Driving behavior evaluation methods, devices, electronic equipment and storage media
By segmenting autonomous driving trajectory data and projecting genetic prototypes, the problem of insufficient expression of driver heterogeneity in existing technologies is solved, and style-aware driver evaluation and traffic flow efficiency are output in parallel.
Patent Information
- Application Number
- CN202611621208.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing autonomous driving behavior evaluation methods are unable to fully express driver heterogeneity, the evaluation results are coupled with driver behavior style, lack a hierarchical aggregation mechanism from the segment level to the driver level, and fail to distinguish the quality of individual controllable behavior.
By acquiring the trajectory data of the target driving object, continuously segmenting it, determining the evaluation results of the behavioral primitive layer, and projecting it onto the preset driving gene prototype, the evaluation results of the scenario-level driver layer are determined based on the matching weight and driving quality score, thus constructing a hierarchical evaluation chain from local behavioral units to high-level driver structures.
It achieves a comprehensive evaluation of drivers with different styles, which is style-aware and can accurately evaluate the driving behavior quality of drivers with different styles. It also outputs individual behavior quality and system traffic flow efficiency in parallel.
Smart Images

Figure CN123115880A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of autonomous driving simulation and driving behavior modeling, and in particular to a driving behavior evaluation method, device, electronic device and storage medium. Background Technology
[0002] The development and validation of autonomous driving systems require testing in numerous traffic scenarios with high interaction intensity and safety risks. Relying solely on real-vehicle road testing is insufficient to efficiently cover such complex situations; therefore, traffic simulation has become an important technical means for the validation of autonomous driving systems.
[0003] In existing technologies, one type of method describes driver behavior using micro-level driving behavior models such as following, lane changing, yielding, and gap acceptance. This type of method has the advantages of clear variable meanings and well-defined computational structures, but it usually expresses driver differences through fixed parameters or coarse-grained driver categories, making it difficult to fully reflect the heterogeneity of driving behavior in interaction-intensive scenarios such as merging, yielding, and overtaking.
[0004] Another type of approach utilizes large-scale natural driving data to directly learn trajectory evolution patterns or behavioral strategies. This type of approach can better maintain the realism of behavior in empirical data, but it usually models directly at the trajectory or strategy layer, lacking a stable intermediate layer representation. This results in insufficient structural interpretability and makes it difficult to directly transfer to a transparent and controllable simulation execution mechanism.
[0005] Furthermore, existing research has attempted to divide the continuous driving process into local behavioral units for scene recognition, or to describe driver differences based on driving style and aggressiveness. However, the former usually remains at the level of local decomposition without constructing a stable driver-level representation, while the latter is mostly based on coarse-grained labels or latent factors without explicitly anchoring high-level differences to observable behavioral units. There is still a lack of a stable and unified connection mechanism between locally interpretable behavioral units and the high-level driver difference structure, and consequently, a driving behavior representation and evaluation method that can simultaneously take into account data foundation, behavioral interpretability, quantitative evaluation capability, and simulation feasibility is lacking.
[0006] While traditional micro-models offer good semantic interpretability, they struggle to fully represent driver heterogeneity. In driving behavior quality assessment, evaluation results are coupled with driver behavior styles, leading to incomparable quality scores for drivers with different styles. Natural driving data-driven methods, while possessing high behavioral realism, lack stable and transparent intermediate layer representation structures. Existing behavior evaluation methods lack a hierarchical aggregation mechanism from the fragment layer to the driver layer and fail to distinguish individual controllable behavioral quality. Summary of the Invention
[0007] In order to solve the above-mentioned technical problems, or at least partially solve the above-mentioned technical problems, this application provides a driving behavior evaluation method, device, electronic device and storage medium.
[0008] Firstly, this application provides a method for evaluating driving behavior, including: The trajectory data of the target driving object is acquired, and the trajectory data is continuously segmented to obtain multiple behavior segments; For each scenario category, the evaluation result of the target driving object's behavior primitive layer is determined based on multiple behavior fragments, and the distribution of the target driving object's behavior primitives is determined; The evaluation results of the behavioral primitive layer are projected onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype. Each driving gene prototype is determined according to the distribution of behavioral primitive composition of each driving object in the group driving data under multiple scene categories. Based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype, the matching weight between the target driving object and each driving gene prototype is determined. The scenario-level driver evaluation result is determined based on the matching weight and the driving quality score.
