Unknown key scene identification method and system based on intersection scene space

Abstract

The invention provides an unknown key scene identification method and system based on an intersection scene space, and relates to the technical field of intelligent traffic, the method comprises the following steps: constructing a road network model based on topological structures of different types of intersection scenes, the road network model comprising the intersection scene space; scene sampling is carried out in the intersection scene space to obtain a plurality of known key scenes, and the known key scenes comprise at least one known risk factor; a plurality of key scene subspaces are constructed, a first center is determined, and each key scene subspace comprises at least one known key scene; and calculating an orthogonal basis of the key scene subspace, taking the first center as a center, extracting an unknown key scene space in the key scene subspace along the direction of the orthogonal basis, and sampling in the unknown key scene space to obtain an unknown key test scene set. The unknown key scene identified by the scheme is used in the test of the automatic driving technology, and the test comprehensiveness can be improved.

Description

Technical Field

[0001] The present invention generally relates to the field of intelligent transportation technology, and in particular to an unknown key scene recognition method and system based on intersection scene space. Background Art

[0002] In the modern process of developing autonomous vehicles, simulation testing plays an important role. Taking the typical V-shaped vehicle development process as an example, the development of various intelligent technologies of autonomous vehicles often requires HIL testing (Chinese name: Hardware-in-the-Loop, English full name: Hardware-in-the-Loop), and this test must be carried out in an autonomous driving test scenario.

[0003] The contents of autonomous driving test scenarios are complex and can be divided into simple scenarios and complex scenarios according to the complexity of their working conditions. Simple scenarios are usually some simple working condition scenarios such as straight driving and highways. Due to their low complexity, the probability of autonomous driving system errors is low and the number of unknown key scenarios is also small. However, for complex scenarios, their complexity is high. In order to improve the effectiveness of autonomous driving system test scenarios, it is necessary to identify and transform unknown key scenarios for complex scenarios.

[0004] In the prior art, the patent with application number CN202110126331.X proposes a method, device and platform for generating an autonomous driving simulation test scenario library. Although it can overcome the limitations of autonomous driving simulation test scenarios, low precision, and difficulty in data processing, it does not analyze and identify unknown key scenarios that may exist in the test scenarios, which will cause the problem of incomplete testing and fail to meet the international standard ISO 34502 for autonomous driving system test scenarios, indirectly leading to the low robustness of the autonomous driving system being tested, and the test results cannot provide accurate guidance for actual production. Summary of the invention

[0005] In view of the above-mentioned defects or deficiencies in the prior art, the present invention aims to provide a method and system for identifying unknown key scenes based on intersection scene space to improve the comprehensiveness of the test.

[0006] A first aspect of the present invention provides an unknown key scene recognition method based on an intersection scene space, comprising the following steps:

[0007] Based on the topological structure of different types of intersection scenes, a road network model is constructed; the road network model includes an intersection scene space;

[0008] Performing scene sampling in the intersection scene space to obtain a plurality of known critical scenes and known non-critical scenes; the known critical scenes include at least one of the known risk factors, and the known non-critical scenes do not include the known risk factors;

[0009] Constructing a plurality of key scene subspaces and determining a first center; each of the key scene subspaces includes at least one known key scene, and the first center is the center of the key scene subspace;

[0010] Calculate an orthogonal basis of the key scene subspace, and extract an unknown key scene space in the key scene subspace along the direction of the orthogonal basis with the first center as the center;

[0011] Unknown key scenes are sampled in the unknown key scene space to obtain a set of unknown key test scenes.

[0012] According to the technical solution provided by the present invention, a road network model is constructed based on the topological structure of different types of intersection scenes, which includes at least the following steps:

[0013] Constructing a first model based on different types of intersection scenes; the types of the intersection scenes at least include T-junction scenes, crossroad scenes and other intersection scenes, and the other intersections are intersections where the roads are not perpendicular to each other;

[0014] Perform path planning modeling based on the first model to obtain a second model; the second model includes a trajectory of vehicle path planning;

[0015] A traffic light scenario and a microscopic traffic flow scenario are added to the second model to obtain the road network model.

[0016] According to the technical solution provided by the present invention, sampling is performed in the intersection scene space to obtain a plurality of known key scenes and known non-key scenes, which at least includes the following steps:

[0017] Acquire all the known risk factors in the intersection scene space; each known risk factor is not limited to one;

[0018] Based on at least one of the known risk factors, a plurality of the known key scenarios are randomly constructed until all of the known risk factors participate in constituting the known key scenarios;

[0019] The scenarios that do not include the known risk factors are determined to be the known non-critical scenarios.

[0020] According to the technical solution provided by the present invention, constructing several key scene subspaces includes at least the following steps:

[0021] Performing feature extraction on the known risk factors to obtain a plurality of risk factor features; each of the known risk factors includes at least one of the risk factor features;

[0022] Based on a decision tree algorithm, a plurality of risk factor characteristics are used as judgment conditions to construct a plurality of key scene subspaces in the intersection scene space.

