involved in the six-parameter overtaking kinematic model.

[0090] As a specific example of this application, the method described in this application can be summarized as follows:

[0091] Step 1: Collect natural driving data and driving simulator data. Collect vehicle interaction-related parameters under natural driving conditions as time series data through onboard sensors or roadside sensors. With the self-vehicle as the target vehicle and other vehicles as background vehicles, use target recognition algorithms and related modules to obtain key driving interaction features, including the lateral and longitudinal speeds, accelerations, headway, vehicle length and width, and lane width of the self-vehicle and surrounding vehicles.

[0092] Step 2: Set typical driving scenario filtering rules. Based on the rules, the scenario is initially filtered. By traversing the existing natural driving dataset, scenario filtering rules are set according to various defined typical driving scenarios, including but not limited to following, lane changing, and other scenarios. Based on the lane and position of the main vehicle, the nearest vehicles in front and behind in the lane, as well as the nearest vehicles in the adjacent lanes and their nearest preceding and following vehicles, are retrieved. The relevant static and dynamic information of the vehicles related to the scenario is extracted from the above vehicles, and the vehicle interaction-related parameters are selected as key driving interaction features, that is, the ideal vehicle interaction time series data is obtained.

[0093] Step 3: Conduct initial scene screening. After completing the scene screening rule settings in Step 2, screen the natural driving data and driving simulator driving data according to the rules to obtain the initial scene data.

[0094] Step 4: Perform initial screening data preprocessing. Preprocess the screened data to reduce noise and eliminate outliers. Outliers include, but are not limited to, noise from the target recognition or data acquisition modules and data with significant errors. After eliminating outliers, noise reduction is achieved through methods such as smoothing shift filtering and symmetric exponential shift filtering, ultimately smoothing the data. The expression for logarithmic shift filtering is:

[0095] ;

[0096] in, For vehicles The original data at sampling time k For the smoothed data, for Total number of sample points Represents the time span of the smoothing operation. Z and i represent the time interval between each frame; D is the smoothing coefficient; and Δ is the time step.

[0097] Step 5: Execute the greedy Gaussian segmentation algorithm to iterate to new nodes. Based on the overtaking kinematics model analysis, select features for the greedy Gaussian segmentation analysis. According to the most basic six-parameter overtaking kinematics model, select velocity, acceleration, and vehicle spacing as time-series features. On this basis, supplement with velocity gradient, acceleration gradient, time-distance feature (thw) for the front-end, and collision time feature (ttc) to construct a feature matrix. Define the preprocessed data features as a multivariate time series X for subsequent processing.

[0098] ;

[0099] This multivariate time series has a total of Given n timestamps, each of which is n-dimensional, this multivariate time series can be divided into points. ,for The segment, the dividing point satisfies:

[0100] ;

[0101] By assuming a multivariate time series Each timestamp vector in the dataset follows a certain multivariate Gaussian distribution, i.e. and mean and covariance The value is only at the split point Changes only occur at the points of division, that is:

[0102] ;

[0103] In the above formula, the subscript Indicates timestamp, superscript Indicates the first Each segment.

[0104] Step 6: Determine the optimal node by calculating the log-likelihood function value.

[0105] Given an initial value for the number of split points K, find the location of the optimal set of split points. and the mean value corresponding to each segment Covariance .

[0106] Based on the above, the original multivariate time series X is modeled, and its total likelihood function is as follows:

[0107] Modeling the original multivariate time series X, its total likelihood function is:

[0108] ;

[0109] ;

[0110] The model iterates over the number of segmentation points K, observing the log-likelihood value as a function of K. The optimal number of segmentation points K and the positions of the segmentation points under this optimal K are determined by the trend of the log-likelihood curve. The optimal segment is then obtained based on the positions of these segmentation points. Through these steps, scene extraction is achieved.