[0009] Optionally, the behavioral primitive layer evaluation result of the target driving object is determined based on multiple behavioral fragments, including: Each behavior segment is classified into scene category and behavior primitive category to obtain the scene category and behavior primitive type corresponding to each behavior segment. For each of the preset multiple evaluation dimensions, the evaluation indicators contained in the evaluation dimension are used to calculate and normalize the indicators of each behavior segment to obtain the segment-level evaluation result of each behavior segment under the evaluation dimension. The fragment-level evaluation results under each evaluation dimension are aggregated according to the scenario category and the behavior primitive type to obtain the behavior primitive layer evaluation results.
[0010] Optionally, the behavioral segments are normalized based on the evaluation indicators included in a plurality of preset evaluation dimensions to obtain segment-level evaluation results for each behavioral segment, including: For evaluation indicators with preset physical thresholds in multiple evaluation dimensions, the calculation results of each evaluation indicator are normalized by using monotonic mapping or piecewise function mapping based on the Sigmoid function to obtain the first normalized indicator value. For evaluation indicators without preset physical thresholds in multiple evaluation dimensions, normalization is performed by scaling to a preset range to obtain the second normalized indicator value. The first normalized index value and the second normalized index value are weighted and summed according to the preset index weights within the evaluation dimension to which the evaluation index belongs, so as to obtain the segment-level dimension score of each behavior segment under each evaluation dimension. The segment-level evaluation result of each behavioral segment is determined based on the segment-level dimension score of each evaluation dimension.
[0011] Optionally, the fragment-level evaluation results under each of the evaluation dimensions are aggregated according to the scene category and the behavior primitive type to obtain the behavior primitive layer evaluation results, including: Fragments with the same scene category and the same behavior primitive type are grouped into the same aggregation group; Within any aggregation group, the segment-level dimension scores in the segment-level evaluation results of each behavioral segment are weighted and aggregated according to the duration of the behavioral segment or by adopting an event-weighted strategy to obtain the aggregation score; The evaluation result of the behavioral primitive layer is determined based on the aggregation scores corresponding to multiple aggregation groups.
[0012] Optionally, the evaluation results of the behavioral primitive layer are projected onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype, including: Based on group driving data under multiple scenario categories, a driving gene prototype corresponding to each scenario category is obtained in advance. The driving gene prototype includes the weights of behavioral primitives corresponding to multiple behavioral primitive types. For each evaluation dimension, based on the weights of the behavioral primitives corresponding to each behavioral primitive type in the driving gene prototype, the evaluation results of the behavioral primitive layer are projected onto multiple driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype.
[0013] Optionally, the driving quality score is the result of a weighted sum of the behavioral primitive layer evaluation results according to the weights of the behavioral primitive composition corresponding to multiple behavioral primitive types in the driving gene prototype.
[0014] Optionally, based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype under each scenario category, the matching weight between the target driving object and each driving gene prototype is determined, including: Calculate the distribution difference measure between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype; The matching weights between the target driving object and each of the driving gene prototypes are determined based on multiple distribution difference measures.
[0015] Optionally, determining the matching weight between the target driving object and each of the driving gene prototypes based on multiple distribution difference measures includes: Each of the aforementioned distribution difference measures is used as the independent variable of the exponential kernel function, and the matching degree is obtained by performing a negative exponential transformation through the exponential kernel function. The matching degrees are normalized to obtain the matching weights between the target driving object and each driving gene prototype.
[0016] Optionally, when the number of driving gene prototypes is less than or equal to a preset number threshold, the temperature parameter in the exponential kernel function is a preset temperature parameter; When the number of driving gene prototypes exceeds a preset threshold, the temperature parameter in the exponential kernel function is determined based on the number of driving gene prototypes.
[0017] Optionally, determining the scenario-level driver evaluation result based on the matching weight and the driving quality score includes: For each evaluation dimension, a weighted aggregation is performed based on the matching weights and driving quality scores corresponding to each driving gene prototype to obtain the scenario-level driver-level evaluation results of the target driving object under multiple evaluation dimensions.
