Modeling and analysis method for triggering conditions of expected functional safety of autonomous driving perception system

Abstract

The present invention discloses a modeling and analysis method for triggering conditions of expected functional safety of an autonomous driving perception system, and belongs to the technical field of autonomous driving perception systems. The present invention ensures that the triggering conditions that cause the expected functional safety of an autonomous driving perception system are scientifically and systematically identified and modeled by analyzing the attributes of complex environmental factors that affect the autonomous driving perception system and the correlation between the attributes through a Bayesian network; the probability distribution of the triggering condition event chain is calculated through a Bayesian network, and the occurrence of the expected functional safety of the autonomous driving perception system is predicted, which effectively promotes the safe development of autonomous driving and promotes the early implementation of autonomous driving products, and is worthy of being promoted and used.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving perception systems, and in particular to a modeling and analysis method for triggering conditions for expected functional safety of autonomous driving perception systems. Background Art

[0002] Tesla's perception algorithm failed to detect the fatal accident of the white truck compartment; Uber's decision-making system failed and hit a pedestrian crossing the road illegally. These accidents have once pushed the safety of autonomous vehicles to the research hotspot of the industry. Unlike traditional vehicles, autonomous vehicles may exhibit unexpected behaviors when there is no failure in the electronic and electrical systems, that is, the problem of expected functional safety.

[0003] The perception system of autonomous vehicles is insufficient in terms of perception ability in extreme weather (rain, snow, fog, wind, etc.) and lighting conditions. For example, harsh environments will cause data to increase and the sensor's field of view to be limited. The ISO21448 Road Vehicles - Intended Functional Safety standard defines trigger conditions as specific scenario conditions that lead to subsequent hazardous behavior of the autonomous driving system.

[0004] However, existing research and analysis on the safety of expected functions of autonomous driving are mostly qualitative analysis of a certain system or a certain function of low-level autonomous driving vehicles; the analysis of trigger conditions is based on the mapping of vehicle-level hazards to perception-decision-execution performance defects, and then from performance defects to trigger conditions. There is a lack of positive and systematic modeling and calculation methods for trigger conditions that cause unexpected behaviors of autonomous driving vehicles. Therefore, a modeling and analysis method for trigger conditions of expected functional safety of autonomous driving perception systems is proposed. Summary of the invention

[0005] The technical problem to be solved by the present invention is: how to solve the problem in the prior art of lacking a method for positively and systematically modeling and calculating the trigger conditions that cause unexpected behaviors of autonomous driving vehicles. A modeling and analysis method for the trigger conditions of the expected functional safety of an autonomous driving perception system is provided. By calculating the Bayesian network of the trigger event chain of the autonomous driving perception system, it is predicted whether the expected functional safety of autonomous driving occurs.

[0006] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0007] S1: Analyze the trigger conditions that lead to expected functional safety issues from the autonomous driving perception system, establish a three-layer structure of input, state, and output based on the correlation between the internal elements of the expected functional safety trigger conditions, describe the trigger variables that lead to the expected functional safety of autonomous driving from the three levels of input, state, and output, use them as variables of the Bayesian network, and establish three sub-event Bayesian networks;

[0008] S2: Determine the correlation between the variables in the sub-event Bayesian networks, merge the three sub-event Bayesian networks to form the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system, and determine the prior probability and conditional probability set of each variable in the network;

[0009] S3: Perform Bayesian network reasoning on the expected functional safety triggering event chain of the autonomous driving perception system to predict whether the expected functional safety of autonomous driving occurs under the triggering condition event chain model.

[0010] Furthermore, in step S1, the three sub-event Bayesian networks are respectively a rain, snow, fog Bayesian network, a low light Bayesian network, and a visual blind spot Bayesian network.

[0011] Furthermore, in step S1, the internal elements of the trigger condition are divided into input elements, state elements and output elements, wherein the input elements refer to a collection of external factors that affect the expected functional safety of the autonomous driving perception system, the state elements are used to describe the properties and characteristics of the trigger condition during the evolution process, and the output elements represent the consequences caused by the trigger condition.