[0111] As a preferred embodiment of this application, considering that the modeling method used in the above embodiments may result in a dimension much larger than the number of samples, i.e., the dimension of the original multivariate time series... It is a relatively large value, and a certain segment length If the value is small, maximum likelihood estimation of the covariance will produce a relatively large error. This can be avoided by adding covariance regularization. The formula for optimizing the objective function is:

[0112] ;

[0113] Among them, among them, It is a regularization term used to penalize the trace of the inverse of the covariance matrix, i.e., the sum of the diagonal elements, thereby indirectly limiting the covariance matrix from having extreme values ​​and avoiding overfitting.

[0114] As a preferred embodiment of this application, the extraction method further includes:

[0115] Calculate the mean and variance of multiple autonomous driving interaction sub-time series elements to obtain the mean information and variance information of multiple autonomous driving interaction sub-time series elements.

[0116] Based on the mean information and variance information of the multiple autonomous driving interaction sub-time series elements, multiple distribution intervals of autonomous driving interaction sub-time series elements are generated; each distribution interval of autonomous driving interaction sub-time series elements includes the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval; the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval correspond one-to-one with the autonomous driving interaction sub-time series information.

[0117] When the minimum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is greater than the left endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, and the maximum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is less than the right endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, then the autonomous driving interaction sub-time sequence information is used as autonomous driving scene extraction information.

[0118] In this embodiment, the autonomous driving interaction sub-time sequence element information can refer to the optimal segmentation sub-segment in the above embodiments. This embodiment employs optimal breakpoint overlay. Rules are used to optimize the range of extracted scene frames. Specifically, based on expert-recommended rule-based extraction methods for driving scenes, the optimal number of segmentation points K and the optimal number of segmentation points used in the above embodiments are combined. Set the breakpoint values ​​and superimpose them. Rule extraction.

[0119] Because of the assumption of multivariate time series Each timestamp vector in the dataset follows a certain multivariate Gaussian distribution. And its mean was obtained through calculation. Covariance , The rule distributes the values ​​based on the obtained mean and covariance. Breakpoint values ​​within this range are considered reliable. If a calculated breakpoint value is outside this range, it is rejected, and a suboptimal breakpoint value that meets the condition is selected instead. This method further validates all obtained results, removing invalid values ​​and improving the accuracy and reliability of the results.

[0120] As a preferred embodiment of this application, this application also provides a method for verifying and evaluating the accuracy of the results obtained by the above method. The evaluation method includes:

[0121] A real-time driving risk assessment method based on artificial potential field theory is constructed as a hazard model, wherein the hazard model is as follows:

[0122] ;

[0123] in, For the risk level model, and These represent the risks associated with interactive vehicles and the risks associated with speed, respectively. These are the weighting coefficients, The decay parameter for the exponential term; in the interactive risk term middle, The introduced virtual energy characterizes the contribution of vehicle size and speed to risk. and These are the vehicle's length and width, respectively. The angle between the directions of travel of the two vehicles; Let be the relative distance vector between the two vehicles. Its Euclidean distance; in the speed risk item middle, The target speed of the main vehicle, The actual speed of the main vehicle;

[0124] Based on the aforementioned hazard model, the evaluation results of the autonomous driving acceleration test scenario extraction method described above are obtained.

[0125] This embodiment proposes a method for assessment and result verification by constructing a hazard model. The method constructs a real-time driving risk assessment method based on artificial potential field theory as the hazard model. This hazard model consists of two parts: the interactive vehicle risk describing the risk of surrounding moving vehicles, and the speed risk describing the difference between the vehicle's actual speed and its expected speed.

[0126] In one embodiment, the autonomous driving acceleration test scenario extraction method described above is evaluated using the hazard model given in this embodiment, and the evaluation result is as follows: Figure 4 As shown in the figure, the 1 / TTC curve represents the reciprocal of the collision time, the horizontal axis Frame represents the time series in the autonomous driving test scenario, and the vertical axis represents the z-score standardized risk, that is, the result after standardizing the output value of the hazard model.