[0018] Optionally, the method further includes: Based on the scenario-level driver evaluation results, the scenario-level driver evaluation results are aggregated according to the preset scenario exposure weights to obtain individual behavior quality scores. Traffic flow efficiency results are determined separately based on system-level traffic flow efficiency evaluation indicators; The traffic flow efficiency results are used as parallel outputs independent of the individual behavior quality scores.
[0019] Optionally, the scene exposure weight is determined according to the proportion of the total duration of the behavioral segments under each scene category to the total duration of all behavioral segments of the target driving object.
[0020] Secondly, this application provides a driving behavior evaluation device, comprising: The first acquisition module is used to acquire trajectory data of the target driving object and continuously segment the trajectory data to obtain multiple behavior segments; The first determining module is used to determine the behavioral primitive layer evaluation result of the target driving object based on multiple behavioral fragments for each scene category, and to determine the behavioral primitive distribution of the target driving object; The projection module is used to project the evaluation results of the behavioral primitive layer onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype. Each driving gene prototype is determined according to the distribution of the behavioral primitive composition of each driving object in the group driving data under multiple scene categories. The second determining module is used to determine the matching weight between the target driving object and each driving gene prototype based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype. The third determining module is used to determine the scenario-level driver evaluation result based on the matching weight and the driving quality score.
[0021] Thirdly, this application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the driving behavior evaluation method described in any of the first aspects.
[0022] Fourthly, this application provides a computer-readable storage medium storing a program for a driving behavior evaluation method, wherein when the program for the driving behavior evaluation method is executed by a processor, it implements the steps of the driving behavior evaluation method described in any of the first aspects.
[0023] The beneficial effects of this application are as follows: This application's embodiments acquire the trajectory data of the target driving object and continuously segment it. For each scenario category, the evaluation results and distribution of the target driving object's behavioral primitive layer are determined based on behavioral fragments. By projecting the evaluation results of the behavioral primitive layer onto multiple driving gene prototypes, the driving quality score of the target driving object from the perspective of multiple driving gene prototypes is obtained. That is, the score corresponding to the driving quality performance of the target driving object when each driving gene prototype is used as the evaluation benchmark. Then, by matching the distribution of the target driving object's behavioral primitives with the distribution of the behavioral primitives of each driving gene prototype, the matching weight of the target driving object and the driving style represented by each driving gene prototype is determined. Finally, the scenario-level driver-level evaluation result is determined based on the driving quality score (performance) and the matching weight (style). This makes the comprehensive evaluation between drivers of different styles style-aware. Even if two drivers obtain the same driving quality score on the same behavioral primitive, they will still obtain different scenario-level driver-level evaluation results due to their different behavioral styles. The driving behavior quality of drivers of different styles can be accurately evaluated under their respective matched driving gene prototypes. Attached Figure Description
[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 A flowchart illustrating a driving behavior evaluation method provided in this application embodiment; Figure 2 for Figure 1 Flowchart of step S102; Figure 3 for Figure 2 Flowchart of step S202; Figure 4 for Figure 2 Flowchart of step S203; Figure 5 for Figure 1 Flowchart of step S103; Figure 6 for Figure 1 Flowchart of step S104; Figure 7 Another flowchart of a driving behavior evaluation method provided in this application embodiment; Figure 8 A structural diagram of a driving behavior evaluation device provided in an embodiment of this application; Figure 9 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0028] While traditional micro-models offer good semantic interpretability, they struggle to fully represent driver heterogeneity. In driving behavior quality evaluation, the results are coupled with driver behavior style, leading to incomparable quality scores for drivers with different styles. Natural driving data-driven methods, while possessing high behavioral realism, lack stable and transparent intermediate layer representation structures. Existing behavior evaluation methods lack a hierarchical aggregation mechanism from the fragment layer to the driver layer and fail to distinguish individual controllable behavioral quality. Therefore, this application provides a driving behavior evaluation method, device, electronic device, and storage medium. Unlike simple driving style recognition schemes, it starts from the aggregation of behavioral primitives and constructs a complete hierarchical evaluation chain from local behavioral units to high-level driver structures through driving gene prototype projection and matching. Unlike simple driving behavior scoring schemes, this application separately calculates behavioral quality projection and behavioral distribution pattern matching, achieving cross-driver comparison based on style perception. It also uses individual behavioral quality and system traffic flow efficiency as parallel outputs rather than a single mixed score. This is used to establish a hierarchical driving behavior representation based on natural driving trajectories, obtaining interpretable, quantifiable driver evaluation results that can be further mapped to the micro-simulation execution layer.