[0012] Furthermore, in step S2, the association relationship between the variables of the sub-event Bayesian networks is determined, and the specific process of merging the three sub-event Bayesian networks is as follows:

[0013] S21: According to the input variable set, state variable set and output variable set information of each sub-event Bayesian network, determine the common event association variable of the rain, snow, fog Bayesian network and the insufficient light Bayesian network as C 12 , the total loss output variable set is L 12 The common event association variable of the insufficient light Bayesian network and the visual blind spot Bayesian network is C 23 ;

[0014] S22: C 12 Assigned to rain, snow, fog Bayesian network, C 23 Assigned to low light network, with C 12 With C 23 The three sub-event Bayesian networks were related for the join points;

[0015] S23: L 12 and L 23 After merging the loss output of rain, snow, fog, insufficient light, and visual blind spot events for the connection point, the overall loss output is LO c , forming a Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system.

[0016] Furthermore, in step S2, the specific process of determining the prior probability and conditional probability set of each variable in the network is as follows:

[0017] S24: Determine the impact of input variables on state variables

[0018] Rain, snow, fog sub-event EN1 is associated with insufficient light sub-event EN2. For any state variable s of EN2 2k ∈S2, if directly connected to the input variable, the probability of occurrence is as follows:

[0019]

[0020] Among them, i 21 ,i 22 ,…,i 2l are the input variables of sub-event EN2, Each common event association variable of sub-event EN1 and sub-event EN2;

[0021] S25: Determine the impact of state variables on state variables

[0022] For any state variable s in sub-event EN2 2k ∈S2, if it is not affected by the input variables, the probability of occurrence is as follows:

[0023] p(s 2k )=∑p(s 21 ,s 22 ,…s 2m )p(s 2k |s 21 ,s 22 ,…s 2m )

[0024] Among them, s 21 ,s 22 ,…s 2m Each state variable of sub-event EN2;

[0025] S26: Determine the impact of state variables on output variables

[0026] S 1， S 2， S 3，…, S n EN is a related trigger event that affects the expected functional safety of the autonomous driving perception system. 1， EN 2， EN 3，…, EN j The set of state variables, O j To trigger the event EN j The output variable set of any event ENj Any output variable O jk ∈O j , there are the following probability relations:

[0027] p(O 2k )=∑p(s 21 ,s 22 ,…s 2m )p(O 2k |s 21 ,s 22 ,…s 2m );

[0028] The overall loss output of the Bayesian network for the expected functional safety triggering event chain of the autonomous driving perception system, And the following probability relationship is satisfied:

[0029]

[0030] in, It represents the overall loss output of the i-th type of loss of the triggering event chain to the external environment, and n represents the number of associated merged sub-event Bayesian networks.

[0031] Furthermore, in step S3, the input variable set I in the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system is assumed to be c and the state variable set S c Some state variables in the equation are evidence variables, S c The remaining state variables and output variables O c is the target variable, and according to the value of the evidence variable, the posterior probability of various values ​​of the target variable is calculated.

[0032] Furthermore, the specific process in step S3 is as follows:

[0033] S31: Calculate the probability of other target variables based on the value information of the evidence variable

[0034] According to the formula in step S24 and step S25 and Determine whether the trigger event occurs. If it occurs, continue to calculate the value probability of other state variables and calculate the value probability of other state variables according to p(O 2k )=∑p(s 21 ,s 22 ,…s 2m )p(O 2k |s 21 ,s 22 ,…s 2m) calculates the probability of the output variable, otherwise stops the calculation, thereby obtaining the posterior probability of each target variable in the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system;

[0035] S32: Summarize the status and output of each trigger event

[0036] According to the posterior probability obtained in step S31, the state variable The state variable set S1, S2, ... S of each sub-event whose probability of taking the value 1 is greater than the probability of taking the value 0 n , and its loss variable output variables LO1, LO2, ... LO n Summarize the variables in

