Automatic driving system ODC boundary construction method

Abstract

The invention provides an automatic driving system ODC boundary construction method, and belongs to the technical field of automatic driving. In order to solve the problem of ODC boundary construction of a high-level automatic driving system or a vehicle-road cooperative automatic driving system, key ODC elements are extracted through a systematic hazard trigger condition identification method, dimensionality reduction is performed on a scene state space, and a test and verification process required by ODC boundary construction is accelerated; on the basis of risk acceptance degree evaluation, the safe scene interval and the unsafe scene interval are classified to define the ODC boundary, so that the accuracy of boundary design is improved, and the automatic driving system can continuously operate in a wider scene state space.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to an ODC boundary construction method for an autonomous driving system. Background Art

[0002] With the rapid development of autonomous vehicles at different levels, autonomous driving systems have been widely used in a variety of scenarios, from closed campuses, ports and airports to highways and even urban traffic scenarios. As a result, the tasks and scenarios that the advanced driving system (ADS) needs to cope with have become increasingly complex, which has brought a series of challenges to the safe operation of the system. Specifically, typical long-tail traffic scenarios, harsh environmental conditions, and abuse by drivers and passengers seriously interfere with the safe and reliable operation of system functions. For example, ADS may encounter rare weather conditions (such as dense fog and blizzards), road anomalies (such as road collapse or construction), and unusual behaviors of other traffic participants or drivers, which are difficult to foresee during the system design phase. Therefore, this requires that the necessary operating conditions of the system be fully considered when designing and developing ADS to ensure that the system maintains safe and effective operation in the face of various unknown and rare events. However, due to the high complexity of the system and the black-box nature of the AI ​​model, traditional safety specifications cannot fully describe and define the safe operating conditions of current autonomous driving systems for advanced autonomous driving systems. Therefore, how to effectively identify, evaluate, and define the capabilities and limitations of advanced ADS is crucial to prevent any functional abuse and promote safe and effective operation.

[0003] At present, the safe scenario boundaries of the system are mainly set by defining the operational design domain (ODD) of the system to ensure the reliable operation of the autonomous driving system. ODD refers to the external environment in which the vehicle can drive safely (for example, road type, weather conditions, traffic participants). By evaluating the size of its available range and the coping scenario conditions, the automation level of the ADS can be reflected. At the L5 level, the autonomous driving system reaches or even exceeds the level of human drivers, and the system can operate in an unlimited ODD. However, a major limitation of the ODD conceptual framework is that it only reflects the external factors that affect the autonomous driving vehicle, while ignoring the equally important in-vehicle conditions. Therefore, based on the concept of ODD, the definition of the operational design condition (ODC) was proposed. The conceptual framework of ODC includes not only ODD, but also the necessary internal vehicle conditions (driver and passenger status, vehicle motion status, system function status, etc.), which comprehensively describes the conditions under which the autonomous driving vehicle can operate effectively. However, there is currently a lack of systematic and process-based ODC boundary design methods. This ultimately limits the promotion and implementation of high-level ADS, resulting in poor system availability and frequent interruptions during the execution of dynamic driving tasks.

[0004] Therefore, it is necessary to provide an ODC boundary construction method for an autonomous driving system to solve the above problems. Summary of the invention

[0005] The present invention provides a method for constructing an ODC boundary of an autonomous driving system. First, the functional objectives and functional specifications of the autonomous driving system are set. Then, hazard identification and risk assessment are performed through a system theory process method to reveal potential deficiencies and hazards of the system, which may be triggered by a specific combination of ODC trigger conditions. Subsequently, the scenario state space to be covered by the test is defined based on the trigger conditions. Then, the system is connected to a simulation framework to carry out simulation testing, and the risk level of the interval is evaluated by collecting test results, and the safety boundary of acceptable risk is defined, thereby solving at least one technical problem involved in the background technology.