[0127] In one embodiment, taking the natural driving data of German highways collected using the highD dataset as the original data as an example, the effectiveness and accuracy of the rule extraction optimization method based on greedy Gaussian segmentation proposed in the above embodiment are illustrated.

[0128] First, overtaking data is filtered from the dataset. The filtering rule is that the target vehicle only performs one consecutive overtaking maneuver, and the surrounding vehicles include the vehicle in front of the target vehicle in its original lane and the vehicles in front of and behind the target vehicle in its own lane, within a distance of 200 meters. A feature matrix for subsequent greedy Gaussian segmentation is set using the overtaking kinematics principle analysis method adopted in the above embodiment. The greedy Gaussian segmentation method adopted in the above embodiment is then used for segmentation. The data features in the natural driving data are defined as multivariate time series. Further processing is performed. This involves calculating the number of segmentation points. The process is iterated, and the log-likelihood value is observed as the number of split points increases. The optimal number of split points is determined by analyzing the changing curve of the log-likelihood function value curve. and the optimal number of split points The breakpoint value. For example... Figure 5 As shown, the theoretical log-likelihood function value will increase with the number of split points. The log-likelihood function value increases with the number of split points. The growth process resembles a logarithmic curve; after a certain point, the growth of the log-likelihood value tends to plateau, and the curve becomes nearly parallel to the x-axis. A straight line along the axis indicates that the log-likelihood value hardly changes with the number of dividing points. The number of split points increases with the increase of [the number of split points], and the turning point is the optimal number of split points. .

[0129] The rule extraction method for overtaking scenarios is optimized by superimposing 3σ rules on breakpoints in the dataset obtained using the greedy Gaussian segmentation method, resulting in a rule extraction method for overtaking scenarios based on greedy Gaussian segmentation. The optimized overtaking scenario extraction process is shown in the table below:

[0130]

[0131] The overtaking scenarios extracted using the optimized rules were compared with those extracted using the unoptimized rules. Figure 6 The optimized hazard model assessment results are shown below. These results are compared with the initial assessment results. Figure 7 As shown, the optimized rules capture a segment at the end of the scene where the interaction risk dominates the final risk and the final risk intensity is extremely high. The original rule-based scene extraction method can only identify scenes that conform to the preset rules, but fails to fully explore the edge scenes in the dataset that are not covered by the rules, and is prone to missing some rare but high-risk scenes that are crucial to the safety assessment of autonomous driving systems. The rule extraction method optimized by greedy Gaussian segmentation can effectively make up for the deficiency of insufficient scene coverage in the original expert-defined rule-based scene extraction method.

[0132] In this embodiment, the optimized risk assessment results for the overtaking scenario are shown in the table below:

[0133]

[0134] A comprehensive analysis of the comparative results clearly verifies the effectiveness of the optimized rules. While the average risk of the captured scenarios remained relatively stable with only a slight increase of 3.04%, the peak risk value of the captured scenarios significantly improved from 2.148 to 2.843, an increase of 32.31%. This indicates that the optimized rules substantially enhance the ability to capture extreme risk scenarios. This improvement is further corroborated by the hazard distribution map, where the probability density curve of the scenarios extracted by the optimized rules shows a wider risk coverage. Notably, the probability density of scenarios extracted by the optimized rules systematically decreases in the medium-risk region, while a clear distribution of enhanced tail-risk regions is observed, achieving accurate capture of high-risk long-tail risks. All of these factors demonstrate that the optimized rules significantly enhance the system's sensitivity in recognizing the boundaries of overtaking scenarios.