[0029] This application provides a driving behavior evaluation method. Starting from the trajectory data of the target driver, it aggregates behavioral primitives, projects and matches driving gene prototypes, and finally obtains scenario-level driver-level evaluation results and individual behavior quality scores. Figure 1 As shown, the driving behavior evaluation method includes: Step S101: Obtain the trajectory data of the target driving object, and continuously segment the trajectory data to obtain multiple behavior segments; In this embodiment of the application, trajectory data refers to the sequence of motion state information of the target driving object as it changes over time during driving, including at least position, speed, acceleration, etc.
[0030] The target driving object can be the driver, and the driver , No. A behavioral fragment Composed of trajectory data within the start and end time interval, it can be represented as:
[0031] in, Index for driver or target driving object, Indicates the sampling index. Indicates driver At sampling time The trajectory data includes at least position, velocity, and acceleration; Indicates the behavior fragment index, and They represent the first The start and end times of each behavioral segment; For the first The duration of each behavioral segment; For behavioral fragments The number of sampling points within.
[0032] The duration of the behavioral segment falls within a preset range. Within this range, the dominant driving state, main interactive objects, local control modes, and external trigger states caused by road environment, traffic risks, or scene constraints remain basically consistent. Each segment is not required to have the same number of sampling points. and These represent the lower and upper limits of the duration of the behavioral segment, respectively. Both can be determined based on the data collection frequency and application scenario.
[0033] The triggering conditions for segment boundaries may further include, but are not limited to: the switching of the main interactive objects (such as the vehicle in front, the vehicle in the adjacent lane, and the conflicting vehicle), the traffic risk indicators (such as TTC and DRAC) crossing the preset threshold, or a significant change in the constraints of the traffic scene environment.
[0034] In this step, continuous trajectory data of the target driving object is acquired from the data acquisition device or storage medium. The trajectory data is divided into multiple behavioral segments with independent semantics according to preset trigger conditions. The control intent and interaction mode of each behavioral segment remain basically consistent within a short time range, and the duration of the behavioral segment is within a preset range.
[0035] Step S102: For each scene category, determine the behavior primitive layer evaluation result of the target driving object based on multiple behavior fragments, and determine the behavior primitive distribution of the target driving object; In this embodiment, the behavioral primitive layer is an intermediate representation layer between the continuous time slice layer and the driving gene prototype layer. The evaluation results of the behavioral primitive layer include the aggregated score matrix of the driver on various behavioral primitives in various scenarios. The distribution of behavioral primitives reflects the proportion of the driver's energy allocation to each behavioral primitive type in each scenario.
[0036] Let the set of scene categories be:
[0037] Where A represents the set of scene categories, (Uppercase) indicates the total number of scenario categories. Scenario categories include, but are not limited to, highway cruising, city driving, intersection driving, and driving in congested areas. Each behavior segment Mapping by scene classification Classified into a scene category (lower case).
[0038] Introducing drivers In scene category The distribution of behavioral primitives reflects the driver's "energy allocation across behavioral primitives" in this scenario, providing a basis for subsequent matching calculations:
[0039] For the driver In scene category Index of Behavioral Primitive Types The distribution of behavioral primitives is determined by the proportion of the cumulative duration of the segment to which the behavioral primitive type belongs; For the driver In scene category The following belongs to the behavioral primitive type index The collection of fragment indices For behavioral fragments Duration; For behavior fragment indexing, The index for summing the primitive types of the behavior in the denominator. For the driver In scene category The following is a summation index of the behavior primitive type. The collection of fragment indices For aggregation group Behavioral fragments for indexing The duration of behavior fragments with the same scene category and the same behavior primitive type index are grouped into the same aggregation group; The total number of behavior primitive types. For the driver In scene category Total number of complete behavioral primitive types under The distribution vector of the dimensional behavior primitives is used for matching calculations of the driving gene prototype layer; The symbol for the empty set; when = At that time, take =0.
[0040] In this step, after classifying each behavior segment into scenarios and behavior primitives, the segment-level evaluation results can be calculated based on multiple preset dimensions, such as safety, comfort, and energy consumption. Then, the segment-level evaluation results are aggregated according to scenario category and behavior primitive type to obtain the behavior primitive layer evaluation results. At the same time, the behavior primitive distribution is determined according to the proportion of the cumulative duration of each behavior primitive type to the total duration of the scenario.