[0006] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0007] A method for constructing an ODC boundary of an autonomous driving system includes the following steps:

[0008] Step S1, defining the functional objectives and functional specifications of the autonomous driving system;

[0009] Step S2, based on the decision logic process of the autonomous driving system, identifying potential hazards of the autonomous driving system, determining triggering conditions of the hazards and analyzing scene elements that cause the hazards, and constructing a scene state subspace for specific hazards based on the scene elements that cause the hazards;

[0010] Step S3, performing simulation test on the autonomous driving system based on the functional objectives and functional specifications defined in step S1 and the scenario state subspace constructed in step S2, and collecting feedback on the autonomous driving system under different test cases;

[0011] Step S4, execute step S2 and step S3 multiple times, continuously discover unknown hazards of the autonomous driving system, until there are no unknown vehicle-level hazards, evaluate the risk level of the autonomous driving system in each hazard-related scene state subspace based on the results of the simulation test, construct the ODC boundary in each scene state subspace according to the preset acceptable risk level, and integrate the ODC boundaries of all scene state subspaces into the ODC boundary of the autonomous driving system.

[0012] As a preferred improvement, the design operating condition ODC of the automatic driving system S is S is a set of working points (u, c), where u∈U represents the functional performance of the autonomous driving system S in the actual operating environment c∈C, and the working point (u, c) should meet the expected design conditions φ=(R, B, E) S , which includes the requirements R that the autonomous driving system should achieve, the constraints B that need to be considered when building the autonomous driving system, and the context E in which the autonomous driving system should operate. Then the ODC of the autonomous driving system S is S It is expressed as:

[0013] ODC S ={(u,c)|(u,c).sat(R,B,E) S ∧c∈C,u∈U}.

[0014] As a preferred improvement, in step S3, the simulation test process specifically includes the following steps:

[0015] The scene state subspace is modeled based on the Scenic scene description language, and a probabilistic modeling language is used to specify the initial state of entity features in the scene state subspace and the distribution of their behavior configuration over time;

[0016] Use a sampling algorithm to search the scene state subspace to ensure that the scene state subspace is fully covered;

[0017] Build a simulation test framework, connect the autonomous driving system to the simulation test framework for simulation testing, and collect feedback from the autonomous driving system under different test cases.

[0018] As a preferred improvement, the sampling algorithm may select a passive sampler or an active sampler using historical test records, and ensure that the scene state subspace is fully covered by equidistant uniform sampling or calculating the required number of scenes based on the Hoffding inequality.

[0019] As a preferred improvement, the simulation test framework includes a simulator, a sampler and a discriminator. The simulator supports the access of the autonomous driving system and the simulation of the corresponding scene state, and is also used to record vehicle state information; the sampler samples and searches the scene state subspace to obtain the initial scene state of the test case; the discriminator determines whether the test case passes based on the defined functional specifications, evaluates the test data recorded by the simulator, and determines the scene label of the test case.

[0020] As a preferred improvement, after step S4, the following step is further included: using a support vector machine to optimize the ODC boundary in each scene state subspace, and the optimization objective function is expressed as:

[0021] Argmin θ μ(T p \f θ (u,c))+w·μ(T n ∩f θ (u,c));

[0022] Where μ(·) represents the risk assessment function, and T n =((u,c)∩Φ) -d They represent true positive cases and true negative cases respectively, d represents the prediction interval, which means the prediction of the scene classification from the current time to the future time d steps; w represents the weight; f θ represents the classifier; θ represents f θ Parameters of the classification function; the objective function aims to optimize the classifier parameters to find a classification boundary function that minimizes the weighted false positive and false negative case classification loss;

[0023] Negative samples are oversampled and the support vector machine is trained to find a hyperplane that maximizes the margin. This hyperplane can effectively distinguish positive and negative samples. The boundary of the hyperplane is directly selected as the ODC boundary of the scene state subspace to minimize the risk.

[0024] As a preferred improvement, the ODC boundary integration of the scene state subspace is determined by taking the intersection of the ODC boundaries of the scene state subspace or performing weighted calculation according to the risk level of the hazardous event.

[0025] The beneficial effects of the present invention are:

[0026] The technical solution proposed in the present invention is mainly aimed at the ODC boundary construction problem of high-level autonomous driving systems or vehicle-road cooperative autonomous driving systems. Through a systematic hazard trigger condition identification method, key ODC elements are extracted, the dimension of the scene state space is reduced, and the ODC test verification process is accelerated; based on the risk acceptability assessment, the safe scene interval and the unsafe scene interval are classified to define the ODC boundary, which improves the accuracy of the boundary design, thereby enabling the autonomous driving system to operate continuously in a wider scene state space. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart showing an ODC boundary construction method for an autonomous driving system provided by the present invention;

[0028] Figure 2 A schematic diagram showing an unprotected left turn scenario;

