An automatic driving scene equivalent construction method for cross-region rule adaptation

Abstract

The application relates to the technical field of traffic control systems and discloses an automatic driving scene equivalent construction method for cross-region rule adaptation, which comprises the following steps: acquiring multi-source road test data and constructing enhanced state data, performing scene slicing based on a running design domain judgment result and an event triggering result to obtain a road test scene segment; performing structured representation on the road test scene segment and writing the road test scene segment into a road scene library to screen representative scenes; generating candidate site working condition parameters according to the representative scenes, calculating a regional rule semantic deviation value, constructing an equivalent target function to solve optimal site working condition parameters, and generating an equivalent test scene for traffic scene reproduction control; and the scheme solves the problem that vehicle trajectory reproduction accuracy and cross-region traffic rule semantic consistency are difficult to be considered in test scene construction.

Description

Technical Field

[0001] This specification relates to the field of traffic control system technology, and in particular to an equivalent construction method for autonomous driving scenarios with cross-regional rule adaptation. Background Technology

[0002] In the research and development and verification of autonomous driving systems, it is necessary to transform typical traffic scenarios from road test data into test scenarios that can be reproduced in closed test sites. Since roads in different regions differ in terms of traffic direction, road signs and markings, right-of-way relationships, and traffic rule semantics, how to construct equivalent test scenarios across regions has become a significant technical challenge in autonomous driving testing.

[0003] In existing autonomous driving scenario construction schemes, some schemes primarily rely on behavioral indicators such as vehicle trajectory, speed, and acceleration for scenario extraction and site reproduction. By adjusting the vehicle's initial position, speed, or travel path, the reproduced scenario is made to closely resemble the kinematics of a real road scenario. However, such schemes typically lack constraints on the semantics of regional traffic rules, easily overlooking the consistency between the direction of travel, yield relationships, the meaning of signs and markings, and right-of-way relationships and the original road scenario. This results in the test scenario, while closely resembling the original road scenario in terms of trajectory, speed, or spacing, deviating from the semantics of traffic rules.

[0004] Meanwhile, overemphasizing regional traffic rules during the site reproduction process may be limited by factors such as the road structure and vehicle layout of the test site, making it difficult to accurately reproduce behavioral indicators such as vehicle trajectory, speed, and relative distance. Therefore, existing methods still face the challenge of simultaneously satisfying the accuracy of vehicle behavior indicator reproduction and the semantic consistency of cross-regional traffic rules during the equivalent construction of road scenarios into test scenarios. There is an urgent need for an equivalent construction method for autonomous driving scenarios that considers both the accuracy of behavioral indicators and the semantic deviation of regional rules. Summary of the Invention

[0005] In view of this, the present invention aims to propose an equivalent construction method for autonomous driving scenarios with cross-regional rule adaptation, in order to solve the problem that the existing road scenario to site test scenario equivalent construction process lacks quantitative constraints on the semantic differences of regional rules, which makes it difficult for the site reproduction scenario to simultaneously meet the requirements of behavioral indicator reproduction and regional rule semantic consistency.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for constructing equivalent autonomous driving scenarios with cross-regional rule adaptation includes: S1: Acquire multi-source road test data, parse and process the multi-source road test data, and construct semantically enhanced state based on the parsed and processed multi-source road test data to obtain enhanced state data containing regional rule semantic information; S2: Determine the operation design domain judgment result and event triggering result based on the enhanced state data, and perform scene slicing on the enhanced state data according to the operation design domain judgment result and event triggering result to obtain road test scene segments; S3: Perform structured representation on the road test scene segment, generate scene index vector and indicator vector, and write the scene index vector and indicator vector into the road scene library; S4: Based on the scene index vector and indicator vector, select representative scenes from the road scene library; S5: Generate candidate site condition parameters based on the representative scenario, and calculate the regional rule semantic deviation value between the candidate site condition parameters and the representative scenario by constructing a regional rule semantic consistency cost term; S6: Construct an equivalent objective function, and under the constraints of site conditions, solve for the optimal site conditions from the candidate site conditions parameters, so that the site reproduction scenario corresponding to the optimal site conditions parameters meets the deviation requirements, and construct an equivalent test scenario based on the optimal site conditions parameters.

[0007] Furthermore, S1 includes: S1-1: Acquire multi-source road test data and preprocess the multi-source road test data to obtain preprocessed multi-source road test data; S1-2: Perform regional rule semantic parsing on the preprocessed multi-source road test data to extract regional rule semantic vectors; S1-3: Perform scene state parsing processing on the preprocessed multi-source road test data to extract the scene state vector; S1-4: Concatenate and splice the scene state vector and the region rule semantic vector at the same time to construct a semantically enhanced state, and obtain enhanced state data containing region rule semantic information.

[0008] Furthermore, S2 includes: S2-1: Based on the scene state vector and regional rule semantic vector in the enhanced state data, determine the operation design domain judgment result; the operation design domain judgment result includes the basic operation design domain judgment result and the regional rule operation design domain judgment result; S2-2: Construct a joint event triggering function based on the scene state change amount and the region rule semantic change amount in the enhanced state data, and determine the event triggering result based on the comparison result of the joint event triggering function and the preset triggering threshold; S2-3: Determine the slice boundary based on the operation design domain judgment result and the event trigger result, and slice the enhanced state data into scene segments based on the slice boundary to obtain road test scene segments.

[0009] Furthermore, in S2-1, the result of the operational design domain determination is determined based on the following formula:

[0010] In the formula, For a moment The results of the operational design domain determination; For a moment The basic operational design domain determination result; For a moment The regional rule operation design domain judgment result; Among them, the regional rule operation design domain determination result It is obtained by determining the allowed values ​​of the semantic components of the region rules, and its calculation logic is as follows:

[0011] In the formula, For a moment The first in the semantic vector of the region rule Semantic components of regional rules; For the first The set of allowed values ​​corresponding to the semantic components of each region rule; Indicates characteristic functions; The number of semantic components of the regional rules that participate in the decision-making of the operational design domain.

[0012] Furthermore, in S2-2, the event triggering result is determined based on the comparison result between the joint event triggering function and the preset triggering threshold, and its calculation logic is as follows:

[0013] In the formula, For the event At any moment The result of the event triggering; For a moment The scene state vector; For a moment The semantic vector of the region rules; This represents the basic event discrimination term obtained based on the scene state vector; This represents the rule semantic event discrimination term obtained based on the region rule semantic vector; Indicates the semantic trigger weight of the rule; Indicates the preset trigger threshold; Indicates characteristic functions; Among them, the basic event discrimination item Determined by the changes in scene state, its calculation logic is as follows:

[0014] In the formula, Indicates the change in velocity; Indicates longitudinal or lateral acceleration; Indicates the change in heading angle; Indicates the collision time; This indicates the change in relative distance; Indicates the corresponding weight coefficient; Among them, the rule semantic event discrimination item The calculation logic for determining the semantic change of the region rules is as follows:

[0015] In the formula, , , , These represent the determination quantities for whether the direction of travel, rule type, semantics of road signs and markings, and right-of-way relationships have changed, respectively. This represents the corresponding weighting coefficient.

[0016] Furthermore, S4 includes: S4-1: Determine the risk score of the road test scene segment based on the indicator vector, and determine the rarity score and coverage score of the road test scene segment based on the scene index vector; S4-2: Determine the semantic coverage score of regional rules based on the semantic information of regional rules in the scene index vector; S4-3: Calculate the comprehensive score of the road test scene segment based on the risk score, rarity score, coverage score, and regional rule semantic coverage score; S4-4: Based on the comprehensive score, candidate representative scenes are selected from the road scene library, and the candidate representative scenes are further selected according to the preset coverage constraints to obtain representative scenes.

[0017] Furthermore, in S4-3, the comprehensive score is determined based on risk score, rarity score, coverage score, and regional rule semantic coverage score, and its calculation logic is as follows:

[0018] In the formula, Indicates the first A comprehensive score for each road test scenario segment; This represents the risk score obtained based on the indicator vector; This represents the rarity score obtained based on the scene index vector; Indicates the first Each road test scenario segment scores the overall scene space coverage. Indicates the first Regional rule semantic coverage scoring for a road test scenario segment; Indicates the first Semantic summary of scene fragment-level region rules corresponding to each road test scene fragment; This represents the corresponding weighting coefficient.

[0019] Furthermore, in S5, the calculation logic for the semantic deviation value of the region rule is as follows:

[0020] In the formula, Indicates the first Candidate site condition parameters corresponding to each representative scenario The semantic deviation value of the region rules between the representative scene and the scene; This is due to a deviation in the consistency of the direction of travel. This is due to a deviation in rule type consistency. This refers to semantic consistency deviations in road signs and markings. This indicates a deviation in the consistency of right-of-way relationships. , , , These are the weighting coefficients for the corresponding deviation terms; The candidate set of site condition parameters is determined based on the semantic deviation value of the regional rules and the site condition constraints. The calculation logic is as follows:

[0021] In the formula, For the first A set of candidate site condition parameters corresponding to each representative scenario; For the first Site operating condition constraints; The number of site condition constraints; This is the semantic deviation threshold for regional rules.

[0022] Furthermore, the equivalent objective function is constructed based on the index vector corresponding to the representative scenario, the index estimation vector corresponding to the candidate site condition parameters, and the semantic deviation value of the regional rules.

[0023] In the formula, Indicates the first The optimal site condition parameters corresponding to each representative scenario; For the first Candidate site condition parameters corresponding to each representative scenario; For the first A vector of indicators corresponding to each scenario. Based on candidate site operating parameters The obtained index estimation vector; This is the indicator weight matrix; Indicates the first The index vector corresponding to each representative scenario and the candidate site condition parameters The index deviation between the corresponding index estimation vectors; This represents the weight coefficient corresponding to the semantic deviation value of the regional rules.

[0024] Furthermore, if the equivalent test scenario constructed from the optimal site condition parameters described in S6 does not meet the deviation requirements, a closed-loop correction is performed on the optimal site condition parameters. The calculation logic is as follows:

[0025] In the formula, Indicates the first The first closed-loop correction of the wheel The corrected optimal site condition parameters correspond to each representative scenario; Indicates the first The first closed-loop correction of the wheel The corrected optimal site condition parameters correspond to each representative scenario; This indicates a parameter update function; Indicates the first The index deviation value under closed-loop correction; Indicates the first Semantic deviation value of regional rules under closed-loop correction.