[0135] As a preferred embodiment of this application, the difference from the above embodiments is that, in the process of covariance regularization, in addition to the regularized log-likelihood formula used in the above embodiments, other regularization methods, including but not limited to Ledoit-Wolf regularization, L1 regularization, and Fused Lasso regularization, can also be used for parameter estimation of the piecewise Gaussian model. The formulas for Ledoit-Wolf regularization, L1 regularization, and Fused Lasso regularization are as follows:

[0136] ;

[0137] ;

[0138] ;

[0139] in, It is the square of its Frobenius norm; For matrix Norm; is the Frobenius norm of the covariance of adjacent segments.

[0140] In this embodiment, optionally, during the risk model assessment in step eight, the total risk is obtained by calculating the superposition value of the interactive vehicle risk and the speed risk. Based on the magnitude, distribution characteristics, and trend of the total risk value, it is determined whether the extracted scene can effectively cover high-risk edge scenes, thus verifying the effectiveness of the scene extraction results.

[0141] Understandably, this application, by introducing a greedy Gaussian segmentation algorithm and modeling the segmentation process of time-series data based on a multivariate Gaussian distribution, effectively solves the problems of subjectivity and insufficient coverage of traditional rule-based methods, as well as the poor interpretability and high computational complexity of machine learning methods. This method significantly improves its adaptability to different datasets by dynamically optimizing segmentation breakpoints and regularizing covariance estimation. It can accurately capture the spatiotemporal features of vehicle interactions in complex traffic scenarios, and is particularly efficient in extracting high-risk edge scenarios missed by traditional rules, such as multi-objective conflict scenarios. Simultaneously, it uses the log-likelihood function curve to determine the optimal segmentation point, reducing the algorithm's dependence on data quality and manual annotation. By superimposing 3σ rules to optimize the scene frame range and verifying the hazard model, it achieves high-coverage and highly interpretable scene extraction, thus providing efficient support for building a more comprehensive and challenging autonomous driving test scenario library, helping to improve testing efficiency and the reliability of system safety assessment.

[0142] like Figure 8 As shown, in one embodiment, an autonomous driving acceleration test scenario extraction device is provided. This device can be integrated into the aforementioned computer device 120, and specifically may include:

[0143] The driving data acquisition module 510 is used to acquire a number of driving data, perform preliminary screening of the driving data based on a preset typical driving scenario, and obtain a time series of driving interaction data that conforms to the preset scenario.

[0144] The feature matrix acquisition module 520 is used to obtain a feature matrix of key driving interaction features based on the time series of the driving interaction data; the key driving interaction features include at least speed gradient features, acceleration gradient features, time headway features, and collision time features.

[0145] The optimal segmentation scheme acquisition module 530 is used to input the feature matrix into the greedy Gaussian segmentation model for segmentation and solve the optimal segmentation scheme based on the total likelihood function estimation criterion.

[0146] The optimal segment acquisition module 540 is used to preset the number of segments for the greedy Gaussian segmentation model, iterate the number of segments, obtain the log-likelihood value during the iteration process, obtain the optimal number of segments corresponding to the abrupt inflection point of the change curve of the log-likelihood value, and obtain the position of the segmentation point and the optimal segment based on the optimal number of segments.

[0147] In the embodiments of this application, the explanation and description of the above-mentioned autonomous driving acceleration test scene extraction device can be referred to the explanation and description of the corresponding method. For the description of the autonomous driving acceleration test scene extraction method, please refer to the above text, which will not be repeated here.

[0148] Figure 9 An internal structural diagram of a computer device in one embodiment is shown. Specifically, this computer device may be... Figure 1 Computer equipment 120. (e.g.) Figure 9 As shown, the computer device includes a processor, memory, network interface, input device, and display screen connected via a system bus. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a method for extracting autonomous driving acceleration test scenarios. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the autonomous driving acceleration test scenario extraction method.

[0149] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0150] In one embodiment, the autonomous driving acceleration test scenario extraction device provided in this application can be implemented as a computer program, which can be implemented in the form of, for example... Figure 9 The device shown operates on this device. The device's memory can store the various program modules that make up the autonomous driving acceleration test scenario extraction device.