[0023] According to the technical solution provided by the present invention, determining the first center includes at least the following steps:

[0024] Acquire the dimension of the key scene subspace, where the dimension of the key scene subspace is the same as that of the intersection scene space;

[0025] Determine a value range of the key scene subspace in one dimension, and calculate a first coordinate of the first center in the dimension; the first coordinate is half of the sum of the maximum and minimum values ​​of the dimension;

[0026] Traverse other dimensions and calculate the first coordinate of the first center in each dimension.

[0027] According to the technical solution provided by the present invention, extracting unknown key scenes in the key scene subspace along the direction of the orthogonal basis comprises at least the following steps:

[0028] The unknown key scene is extracted by extending along the positive direction of the orthogonal basis and the negative direction of the orthogonal basis respectively.

[0029] According to the technical solution provided by the present invention, sampling unknown key scenes in the unknown key scene space to obtain a set of unknown key test scenes includes at least the following steps:

[0030] Setting a sampling step length; the sampling step length is any length less than a first distance, where the first distance is the length between two vertices closest to each other in the area formed by the unknown key scene space;

[0031] Randomly select a sampling starting point, and perform sampling once every sampling step along one direction of the unknown key scene space to obtain an unknown key scene; the direction of the unknown key scene space is the positive direction of the orthogonal basis and the negative direction of the orthogonal basis;

[0032] Traverse all directions of the unknown key scenes to obtain a plurality of the unknown key scenes; and construct the unknown key test scene set according to the plurality of the unknown key scenes.

[0033] According to the technical solution provided by the present invention, after constructing a road network model based on the topological structure of different types of intersection scenes, before performing scene sampling in the intersection scene space to obtain a number of known key scenes and known non-key scenes, the following steps are also included:

[0034] The road network model is numerically processed based on vehicle mechanical parameters, physical constraints and traffic rules.

[0035] According to the technical solution provided by the present invention, the known risk factors include at least: following vehicle risk factors, lane changing risk factors, turning risk factors and roadblock risk factors.

[0036] A second aspect of the present invention provides an unknown key scene recognition system based on an intersection scene space, comprising:

[0037] A model building module, wherein the model building module is configured to build a road network model based on the topological structure of different types of intersection scenes; the road network model includes an intersection scene space;

[0038] A first sampling module, wherein the scene sampling module is configured to perform scene sampling in the intersection scene space to obtain a plurality of known critical scenes and known non-critical scenes; the known critical scenes include at least one of the known risk factors, and the known non-critical scenes do not include the known risk factors;

[0039] A subspace construction module, wherein the subspace construction module is configured to construct a plurality of key scene subspaces and determine a first center; each of the key scene subspaces includes at least one known key scene, and the first center is the center of the key scene subspace;

[0040] A scene generation module, the scene generation module is configured to calculate an orthogonal basis of the key scene subspace, and extract an unknown key scene space in the key scene subspace along the direction of the orthogonal basis with the first center as the center;

[0041] A second sampling module is configured to sample unknown key scenes in the unknown key scene space to obtain a set of unknown key test scenes.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention targets typical and common complex scenes such as intersections, identifies known key scenes based on known risk factors, constructs a key scene subspace based on the identified known key scenes, constructs an unknown key scene space through the key scene subspace, and then collects unknown key scenes in the unknown key scene space, and uses the collected unknown key scenes for the testing of autonomous driving technology; the present invention obtains unknown key scenes through known key scenes, and tests the known key scenes and unknown key scenes together as test scenes, thereby improving the comprehensiveness of the test, meeting the international standard ISO 34502 for autonomous driving system test scenes, and thereby improving the robustness of the tested autonomous driving system, and the test results can provide accurate guidance for actual production. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of the steps of the unknown key scene recognition method based on the intersection scene space provided by an embodiment of the present invention;

[0044] Figure 2 It is a schematic diagram of the spatial topology of the intersection scene in the present invention;

[0045] Figure 3 A schematic diagram of the structure of an unknown key scene space provided by an embodiment of the present invention;

[0046] Figure 4 A schematic diagram of the structure of a decision tree algorithm provided by an embodiment of the present invention;

[0047] Figure 5 An inclusion relationship diagram among the intersection scene space, key scene subspace, and known key scenes provided in an embodiment of the present invention.

[0048] The text annotations in the figure represent:

[0049] 1. Intersection scene space; 2. Known non-critical scene space; 3. Critical scene subspace; 4. Known critical scene. DETAILED DESCRIPTION

[0050] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0051] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0052] Example 1

[0053] Please refer to Figure 1 This embodiment provides an unknown key scene recognition method based on intersection scene space, please refer to Figure 1 As shown, the following steps are included:

[0054] S100, constructing a road network model based on the topological structure of different types of intersection scenes; the road network model includes an intersection scene space;

[0055] S200, performing scene sampling in the intersection scene space to obtain a plurality of known critical scenes and known non-critical scenes; the known critical scenes include at least one of the known risk factors, and the known non-critical scenes do not include the known risk factors;

[0056] S300, constructing a plurality of key scene subspaces and determining a first center; each of the key scene subspaces includes at least one known key scene, and the first center is the center of the key scene subspace;

[0057] S400, calculating an orthogonal basis of the key scene subspace, and extracting an unknown key scene space in the key scene subspace along the direction of the orthogonal basis with the first center as the center;

[0058] S500: Sampling unknown key scenes in the unknown key scene space to obtain a set of unknown key test scenes.