[0026] The principles and advantages of this invention are: 1. By parsing and processing multi-source road test data, enhanced state data containing regional rule semantic information is constructed, enabling regional rule semantics such as traffic direction, signs and markings, and right-of-way relationships to participate in subsequent scene slicing and equivalent construction together with the road scene state, thus solving the problem that regional rule semantics are difficult to effectively express and utilize in existing methods.

[0027] 2. By slicing enhanced state data into scenarios based on the results of the operation design domain judgment and the event triggering results, and by performing structured representation of road test scenario fragments, scenario index vectors and indicator vectors are generated and written into the road scenario library. This enables the formation of a searchable and filterable scenario data foundation from the road test data, providing support for subsequent representative scenario selection and site condition parameter generation.

[0028] 3. By constructing a regional rule semantic consistency cost term, the regional rule semantic difference between the candidate site condition parameters and the representative scenario is quantified into a regional rule semantic deviation value. This regional rule semantic deviation value and the index deviation are jointly incorporated into the equivalent objective function. Under the site condition constraints, the optimal site condition parameters are solved, thereby constructing an equivalent test scenario that meets the deviation requirements. This avoids the problem of rule semantic distortion caused by only pursuing behavior index matching, or the problem of decreased behavior index reproduction accuracy caused by only maintaining rule semantic consistency. Attached Figure Description

[0029] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is a flowchart illustrating an equivalent construction method for autonomous driving scenarios that adapts to cross-regional rules. Figure 2 This is an example graph showing the variation of the weight coefficients of the basic event discrimination term under different validation data partitions. Detailed Implementation

[0030] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0031] As indicated in this specification and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0032] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0033] The following detailed explanation illustrates the specific implementation methods: like Figure 1As shown, an equivalent construction method for autonomous driving scenarios with cross-regional rule adaptation includes: S1: Obtain multi-source road test data, parse and process the multi-source road test data, and construct semantically enhanced state based on the parsed and processed multi-source road test data to obtain enhanced state data containing regional rule semantic information.

[0034] S1 includes: S1-1: Acquire multi-source road test data and preprocess the multi-source road test data to obtain preprocessed multi-source road test data.

[0035] Multi-source road test data refers to relevant data collected by autonomous vehicles during road testing that can characterize the road environment, vehicle operating status, and interactions among traffic participants.

[0036] For example, multi-source road test data may include one or more of the following: vehicle positioning data, vehicle speed data, vehicle acceleration data, heading angle data, lane line data, road sign and marking data, traffic participant location data, traffic participant speed data, vehicle relative distance data, and road scene timestamp data.

[0037] In this embodiment, multi-source road test data can be obtained in a variety of ways.

[0038] For example, data can be collected through one or more of the following: cameras, millimeter-wave radar, lidar, integrated navigation equipment, onboard controllers, and vehicle-to-everything (V2X) devices installed on autonomous vehicles. Alternatively, it can be obtained from one or more of the following: road test logs, vehicle operation records, map data files, and road test annotation files.

[0039] In this embodiment, after obtaining multi-source road test data, in order to reduce the impact of differences in acquisition frequency, timestamps and data formats of different data sources on the subsequent analysis results, the multi-source road test data is preprocessed.

[0040] In this embodiment, preprocessing includes time synchronization, outlier removal, and data format unification of multi-source road test data, so that data from different sources can be subsequently analyzed under the same time reference and spatial coordinate system.

[0041] First, time synchronization processing is performed on the multi-source road test data. Specifically, the road scene timestamps can be used as a unified time reference to establish a target time series.

[0042] The target time series refers to the time sequence obtained by arranging the start timestamp to the end timestamp of the road test data according to a preset time interval, and is used as a unified time axis after aligning multi-source road test data.

[0043] For example, the start timestamp in multi-source road test data is recorded as The preset time interval is denoted as Then the target time series can be represented as ( , ,…, );in, The timestamp should not exceed the end timestamp corresponding to the multi-source road test data.

[0044] in, The time resolution can be preset based on the acquisition cycle of multi-source road test data or the time resolution required for scene slices. For example, when the acquisition cycle of the main vehicle state data in the multi-source road test data is 0.1s, and the scene slices need to identify vehicle acceleration / deceleration, steering, or relative distance changes at the 0.1s level, the time resolution can be preset. Set to 0.1s; when the collection period for the main vehicle status data is 0.05s, it can also be set to 0.1s. Set to 0.05s.

[0045] After establishing the target time series, for continuous data such as vehicle speed, vehicle acceleration, heading angle, vehicle positioning data, and traffic participant positions and speeds, linear interpolation is used to align them to the target time series. For discrete semantic data such as lane line data, road sign and marking data, and road rule area identifiers, nearest neighbor matching is used to match the valid semantic data closest to the target timestamp to the corresponding time. This ensures that multi-source road test data from different collection frequencies and sources correspond under the same time reference.

[0046] Specifically, for nearest neighbor matching of discrete semantic data, the absolute value of the time difference between the collection timestamp of each discrete semantic data and the target timestamp can be calculated, and the valid semantic data with the smallest absolute value of the time difference can be selected as the matching result at the corresponding time. When there are multiple candidate data with the same absolute value of time difference, the data with the earlier collection time is selected first. When the time difference exceeds the preset allowable range, the corresponding semantic data is marked as invalid.

[0047] Secondly, outlier removal is performed. Specifically, for kinematic data such as vehicle speed, vehicle acceleration, heading angle, relative distance, and speed of traffic participants, a combination of physical threshold verification and sliding window statistical verification is used.

[0048] For example, the mean and standard deviation of the data are calculated within a preset sliding window. When the current data exceeds the range of the mean plus or minus a preset multiple of the standard deviation, it is marked as an outlier. For data marked as outliers, they can be replaced by linear interpolation results of adjacent valid data, the previous valid value, or the next valid value. For discrete semantic data such as lane line data, road sign and marking data, and traffic participant categories, outlier removal can be performed through continuity checks. If the semantic data of the current frame undergoes a sudden change that does not conform to the continuity of vehicle motion compared to the semantic data of the previous frame, the semantic data of the current frame is marked as an outlier, and it is replaced by the semantic data of adjacent valid frames or the previous valid semantic data is retained.

[0049] The preset sliding window is a local time window composed of multiple consecutive target time points in the target time series. The length of the preset sliding window is based on a preset time interval of the target time series. Settings. For example, when the target time series has a preset time interval. When set to 0.1s, the current time and the two target time points before and after it can be combined into a preset sliding window, that is, the preset sliding window contains 5 consecutive target time points, and the corresponding window duration is 0.5s.

[0050] After time synchronization and outlier removal, the data format is then standardized.

[0051] Specifically, for spatial data, vehicle positioning data, lane line data, road sign and marking data, and traffic participant location data are uniformly converted to the same coordinate system; for example, they can be first converted to a map coordinate system, and then converted to a vehicle body coordinate system with the vehicle as the reference based on the vehicle's current position and heading angle.

[0052] For physical quantity data, vehicle speed is standardized to meters per second, vehicle acceleration is standardized to meters per second squared, heading angle is standardized to radians or degrees, and distance is standardized to meters.

[0053] For semantic data, lane line types, road sign and marking types, traffic participant categories, and road rule area identifiers are converted into unified codes according to preset coding rules. Preset coding rules refer to the pre-established correspondence rules between semantic categories and coded values.

[0054] Specifically, corresponding coding fields can be set for lane line types, road sign and marking types, traffic participant categories, and road rule area identifiers, and an enumeration coding method can be used to encode each semantic category. The same semantic category corresponds to a unique code, and different semantic categories correspond to different codes.

[0055] As an example only, lane line types can be encoded as follows: 0 represents an unknown lane line, 1 represents a solid line, 2 represents a dashed line, 3 represents double solid lines, and 4 represents a dashed-solid line; road sign and marking types can be encoded as follows: 0 represents an unknown sign or marking, 1 represents a yield sign, 2 represents a stop sign, 3 represents a speed limit sign, 4 represents a lane arrow, and 5 represents a no-entry sign; traffic participant categories can be encoded as follows: 0 represents an unknown traffic participant, 1 represents a motor vehicle, 2 represents a non-motorized vehicle, and 3 represents a pedestrian; road rule area identifiers can be encoded as follows: 0 represents an unknown area, 1 represents a right-hand traffic area, 2 represents a left-hand traffic area, 3 represents a signal-controlled area, 4 represents an unsignalized intersection area, and 5 represents a roundabout area. Thus, semantic data can be converted into a unified encoding format, facilitating subsequent semantic parsing of area rules, extraction of semantic components, and construction of semantic vectors for area rules.

[0056] This yields preprocessed multi-source road test data that maintains consistency in time reference, spatial coordinate system, physical units, and field format.

[0057] S1-2: Perform regional rule semantic parsing on the preprocessed multi-source road test data to extract the regional rule semantic vector.

[0058] The regional rule semantic vector is a vectorized representation of the semantic information of regional rules, used to characterize one or more regional traffic rule features in a road scenario, such as traffic direction, rule type, road sign and marking semantics, and right-of-way relationship.

[0059] Regional rule semantic parsing processing refers to the process of extracting semantic information related to regional traffic rules from preprocessed multi-source road test data and converting the semantic information into computable vector data.

[0060] In this embodiment, the semantic parsing process for region rules includes: A-1. Based on the map road attributes, vehicle positioning data, and road scene timestamps in the multi-source road test data, determine the area information corresponding to the current road scene.

[0061] Among them, map road attributes refer to map attribute information related to the spatial location of roads and road traffic rules; map road attributes may include one or more of the following: road boundaries, lane center lines, lane numbers, lane travel directions, road types, administrative regions to which the roads belong, road speed limits, lane turning attributes, road sign and marking attributes, and road rule area identifiers.

[0062] During matching, the target time to be parsed can be determined first based on the timestamp of the road scene, and the vehicle positioning data at the target time can be converted to the map coordinate system to obtain the current location of the vehicle. Then, the spatial relationship between the current location of the vehicle and the road boundary, lane center line or lane area in the map can be calculated, and the road or lane that is closest to the vehicle and meets the condition that the vehicle's heading angle is consistent with the lane's travel direction can be determined as the road or lane where the vehicle is located. Then, the administrative region to which the road belongs, the road rule area identifier, the lane travel direction and the road sign and marking attributes corresponding to the road or lane can be read to determine the area information corresponding to the current road scene.