[0029] Figure 3 A decision logic flow chart showing the unprotected left turn collision avoidance function;

[0030] Figure 4 A schematic diagram showing a hazard causal model of an unprotected left turn collision avoidance function;

[0031] Figure 5 A schematic diagram showing the structure of the simulation test framework;

[0032] Figure 6 Scatter plot showing the state space test cases for the unprotected left turn scenario;

[0033] Figure 7 represents the ODC boundary constructed based on the sub-interval segmented risk acceptability assessment;

[0034] Figure 8 Represents the ODC bound based on risk acceptability oversampling support vector machine. DETAILED DESCRIPTION

[0035] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] Please refer to Figure 1-Figure 8 The present invention provides an ODC boundary construction method for an autonomous driving system, comprising the following steps:

[0037] Step S1, defining the functional objectives and functional specifications of the autonomous driving system.

[0038] In order to define the ODC of an autonomous driving system, it is first necessary to determine the expected functional goals and functional specifications of the autonomous driving system. Specifically, the following three types of conditions need to be considered, namely: the requirements R that the autonomous driving system should achieve, the constraints B that need to be considered when building the autonomous driving system, and the contextual environment E in which the autonomous driving system should operate.

[0039] Specifically, the design operating conditions of the autonomous driving system S are ODC S is a set of operating points (u, c), where u∈U represents the functional performance of the autonomous driving system under the actual operating environment c∈C, and the operating point (u, c) should meet the expected design conditions φ=(R, B, E) S , then the ODC of the autonomous driving system S S It is expressed as:

[0040] ODC S ={(u,c)|(u,c).sat(R,B,E) S ∧c∈C,u∈U}.

[0041] The design condition φ reflects the functional objectives and functional specifications of the autonomous driving system, and can be used as a criterion for determining the feasibility of the autonomous driving system in a specified scenario: if the working point (u, c) is such that the autonomous driving system operates in the expected contextual environment E under the premise of considering constraint B and can meet the functional objective R, then the autonomous driving system is determined to be feasible in the specified scenario; otherwise, it is determined to be infeasible in the specified scenario.

[0042] Specifically, Figure 2 As shown in the figure, taking the unprotected left turn scenario at the intersection as an example, the functional goal of the autonomous driving system is to avoid collision between the left-turning vehicle and the straight-moving vehicle. The trajectory of the left-turning vehicle is composed of a set of data points T1 = {(t 1,1 ,p 1,1 ),(t 1,2 ,p 1,2 ),...(t 1,k ,p 1,k ),...(t 1,n ,p 1,n )}, k represents the index of the trajectory data in T1, and n represents the number of trajectory data; the trajectory of the straight-moving vehicle is composed of a set of data points T2 = {(t 2,1 ,p 1,1 ),(t 2,2 ,p 2,2 ),...(t 2,l ,p 2,l ),...(t 2,n ,p 2,n )} represents, l represents the index of trajectory data in T2.

[0043] In order to quantify whether a left-turning vehicle collides with a straight-moving vehicle, it is necessary to define a time window Δt and a distance threshold d Θ , within the time window Δt, if the spatial distance between the two vehicles is less than the distance threshold d Θ , then it is considered that they may have collided. For each pair of time points (t 1,k ,t 2,l ), if condition Φ1 is met, it is considered that the two vehicles may have collided:

[0044] Φ1:||p 1,k -p 2,l ||≤d Θ And|t 1,k -t 2,l |≤Δt;

[0045] In the formula, ||.|| represents the trajectory point p 1,k With p 2,l The Euclidean distance between them. For a specified scenario, if at least one pair of time points satisfying condition Φ1 is found in the trajectory data, it can be considered that the two vehicles have collided.

[0046] In addition to considering the collision constraints of the two vehicles, it is also necessary to consider whether the two vehicles have emergency braking. Emergency braking means that the system activates the minimum risk strategy and will also exit the autonomous driving system at this time. 2 When the vehicle has emergency braking, the condition Φ2 is obtained:

[0047] Φ2: a l ≤-12m / s 2 ∩a k ≤-12m / s 2 ;

[0048] In the formula, a l represents the acceleration of the vehicle moving straight; a k Indicates the acceleration of the vehicle turning left.