[0037] The state of the triggering event chain is the set of state variables of each sub-event on the chain, and its probability distribution is:

[0038] p(S1,S2,…S n )=p(s 11 ,s 12 ,…,s 21 ,s 22 ,…s n1 ,s n2 ,…,s nm )

[0039] Among them, s 11 ,s 12 ,…are the state variables of sub-event EN1, s 21 ,s 22 ,…are the state variables of sub-event EN2, s n1 ,s n2 ,…,s nm For sub-event EN n The various state variables;

[0040] Combine the status of each trigger event and the formula in step S6 to obtain the overall loss output of the trigger event chain That is, the overall loss output of the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system;

[0041] S33: Prediction

[0042] Based on the set of state variables of the trigger condition event chain that causes the expected functional safety of the autonomous driving perception system and the value probability of the overall loss output variable set, the key states of other trigger conditions associated with the initial trigger condition and the expected functional safety issues brought to the autonomous driving vehicle by the trigger condition event chain are predicted.

[0043] Compared with the prior art, the present invention has the following advantages: the modeling and analysis method of the triggering conditions for the expected functional safety of the autonomous driving perception system ensures that the triggering conditions that cause the expected functional safety of the autonomous driving perception system are scientifically and systematically identified and modeled through the Bayesian network analysis of the attributes of the complex environmental factors that affect the autonomous driving perception system and the correlation between the attributes; the probability distribution of the triggering condition event chain is calculated through the Bayesian network, and the occurrence of the expected functional safety of the autonomous driving perception system is predicted, which effectively promotes the safe development of autonomous driving and promotes the early implementation of autonomous driving products, and is worthy of popularization and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a modeling and analysis method for the expected functional safety triggering conditions of the autonomous driving perception system of the present invention. DETAILED DESCRIPTION

[0045] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0046] like Figure 1 As shown, this embodiment provides a technical solution: a modeling and analysis method for the expected functional safety triggering conditions of an autonomous driving perception system, comprising the following steps:

[0047] First, the trigger conditions from the autonomous driving perception system that may lead to expected functional safety issues are analyzed, and a Bayesian network of rain, snow, fog, insufficient light, and visual blind spot trigger event chains is constructed. For example, in extreme weather conditions (rain, snow, and fog), both vision-based perception systems and LiDAR-based perceptions have perception difficulties. Thin snow covering the road will cause road markings to disappear completely, splashing rain will affect the reflection effect of the radar, and fog may block the camera's line of sight. The above external conditions will affect the accuracy of the perception results, which will lead to unexpected functional safety issues in the autonomous driving system.

[0048] According to the theory of system engineering, the triggering conditions that cause the expected functional safety of the autonomous driving perception system are regarded as a system formed by the interaction of various elements. The elements related to the triggering conditions can be divided into: input elements, state elements and output elements. Input elements refer to the set of external factors that affect the expected functional safety of the autonomous driving perception system, mainly road structures, traffic facilities, temporary events, participants, environmental information and vehicle status (i.e., the six-layer model of environmental information). State elements are used to describe the nature and characteristics of the triggering conditions in the evolution process, including the attributes of the above six-layer model and the relationship between the attributes. Output elements represent the consequences caused by the triggering conditions, including aspects of the perception process such as distortion, loss and noise of the original data; aspects of the cognitive process such as failure to identify the target object, incorrect classification of the target object and parameter errors.

[0049] The trigger variables that may lead to the safety of the expected functions of autonomous driving are described from the three levels of input, state and output as Bayesian network variables, and three sub-event Bayesian networks are established. Then, according to the association relationship between the three, the three sub-event Bayesian networks are merged to form a trigger event chain Bayesian network of rain, snow, fog-lack of light-blind spot, that is, a trigger event chain Bayesian network. The variable description of the three sub-event Bayesian networks of rain, snow, fog, lack of light and blind spot is shown in Table 1 below, where bold fonts represent common event association variables of the sub-event Bayesian network, and italics represent common loss output variables.