[0063] A-2. Retrieve the corresponding regional rule semantic mapping table based on the regional information. The regional rule semantic mapping table is a pre-established structured mapping table that pre-stores the correspondence between regional information and traffic direction, rule type, meaning of road signs and markings, and right-of-way relationships.

[0064] Specifically, the regional rule semantic mapping table uses regional information as the index and traffic direction, rule type, meaning of road signs and markings, and right-of-way relationship as the mapping result.

[0065] Among them, the direction of travel is used to represent the left-hand or right-hand traffic rules corresponding to the current area; the rule type is used to represent the yield rule, priority rule, signal control rule, roundabout rule or pedestrian priority rule applicable to the corresponding road scenario under the current area; the meaning of road signs and markings is used to represent the meaning of road signs and markings under the rules of the current area; the right-of-way relationship is used to represent the priority relationship between the vehicle and traffic participants under the rules of the current area.

[0066] In this embodiment, right-of-way relationships are represented by priority values. The larger the value, the higher the priority. By comparing the priority values ​​of the vehicle and other road users, the right-of-way relationship between them can be determined.

[0067] For example, in signalized intersections: vehicles traveling on a green light have a priority of 3, vehicles on a red light have a priority of 1, and pedestrians on a green light have a priority of 4. In unsignalized intersections: vehicles going straight on main roads have a priority of 3, vehicles on side roads have a priority of 1, and vehicles turning left have a priority of 2, with straight traffic taking precedence over left turns. In roundabouts: vehicles traveling within the roundabout have a priority of 3, and vehicles entering the roundabout have a priority of 1. In pedestrian crossings: pedestrians within the crossing have a priority of 4, and motor vehicles have a priority of 1. By comparing the priority values ​​of the vehicle and other road users, the right-of-way relationship between them can be determined.

[0068] Therefore, by calling the regional rule semantic mapping table, the regional information corresponding to the current road scene can be converted into the traffic direction, rule type, meaning of road signs and markings, and right-of-way relationship required for the subsequent construction of regional rule semantic vectors.

[0069] In this embodiment, the area information may include the area code and the road scene element type.

[0070] For illustrative purposes only, the regional rule semantic mapping table is a three-level structured mapping table. It uses administrative division codes (GB / T2260 standard) as the first-level index, road scene element types as the second-level index, and traffic direction, rule type, meaning of road signs and markings, and right-of-way relationships as the mapping results. The core table structure is shown in the table below:

[0071] A-3. Based on the regional rule semantic mapping table, semantic recognition is performed on road elements and traffic participation relationships in the current road scene to obtain traffic direction, rule type, road sign and marking semantics and right-of-way relationships.

[0072] Specifically, the current road scene's traffic direction is determined based on the lane's direction of travel on the road or lane where the vehicle is located; the rule type corresponding to the current road scene is determined based on the road type, lane turning attributes, and road scene location corresponding to the road or lane where the vehicle is located; the semantic meaning of road signs and markings under the current regional rules is determined based on the attributes of road signs and markings and in conjunction with the regional rule semantic mapping table; and the right-of-way relationship between the vehicle and other traffic participants is determined based on the relative positional relationship and movement sequence between the vehicle and other traffic participants on the target road, target lane, or conflict area and in conjunction with the regional rule semantic mapping table.

[0073] A-4. According to the preset coding rules, the identified traffic direction, rule type, road sign and marking semantics and right-of-way relationship are converted into corresponding semantic components, and the semantic components are combined to form a regional rule semantic vector.

[0074] The preset encoding rules can be enumeration encoding, one-hot encoding, binary encoding, numerical level encoding, or a combination thereof.

[0075] As an example only, the semantic vector of a region rule is represented as follows:

[0076] in, Indicates time The corresponding region rule semantic vector; This represents a regional semantic component, used to characterize the region or specific jurisdiction to which the current road scene belongs; The semantic component representing the direction of travel is used to indicate whether travel is on the left or on the right. Represents the semantic component of the rule type, used to characterize priority passage rules, yielding rules, turning rules, or parking control rules; It represents the semantic components of road signs and markings, used to characterize the meaning of road facilities such as lane lines, guide lines, no-entry signs, lane arrows or text prompts under the rules of the corresponding area; This represents the semantic component of right-of-way relationship, used to characterize the priority relationship between this vehicle and other traffic participants in a specific traffic scenario.

[0077] S1-3: Perform scene state parsing on the preprocessed multi-source road test data to extract the scene state vector.

[0078] A scene state vector is a vector related to the vehicle's kinematic state. For example, a scene state vector may include the vehicle's speed, acceleration, heading angle, target vehicle's speed, and target vehicle's acceleration.

[0079] The target vehicle can be determined based on factors such as its relative distance from the vehicle, its relative speed, whether it is located on the vehicle's planned path or within a conflict area, and whether it is involved in triggering the current event.

[0080] After obtaining the preprocessed multi-source road test data, the scene state parsing process for the data related to vehicle motion state includes: extracting the motion state data of the vehicle and the target vehicle from the preprocessed multi-source road test data according to a unified timestamp, and associating the motion state data of the vehicle and the target vehicle at the same time with the timestamp; then, performing numerical processing and normalization on the extracted motion state data, and arranging them according to the preset state quantity order to obtain the scene state vector at the corresponding time.

[0081] The preset state quantity sequence can be set according to the vehicle kinematic feature extraction requirements, and this manual does not impose specific restrictions on it.

[0082] S1-4: Concatenate and splice the scene state vector and the region rule semantic vector at the same moment to construct the semantically enhanced state and obtain the enhanced state data containing the region rule semantic information.

[0083] Enhanced state data refers to the data formed by fusing the scene state vector and the region rule semantic vector at the same moment.

[0084] Cascading splicing is the process of connecting the scene state vector and the region rule semantic vector at the same moment in a preset splicing order to form a semantically enhanced state that simultaneously contains vehicle kinematic state and region rule semantic information.

[0085] As an example only, let time be... The corresponding scene state vector is The semantic vector of the region rule is Then enhance the state data This can be expressed by the vector concatenation formula as follows:

[0086] In this embodiment, this step provides a data foundation that simultaneously includes vehicle kinematics state and regional rule semantic information for subsequent operation design domain determination, event trigger determination, and scene slicing, enabling subsequent processing to take into account both the road scene state and the regional rule semantic features.

[0087] S2: Determine the operation design domain judgment result and event trigger result based on the enhanced state data, and slice the enhanced state data into scene segments according to the operation design domain judgment result and event trigger result to obtain road test scene segments.

[0088] S2 includes: S2-1: Scene state vectors based on enhanced state data and region rule semantic vector Determine the results of the operational design domain judgment.

[0089] The operational design domain refers to the range of road scenarios in which an autonomous driving system is designed to operate normally.

[0090] The runtime design domain determination result is a numerical determination result used to determine whether the current road scene belongs to the road scene range, and is used to provide an effective scene determination basis for subsequent scene slicing.

[0091] In this embodiment, the operation design domain determination result includes the basic operation design domain determination result and the regional rule operation design domain determination result.

[0092] The basic operation design domain determination result is used to characterize whether the current road scenario meets the requirements of the operation design domain in terms of vehicle kinematics.

[0093] The regional rule operation design domain determination result is used to characterize whether the current road scene meets the requirements of the operation design domain in terms of regional rule semantics.

[0094] In S2-1, the result of the design domain determination is based on the following formula:

[0095] In the formula, For a moment The results of the operational design domain determination; For a moment The basic operational design domain determination result; For a moment The regional rules are used to determine the domain decision results.

[0096] In this embodiment, All can be represented by binary numerical values, where a value of 1 indicates that the corresponding operational design domain requirements are met, and a value of 0 indicates that the corresponding operational design domain requirements are not met.

[0097] Specifically, when and When both are 1, A value of 1 indicates that the current road scene belongs to the operational design domain, and the enhanced state data corresponding to that moment will be used as valid data in subsequent event triggering judgments and scene slicing.

[0098] When any one of them is 0 A value of 0 indicates that the current road scene does not belong to the operational design domain. The enhanced state data corresponding to that moment is marked as data outside the operational design domain, which will serve as the basis for subsequent scene slice boundary judgment or invalid scene removal.

[0099] Among them, the basic operation design domain determination result can be based on the scenario state vector. Whether the vehicle's kinematic state variables meet the corresponding preset allowable range is determined.

[0100] Specifically, the scene state vector One or more state variables among velocity, acceleration, heading angle, relative distance, and relative velocity are compared with their corresponding preset allowable ranges; when all state variables meet their corresponding preset allowable ranges, the basic operation design domain judgment result is determined. Set to 1; when any state variable does not meet the corresponding preset allowable range, determine the basic operation design domain judgment result. The value is 0. The preset allowable range can be preset according to the design and operating conditions of the autonomous driving system, the requirements of the road test task, the type of test scenario, or the vehicle's safe operation requirements. This manual does not impose specific restrictions on this.

[0101] Among them, the regional rule operation design domain judgment result It is obtained by determining the allowed values ​​of the semantic components of the region rules, and its calculation logic is as follows:

[0102] In the formula, For a moment The first in the semantic vector of the region rule Semantic components of regional rules; For the first The set of allowed values ​​corresponding to the semantic components of each region rule; Indicates characteristic functions; The number of semantic components of the regional rules that participate in the decision-making of the operational design domain.

[0103] In this embodiment, the set of allowed values ​​is... It can be determined jointly in the following ways.

[0104] For example, it can be determined based on a region rule semantic mapping table or an autonomous driving test rule base. Specifically, it is based on the region information corresponding to the current road scenario and the first... For each region rule semantic component, rule values ​​matching the region information and the region rule semantic component are selected from the region rule semantic mapping table or the autonomous driving test rule base, and the selected rule values ​​are used as the first... The set of allowed values ​​corresponding to the semantic components of each region rule.

[0105] Alternatively, based on the labeled or verified samples in historical road test data, the effective value range or high-frequency value set of the rule semantic components in each region can be statistically obtained.

[0106] In practical applications, the above information can be intersected or weighted and fused to obtain the set of allowed values ​​corresponding to the semantic components of each region's rules.

[0107] Based on the above formula, time... semantic components of rules in each region Each with the corresponding set of allowed values Perform a matching judgment; when all semantic components of the regional rules participating in the operational design domain determination belong to the corresponding allowed value set, the value of each indicator function is 1, and the regional rule operational design domain determination result is... A value of 1 indicates that the current road scenario meets the operational design domain requirements in terms of regional rule semantics; when any regional rule semantic component does not belong to the corresponding set of allowed values, the corresponding indicator function is 0, and the regional rule operational design domain determination result is... A value of 0 indicates that the current road scenario does not meet the requirements of the operational design domain in terms of regional rule semantics.