[0049] For the unprotected left turn collision avoidance function, the design operating condition of the automatic system is ODC S Should also meet and Condition, expressed as:

[0050]

[0051] Step S2, based on the decision logic process of the autonomous driving system, identifies potential hazards of the autonomous driving system, determines the triggering conditions of the hazards and analyzes the scene elements that cause the hazards, and constructs a scene state subspace for specific hazards based on the scene elements that cause the hazards.

[0052] Hazards reflect the expected functional deficiencies of the automated driving system in the actual operating environment, that is, the automated driving system fails to achieve the expected functional goals or violates the functional specifications. Hazards are caused by one or more specific triggering conditions, and ultimately lead to hazardous behavior or the inability to prevent, detect and mitigate reasonably foreseeable indirect misuse.

[0053] In order to systematically identify potential hazards, it is necessary to trigger analysis from the decision-making logic process of the autonomous driving system. Figure 3 As shown in the figure, taking the unprotected left turn scenario at the intersection as an example, in order to achieve the unprotected left turn collision avoidance function, it is necessary to install an RSU (Road Side Unit) including sensors and communication equipment on the roadside. The RSU provides real-time traffic information to the autonomous driving vehicle to assist it in making a safer left turn decision. In this scenario, the left-turning vehicle is tracked according to the planned trajectory of the straight-moving vehicle. The left-turning vehicle continuously sends the straight-moving vehicle positioning information and left-turn intention to the RSU. When the left-turning vehicle enters the warning area, the RSU sends a warning signal to the straight-moving vehicle at the intersection. The straight-moving vehicle determines whether to accelerate or decelerate to give way based on the distance from the straight-moving vehicle to the intersection, thereby achieving the right of way transfer and avoiding collisions.

[0054] In this scenario, the conditions for triggering the hazard are the speed of the straight-moving vehicle and the communication delay. If the speed of the straight-moving vehicle is too high, the braking distance will increase, which may easily cause a collision hazard between the left-turning vehicle and the straight-moving vehicle. The communication delay will affect the uploading and sending of data, causing a lag in the execution of control commands of the autonomous driving system, which may also easily cause a collision hazard between the left-turning vehicle and the straight-moving vehicle.

[0055] Furthermore, in this scenario, if the distance between the straight-moving vehicle and the intersection is in the right-of-way conflict zone with the left-turning vehicle, the exposure rate of collision and emergency braking increases. The controllability of the hazardous event is used to determine whether it will turn into injury. If the driver cannot avoid the collision accident by braking, that is, the hazardous event is uncontrollable, it is judged that the hazardous event will turn into injury and will cause personal injury.

[0056] By analyzing the triggering conditions of the hazard, we can obtain the scene elements that cause the hazard. In this scenario, the scene elements that cause the hazard are: communication delay, the distance between the straight vehicle and the intersection, and the speed of the straight vehicle.

[0057] Step S3, based on the functional objectives and functional specifications defined in step S1 and the scenario state subspace constructed in step S2, a simulation test is performed on the autonomous driving system to collect feedback on the autonomous driving system under different test cases.

[0058] The simulation test process specifically includes the following steps:

[0059] The scene state subspace is modeled based on the Scenic scene description language, and a probabilistic modeling language is used to specify the initial state of entity features in the scene state subspace and the distribution of their behavior configuration over time;

[0060] Use a sampling algorithm to search the scene state subspace to ensure that the scene state subspace is fully covered;

[0061] Build a simulation test framework, connect the autonomous driving system to the simulation test framework for simulation testing, and collect feedback from the autonomous driving system under different test cases.

[0062] Taking the unprotected left turn scenario at an intersection as an example, the speed of the straight-moving vehicle is sampled from the speed range of [20km / h, 50km / h], the delay is sampled from the range of [100ms, 200ms], and the initial position of the straight-moving vehicle is randomly sampled from the front section of the intersection road, which will determine the initial distance of the straight-moving vehicle from the intersection. In addition, other non-hazard-related ODC elements, such as friction coefficient, packet loss rate, communication range, packet loss rate, etc., are set to fixed values.

[0063] The sampling algorithm can be selected as a passive sampler, such as uniform sampling, low-discrepancy (Halton) sampling, and an active sampler using historical test records, such as cross entropy sampling, policy gradient sampling, etc. The scene state subspace can be fully covered by uniform sampling at equal intervals or by calculating the number of scenes to be sampled based on the Hoffding inequality.