[0050] Table 1 Variables in the Bayesian network of the autonomous driving perception system

[0051]

[0052] As can be seen from the table above, the common event association variable of the rain, snow, fog Bayesian network and the insufficient light Bayesian network is C 12 = {O 11}, the total loss output variable set is L 12 = {O 14、 O 15、 O 16、 O 17}, the common event association variable of the insufficient light Bayesian network and the visual blind spot Bayesian network is C 23 = {O 22}, the total loss variable set is L 23 =

[0053] {O 24、 O 25、 O 26、}, C 12 Assigned to rain, snow, fog Bayesian network, C 23 Assigned to low light network, with C 12 With C23 As the connection point, the three sub-events of rain, snow, fog, insufficient light and visual blind spot are associated with the Bayesian network; L 12 and L 23 After merging rain, snow, and fog for the connection point and outputting the loss of insufficient light and visual blind spots, the overall loss output is A Bayesian network forming the rain, snow, fog-lack of light-blind spot event chain.

[0054] Secondly, based on conditional probability, the causal relationship between the input variables and state variables, state variables and state variables, and state variables and outcome variables in the Bayesian network of the event chain triggered by the external conditions of the autonomous driving perception system (rain, snow, fog, insufficient light, and visual blind spots) is analyzed;

[0055] 1) The impact of input on the state. Sub-event EN1 rain, snow, fog is associated with sub-event EN2 insufficient light. For any state variable s of EN2 2k ∈S2, if directly connected to the input variable, the probability of occurrence is as follows:

[0056]

[0057] 2) Impact of state on state. For any state variable s in sub-event EN2 2k ∈S2, if it is not affected by the input variables, the probability of occurrence is as follows:

[0058] p(s 2k )=∑p(s 21 ,s 22 ,…s 2m )p(s 2k |s 21 ,s 22 ,…s 2m ) (2)

[0059] 3) The impact of state on output. 1， S 2， S 3，…, S n EN is a related trigger event that affects the expected functional safety of the autonomous driving perception system. 1， EN 2， EN 3，…, EN j The set of state variables, O j To trigger the event EN j The output variable set of any event EN j Any output variable O 2k ∈O j , there are the following probability relations:

[0060] p(O 2k )=∑p(s 21 ,s 22 ,…s 2m )p(O 2k |s 21 ,s 22 ,…s 2m ) (3)

[0061] EN j The output variable set O j It contains two parts, one part becomes the input variable of other events, and the other part is the sub-event EN j If the loss output variable is also the common loss output variable of other events, it can be used as the overall loss output variable set LO of the trigger event chain Bayesian network c The variables in .

[0062] To trigger the overall loss output of the event chain Bayesian network, And the following probability relationship is satisfied:

[0063]

[0064] in, It represents the overall loss output of the i-th type of loss of the triggering event chain to the external environment, and n represents the number of associated merged sub-event Bayesian networks.

[0065] Finally, the Bayesian network reasoning of the expected functional safety triggering event chain of the autonomous driving perception system is carried out to predict whether the expected functional safety of autonomous driving occurs under the triggering condition event chain model.

[0066] Assume that the input variable set I in the trigger event chain Bayesian network CN c and the state variable set S c Some state variables in the equation are evidence variables, S c The remaining state variables and output variables O c is the target variable. According to the value of the evidence variable, the posterior probability of various values ​​of the target variable can be calculated.

[0067] 1) Calculate the probability of other target variables based on the value information of the evidence variable

[0068] According to formula (1), (2) and Determine whether the trigger event occurs. If it occurs, continue to calculate the value probabilities of other state variables and calculate the value probability of the output variable according to formula (3). Otherwise, stop the calculation to obtain the posterior probability of each target variable in the trigger event chain Bayesian network.