[0108] S2-2: Construct a joint event triggering function based on the scene state change and regional rule semantic change in the enhanced state data, and determine the event triggering result based on the comparison result between the joint event triggering function and the preset triggering threshold.

[0109] Event trigger result refers to the judgment result used to determine whether a key event requiring scene slicing has occurred at the current moment.

[0110] In S2-2, the event triggering result is determined based on the comparison between the joint event triggering function and the preset triggering threshold. The joint event triggering function is as follows:

[0111] In the formula, For the event At any moment The result of the event triggering; For a moment The scene state vector; For a moment The semantic vector of the region rules; This represents the basic event discrimination term obtained based on the scene state vector; This represents the rule-based semantic event discrimination term obtained based on the region rule semantic vector; Indicates the semantic trigger weight of the rule; Indicates the preset trigger threshold; Represents the characteristic function.

[0112] Among them, rule semantic trigger weight and preset trigger threshold The values ​​can be determined based on historical labeled event samples, expert experience rules, and validation set parameter tuning results. Preferably, initial values ​​can be given first based on existing testing experience, and then calibration can be performed using labeled slice samples through cross-validation or quantile statistics to ensure that the event triggering results are consistent with the manually labeled results.

[0113] According to this formula, in determining the basic event discrimination terms and rule semantic event discrimination item Next, both are normalized to the same numerical range, and then the weights are triggered according to the semantic rules. The weighted values ​​are then added to the basic event criteria to obtain the event. At any moment The combined event trigger value is compared with a preset trigger threshold. When the combined event trigger value is greater than or equal to the preset trigger threshold... When the event is triggered, the result is... A value of 1 indicates an event. At any moment Triggered; when the combined event trigger value is less than the preset trigger threshold. When the event is triggered, the result is... A value of 0 indicates an event. At any moment Not triggered.

[0114] Basic event discrimination items Determined by the amount of change in the scene state.

[0115] The scene state change refers to the changes in kinematic state quantities such as vehicle speed, acceleration, heading angle, relative distance, and relative velocity in the scene state vector within adjacent moments or a preset time window. The preset time window is set according to actual needs or experience.

[0116] In this embodiment, the scene state changes include velocity changes, longitudinal or lateral acceleration, heading angle changes, collision time, and relative distance changes.

[0117] Among them, changes in velocity, acceleration, and heading angle can reflect abrupt changes in the vehicle's motion state, such as acceleration, braking, and steering, while changes in relative distance and collision time can reflect the degree of proximity between the vehicle and the target vehicle and the potential risk of conflict. Therefore, selecting the above state variables as scene state change variables is beneficial for identifying key events that affect scene slices.

[0118] Basic event discrimination items The calculation logic is as follows:

[0119] In the formula, Indicates the change in velocity; Indicates longitudinal or lateral acceleration; Indicates the change in heading angle; Indicates the collision time; It represents the change in relative distance.

[0120] After calculating the parameters, these parameters are normalized or dimensionless, then substituted into the formula, and weighted summation is performed according to the corresponding weight coefficients to obtain the basic event discrimination term. This is used to characterize the degree of influence of the change in the vehicle's kinematic state at the current moment on the event triggering result.

[0121] In this embodiment, This represents the corresponding weight coefficient, used to characterize the degree of influence of the above parameters on the basic event discrimination item, satisfying... .

[0122] In this embodiment, the weight coefficients can be obtained as follows: First, multiple road test samples are acquired, and based on the vehicle motion state data in the road test samples, the normalized features corresponding to the speed change, acceleration change, heading angle change, reciprocal of collision time, and relative distance change are calculated respectively; then, the event triggering results corresponding to the road test samples are used as calibration targets, and the weight coefficients corresponding to each state change are determined through constraint optimization.

[0123] Specifically, 77 road test samples were obtained through the Automatic Emergency Braking (AEB) function testing process. First, the obtained road test samples underwent an integrity check, resulting in 74 valid road test samples that met the data integrity requirements. Then, based on the event triggering results in the samples, the basic event discrimination items were... The weights of each parameter in the algorithm are calibrated. Different road test samples are grouped for verification, and the weight coefficients are adjusted based on the verification results of each group. Finally, the weight coefficients corresponding to the basic event discrimination items are obtained as follows: =0.185957、 =0.299148、 =0.061288、 =0.443607、 =0.010000.

[0124] In this embodiment, to verify the stability of the weight coefficients in the basic event discrimination items, a five-fold cross-validation method is used to divide the road test data. Specifically, the road test data that meets the data integrity requirements is divided into 5 data subsets. Each time, 4 of the data subsets are selected for weight coefficient calibration, and the remaining data subset is used for verification. This process is repeated 5 times to obtain 5 sets of weight coefficients for the basic event discrimination items.

[0125] like Figure 2 As shown, the horizontal axis represents the data partition number corresponding to the five verification processes, and the vertical axis represents the value of the corresponding weight coefficient. Different colored curves represent the weight coefficients corresponding to the changes in velocity, acceleration, heading angle, reciprocal of the collision time, and relative distance, respectively. Figure 2 It can be seen that the variation range of each weight coefficient is small under different data partitioning conditions. In different validation groups, the weight coefficient corresponding to the reciprocal of the collision time term is relatively large. The weighting coefficient corresponding to the change in acceleration is always kept at a high level. The fact that the reciprocal of the collision time and the change in acceleration have a significant impact on the triggering result of the basic event indicates that the collision time reciprocal and the change in acceleration have a significant impact on the triggering result of the basic event.

[0126] Therefore, the weight calibration process based on road test samples described above can reduce the deviation caused by manually setting weight parameters, enabling the basic event discrimination items to more accurately reflect the characteristics of vehicle motion state changes in road test scenarios. In this embodiment, vehicle speed, longitudinal or lateral acceleration, and heading angle can be read from the vehicle controller, integrated navigation device, or road test log; relative distance can be calculated based on vehicle positioning data.

[0127] Furthermore, the changes in velocity, heading angle, and relative distance can be calculated based on the differences between the corresponding state quantities at adjacent moments.

[0128] Collision time It can be calculated based on the relative distance and relative speed between the vehicle and the target vehicle. When the time is less than the preset minimum threshold, it can be The minimum time threshold is set to 0. When there is no tendency for the vehicle to collide with the target vehicle, the reciprocal of the collision time can be set to 0 or a preset low-risk value.

[0129] Rule-based semantic event discrimination item Determined by the semantic changes in regional rules.

[0130] Among them, the semantic change of regional rules refers to the change of semantic components in the semantic vector of regional rules within adjacent time moments or preset time windows.

[0131] In this embodiment, the semantic change of regional rules includes the determination of whether the direction of travel, rule type, road sign and marking semantics, and right-of-way relationship have changed.

[0132] Changes in traffic direction, rule type, semantic changes in road signs and markings, and changes in right-of-way relationships can reflect semantic shifts or rule constraint changes in road scenarios at the regional traffic rule level. Therefore, selecting the changes in the above semantic components as the semantic changes in regional rules is beneficial for identifying rule semantic events that affect scene slices.

[0133] Rule-based semantic event discrimination item The calculation logic is as follows:

[0134] In the formula, , , , These represent the determination quantities for whether the direction of travel, rule type, semantics of road signs and markings, and right-of-way relationships have changed.

[0135] This represents the corresponding weight coefficient, satisfying... =1. For example, For 0.20, For 0.30, For 0.20, It is 0.30.

[0136] in, Due to semantic variations in their corresponding categories, this embodiment employs a predefined 1-5 level engineering impact scale for weight allocation, and constrains its values ​​through normalization and perturbation verification. The levels only compare the relative impact within the same weighted group and do not represent the probability of occurrence.

[0137]

[0138] The engineering impact scales at levels 1 to 5 are used to evaluate the degree of impact of different rule semantic changes on the traffic decisions and rule semantic event triggering results of autonomous vehicles. The higher the level, the more direct the impact of the corresponding rule semantic change on the vehicle's decision-making logic, traffic constraints, and scene slice boundaries.

[0139] Specifically, Level 1 indicates that the semantic change mainly affects the road scene description information and does not directly change the vehicle's current traffic decision; Level 2 indicates that the semantic change mainly affects road facilities or signage information and needs to be judged in conjunction with other semantics; Level 3 indicates that the semantic change may affect the vehicle's driving path, direction of travel, or expected interaction; Level 4 indicates that the semantic change will directly change the traffic rules or control logic that the vehicle needs to follow; Level 5 indicates that the semantic change directly changes the priority relationship between the vehicle and other traffic participants, and has a direct impact on the vehicle's decision to yield, wait, or avoid.

[0140] Therefore, in determining When determining the weighting coefficients, we can first score the changes in traffic direction, rule type, semantic changes in road signs and markings, and changes in right-of-way relationships according to the above engineering impact scale, and then normalize the scoring results to obtain the corresponding weighting coefficients.

[0141] In this embodiment, the traffic direction, rule type, signs and markings, and right-of-way are assigned influence levels [3, 4, 2, 5] in sequence. Specifically, the engineering influence levels corresponding to the semantic changes of each rule are normalized by dividing each level score by the sum of all level scores, where the sum of the level scores is 14, to obtain the corresponding weight coefficients. =[3 / 14,4 / 14,2 / 14,5 / 14]=[0.214,0.286,0.143,0.357].

[0142] Among these, changes in right-of-way relationships received the highest score because they directly determine the priority relationship between vehicles and other traffic participants, directly impacting decisions regarding yielding, waiting, and avoiding. Changes in rule types received the second highest score because they may alter the traffic rules vehicles must follow, such as switching from ordinary traffic rules to yielding or priority rules. Changes in traffic direction primarily affect vehicle direction and road topology understanding, but under the same traffic conflict, they do not necessarily change the priority subjects among vehicles. Changes in the semantics of road signs and markings, after unified semantic parsing, are mainly used to assist in road element recognition and scene configuration, therefore their impact level is relatively low.

[0143] In this embodiment, , , , As change discrimination quantities, since the semantics of regional rules are discrete topological attributes rather than continuous physical quantities such as speed and acceleration, changes in regional rules in autonomous driving scenarios are essentially state transitions in a discrete state machine (e.g., from "arterial road priority" to "roundabout priority"). Therefore, each specific change discrimination quantity is directly determined by comparing whether the semantic components of regional rules change in adjacent moments or within a preset time window. That is, when the semantic components of traffic direction, rule type, road signs and markings, or right-of-way relationships change in adjacent moments or within a preset time window, the corresponding discrimination quantity is set to 1; when there is no change, the corresponding discrimination quantity is set to 0.