[0151] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, causes the processor to perform the steps of the autonomous driving acceleration test scenario extraction method described above.

[0152] In the embodiments of this application, please refer to the above description of the method for extracting autonomous driving acceleration test scenarios, which will not be repeated here.

[0153] In one embodiment, an autonomous driving acceleration test scenario extraction system is provided. The system includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the autonomous driving acceleration test scenario extraction method as described above.

[0154] In this embodiment, the system can be a computer hardware system that executes its corresponding method when the system is running. For a description of the above-mentioned method for extracting autonomous driving acceleration test scenarios, please refer to the preceding text; it will not be repeated here.

[0155] It should be understood that although the steps in the flowcharts of the various embodiments of this application are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0156] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Furthermore, any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for extracting autonomous driving acceleration test scenarios, characterized in that, The extraction method includes: Acquire a number of driving data points, and perform preliminary screening of the driving data based on preset typical driving scenarios to obtain a time series of driving interaction data that conforms to the preset scenarios. Based on the time series of the driving interaction data, a feature matrix of key driving interaction features is obtained; the key driving interaction features include at least speed gradient features, acceleration gradient features, time headway features, and collision time features. The feature matrix is ​​input into a greedy Gaussian segmentation model for segmentation, and the optimal segmentation scheme is solved based on the total likelihood function estimation criterion. A preset segmentation number value is set for the greedy Gaussian segmentation model, and the segmentation number value is iterated to obtain the log-likelihood value during the iteration process. The optimal segmentation number is obtained corresponding to the abrupt inflection point of the log-likelihood value change curve. Based on the optimal segmentation number, the position of the segmentation point and the optimal segmentation sub-segment are obtained. The feature matrix is ​​input into a greedy Gaussian segmentation model for segmentation based on the following formula: ； ； in: This is the characteristic matrix, which has a total of There are n timestamps, each with a dimension of n; Let t be the t-th element in the matrix; N represents a multivariate Gaussian distribution; The segmentation points of the feature matrix are: It is divided into 100 parts. The segment, the dividing point satisfies: ； This represents the mean. Describing covariance, Represents timestamp, mean and covariance The value is only at the split point Changes only occur at the branching points: and These represent the upper and lower limits of the values ​​that the dividing point can take; ； superscript in the above formula Indicates the first Each segment This represents the mean value corresponding to each segment. The covariance of each segment; The extraction method further includes: Calculate the mean and variance of multiple autonomous driving interaction sub-time series elements to obtain the mean information and variance information of multiple autonomous driving interaction sub-time series elements. Based on the mean information and variance information of the multiple autonomous driving interaction sub-time series elements, multiple distribution intervals of autonomous driving interaction sub-time series elements are generated; each distribution interval of autonomous driving interaction sub-time series elements includes the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval; the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval correspond one-to-one with the autonomous driving interaction sub-time series information. When the minimum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is greater than the left endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, and the maximum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is less than the right endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, then the autonomous driving interaction sub-time sequence information is used as autonomous driving scene extraction information.

2. The method for extracting autonomous driving acceleration test scenarios according to claim 1, characterized in that, Before the initial screening of the driving data, the process includes preprocessing the driving data. The preprocessing method is as follows: Outliers in the driving data are removed to obtain the data to be denoised; The data to be denoised is then subjected to denoising and smoothing processing based on the following formula: ； in, For vehicles The original data at sampling time k For the smoothed data, for Total number of sample points Represents the time span of the smoothing operation. Representing the time interval between each frame, Z, i, D, and All are parameters.

3. The method for extracting autonomous driving acceleration test scenarios according to claim 1, characterized in that, The total likelihood function is: ； ； in, Let be the objective function. For a sub-item in the objective function, All of these are the optimal segmentation point positions.