[0069] 2) Summarize the status and output of each trigger event

[0070] According to the posterior probability obtained in step (1), the state variable The state variable set S1, S2, ... S of each sub-event whose probability of taking the value 1 is greater than the probability of taking the value 0 n , and its loss variable output variables LO1, LO2, ... LO n Summarize the variables in

[0071] The state of the triggering event chain is the set of state variables of each sub-event on the chain, and its probability distribution is:

[0072] p(S1,S2,…S n )=p(s 11 ,s 12 ,…,s 21 ,s 22 ,…s n1 ,s n2 ,…,s nm ) (4)

[0073] Combining the states of each trigger event with formula (3) and formula (4), the overall loss output of the trigger event chain can be obtained:

[0074] Based on the set of state variables of the trigger condition event chain that may cause the expected functional safety of the autonomous driving perception system and the value probability of the overall loss output variable set, it is possible to predict the key states of other trigger conditions associated with the initial trigger condition and the expected functional safety issues brought to the autonomous driving vehicle by the trigger condition event chain.

[0075] It should be noted that, in the present invention, the expected functional safety refers to the hazards and risks caused by insufficient functions or reasonably foreseeable human misuse.

[0076] In summary, the modeling and analysis method of the trigger conditions for the expected functional safety of the autonomous driving perception system in the above-mentioned embodiment analyzes the trigger conditions positively from the functional deficiencies and performance limitations that may cause the expected functional safety of the autonomous driving perception system. This is consistent with the forward development process of the vehicle during the vehicle system safety development process, and the analysis and modeling method can be incorporated into the V-shaped development process of the vehicle. The trigger events of the perception system and the correlation between events are analyzed with the idea of ​​event chain analysis, which can effectively predict the chain reactions and harmful consequences of the events, and realize scientific and quantitative prediction and analysis of the trigger events that may cause the expected functional safety of the autonomous driving perception system, which is worthy of promotion and use.

[0077] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A modeling and analysis method for the expected functional safety triggering conditions of an autonomous driving perception system, characterized in that: The following steps are involved: S1: Analyze the trigger conditions that lead to expected functional safety issues from the autonomous driving perception system, establish a three-layer structure of input, state, and output based on the correlation between the internal elements of the expected functional safety trigger conditions, describe the trigger variables that lead to expected functional safety issues of autonomous driving from the three levels of input, state, and output, use them as variables of the Bayesian network, and establish three sub-event Bayesian networks; In step S1, the three sub-event Bayesian networks are respectively a rain, snow, fog Bayesian network, a low light Bayesian network, and a visual blind spot Bayesian network; S2: Determine the correlation between the variables in the sub-event Bayesian networks, merge the three sub-event Bayesian networks to form the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system, and determine the prior probability and conditional probability set of each variable in the network; In step S2, the association relationship between the variables of the sub-event Bayesian networks is determined, and the specific process of merging the three sub-event Bayesian networks is as follows: S21: According to the input variable set, state variable set and output variable set information of each sub-event Bayesian network, determine the common event association variable of the rain, snow, fog Bayesian network and the insufficient light Bayesian network as C 12 , the total loss output variable set is L 12 The common event association variable of the insufficient light Bayesian network and the visual blind spot Bayesian network is C 23 ; S22: C 12 Assigned to rain, snow, fog Bayesian network, C 23 Assigned to low light network, with C 12 With C 23 The three sub-event Bayesian networks were related for the join points; S23: L 12 and L 23 After merging the loss output of rain, snow, fog, insufficient light, and visual blind spot events for the connection point, the overall loss output is LO c , forming a Bayesian network of expected functional safety triggering events of the autonomous driving perception system, where L 23 is a set of common loss variables; In step S2, the specific process of determining the prior probability and conditional probability set of each variable in the network is as follows: S24: Determine the impact of input variables on state variables Rain, snow, fog sub-event EN1 is associated with insufficient light sub-event EN2. For any state variable s of EN2 2k ∈S2, if it is directly connected to the input variable, the probability of occurrence is as follows: Among them, i 21 ,i 22 ,…,i 2l are the input variables of sub-event EN2, Each common event association variable of sub-event EN1 and sub-event EN2; S25: Determine the impact of state variables on state variables For any state variable s in sub-event EN2 2k ∈S2, if it is not affected by the input variables, the probability of occurrence is as follows: p(s 2k )=∑p(s 21 ,s 22 ,…s 2m )p(s 2k |s 21 ,s 22 ,…s 2m ) Among them, s 21 ,s 22 ,…s 2m Each state variable of sub-event EN2; S26: Determine the impact of state variables on output variables S1,S2,S3,…,S n The related triggering events EN1, EN2, EN3, ..., EN3 affect the expected functional safety of the autonomous driving perception system. j The set of state variables, O j To trigger the event EN j The output variable set of any event EN j Any output variable O jk ∈O j , there are the following probability relations: p(O 2k )=∑p(s 21 ,s 22 ,…s 2m )p(O 2k |s 21 ,s 22 ,…s 2m ); The overall loss output of the Bayesian network for the expected functional safety triggering event chain of the autonomous driving perception system, And the following probability relationship is satisfied: in, represents the overall loss output of the trigger event chain to the i-th type of loss of the external environment, n represents the number of associated merged sub-event Bayesian networks, O c is the output variable; S3: Perform Bayesian network reasoning on the expected functional safety triggering event chain of the autonomous driving perception system to predict whether the expected functional safety problem of autonomous driving will occur under the triggering condition event chain model.