[0059] In an optional embodiment, constructing a road network model based on the topological structure of different types of intersection scenes includes at least the following steps:

[0060] S110, constructing a first model based on different types of intersection scenes; the types of the intersection scenes at least include T-junction scenes, crossroad scenes and other intersection scenes, and the other intersections are intersections where the roads are not perpendicular to each other;

[0061] Specifically, constructing the first model based on different types of intersection scenarios includes three steps:

[0062] Step 1: Road network modeling based on T-junction topology structure;

[0063] A T-junction is a typical intersection scene, mainly composed of two perpendicular roads. In this road scene, the road network structure presents a three-way "T" shape, and the road network structure mainly includes the following elements: entrance road straight lane, entrance road left turn lane, entrance road right turn lane, entrance road straight left mixed lane, entrance road straight right mixed lane, widening gradient section, corner island, sidewalk, bicycle lane and exit road. When establishing a T-junction road network, first set the number of lanes of the entrance and exit roads, set the parameters of each lane of the entrance and exit roads according to the provisions of the standard at the T-junction, and then add the widening gradient section, corner island, sidewalk and bicycle lane to form a complete T-junction road network.

[0064] Step 2: Road network modeling based on intersection topology;

[0065] Crossroads are the most common and typical intersection scenes in life. They also have many similarities with the T-junctions described in Section 1. Their road network structure also mainly includes the following elements: entrance road straight lanes, entrance road left turn lanes, entrance road right turn lanes, entrance road straight left mixed lanes, entrance road straight right mixed lanes, widening gradient sections, corner islands, sidewalks, bicycle lanes, and exit roads. When establishing a crossroads network, the number of lanes of the entrance and exit roads is also set first, and the parameters of each lane of the entrance and exit roads are set according to the provisions of the standard. Then, widening gradient sections, corner islands, sidewalks, and bicycle lanes are added to form a complete crossroads network.

[0066] Step 3: Road network modeling based on other intersection topology structures;

[0067] The two most common intersection scenarios in life are described in the first and second parts. Both of these scenarios have perpendicular relationships between roads, which makes the road network modeling of these two intersections relatively simple and clear. However, in actual driving scenarios, there are still some multi-fork intersections or three-way intersections and intersections where the roads are not perpendicular to each other. The road network structure of these scenarios is more complex, and the elements included are the same as the first two parts: entrance road straight lane, entrance road left turn lane, entrance road right turn lane, entrance road straight left mixed lane, entrance road straight right mixed lane, widening gradient section, corner island, sidewalk, bicycle lane and exit lane. However, when establishing other types of intersection road networks, the orientation of each lane should be determined according to the direction of the road, and then the entrance and exit lanes and their parameters should be set for each lane, and widening gradient sections, corner islands, sidewalks and bicycle lanes should be added to form a complete intersection road network.

[0068] After the above three steps are completed, the first model is obtained, which is an intersection scene model including a T-junction scene, a crossroad scene and other intersection scenes. The first model also includes at least a main vehicle and an environmental vehicle, and the environmental vehicle is a vehicle other than the main vehicle. The obtained intersection scene model is verified according to the requirements of GB50647, including but not limited to the sight triangle requirements, to ensure that the constructed intersection scene has a high practical value, and the construction of the entire intersection scene model is completed.

[0069] S120, performing path planning modeling based on the first model to obtain a second model; the second model includes a trajectory of vehicle path planning;

[0070] Specifically, in the intersection scene, assuming that all traffic vehicles strictly abide by traffic rules, the path of the main vehicle can be modeled according to its driving intention. The driving intention here needs to be manually input, such as the coordinates of the path starting point, path end point and waypoints. The path starting point is the entrance lane at the initial moment, and the end point is the target exit lane. It is judged whether it can reach the end point according to the lane type of the lane (entry straight lane / entry left turn lane / entry right turn lane / entry straight left mixed lane / entry straight right mixed lane). According to the judgment result, it is selected whether it is necessary to change lanes before entering the intersection. If it is necessary to change lanes or turn in advance, the path is planned based on the dynamic programming algorithm according to the real-time position of the surrounding vehicles. The straight track, possible lane change, possible turn and straight track after entering the exit lane are spliced ​​together to form the trajectory formed by the vehicle according to the driving intention. On the basis of the first model obtained in step S110, the trajectory of the vehicle's path planning is added to obtain the second model. The second model is the model after the first model in step S110 introduces the new dimension of the trajectory of the vehicle's path planning.