[0064] The simulation test framework includes a simulator, a sampler and a discriminator. The simulator supports the access of the autonomous driving system and the simulation of the corresponding scene state, and is also used to record vehicle state information; the sampler samples and searches the scene state subspace to obtain the initial scene state of the test case; the discriminator determines whether the test case passes based on the defined functional specifications, evaluates the test data recorded by the simulator, and determines the scene label of the test case.

[0065] like Figure 5 As shown in the figure, the test case results show that in the delay interval of [100ms, 200ms], the delay has no significant impact on whether the test passes. Therefore, in this scenario, the ODC state space is actually composed of (the distance between the straight-moving vehicle and the intersection during the warning, and the speed of the straight-moving vehicle during the warning).

[0066] Step S4, execute step S2 and step S3 multiple times, continuously discover unknown hazards of the autonomous driving system, until there are no unknown vehicle-level hazards, evaluate the risk level of the autonomous driving system in each hazard-related scene state subspace based on the results of the simulation test, construct the ODC boundary in each scene state subspace according to the preset acceptable risk level, and integrate the ODC boundaries of all scene state subspaces into the ODC boundary of the autonomous driving system.

[0067] The scenario use cases are classified in the scenario state subspace according to the preset acceptable risk threshold. Taking the unprotected left turn scenario as an example, since there is not enough reaction time in the left turn scenario, the driver cannot take over the vehicle in time, and the degree of loss of control is defined as 1, that is, the system is completely uncontrollable. At the same time, whether a collision or emergency braking occurs at the intersection, considering the complex traffic conditions at the intersection and the subsequent impact on traffic, the severity is defined as 1, that is, the consequences of the hazard are very serious; Figure 7 As shown, the acceptable risk exposure rate is set to 2% at this time, the scenario state space is segmented towards the ODC element interval, and the risk exposure rate of each interval is evaluated to define the available interval on the scenario state space. At this time, the ODC boundary can be expressed as:

[0068] ODC S ={(u 车速 ,c 距离 )|(20≤u 车速 ≤50, c 距离 ≤6)∪(30≤u 车速 ≤40,0≤c 距离 ≤11)}.

[0069] Furthermore, in order to obtain a more accurate ODC boundary, the present invention uses a support vector machine to optimize the constructed ODC boundary, and the optimization objective function is expressed as:

[0070] Argmin θ μ(T p \f θ (u,c))+w·μ(T n ∩f θ (u,c));

[0071] Where μ(·) represents the risk assessment function, and T n =((u,c)∩Φ) -d They represent true positive cases and true negative cases respectively, d represents the prediction interval, which means the prediction from the current time to the future time d steps; w represents the weight; f θ represents the classifier; θ represents f θParameters of the classification function; this objective function aims to optimize the classifier parameters to find the classification boundary function that minimizes the weighted false positive and false negative case classification loss.

[0072] The left side of the objective function defines the measure of the false positive part, while the false negative part on the right side is weighted by w to represent the acceptable degree of risk (the acceptable degree of positive samples being classified as negative samples and the acceptable degree of negative samples being classified as positive samples). By optimizing the objective function to minimize the false positive and weighted false negative losses of the boundary classifier, the accuracy of the boundary design is improved.

[0073] Based on this objective function, negative samples are oversampled, and the support vector machine that minimizes the weighted false positive and false negative case classification loss is trained to find a hyperplane that maximizes the margin. This hyperplane can effectively distinguish positive and negative samples. In the unprotected left turn scenario, because the acceptable risk exposure rate is set to 2%, the value of w is defined as 49, and negative samples are oversampled. Figure 8 As shown in the figure, the support vector machine effectively forms a more flexible and accurate scenario classification boundary, that is, the support vector machine hyperplane defines an approximately optimal ODC boundary for the risk acceptability function. Compared with the ODC boundary design method based on sub-interval segmented risk acceptability assessment, the ODC continuity level of ADS operation is improved due to the larger hyperspace volume.

[0074] Finally, other hazard triggering conditions are continuously explored and their ODC boundaries are defined. The ODC boundary subspaces of multiple hazard definitions are integrated to determine the ODC boundary of the autonomous driving system. Optionally, the boundary integration can be determined by taking the intersection of multiple subspace boundaries or by weighted calculation according to the risk level of the hazard event.