2. The modeling and analysis method for triggering conditions of expected functional safety of an autonomous driving perception system according to claim 1, characterized in that: In step S1, the internal elements of the trigger condition are divided into input elements, state elements and output elements, wherein the input elements refer to a collection of external factors that affect the expected functional safety of the autonomous driving perception system, the state elements are used to describe the properties and characteristics of the trigger condition during the evolution process, and the output elements represent the consequences caused by the trigger condition.

3. The modeling and analysis method for triggering conditions of expected functional safety of an autonomous driving perception system according to claim 1, characterized in that: In step S3, it is assumed that the input variable set I in the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system c and the state variable set S c Some state variables in the equation are evidence variables, S c The remaining state variables and output variables O c is the target variable, and according to the value of the evidence variable, the posterior probability of various values ​​of the target variable is calculated.

4. The modeling and analysis method for triggering conditions of expected functional safety of an autonomous driving perception system according to claim 3, characterized in that: The specific process in step S3 is as follows: S31: Calculate the probability of other target variables based on the value information of the evidence variable According to the formula in step S24 and step S25 and Determine whether the trigger event occurs. If it occurs, continue to calculate the value probability of other state variables and calculate the value probability of other state variables according to p(O 2k )=∑p(s 21 ,s 22 ,…s 2m )p(O 2k |s 21 ,s 22 ,…s 2m ) calculates the probability of the output variable, otherwise stops the calculation, thereby obtaining the posterior probability of each target variable in the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system; S32: Summarize the status and output of each trigger event According to the posterior probability obtained in step S31, the state variable The state variable set S1, S2, ... S of each sub-event whose probability of taking the value 1 is greater than the probability of taking the value 0 n , and its loss variable output variables LO1, LO2, ... LO n Summarize the variables in The state of the triggering event chain is the set of state variables of each sub-event on the chain, and its probability distribution is: p(S1,S2,…S n )=p(s 11 ,s 12 ,…,s 21 ,s 22 ,…s n1 ,s n2 ,…,s nm ) Among them, s 11 ,s 12 ,…are the state variables of sub-event EN1, s 21 ,s 22 ,…are the state variables of sub-event EN2, s n1 ,s n2 ,…,s nm For sub-event EN n The various state variables; Combine the status of each trigger event and the formula in step S6 to obtain the overall loss output of the trigger event chain That is, the overall loss output of the Bayesian network of the expected functional safety triggering event chain of the autonomous driving perception system; S33: Prediction Based on the set of state variables of the trigger condition event chain that causes the expected functional safety of the autonomous driving perception system and the value probability of the overall loss output variable set, the key states of other trigger conditions associated with the initial trigger condition and the expected functional safety issues brought to the autonomous driving vehicle by the trigger condition event chain are predicted.

Images