[0144] Specifically, in actual road tests, there are cases where all change discriminants are simultaneously zero. This means that within adjacent timeframes or a preset time window, the semantic environment of the area where the vehicle is located remains stable, and no rule switching occurs. When all four types of discriminants are simultaneously zero, the rule semantic event discriminant term... When the value is 0, the joint event triggering function is determined solely by the basic event criteria. This is determined by the normal triggering logic for regular driving scenarios with no irregular changes, and requires no additional special handling.

[0145] After determining each decision quantifier, according to the corresponding weight coefficients We perform a weighted summation to obtain the rule semantic event discriminant. .

[0146] S2-3: Determine the slice boundary based on the results of the operation design domain judgment and the event triggering results, and slice the enhanced state data into scene segments based on the slice boundary to obtain road test scene segments.

[0147] Slice boundaries refer to the start and end times used to divide road test scene segments.

[0148] When the result of the running design domain changes from 0 to 1, the corresponding moment is determined as the starting boundary of the road test scene segment; when the result of the running design domain changes from 1 to 0, the corresponding moment is determined as the ending boundary of the road test scene segment; when the event trigger result is 1, the corresponding moment is determined as the event boundary, and the enhanced state data is segmented according to the event boundary.

[0149] After determining the slice boundaries, the enhanced state data between adjacent slice boundaries is treated as a road test scenario segment. Thus, each road test scenario segment consists of enhanced state data from consecutive time points, and retains the vehicle kinematics and regional rule semantic information within that segment.

[0150] In this embodiment, the scene slices are no longer divided solely based on changes in the kinematic state such as vehicle acceleration / deceleration, steering, or approaching a target vehicle, but can also form new road test scene segments when the semantics of the regional rules change.

[0151] For example, when a vehicle enters an area with different traffic direction requirements, enters an intersection with different yield rules or priority rules, the semantics of road signs and markings change, or the right-of-way relationship between the vehicle and pedestrians, bicycles, or oncoming vehicles undergoes a substantial change, corresponding slice boundaries can be formed. This allows the obtained road test scene segments to simultaneously reflect changes in vehicle motion state and semantic changes in area rules, improving the consistency between the scene slicing results and the behavioral boundaries of real road scenes.

[0152] S3: Perform structured representation of road test scene segments, generate scene index vectors and indicator vectors, and write the scene index vectors and indicator vectors into the road scene library.

[0153] A scene index vector is a vector used to represent the searchable features of a road test scene segment. These searchable features include basic scene index features and semantic summaries of region rules.

[0154] Regional rule semantic summary refers to the summary information obtained by statistically analyzing, aggregating, or encoding the regional rule semantic vector sequence within a road test scenario segment. It is used to characterize the overall features of the road test scenario segment in terms of regional rule semantics.

[0155] An indicator vector is a vector used to characterize the evaluation indicator features of a road test scenario segment. These evaluation indicator features include basic indicator features and regional rule semantic indicator features.

[0156] This embodiment uses structured representation to generate vectorized labels containing basic traffic behavior features and regional rule semantic features for each road test scenario segment, and writes the vectorized labels into the road scenario library for subsequent scenario retrieval, similar scenario comparison and representative scenario selection.

[0157] Each segment of the road test scene obtained is given a structured representation and stored in a road scene library. Specifically: B-1, regarding the first A road test scenario clip Based on the regional rule semantic vectors within the corresponding time interval of this scene segment Sequence construction of region rule semantic summary Its definition is:

[0158] in, For the first Semantic summary of region rules corresponding to each road test scenario segment. Indicates the first A road test scenario segment; Indicates time Located in the Within the time interval of each road test scenario segment.

[0159] This represents a mapping function that aggregates and encodes a sequence of semantic vectors representing regional rules. It is used to encode the first... Semantic vectors of regional rules corresponding to multiple moments within a road test scenario segment Convert into a region rule semantic summary .

[0160] Specifically, For the A road test scenario clip When performing aggregation encoding on the sequence of region rule semantic vectors within a road test scene segment, the region rule semantic vectors at each time point can be read in chronological order first. Then, the semantic components of the region, traffic direction, rule type, road sign and marking, and right-of-way relationship are extracted. For any semantic component, the semantic component at the [missing information]th ... The values ​​at each time point within a road test scenario segment are analyzed, and the value with the most occurrences or the longest duration is determined as the summary component of that semantic component. When the semantic component changes within the road test scenario segment, the number of changes or the determination of whether a change has occurred can be used as auxiliary summary information for that semantic component. Thus, semantic summary components for the region, traffic direction, rule type, road signs and markings, and right-of-way relationships are obtained, and these summary components together form the regional rule semantic summary. .

[0161] Among them, extracting each semantic component can be done by reading the semantic vector of the region rules. This can be achieved by using preset positions or preset fields.

[0162] As an example only, semantic summary of regional rules It can be represented as:

[0163] in, , They represent the first The semantic summary components for each road test scenario segment include the region semantic summary component, the traffic direction semantic summary component, the rule type semantic summary component, the road sign and marking semantic summary component, and the right-of-way relationship semantic summary component.

[0164] After obtaining the semantic summary of the region rules Then, from the first Extracting basic scene index features from the enhanced state data corresponding to each road test scene segment .

[0165] Basic Scene Index Features These are the basic retrieval attributes used to describe road test scenario segments, such as road type, traffic participant type, dynamic behavior pattern, speed range, etc.

[0166] Among them, basic scene index features It can be done by the first The enhanced state data and map road attributes corresponding to each road test scene segment are statistically analyzed or encoded. Specifically, road type and lane type can be extracted based on map road attributes, traffic participant type and number can be extracted based on traffic participant identification results, dynamic behavior patterns can be determined based on changes in speed, acceleration, heading angle, and relative distance in the scene state vector sequence, and speed ranges can be determined based on maximum, minimum, or average speed values, thereby obtaining the basic scene index features. .

[0167] Then, it is combined with the semantic summary of region rules. By concatenating the elements, we obtain the scene index vector. .

[0168] It is represented as:

[0169] in, Indicates the first The scene index vector corresponding to each road test scene segment.

[0170] By using the above splicing method, regional rule semantic information can be incorporated into the scene index vector while preserving the basic scene index features.

[0171] Similarly, the basic indicator characteristics can be... Semantic index features of regional rules By concatenating the data, we obtain the index vector. It is represented as:

[0172] in, It represents the basic indicator characteristics and is used to describe one or more evaluation indicators among velocity change, acceleration change, trajectory deviation, interaction gap, collision risk, and scene duration. This represents the semantic features of regional rules, used to describe the semantic coverage, semantic changes, or changes in right-of-way relationships of regional rules in this road test scenario segment.

[0173] In this embodiment, basic indicator features It can be done through the first The sequence of scene state vectors in each road test scene segment is obtained through statistical calculation.

[0174] For example, the maximum, minimum, mean, change, or duration of state variables such as speed, acceleration, heading angle, relative distance, relative velocity, and collision time within the road test scenario segment can be statistically analyzed to obtain basic index characteristics. .

[0175] Semantic Indicator Features of Regional Rules It can be done by the first The sequence of semantic vectors representing regional rules in a road test scenario segment is obtained through statistical calculation.

[0176] For example, the occurrence, number of changes, or number of coverage categories of semantic components such as traffic direction, rule type, road sign and marking semantics, and right-of-way relationships within a road test scenario segment can be statistically analyzed to obtain the semantic index features of regional rules. .

[0177] Subsequently, the first Scene index vector corresponding to each road test scene segment and indicator vector The start and end times, data source information, and other related information are stored together and written into the road scene library.

[0178] In this embodiment, the semantic summary of regional rules is used. Incorporate scene index vector and the semantic index features of regional rules Inclusion of indicator vector This enables road test scene fragments in the road scene database to simultaneously possess basic traffic behavior representation capabilities and regional rule semantic representation capabilities during subsequent retrieval, similar scene comparison, representative scene selection, and equivalent construction analysis.

[0179] S4: Based on scene index vector and indicator vector Representative scenes are selected from the road scene library.

[0180] A representative scenario refers to a scenario selected from a road scenario library that can characterize the main features of a class of road test scenario segments.

[0181] In particular, in this embodiment, representative scenarios are not only selected based on risk level, frequency of occurrence, or similarity of basic traffic behavior, but also further selected for road test scenario segments that have a low frequency of occurrence but have obvious regional rule semantic features.

[0182] Among them, road test scenario fragments with obvious regional rule semantic features can include special scenario fragments formed due to differences in traffic direction, yield rules, priority rules, road sign and marking semantics, or right-of-way relationships in different regions. By including these low-frequency but rule-specific road test scenario fragments in representative scenarios, it is possible to avoid the representative scenario set only covering common traffic behaviors and omitting rule semantic difference scenarios that have an important impact on cross-regional equivalent testing.

[0183] The filtering method is as follows, S4 includes: S4-1: Based on index vectors Determine the risk score of road test scenario segments and based on scene index vector Determine the rarity score of road test scene segments and coverage score .

[0184] In this embodiment, risk scoring According to the The index vector corresponding to each road test scenario segment Sure.

[0185] Specifically, this can be achieved from the indicator vector. Risk-related indicators are extracted to obtain a risk indicator subvector. The risk-related indicators may include one or more of the following: collision risk indicators, comfort risk indicators, rule conflict risk indicators, and abnormal behavior risk indicators. Then, the risk-related indicators are normalized and weighted and summed according to preset risk weights to obtain the risk score of the road test scenario segment.

[0186] Among them, the preset risk weights can be pre-set based on the impact of each risk-related indicator on the degree of danger of the scenario, the statistical results of historical road test data, or the requirements of autonomous driving test tasks.

[0187] The initial risk weights are set using an impact level of 1 to 5. The impact level is determined comprehensively based on the degree of influence of each risk indicator on the safety of autonomous driving scenarios, vehicle decision-making behavior, and subsequent testing and handling. Specifically, if a risk indicator directly affects the vehicle's collision safety outcome or causes a change in the vehicle's passage decision, a higher impact level is set; if a risk indicator mainly affects vehicle operating comfort or is used for auxiliary evaluation, a relatively lower impact level is set. Based on the above impact level evaluation results, each risk indicator is assigned a level value, and the corresponding risk weight is obtained through normalization, as shown in the table below:

[0188] As shown in the table above, different risk items are assigned different impact levels based on their influence on the safety of autonomous driving scenarios and vehicle decision-making, and the corresponding risk weights are obtained through normalization.