[0071] S130: Add a traffic light scene and a microscopic traffic flow scene to the second model to obtain the road network model.

[0072] Specifically, the specific content of adding the traffic light scene is as follows:

[0073] There is usually another important element in the intersection scene, the traffic light element. The traffic light is mainly used to regulate the traffic flow at the intersection. By lighting up the red, yellow or green lights at a certain time, vehicles in the specified direction are released to regulate the traffic flow. This can not only greatly improve the traffic efficiency of the intersection, but also greatly improve the safety of the intersection. Therefore, it has been widely used in most intersections. Furthermore, in this embodiment, the traffic light can be simplified into two parts: the initial color and the cycle time. First, the cycle time of the traffic light is specified (a cycle should include the process of the green light changing to the yellow light, then to the red light, and finally to the green light. The three processes are independently designed in time). Then, by setting the time step, the initial moment of the cycle is selected according to the traversal sampling method, so as to ensure that the scene has a high coverage. On the basis of the second model obtained in step S120, the traffic light scene is added to obtain the road network model. The road network model is the model after the second model in step S120 introduces the new dimension of the traffic light scene.

[0074] Specifically, the specific content of adding the microscopic traffic flow scene is as follows:

[0075] my country's intelligent vehicle autonomous driving technology is developing rapidly. According to the current industrialization process of my country's intelligent vehicles and the complexity of my country's traffic, it is inferred that the phenomenon of human-machine co-driving in urban structured roads will continue to exist for at least 10 years from 2020 to 2030, and this phenomenon will continue for more than 10 years in towns and villages. Therefore, in this embodiment, the intelligent agent node control of different types of environmental vehicles is modeled to form a complete micro traffic flow. The intelligent agent node control includes but is not limited to the following models:

[0076] Model 1: Driver model. A portion of the vehicles in the setting environment are traditional cars operated by real drivers. In the simulation environment, these vehicles are completed by setting the driver model, which is represented by the following driver model:

[0077]

[0078]

[0079] In the formula, s 0,i represents the minimum expected distance between the i-th vehicle and the i-1-th vehicle; v 0,i represents the maximum expected speed of the i-th vehicle; v i represents the speed of the i-th vehicle; δ represents the acceleration index, which controls the "smoothness" of the acceleration; T i represents the reaction time of the driver of the i-th vehicle; a i represents the maximum acceleration of the i-th vehicle; b irepresents the comfortable deceleration of the i-th vehicle; s* represents the actual expected distance between the i-th vehicle and the i-1-th vehicle.

[0080] Model 2: Autonomous driving control model. A portion of the vehicles in the environment are set as autonomous driving vehicles. In the simulation environment, these vehicles are completed by setting an autonomous driving control algorithm. This patent is represented by the following PID control:

[0081]

[0082] In the formula, k p represents proportional gain; k i represents the integral gain; k d represents differential gain; e(t) represents error, which is the deviation between the set value and the actual value; t represents time.

[0083] Model 3: Custom model. There are some special vehicles in the setting environment, including the main vehicle or the fleet, which are different from the first two types of environmental vehicles. The main vehicle refers to the self-vehicle; these vehicles play a decisive role in the scene and therefore need to be handled separately. For these special vehicles, it can be achieved by connecting to a driving simulator or the autonomous driving algorithm to be tested.

[0084] At this point, the road network model constructed in this embodiment can realistically simulate the traffic conditions of intersection scenarios and provide data support for identifying and extracting unknown key scenarios.

[0085] Furthermore, step S100 also includes the following steps:

[0086] After the road network model is obtained, the intersection scene of the road network model is parameterized and numerically modeled to determine the intersection scene space.

[0087] Specifically, in order to effectively extract unknown key scenes in subsequent processing, the established intersection scene is first parameterized so that it finally forms a high-dimensional scene space. The specific steps are: mainly by extracting the coordinates and headings of each vehicle in the scene, as well as the lane line curvature of the road environment and other scene data, and then parameterizing the prime scene data, and constructing the scene space through the parameterized scene data. The dimensional independent variables of the scene space are mainly the time of the scene and the position information of the main vehicle, and the dependent variables are mainly safety-related indicators related to the safety of the main vehicle, among which the safety-related indicators include at least the collision time TTC and the headway THW, etc. The constructed intersection scene space can provide direct and effective data support for the subsequent unknown key scene sampling process.

[0088] Optionally, TTC is the following formula: THW is as follows: In the formula, Δx represents the distance between the two vehicles (the rear vehicle and the front vehicle), and Represent the speed of the rear vehicle and the front vehicle respectively. It should be noted that the TTC and THW of the scene refer to the smallest value among the TTC and THW values ​​of multiple groups of vehicles in the scene.

[0089] Specifically, a critical scenario refers to an autonomous driving test scenario that contains one or more risk factors that, once occurring, will increase the probability of a hazard occurring, the severity of the hazard, or both. Once the risk factors exist, they will increase the probability of a hazard occurring and / or the severity of the hazard. Furthermore, a critical scenario can be defined as a scenario that includes one or more risk factors.