[0075] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for constructing an ODC boundary of an autonomous driving system, characterized in that: The steps include: Step S1, defining the functional objectives and functional specifications of the autonomous driving system; Step S2, based on the decision logic process of the autonomous driving system, identifying potential hazards of the autonomous driving system, determining triggering conditions of the hazards and analyzing scene elements that cause the hazards, and constructing a scene state subspace for specific hazards based on the scene elements that cause the hazards; Step S3, performing simulation test on the autonomous driving system based on the functional objectives and functional specifications defined in step S1 and the scenario state subspace constructed in step S2, and collecting feedback on the autonomous driving system under different test cases; Step S4, execute step S2 and step S3 multiple times, continuously discover unknown hazards of the autonomous driving system, until there are no unknown vehicle-level hazards, evaluate the risk level of the autonomous driving system in each hazard-related scene state subspace based on the results of the simulation test, construct the ODC boundary in each scene state subspace according to the preset acceptable risk level, and integrate the ODC boundaries of all scene state subspaces into the ODC boundary of the autonomous driving system.

2. The method for constructing an ODC boundary of an autonomous driving system according to claim 1, characterized in that: Design operating conditions of the automated driving system S ODC S is a set of working points (u, c), where u∈U represents the functional performance of the autonomous driving system S in the actual operating environment c∈C, and the working point (u, c) should meet the expected design conditions φ=(R, B, E) S , which includes the requirements R that the autonomous driving system should achieve, the constraints B that need to be considered when building the autonomous driving system, and the context E in which the autonomous driving system should operate. Then the ODC of the autonomous driving system S is S It is expressed as: ODC S ={(u,c)|(u,c).sat(R,B,E) S ∧c∈C,u∈U}。 3. The method for constructing an ODC boundary of an autonomous driving system according to claim 1, characterized in that: In step S3, the simulation test process specifically includes the following steps: The scene state subspace is modeled based on the Scenic scene description language, and a probabilistic modeling language is used to specify the initial state of entity features in the scene state subspace and the distribution of their behavior configuration over time; Use a sampling algorithm to search the scene state subspace to ensure that the scene state subspace is fully covered; Build a simulation test framework, connect the autonomous driving system to the simulation test framework for simulation testing, and collect feedback from the autonomous driving system under different test cases.

4. The method for constructing an ODC boundary of an autonomous driving system according to claim 3, characterized in that: The sampling algorithm can select a passive sampler or an active sampler using historical test records, ensuring that the scene state subspace is fully covered by equidistant uniform sampling or calculating the required number of scenes based on the Hoffding inequality.

5. The method for constructing an ODC boundary of an autonomous driving system according to claim 3, characterized in that: The simulation test framework includes a simulator, a sampler and a discriminator. The simulator supports the access of the autonomous driving system and the simulation of the corresponding scene state, and is also used to record vehicle state information; the sampler performs sampling and search on the scene state subspace to obtain the initial scene state of the test case; The discriminator determines whether the test case passes based on the defined functional specifications, evaluates the test data recorded by the simulator, and determines the scenario label of the test case.

6. The method for constructing an ODC boundary of an autonomous driving system according to claim 1, characterized in that: After step S4, the following steps are also included: using a support vector machine to optimize the ODC boundary in each scene state subspace, and the optimization objective function is expressed as: Argmin θ μ(T p \f θ (u,c))+w·μ(T n ∩f θ (u,c)); Where μ(·) represents the risk assessment function, and T n =((u,c)∩Φ) -d They represent true positive cases and true negative cases respectively, d represents the prediction interval, which means the prediction of the scene classification from the current time to the future time d steps; w represents the weight; f θ represents the classifier; θ represents f θ Parameters of the classification function; the objective function aims to optimize the classifier parameters to find a classification boundary function that minimizes the weighted false positive and false negative case classification loss; Negative samples are oversampled and the support vector machine is trained to find a hyperplane that maximizes the margin. This hyperplane can effectively distinguish positive and negative samples. The boundary of the hyperplane is directly selected as the ODC boundary of the scene state subspace to minimize the risk.

7. The method for constructing an ODC boundary of an autonomous driving system according to claim 1, characterized in that: The ODC boundary integration of the scene state subspace is determined by taking the intersection of the ODC boundary of the scene state subspace or is weighted according to the risk level of the hazard event.

Images