[0189] In this embodiment, the impact levels of collision risk-related indicators, comfort risk-related indicators, rule conflict risk-related indicators, and abnormal behavior risk-related indicators can be set to 5, 2, 4, and 3, respectively, with a total level of 14. After normalization, the corresponding risk weights are 0.357, 0.143, 0.286, and 0.214, respectively. In the above weight allocation, collision risk-related indicators are assigned a higher weight because they directly reflect the probability of a vehicle collision and its potential safety consequences; rule conflict risk-related indicators are assigned a higher weight because they involve violations of traffic rules, incorrect priority relationships, and incorrect decision-making in conflict areas; comfort risk-related indicators mainly reflect the smoothness of vehicle operation, and their direct impact on safety decisions is relatively low, therefore they are assigned a lower weight; abnormal behavior risk-related indicators describe unexpected behaviors, and their impact needs to be judged in conjunction with the specific road test scenario. Furthermore, when calculating risk scores, if there is no valid data for a certain risk-related indicator, the indicator is not directly assigned a value of zero. Instead, the missing status is retained based on the data integrity judgment result. If the data corresponding to each risk-related indicator has been obtained, the risk score corresponding to the road test scenario segment is obtained by weighting the data according to the normalized risk weights mentioned above.

[0190] In this embodiment, rarity scoring According to the Scene index vector corresponding to each road test scene segment Sure.

[0191] Specifically, it can be based on the scene index vector. The scene type, road type, traffic participant type, event type, and regional rule semantic features are used to retrieve historical scene fragments that are the same as or similar to the road test scene fragment in the road scene database, and the corresponding occurrence frequency, cluster density, or number of similar samples is counted. Then, the occurrence frequency, cluster density, or number of similar samples is normalized, and the rarity score is determined based on the normalization result. .

[0192] Rarity rating Normalized numerical representations are used, with values ​​ranging from 0 to 1. Larger values ​​indicate rarer road test scene segments. Specifically, when compared with the... When the number of identical or similar historical scene segments is small, or when the cluster density of a road test scene segment in the road scene library is low, the rarity score corresponding to the road test scene segment is determined to be a value close to 1; when the number of identical or similar historical scene segments is large, or when the cluster density of a road test scene segment in the road scene library is high, the rarity score corresponding to the road test scene segment is determined to be a value close to 0.

[0193] In this embodiment, coverage scoring According to the The contribution of each road test scenario segment to the overall scene space coverage is determined after it is included in the representative scene set.

[0194] Specifically, it can be based on the scene index vector. The road test scenario fragment is characterized by road type, traffic participant type, event type, dynamic behavior pattern, and regional rule semantic features. This is used to determine whether the fragment covers scenario categories not yet covered or insufficiently covered by the currently selected representative scenario set. When the road test scenario fragment can supplement new scenario categories, rule semantic types, or traffic interaction patterns, its coverage score is calculated. A larger value indicates a lower coverage score when the road test scene segment is highly similar to the selected representative scene and contributes little to the overall new coverage of the scene space. The value is relatively small.

[0195] Among them, coverage score Normalized numerical representation can be used, with values ​​ranging from 0 to 1. A larger value indicates the [number of]th [unit]. The greater the contribution of each road test scenario segment to the increase in overall scene space coverage.

[0196] S4-2: Based on Scene Index Vector The semantic information of the region rules determines the semantic coverage score of the region rules. .

[0197] In this embodiment, due to the scene index vector Includes semantic summary of region rules Therefore, semantic summarization based on region rules can be achieved. Calculate the semantic coverage score of the region rules .

[0198] Region rule semantic coverage score Used to characterize the The contribution of each road test scenario fragment to the coverage of the representative scenario set in terms of regional rule semantics.

[0199] Specifically, the semantic coverage score of regional rules The rarity of region semantics, traffic direction semantics, rule type semantics, and right-of-way relation semantics can be determined, and the calculation logic is as follows: in, A rarity score representing the semantics of a region; A rarity score representing the semantics of the direction of travel; A rarity score representing the semantics of the rule type; A rarity score representing the semantics of right-of-way relationships. For the corresponding weight coefficients, satisfying .

[0200] The rarity score of each semantic component is determined based on the frequency of occurrence of the corresponding semantic category in the road scene database. Specifically, it is obtained by counting the number of road test scene fragments corresponding to each semantic category and normalizing the statistical results. The lower the frequency of occurrence or the fewer the number of similar samples, the less common the semantic feature is, and the higher its corresponding rarity score is.

[0201] Specific weighting coefficients Initial values ​​can be assigned based on expert experience and then calibrated by combining historical test data, representative scenario screening results, and validation set coverage results; preferably, methods such as the analytic hierarchy process (AHP), entropy weighting, and Bayesian optimization can be used to determine the values. For example, 0.15 0.22, 0.28 It is 0.35.

[0202] Specifically, since regional semantics is used to characterize the area to which the road test scenario belongs, it is mainly used to ensure that samples from different areas can be retained; traffic direction semantics is used to characterize the difference between left-hand and right-hand traffic, which affects lane organization and driving direction; rule type semantics is used to characterize the differences in yielding rules, priority rules, signal control rules, roundabout rules, or pedestrian priority rules, which affects vehicle traffic decisions; and right-of-way semantics is used to characterize the priority relationship between the vehicle and other traffic participants, which directly affects whether the vehicle yields, waits, or has priority. Therefore, right-of-way semantics and rule type semantics have a relatively high impact on the representative scenario selection results, followed by traffic direction semantics, and regional semantics has a relatively low impact.

[0203] Experts in traffic engineering and autonomous driving testing can determine the semantic coverage impact scores by comparing the influence of region semantics, traffic direction semantics, rule type semantics, and right-of-way relationship semantics on the coverage results of representative scenarios and subsequent vehicle traffic decisions. The semantic coverage impact scores are then normalized to obtain initial weights. For example, the initial scores for region semantics, traffic direction semantics, rule type semantics, and right-of-way relationship semantics can be set to 2, 3, 4, and 5, respectively. After normalizing the scores, the initial weights are approximately 0.14, 0.21, 0.29, and 0.36.

[0204] Following this, the initial weights are calibrated by combining the representative scenario selection results and the validation set coverage results. Specifically, different weight combinations can be used to select representative scenarios from road test scenario fragments in the road scenario library, and the coverage of the selected representative scenario set in terms of regional semantics, traffic direction semantics, rule type semantics, and right-of-way relationship semantics can be statistically analyzed. When a certain weight combination can retain high-risk scenarios while ensuring that low-frequency regional rule semantic scenarios, different traffic direction scenarios, different rule type scenarios, and different right-of-way relationship scenarios all meet the preset coverage requirements, this weight combination is used as the calibrated weight combination.

[0205] In this embodiment, after calibration of the representative scene selection effect, it can be Set to 0.15 Set to 0.22 Set to 0.28 The value is set to 0.35. Therefore, in the calculation of regional rule semantic coverage score, right-of-way relationship semantics and rule type semantics can have a high influence on the representative scene selection results, while also taking into account the coverage requirements of traffic direction semantics and regional semantics.

[0206] In this embodiment, each rarity score can be based on the semantic summary of the corresponding region rules. The occurrence of the corresponding semantic components in the road scene database is determined.

[0207] As an example only, each rarity score can be represented by a normalized numerical value between 0 and 1. The closer the value is to 1, the rarer the corresponding semantic feature is, and the closer the value is to 0, the more common the corresponding semantic feature is.

[0208] S4-3: Based on risk score Rarity rating Coverage score semantic coverage score of region rules The overall score for each road test scenario segment is calculated. The calculation logic is as follows:

[0209] In the formula, Indicates the first A comprehensive score for each road test scenario segment; This represents the corresponding weight coefficient, satisfying... .

[0210] Corresponding weight coefficient The importance of the target selection for representative scenarios can be determined through different evaluation indicators.

[0211] Among them, the risk score is used to reflect the degree of potential safety risks in road test scene segments, which directly affects the priority of subsequent test scenes, so it is set with a relatively high impact level; the coverage score is used to ensure that the selected representative scenes can cover the main scene types in the road scene library, avoiding the screening results from being overly concentrated in a few similar scenes, so it has a high impact level; the regional rule semantic coverage score is used to ensure that the selected representative scenes can cover the semantic features of different regional rules, which plays an important role in cross-regional rule adaptation; the rarity score is mainly used to increase the screening probability of low-frequency special scenes, which has an auxiliary role in discovering rare scenes.

[0212] Based on the aforementioned levels of impact, we can first evaluate the engineering impact level of risk score, rarity score, coverage score, and regional rule semantic coverage score, and then normalize the evaluation results. For example, the importance levels of the four scoring indicators can be set to (5:2:3:4), and the corresponding weight coefficients can be obtained after normalization: [ =[0.357,0.143,0.214,0.286].

[0213] Furthermore, in specific testing tasks, the above weights can be adjusted according to the testing objectives, and calibrated in combination with the coverage effect and verification results after the representative scenario screening, so as to ensure that the comprehensive scoring results can take into account the priority discovery of high-risk scenarios, the completeness of scenario coverage, and the semantic coverage capability of regional rules.

[0214] S4-4: Based on the comprehensive score, select candidate representative scenes from the road scene library.

[0215] In this embodiment, after calculating the comprehensive score of each road test scene segment, the road test scene segments in the road scene library can be sorted from high to low according to the comprehensive score, and candidate representative scenes can be selected from the sorted road test scene segments according to the preset comprehensive score threshold or preset screening quantity.

[0216] For example, road test scene segments with a comprehensive score greater than or equal to a preset comprehensive score threshold can be used as candidate representative scenes; alternatively, a preset number of road test scene segments with the highest comprehensive scores can be used as candidate representative scenes. The preset comprehensive score threshold or the preset number of segments to be selected can be set based on historical screening experience.

[0217] Furthermore, representative scenarios are obtained by filtering candidate representative scenarios based on preset coverage constraints.

[0218] In this embodiment, the preset coverage constraint is used to ensure that the selected representative scene set has sufficient coverage in the semantic dimension of regional rules, so as to avoid the representative scene set being concentrated only in high-frequency regions or high-frequency rule types, while omitting low-frequency road test scene segments that have semantic differences in regional rules.