[0090] Specifically, in the intersection scenario of the present invention, risk factors can be divided into following risk factors, lane-changing risk factors, turning risk factors and roadblock risk factors; following risk factors refer to risk factors caused by improper distance control between the following vehicle and the preceding vehicle during the straight-ahead process; lane-changing risk factors refer to risk factors caused by improper position control between the vehicle and the vehicle in the adjacent lane during the process of changing lanes due to the vehicle being in a lane that does not match its driving intention; turning risk factors refer to potential dangerous factors between vehicles in different directions during the turning process; roadblock risk factors refer to risk factors that may occur due to physical interference between the vehicle and elements such as fences or green belts in the static road in the driving scenario. These risk factors can also be called known risk factors.

[0091] Known scenarios are divided into known critical scenarios and known non-critical scenarios. Known critical scenarios refer to scenarios that are included in the intersection scenario space and have been identified as having known risk factors; known non-critical scenarios refer to scenarios that are included in the intersection scenario space and have been identified as not having known risk factors. Furthermore, the risk factors present in the known critical scenarios are known risk factors, and each of the known critical scenarios includes at most one of each risk factor.

[0092] In an optional embodiment, scene sampling is performed in the intersection scene space to obtain a plurality of known key scenes and known non-key scenes, which at least includes the following steps:

[0093] S201, obtaining all the known risk factors in the intersection scene space; each known risk factor is not limited to one;

[0094] Specifically, for example, there are ten known risk factors in the intersection scene space, including one risk factor for following a vehicle, two risk factors for lane changing, three risk factors for turning, and four risk factors for roadblocks. In actual operation, the number of known risk factors is much greater than ten.

[0095] S202, based on at least one of the known risk factors, randomly constructing a plurality of the known key scenarios until all of the known risk factors participate in constituting the known key scenarios;

[0096] S203: Determine that a scenario that does not include the known risk factor is the known non-critical scenario.

[0097] Specifically, several scenes are collected in the intersection scene space, and an example is given here including M scenes. Then the risk factors obtained in step S201 are randomly distributed to the M scenes, each scene includes zero to four risk factors, and it is ensured that there is at most one of each risk factor in each scene; after the known risk factors are distributed, the several scenes are divided into multiple known critical scenes and multiple known non-critical scenes.

[0098] In an optional embodiment, constructing a plurality of key scene subspaces includes at least the following steps:

[0099] S301, extracting features from the known risk factors to obtain a plurality of risk factor features; each of the known risk factors includes at least one risk factor feature;

[0100] S302: Based on a decision tree algorithm, a plurality of risk factor characteristics are used as judgment conditions to construct a plurality of key scene subspaces in the intersection scene space.

[0101] Specifically, the known risk factors are extracted in the intersection scene space by using a machine learning method, which can be understood as all the known risk factors obtained in the above step S201, and then used for the derivation of unknown key scenes. The specific method is: the known key scenes obtained in step S202 are used as the input of the risk factor feature extraction model, and the risk factor feature extraction model outputs a number of risk factor features. Each of the known risk factors is composed of a plurality of risk factor features, and the risk factor feature extractor is a trained neural network model.

[0102] Furthermore, since supervised machine learning algorithms are highly dependent on training data, in order to avoid systematic deviations caused by poor data quality, the risk factor feature extraction model obtained above should be verified, and if systematic deviations occur, they should be manually adjusted. The verification method is described as follows:

[0103] First, some scenes are found in the bounded intersection scene space by uniform sampling, and then these scenes are input into the risk factor feature extractor to obtain two types of classification results: critical scenes and non-critical scenes. The classification results are manually verified, and the verification results are recorded through a confusion matrix. Finally, according to the verification results, if it is verified that the risk factor feature extraction model has no systematic deviation or the deviation is extremely small, there is no need to adjust it. If some key scene subspaces in the risk factor feature extraction model have certain systematic deviations, then the classification boundaries of these key scene subspaces are retrained and adjusted according to the sampling results, so as to obtain an effective risk factor feature extraction model. Through the verified risk factor feature extraction model, the intersection scene space can be divided into two categories: critical scenes and non-critical scenes, thereby realizing the derivation of key scenes based on risk factors.

[0104] Specifically, the present invention divides the key scene subspace by taking the decision tree algorithm as an example. In the actual implementation process, other classification algorithms can be used according to the situation. The decision tree algorithm is a greedy classification algorithm that solves the classification problem by repeatedly searching for the classification boundary with the largest information entropy among all classification types. Its information entropy can be defined as the following formula:

[0105]

[0106] In the formula, p k It represents the proportion of samples of the kth category in the set D, k = 1, 2, 3, …, |y|.