4. The method for extracting autonomous driving acceleration test scenarios according to claim 3, characterized in that, The extraction method further includes: During the iterative process of calculating the number of segments, the covariance is constrained from overfitting based on the following formula: ； in, For regularization terms, This is the optimized objective function.

5. A method for evaluating autonomous driving acceleration test scenarios, characterized in that, The evaluation method includes: A real-time driving risk assessment method based on artificial potential field theory is constructed as a hazard model, wherein the hazard model is as follows: ； in, For the risk level model, and These represent the risks associated with interactive vehicles and the risks associated with speed, respectively. , These are the weighting coefficients, The decay parameter for the exponential term; in the interactive risk term middle, The introduced virtual energy characterizes the contribution of vehicle size and speed to risk. and These are the vehicle's length and width, respectively. The angle between the directions of travel of the two vehicles; Let be the relative distance vector between the two vehicles. Its Euclidean distance; in the speed risk item middle, The target speed of the main vehicle, The actual speed of the main vehicle; Based on the aforementioned hazard model, an evaluation result is obtained for the autonomous driving acceleration test scenario extraction method as described in any one of claims 1-4.

6. An autonomous driving acceleration test scenario extraction device, characterized in that, The device includes: The driving data acquisition module is used to acquire a number of driving data, and to perform preliminary screening of the driving data based on a preset typical driving scenario to obtain a time series of driving interaction data that conforms to the preset scenario. The feature matrix acquisition module is used to obtain a feature matrix of key driving interaction features based on the time series of the driving interaction data; the key driving interaction features include at least speed gradient features, acceleration gradient features, time headway features, and collision time features. The optimal segmentation scheme acquisition module is used to input the feature matrix into the greedy Gaussian segmentation model for segmentation and solve the optimal segmentation scheme based on the total likelihood function estimation criterion. The optimal segment acquisition module is used to preset the number of segments for the greedy Gaussian segmentation model, iterate the number of segments, obtain the log-likelihood value during the iteration process, obtain the optimal number of segments corresponding to the abrupt inflection point of the change curve of the log-likelihood value, and obtain the position of the segmentation point and the optimal segment based on the optimal number of segments. The feature matrix is ​​input into a greedy Gaussian segmentation model for segmentation based on the following formula: ； ； in: This is the characteristic matrix, which has a total of There are n timestamps, each with a dimension of n; Let t be the t-th element in the matrix; N represents a multivariate Gaussian distribution; The segmentation points of the feature matrix are: It is divided into 100 parts. The segment, the dividing point satisfies: ； This represents the mean. Describing covariance, Represents timestamp, mean and covariance The value is only at the split point Changes only occur at the branching points: and These represent the upper and lower limits of the values ​​that the dividing point can take; ； superscript in the above formula Indicates the first Each segment This represents the mean value corresponding to each segment. The covariance of each segment; Extraction methods also include: Calculate the mean and variance of multiple autonomous driving interaction sub-time series elements to obtain the mean information and variance information of multiple autonomous driving interaction sub-time series elements. Based on the mean information and variance information of the multiple autonomous driving interaction sub-time series elements, multiple distribution intervals of autonomous driving interaction sub-time series elements are generated; each distribution interval of autonomous driving interaction sub-time series elements includes the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval; the left endpoint information and the right endpoint information of the autonomous driving interaction sub-time series element distribution interval correspond one-to-one with the autonomous driving interaction sub-time series information. When the minimum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is greater than the left endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, and the maximum value of multiple autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information is less than the right endpoint information of the distribution interval of the autonomous driving interaction sub-time sequence element information corresponding to the autonomous driving interaction sub-time sequence information, then the autonomous driving interaction sub-time sequence information is used as autonomous driving scene extraction information.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the steps of the autonomous driving acceleration test scenario extraction method as described in any one of claims 1 to 4.

8. A system for extracting autonomous driving acceleration test scenarios, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the autonomous driving acceleration test scenario extraction method as described in any one of claims 1 to 4.

Images