[0219] The preset coverage constraints are as follows:

[0220] in, Represents the set of candidate representative scenarios The degree of coverage in the semantic dimension of the region; Represents the set of candidate representative scenarios The degree of coverage in the directional semantic dimension; Represents the set of candidate representative scenarios The degree of coverage in the semantic dimension of rule types. , , These represent the preset lower limit of coverage for the corresponding dimension.

[0221] In the specific screening process, candidate representative scenarios can be selected first according to the comprehensive score from high to low; if the current set of candidate representative scenarios... If the above coverage constraints are not met, road test scene fragments that can supplement the semantics of uncovered areas, traffic direction, or rule type from the road scene library will be selected and added to the candidate representative scene set until the candidate representative scene set meets the preset coverage constraints.

[0222] When the candidate representative scenario set After satisfying the preset coverage constraints, the candidate representative scene set is... The road test scene segments in the data were identified as representative scenes, resulting in a set of representative scenes.

[0223] The preset coverage lower limit can be pre-set based on the requirements of autonomous driving test tasks, the distribution of target test areas, the degree of semantic differences in regional rules, and the number of samples in the road scene library.

[0224] Preferably, when the target test range covers multiple areas, each area retains at least one or more representative scenarios; when the road scenario library contains both left-hand and right-hand traffic scenarios, both types of traffic directions meet the preset minimum coverage ratio; when the road scenario library contains different rule types such as forced yielding, signal control, roundabout priority, and unprotected turning, each rule type meets the preset minimum coverage quantity or minimum coverage ratio.

[0225] S5: Generate candidate site condition parameters based on representative scenarios, and calculate the regional rule semantic deviation value between the candidate site condition parameters and the representative scenarios by constructing a regional rule semantic consistency cost term.

[0226] In this embodiment, generating candidate site condition parameters based on representative scenarios means extracting scenario elements for reproduction in the test site from the representative scenarios and combining them with the available road structure and vehicle layout conditions of the test site to generate a parameter combination for constructing the site reproduction scenario.

[0227] During the specific conversion process, a matching road area can be selected from the available road structures in the test site based on the road type and number of lanes in the representative scenario; the initial positions of the vehicle and the target vehicle in the test site can be determined based on the relative positions, relative distances, and lane relationships of the vehicles in the representative scenario; the initial speeds of the vehicle and the target vehicle can be determined based on the vehicle speeds and relative speeds in the representative scenario; the trigger time for the target vehicle's action can be determined based on the event occurrence time or interaction sequence in the representative scenario; and the configuration of road signs and markings and the yielding or priority control logic in the test site can be determined based on the semantics of road signs and markings and the right-of-way relationships in the representative scenario. Thus, different candidate site condition parameters can be generated for different representative scenarios, resulting in preliminary candidate site condition parameters.

[0228] Furthermore, the preliminary candidate site operating parameters are further screened.

[0229] In S5, the semantic consistency cost term for regional rules is constructed using the following formula to quantify scene differences; its calculation logic is as follows: In the formula, Indicates the first Candidate site condition parameters corresponding to each representative scenario Semantic deviation value of the region rules from the representative scene; This is due to a deviation in the consistency of the direction of travel. This is due to a deviation in rule type consistency. This refers to semantic consistency deviations in road signs and markings. This indicates a deviation in the consistency of right-of-way relationships. , , , The weighting coefficients for the corresponding deviation terms satisfy the following conditions: .

[0230] Corresponding weight coefficient , , , The weights are set based on the degree of impact of different semantic terms on security and reproducibility, and calibrated by combining historical scenario reproduction results. Preferably, initial weights can be set based on expert experience, and then adjusted through validation set optimization or sensitivity analysis based on the reproduction errors and rule distortion in verified scenarios. For example, 0.25 0.25 0.15 It is 0.35.

[0231] In this embodiment, each consistency deviation term can be determined by comparing the candidate site operating parameters. Semantic summary of regional rules representing the scene Consistency is achieved on the corresponding semantic terms.

[0232] When the direction of travel, rule type, road sign and marking semantics, or right-of-way relationship are completely consistent, the corresponding consistency deviation term takes a value of 0; when there is inconsistency, the corresponding consistency deviation term takes a positive value, and this positive value can be determined based on the discrete semantic distance, the number of rule conflicts, or the normalized inconsistency rate.

[0233] Obtain the semantic deviation value of the region rule Then, based on the semantic deviation value of the region rules The set of candidate site condition parameters is determined by the site condition constraints.

[0234] The set of candidate site condition parameters consists of candidate site condition parameters that meet both the threshold requirement and the site condition constraints.

[0235] Site condition constraints are used to limit the executable range of candidate site condition parameters in the test site, so as to ensure that the site reproduction scenario corresponding to the candidate site condition parameters can be safely and effectively constructed.

[0236] Its calculation logic is as follows:

[0237] In the formula, For the first A set of candidate site condition parameters corresponding to each representative scenario; Indicates the first Each representative scenario corresponds to any candidate site condition parameter; For the first Site operating condition constraints; This refers to the sequence number of the site condition constraints. The number of site condition constraints; This is the semantic deviation threshold for regional rules.

[0238] Specifically, the site conditions constraints may include one or more of the following: test site geometry constraints, target placement constraints, vehicle speed range constraints, timing constraints, and equipment capability constraints.

[0239] If a candidate site condition parameter satisfies all site condition constraints simultaneously, and its regional rule semantic deviation value is not greater than the regional rule semantic deviation threshold, then the candidate site condition parameter is retained in the candidate site condition parameter set.

[0240] The semantic deviation threshold for regional rules can be determined jointly based on the test task requirements, the statistical results of historical successful reproduction samples, and the results of manual equivalence judgment.

[0241] Specifically, initial thresholds can be set based on the importance and security impact of representative scenarios; further, the semantic deviation values ​​of regional rules can be statistically analyzed based on historically verified equivalent site reproduction scenario samples. The distribution of the region rules is determined, and a preset quantile, the maximum acceptable value, or the human consistency judgment boundary is selected as the semantic deviation threshold of the region rules. This ensures that the threshold meets the requirements of autonomous driving test tasks and is consistent with historical reproduction experience.

[0242] In this embodiment, by setting site condition constraints, the retained candidate site condition parameters can be safely and executablely reproduced in the actual test site.

[0243] S6: Construct an equivalent objective function, and under the constraints of site conditions, solve for the optimal site conditions from the candidate site conditions, so that the site reproduction scenario corresponding to the optimal site conditions meets the deviation requirements, and construct an equivalent test scenario based on the optimal site conditions.

[0244] The equivalent objective function is based on the index vector corresponding to the representative scenario. The index estimation vector corresponding to the candidate site working condition parameters And the construction of semantic deviation values ​​for region rules:

[0245] In the formula, Indicates the first The optimal site condition parameters corresponding to each representative scenario; This is the indicator weight matrix; Indicates the first The index vector corresponding to each representative scenario and the candidate site condition parameters Corresponding index estimation vector The deviation value between the indicators; The weight coefficient corresponding to the semantic deviation value of the region rule; For index vectors; This is the index estimation vector.

[0246] Indicator Vector It is a vector used to characterize the evaluation index features of road test scenario segments, that is, the real index in the original road test scenario segments.

[0247] Indicator estimation vector It is to use a set of candidate site condition parameters After reproducing the phenomenon in a test environment or simulation platform, the corresponding predicted indicators are obtained. This is achieved by vectorizing the indicators. With index estimation vector By comparing the parameters, the operating conditions of the candidate sites can be determined. The index deviation between the corresponding site reproduction scenario and the representative scenario provides a basis for solving the optimal site condition parameters for the equivalent objective function.

[0248] In this embodiment, the index estimation vector Based on candidate site operating parameters The corresponding initial vehicle position, initial vehicle speed, target vehicle deployment parameters, action timing parameters, and control logic parameters are used to configure the corresponding site reproduction scenario. Then, based on the existing vehicle kinematics model or site reproduction simulation model, according to the index vector corresponding to the representative scenario, the scenario is recreated. The same indicators are arranged in order.

[0249] Indicator weight matrix This is used to weight the importance of different indicator deviation terms, thereby adjusting the influence of indicators such as velocity change, acceleration change, trajectory deviation, collision risk, and scene duration in the equivalent objective function. The indicator weight matrix... It can be preset based on historical road test data statistics or expert experience.

[0250] This is the weighting coefficient corresponding to the semantic deviation value of the regional rules, used to adjust the influence of the semantic deviation value of the regional rules in the equivalent objective function.

[0251] Specifically, when it is necessary to prioritize ensuring consistency between the recreated and representative scenarios in terms of regional rule semantics such as traffic direction, rule type, road sign and marking semantics, and right-of-way relationships, the [limitation / increase] can be increased. The value of can be reduced when prioritizing the reproduction accuracy of basic indicators such as vehicle trajectory, speed, acceleration, and interaction gaps. The value of . The aforementioned It can be preset according to the requirements of autonomous driving test tasks, the semantic consistency requirements of regional rules, or the test verification results. This manual does not impose specific restrictions on this.

[0252] In this embodiment, when solving for the optimal site condition parameters, the candidate site condition parameter set can be considered. The equivalent objective function value is calculated for each candidate site condition parameter, and the candidate site condition parameter with the smallest equivalent objective function value is determined as the optimal site condition parameter.

[0253] After obtaining the optimal site working parameters Subsequently, scenario configuration data is generated based on the optimal site condition parameters. This scenario configuration data may include one or more of the following: test site road geometry, lane layout, road sign and marking configuration, initial position of the vehicle, initial speed of the vehicle, location of target vehicles or objects, speed of target vehicles, action trigger times, signal control parameters, and yield or priority passage control logic. After obtaining the scenario configuration data, a corresponding site reproduction scenario is constructed in a closed test site, simulation test platform, or scenario editing system based on the scenario configuration data.

[0254] After constructing the site reproduction scenario, the index deviation value can be used as a basis. semantic deviation value of region rules An equivalence assessment is performed on the corresponding site reproduction scenarios.

[0255] When the index deviation value does not exceed the preset index deviation threshold and the regional rule semantic deviation value does not exceed the preset regional rule semantic deviation threshold, the site reproduction scenario is determined to meet the equivalence requirements, and the site reproduction scenario is determined to be the equivalent test scenario; otherwise, the site reproduction scenario is determined to not meet the requirements.