[0107] Through multiple binary classifications based on information entropy minimization, the decision tree can eventually divide the intersection scene space into several mutually exclusive subspaces. The key scene subspace can be confirmed according to the classification conditions of the decision tree. The classification conditions on the decision tree are the risk factor feature representations of the risk factors corresponding to the subspace in the intersection scene space, such as Figure 4 An example of a critical scenario subspace based on a risk factor classification decision tree and its corresponding scenario space schematic diagram is shown, where 0 or 1 represents a critical scenario or a non-critical scenario, Y indicates that the condition is met, and N indicates that the condition is not met, and the classification condition includes at least the risk factor characteristics.

[0108] In an optional embodiment, determining the first center includes at least the following steps:

[0109] S303, obtaining the dimension of the key scene subspace 3, where the dimension of the key scene subspace is the same as the dimension of the intersection scene space;

[0110] Specifically, the dimension of the intersection scene space is the number of types of parameters added in constructing the road network model. Taking the present invention as an example, the present invention adds three parameters, namely, the trajectory of vehicle path planning, the traffic light scene and the microscopic traffic flow scene, in constructing the road network model. Therefore, the dimension of the intersection scene space is three-dimensional. Figure 2 As shown, the directions of the three dimensions are only for illustration, and the three dimensions can also be considered as three non-parallel and intersecting straight lines in the same plane.

[0111] S304, determining the value range of the key scene subspace 3 in one dimension, and calculating the first coordinate of the first center in the dimension; the first coordinate is half of the sum of the maximum and minimum values ​​of the dimension;

[0112] S305 , traverse other dimensions and calculate the first coordinate of the first center in each dimension.

[0113] Specifically, the unknown key scene space is constructed at the center of each key scene subspace. If the intersection scene space is an m-dimensional space (n is less than m), the first coordinate of the first center in the nth dimension can be understood as: if the value range of the nth dimension direction of the key scene subspace is [a n ,b n ](or [a n ,b n ) / (a n ,b n ] / (a n ,b n )), then the first coordinate of the first center in this dimension is (a n +b n ) / 2, and the first coordinate of the first center in each other dimension is calculated in the same way, thereby determining the first center. Taking three dimensions as an example, which can also be considered as three non-parallel and intersecting straight lines in the same plane, the value range of the key scene subspace in a certain dimension can be understood as the distance of one end on the straight line corresponding to the dimension, and the threshold of the value range can be understood as the point on the straight line corresponding to the dimension.

[0114] In an optional embodiment, extracting unknown key scenes in the key scene subspace 3 along the direction of the orthogonal basis comprises at least the following steps:

[0115] The unknown key scene is extracted by extending along the positive direction of the orthogonal basis and the negative direction of the orthogonal basis respectively.

[0116] Specifically, if the dimension of the intersection scene space is m, then in the key scene subspace, there are at most m mutually linearly independent orthogonal bases. Orthogonal bases have many special properties. One important property is that their directions are perpendicular to each other, which makes it possible to quickly derive sampling in the scene space according to their directions with the minimum sampling amount.

[0117] Therefore, extracting the unknown key scene space according to the direction of the orthogonal basis can minimize the number of unknown key scenes in the unknown key test scene set while ensuring its efficient coverage of all directions of the key scene subspace.

[0118] As can be seen from the above, an important step in this embodiment is to extract a set of m mutually linearly independent orthogonal bases in the m-dimensional key scene subspace. The specific steps are as follows:

[0119] First, a unit vector x1 is randomly generated in the key scene subspace, and its direction is represented as a column vector x1 = (a1, b1, ..., m1) T According to the linear independence of orthogonal bases, if x2=(a2,b2,...,m2) T , then according to and We get a quadratic equation system with m variables. When m>2, the equation system has countless solutions. We can get x2 by picking any one of the countless solutions. Then let x3=(a3,b3,...,m3) T , then according to and We get a quadratic equation system with m variables. When m>3, the equation system has countless solutions. We can get x3 by taking any one of the countless solutions. And so on, until we make the last x m =(a m ,b m ,...,m m ) T , at this time by and Construct a quadratic equation system with two unique solutions. The two solutions obtained by solving the equations are in opposite directions. Similarly, any one of them can get the final x m , so far, {x1, x2, …, x m} is a set of required orthogonal bases. By extending along the positive direction of this set of orthogonal bases and the negative direction of the orthogonal bases, the unknown key scene space can be extracted in the key scene subspace, such as Figure 3 The polygon shown can be regarded as the unknown key scene space. The polygon includes multiple vertices, and each line segment in the figure is an orthogonal basis.

[0120] Specifically, the relationship between the intersection scene space, the known key scene subspace and the known key scene is as follows: Figure 5 As shown, the largest circular range in the figure represents the intersection scene space 1, the points in the figure represent the known key scenes 4, the three circled areas in the figure represent the key scene subspaces 3, each of the key scene subspaces 3 includes multiple known key scenes 4, and the area in the intersection scene space 1 excluding the key scene subspaces 3 represents a non-critical scene space including multiple known non-critical scenes.