[0256] If not, further in this embodiment, if the equivalent test scenario constructed from the optimal site condition parameters in S6 does not meet the deviation requirements, the optimal site condition parameters are corrected in a closed loop; the calculation logic is as follows:

[0257] In the formula, Indicates the first The first closed-loop correction of the wheel The corrected optimal site condition parameters correspond to each representative scenario; Indicates the first The first closed-loop correction of the wheel The corrected optimal site condition parameters correspond to each representative scenario; This indicates a parameter update function; Indicates the first The index deviation value under closed-loop correction; Indicates the first Semantic deviation value of regional rules under closed-loop correction.

[0258] In this embodiment, according to the formula, in the first... During the closed-loop correction process, based on the current corrected optimal site condition parameters... The corresponding index deviation value and regional rule semantic deviation value are used to update the site working parameters, resulting in the first... Optimal site condition parameters after wheel correction When the corrected index deviation value does not exceed the preset index deviation threshold, and the regional rule semantic deviation value does not exceed the preset regional rule semantic deviation threshold, the closed-loop correction stops; or, when the number of iterations reaches the preset maximum number of iterations, the closed-loop correction stops.

[0259] Among them, the parameter update function Specifically, it is a function used to perform closed-loop correction of site condition parameters based on index deviation values ​​and regional rule semantic deviation values. Specifically, it can adjust road sign and marking configuration parameters, traffic direction configuration parameters, and yield or priority traffic control logic parameters based on regional rule semantic deviation values, and adjust vehicle initial position, vehicle initial speed, target placement position, action timing parameters, or vehicle trajectory parameters based on index deviation values, thereby obtaining the site condition parameters for the next iteration.

[0260] By setting up an equivalent correction process, when the site reproduction scenario corresponding to the optimal site condition parameters does not meet the deviation requirements, the site condition parameters can be iteratively adjusted based on the index deviation value and the semantic deviation value of the regional rules. This avoids insufficient reproduction results caused by test site constraints, vehicle layout errors, or semantic deviations in the site condition parameters obtained from a single solution.

[0261] Therefore, the vehicle motion state, traffic rule semantics, and right-of-way control relationships in real road traffic scenarios can be transformed into traffic scenario reproduction control parameters that can be executed in the test site, thereby realizing the construction of equivalent control for traffic scenarios for autonomous driving testing.

[0262] It should be noted that this embodiment includes subscripts. The parameters are all used to represent the first The parameters or features corresponding to each scene object. Specifically, in the road test scene fragment processing stage, This indicates the sequence number of the road test scenario segment; it represents the scenario selection and site condition parameter generation stage. This indicates the sequence number of the scene, or the sequence number of the scene. The index is the sequence number corresponding to the representative scene obtained by filtering road test scene segments. It is only used to distinguish different scene objects, and is not used to limit the order of processing between scene objects.

[0263] It should be noted that the weight coefficients in this embodiment are used to characterize the relative importance of different evaluation items in the corresponding calculation process, and the specific values ​​of the weight coefficients do not affect the implementation logic of the method described in this embodiment.

[0264] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for constructing equivalent autonomous driving scenarios with cross-regional rule adaptation, characterized in that, include: S1: Acquire multi-source road test data, parse and process the multi-source road test data, and construct semantically enhanced state based on the parsed and processed multi-source road test data to obtain enhanced state data containing regional rule semantic information; S2: Determine the operation design domain judgment result and event triggering result based on the enhanced state data, and perform scene slicing on the enhanced state data according to the operation design domain judgment result and event triggering result to obtain road test scene segments; S3: Perform structured representation on the road test scene segment, generate scene index vector and indicator vector, and write the scene index vector and indicator vector into the road scene library; S4: Based on the scene index vector and indicator vector, select representative scenes from the road scene library; S5: Generate candidate site condition parameters based on the representative scenario, and calculate the regional rule semantic deviation value between the candidate site condition parameters and the representative scenario by constructing a regional rule semantic consistency cost term; S6: Construct an equivalent objective function, and under the constraints of site conditions, solve for the optimal site conditions from the candidate site conditions parameters, so that the site reproduction scenario corresponding to the optimal site conditions parameters meets the deviation requirements, and construct an equivalent test scenario based on the optimal site conditions parameters.

2. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 1, characterized in that, S1 includes: S1-1: Acquire multi-source road test data and preprocess the multi-source road test data to obtain preprocessed multi-source road test data; S1-2: Perform regional rule semantic parsing on the preprocessed multi-source road test data to extract regional rule semantic vectors; S1-3: Perform scene state parsing processing on the preprocessed multi-source road test data to extract the scene state vector; S1-4: Concatenate and splice the scene state vector and the region rule semantic vector at the same time to construct a semantically enhanced state, and obtain enhanced state data containing region rule semantic information.

3. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 2, characterized in that, S2 include: S2-1: Based on the scene state vector and regional rule semantic vector in the enhanced state data, determine the operation design domain judgment result; the operation design domain judgment result includes the basic operation design domain judgment result and the regional rule operation design domain judgment result; S2-2: Construct a joint event triggering function based on the scene state change amount and the region rule semantic change amount in the enhanced state data, and determine the event triggering result based on the comparison result of the joint event triggering function and the preset triggering threshold; S2-3: Determine the slice boundary based on the operation design domain judgment result and event trigger result, and slice the enhanced state data into scene segments based on the slice boundary to obtain road test scene segments.

4. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 3, characterized in that, In S2-1, the result of the operational design domain determination is determined based on the following formula: In the formula, For a moment The results of the operational design domain determination; For a moment The basic operational design domain determination result; For a moment The regional rule operation design domain judgment result; Among them, the regional rule operation design domain determination result It is obtained by determining the allowed values ​​of the semantic components of the region rules, and its calculation logic is as follows: In the formula, For a moment The first in the semantic vector of the region rule Semantic components of regional rules; For the first The set of allowed values ​​corresponding to the semantic components of each region rule; Indicates characteristic functions; The number of semantic components of the regional rules that participate in the decision-making of the operational design domain.

5. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 3, characterized in that, In S2-2, the event triggering result is determined based on the comparison result between the joint event triggering function and the preset triggering threshold, and its calculation logic is as follows: In the formula, For the event At any moment The result of the event triggering; For a moment The scene state vector; For a moment The semantic vector of the region rules; This represents the basic event discrimination term obtained based on the scene state vector; This represents the rule semantic event discrimination term obtained based on the region rule semantic vector; Indicates the semantic trigger weight of the rule; Indicates the preset trigger threshold; Indicates characteristic functions; Among them, the basic event discrimination item Determined by the changes in scene state, its calculation logic is as follows: In the formula, Indicates the change in velocity; Indicates longitudinal or lateral acceleration; Indicates the change in heading angle; Indicates the collision time; This indicates the change in relative distance; This represents the corresponding weight coefficient; Among them, the rule semantic event discrimination item The calculation logic for determining the semantic change of the region rules is as follows: In the formula, , , , These represent the determination quantities for whether the direction of travel, rule type, semantics of road signs and markings, and right-of-way relationships have changed, respectively. This represents the corresponding weighting coefficient.

6. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 1, characterized in that, S4 include: S4-1: Determine the risk score of the road test scene segment based on the indicator vector, and determine the rarity score and coverage score of the road test scene segment based on the scene index vector; S4-2: Determine the semantic coverage score of regional rules based on the semantic information of regional rules in the scene index vector; S4-3: Calculate the comprehensive score of the road test scene segment based on the risk score, rarity score, coverage score, and regional rule semantic coverage score; S4-4: Based on the comprehensive score, candidate representative scenes are selected from the road scene library, and the candidate representative scenes are further selected according to the preset coverage constraints to obtain representative scenes.

7. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 6, characterized in that, In S4-3, the comprehensive score is determined based on risk score, rarity score, coverage score, and regional rule semantic coverage score, and its calculation logic is as follows: In the formula, Indicates the first A comprehensive score for each road test scenario segment; This represents the risk score obtained based on the indicator vector; This represents the rarity score obtained based on the scene index vector; Indicates the first Each road test scenario segment scores the overall scene space coverage. Indicates the first Regional rule semantic coverage scoring for a road test scenario segment; Indicates the first Semantic summary of scene fragment-level region rules corresponding to each road test scene fragment; This represents the corresponding weighting coefficient.

8. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 7, characterized in that, In S5, the calculation logic for the semantic deviation value of the region rule is as follows: In the formula, Indicates the first Candidate site condition parameters corresponding to each representative scenario The semantic deviation value of the region rules between the representative scene and the scene; This is due to a deviation in the consistency of the direction of travel. This is due to a deviation in rule type consistency. This refers to semantic consistency deviations in road signs and markings. This indicates a deviation in the consistency of right-of-way relationships. , , , These are the weighting coefficients for the corresponding deviation terms; The candidate set of site condition parameters is determined based on the semantic deviation value of the regional rules and the site condition constraints. The calculation logic is as follows: In the formula, For the first A set of candidate site condition parameters corresponding to each representative scenario; For the first Site operating condition constraints; The number of site condition constraints; This is the semantic deviation threshold for regional rules.

9. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 8, characterized in that, The equivalent objective function is constructed based on the index vector corresponding to the representative scenario, the index estimation vector corresponding to the candidate site condition parameters, and the semantic deviation value of the regional rules. In the formula, Indicates the first The optimal site condition parameters corresponding to each representative scenario; For the first Candidate site condition parameters corresponding to each representative scenario; For the first A vector of indicators corresponding to each scenario. Based on candidate site operating parameters The obtained index estimation vector; This is the indicator weight matrix; Indicates the first The index vector corresponding to each representative scenario and the candidate site condition parameters The index deviation between the corresponding index estimation vectors; This represents the weight coefficient corresponding to the semantic deviation value of the regional rules.

10. The method for constructing an equivalent autonomous driving scenario with cross-regional rule adaptation according to claim 9, characterized in that, If the equivalent test scenario constructed using the optimal site condition parameters described in S6 does not meet the deviation requirements, a closed-loop correction is performed on the optimal site condition parameters. The calculation logic is as follows: In the formula, Indicates the first The first closed-loop correction of the wheel The corrected optimal site condition parameters correspond to each representative scenario; Indicates the first The first closed-loop correction of the wheel The corrected optimal site condition parameters correspond to each representative scenario; This indicates a parameter update function; Indicates the first The index deviation value under closed-loop correction; Indicates the first Semantic deviation value of regional rules under closed-loop correction.

Images