[0121] In an optional embodiment, sampling unknown key scenes in the unknown key scenes to obtain a set of unknown key test scenes includes at least the following steps:

[0122] Setting a sampling step length; the sampling step length is any length less than a first distance, where the first distance is the length between two vertices closest to each other in the area formed by the unknown key scene space;

[0123] Randomly select a sampling starting point, and perform sampling once every sampling step along one direction of the unknown key scene space to obtain an unknown key scene; the direction of the unknown key scene space is the positive direction of the orthogonal basis and the negative direction of the orthogonal basis;

[0124] Traverse all directions of the unknown key scenes to obtain a plurality of the unknown key scenes; and construct the unknown key test scene set according to the plurality of the unknown key scenes.

[0125] Specifically, a sampling starting point is randomly selected in the unknown key scene space, and the unknown key scene sampling space can be obtained along different directions of the unknown key scene. Sampling is performed once every sampling step in the unknown key scene sampling space until the scene boundary is reached. When all directions of all unknown key scenes are traversed, the scene set obtained is the unknown key test scene set obtained by sampling.

[0126] In an optional embodiment, after constructing a road network model based on the topological structure of different types of intersection scenes, before performing scene sampling in the intersection scene space to obtain a plurality of known key scenes 4 and known non-key scenes, the following steps are also included:

[0127] The road network model is numerically processed based on vehicle mechanical parameters, physical constraints and traffic rules.

[0128] Specifically, numerical boundary processing refers to limiting the originally unbounded road network model to a bounded space according to various constraints existing in the intersection scene, such as vehicle mechanical parameters, physical constraints and traffic rules. The space within the numerical boundary of the road network model is the road network model after numerical boundary processing. The purpose of this is to exclude invalid parts in the scene space, facilitate subsequent implementation processes (such as sampling), greatly reduce the amount of calculation, and thus improve the practicality of this method. The specific method of numerical boundary processing here is the prior art. For example, the vehicle mechanical parameter is the minimum turning radius of the vehicle, and the physical constraint is the current speed of the vehicle.

[0129] Example 2

[0130] Based on Example 1, this example proposes how to use the unknown key scenes identified in Example 1 in the test of the autonomous driving scene. First, the unknown key scenes obtained in Example 1 should be integrated and processed. The specific steps are as follows:

[0131] Step 1: Verify safety assessment objectives. The safety test objectives represent the safety performance of the autonomous driving system as shown by a set of tests. Specifically, the safety test objectives refer to the acceptable and verifiable incidence boundary values ​​of safety-related quantifiable behaviors during the operation of the autonomous driving system (or its components) within the operating range. Safety assessment objectives usually come from general risk acceptance criteria in various standards, including but not limited to "as low as reasonably possible", "lowest endogenous mortality rate" and "positive risk balance". During the implementation of this step, these safety assessment objectives are quantitatively expressed. For example, in a given scenario, the relative frequency of unexpected behaviors is lower than x. The significance of this step is that the unknown key scenarios can be directly used for the testing of the autonomous driving system after identification, which has high practical value.

[0132] Step 2: Integration of key scenario test cases. In order to prevent test failures caused by the randomness of the experiment, this step is added when integrating key scenario test cases. According to the actual test situation, a set of unknown key scenarios obtained by sampling in the past is combined with another set or multiple sets (usually two sets are completely sufficient) of different unknown key scenarios, and the number of test scenarios is appropriately expanded, thereby greatly improving the effectiveness of the test. The integrated set of unknown key test scenarios with security assessment objectives is the final required set of test cases containing unknown key scenarios.

[0133] Example 3

[0134] This embodiment provides an unknown key scene recognition system based on intersection scene space, including:

[0135] A model building module, wherein the model building module is configured to build a road network model based on the topological structure of different types of intersection scenes; the road network model includes an intersection scene space;

[0136] A first sampling module, wherein the scene sampling module is configured to perform scene sampling in the intersection scene space to obtain a plurality of known critical scenes and known non-critical scenes; the known critical scenes include at least one of the known risk factors, and the known non-critical scenes do not include the known risk factors;

[0137] A subspace construction module, wherein the subspace construction module is configured to construct a plurality of key scene subspaces and determine a first center; each of the key scene subspaces includes at least one known key scene, and the first center is the center of the key scene subspace;

[0138] A scene generation module, wherein the scene generation module is configured to calculate an orthogonal basis of the key scene subspace, and extract unknown key scenes in the key scene subspace along the direction of the orthogonal basis with the first center as the center;

[0139] A second sampling module is configured to sample unknown key scenarios in the unknown key scenarios to obtain a set of unknown key test scenarios.

[0140] The above description is only a preferred embodiment of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present invention is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present invention (but not limited to) to form a technical solution.

Claims

1. A method for identifying unknown key scenes based on intersection scene space, characterized in that: The steps include: Based on the topological structure of different types of intersection scenes, a road network model is constructed; the road network model includes an intersection scene space; Performing scene sampling in the intersection scene space to obtain a plurality of known critical scenes and known non-critical scenes; the known critical scenes include at least one known risk factor, and the known non-critical scenes do not include the known risk factor; Constructing a plurality of key scene subspaces and determining a first center; each of the key scene subspaces includes at least one known key scene, and the first center is the center of the key scene subspace; Calculate an orthogonal basis of the key scene subspace, and extract an unknown key scene space in the key scene subspace along the direction of the orthogonal basis with the first center as the center; Unknown key scenes are sampled in the unknown key scene space to obtain a set of unknown key test scenes.

2. The unknown key scene recognition method based on intersection scene space according to claim 1 is characterized in that: Based on the topological structure of different types of intersection scenarios, a road network model is constructed, which includes at least the following steps: Constructing a first model based on different types of intersection scenes; the types of the intersection scenes at least include T-junction scenes, crossroad scenes and other intersection scenes, and the other intersections are intersections where the roads are not perpendicular to each other; Perform path planning modeling based on the first model to obtain a second model; the second model includes a trajectory of vehicle path planning; A traffic light scenario and a microscopic traffic flow scenario are added to the second model to obtain the road network model.

3. The unknown key scene recognition method based on intersection scene space according to claim 2 is characterized in that: Sampling is performed in the intersection scene space to obtain a plurality of known key scenes and known non-key scenes, which at least includes the following steps: Acquire all the known risk factors in the intersection scene space; each known risk factor is not limited to one; Based on at least one of the known risk factors, a plurality of the known key scenarios are randomly constructed until all of the known risk factors participate in constituting the known key scenarios; The scenarios that do not include the known risk factors are determined to be the known non-critical scenarios.

4. The unknown key scene recognition method based on intersection scene space according to claim 3 is characterized in that: Constructing several key scene subspaces includes at least the following steps: Performing feature extraction on the known risk factors to obtain a plurality of risk factor features; each of the known risk factors includes at least one of the risk factor features; Based on a decision tree algorithm, a plurality of risk factor characteristics are used as judgment conditions to construct a plurality of key scene subspaces in the intersection scene space.

5. The unknown key scene recognition method based on intersection scene space according to any one of claims 1 to 4, characterized in that: Determining the first center includes at least the following steps: Acquire the dimension of the key scene subspace, where the dimension of the key scene subspace is the same as that of the intersection scene space; Determine a value range of the key scene subspace in one dimension, and calculate a first coordinate of the first center in the dimension; the first coordinate is half of the sum of the maximum and minimum values ​​of the dimension; Traverse other dimensions and calculate the first coordinate of the first center in each dimension.

6. The unknown key scene recognition method based on intersection scene space according to any one of claims 1 to 4, characterized in that: Extracting unknown key scenes in the key scene subspace along the direction of the orthogonal basis comprises at least the following steps: The unknown key scene is extracted by extending along the positive direction of the orthogonal basis and the negative direction of the orthogonal basis respectively.

7. The unknown key scene recognition method based on intersection scene space according to any one of claims 1 to 4 is characterized in that: Sampling unknown key scenes in the unknown key scene space to obtain a set of unknown key test scenes includes at least the following steps: Setting a sampling step length; the sampling step length is any length less than a first distance, where the first distance is the length between two vertices closest to each other in the area formed by the unknown key scene space; Randomly select a sampling starting point, and perform sampling once every sampling step along one direction of the unknown key scene space to obtain an unknown key scene; the direction of the unknown key scene space is the positive direction of the orthogonal basis and the negative direction of the orthogonal basis; Traverse all directions of the unknown key scenes to obtain a plurality of the unknown key scenes; and construct the unknown key test scene set according to the plurality of the unknown key scenes.

8. The unknown key scene recognition method based on intersection scene space according to claim 1 is characterized in that: After constructing the road network model based on the topological structure of different types of intersection scenes, before performing scene sampling in the intersection scene space to obtain a number of known key scenes and known non-key scenes, the following steps are also included: The road network model is numerically processed based on vehicle mechanical parameters, physical constraints and traffic rules.

9. The unknown key scene recognition method based on intersection scene space according to claim 1 is characterized in that: The known risk factors include at least: following vehicle risk factors, lane changing risk factors, turning risk factors and roadblock risk factors.

10. An unknown key scene recognition system based on intersection scene space, characterized in that: include: A model building module, wherein the model building module is configured to build a road network model based on the topological structure of different types of intersection scenarios; The road network model includes an intersection scene space; A first sampling module, wherein the scene sampling module is configured to perform scene sampling in the intersection scene space to obtain a plurality of known critical scenes and known non-critical scenes; the known critical scenes include at least one known risk factor, and the known non-critical scenes do not include the known risk factor; A subspace construction module, wherein the subspace construction module is configured to construct a plurality of key scene subspaces and determine a first center; each of the key scene subspaces includes at least one known key scene, and the first center is the center of the key scene subspace; A scene generation module, the scene generation module is configured to calculate an orthogonal basis of the key scene subspace, and extract an unknown key scene space in the key scene subspace along the direction of the orthogonal basis with the first center as the center; A second sampling module is configured to sample unknown key scenes in the unknown key scene space to obtain a set of unknown key test scenes.

Images