[0041] Step S103: Project the evaluation results of the behavioral primitive layer onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype. Each driving gene prototype is determined according to the distribution of behavioral primitive composition of each driving object in the group driving data under multiple scene categories. In this embodiment, the driving gene prototype is a shared behavioral pattern learned from a group dataset containing multiple drivers. Each driving gene prototype is described by a weighted description of behavioral primitives under a set of scene categories. The weights describe which behavioral primitives the driving gene prototype prefers and their relative proportions.
[0042] In this step, a fixed number of driving gene prototypes are pre-learned based on group driving data. Each driving gene prototype has a composition weight for each behavioral primitive type in each scenario. According to the scenario category, the scores of each behavioral primitive in the behavioral primitive layer evaluation results of the corresponding scenario are aggregated, and the driving quality score corresponding to the driving gene prototype is obtained by weighted summation based on the composition weight of the behavioral primitives of the driving gene prototype in that scenario.
[0043] Step S104: Based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype, determine the matching weight between the target driving object and each driving gene prototype. In this embodiment, the matching weight is a normalized weight value determined based on the proximity between the distribution of the driver's actual behavior primitives and the distribution of the driving gene prototype composition, and is used to characterize the degree of matching between the driver's behavior style and each driving gene prototype.
[0044] In this step, the distribution difference measure between the distribution vector of the target driving object's behavioral primitives in each scenario and the weight vector composed of the behavioral primitives of each driving gene prototype in the same scenario is calculated. The difference measure is converted into a matching degree through an exponential kernel function. Then, multiple matching degrees in the same scenario are normalized to obtain the matching weight corresponding to each driving gene prototype.
[0045] Step S105: Determine the scenario-level driver evaluation result based on the matching weight and the driving quality score.
[0046] In this embodiment of the application, the scenario-level driver-level evaluation result is the core output of the driver-level in the driving behavior evaluation framework, representing the final evaluation score of the target driving object in the safety, comfort and energy consumption dimensions under each scenario category.
[0047] In this step, for each evaluation dimension, a weighted aggregation can be performed based on the matching weights corresponding to each driving gene prototype and the driving quality score to obtain the scene-level driver evaluation results of the target driving object under multiple evaluation dimensions. In other words, the matching weights corresponding to each driving gene prototype under the same scenario can be multiplied by the corresponding driving quality scores and then summed to obtain the scene-level driver evaluation results of the target driving object under each evaluation dimension in that scenario.
[0048] This application embodiment acquires the trajectory data of the target driving object and performs continuous segmentation. For each scenario category, it determines the behavioral primitive layer evaluation result and behavioral primitive distribution of the target driving object based on behavioral fragments. By projecting the behavioral primitive layer evaluation result onto multiple driving gene prototypes, it obtains the driving quality score of the target driving object from the perspective of multiple driving gene prototypes. That is, the score corresponding to the driving quality performance of the target driving object when each driving gene prototype is used as the evaluation benchmark. Then, by matching the behavioral primitive distribution of the target driving object with the behavioral primitive composition distribution of each driving gene prototype, it determines the matching weight of the driving style corresponding to each driving gene prototype. Finally, it comprehensively determines the scenario-level driver layer evaluation result based on the driving quality score (performance) and the matching weight (style). This makes the comprehensive evaluation between drivers of different styles style-aware. Even if two drivers obtain the same driving quality score on the same behavioral primitive, they will still obtain different scenario-level driver layer evaluation results due to their different behavioral styles. Drivers of different styles can obtain accurate evaluations under their respective matched driving gene prototypes.
[0049] In another embodiment of this application, step S102 determines the behavioral primitive layer evaluation result of the target driving object based on multiple behavioral fragments, such as... Figure 2 As shown, it includes: Step S201: Perform scene classification and behavior primitive classification on each behavior segment to obtain the scene category and behavior primitive type corresponding to each behavior segment; In this embodiment of the application, the scene category is used to describe the traffic environment type in which the behavior segment occurs, and the behavior primitive type is used to describe the local driving control intention and interaction mode of the behavior segment within a short time range.
[0050] A behavioral primitive is defined as a local driving behavior unit whose control intent and interaction pattern remain largely consistent within a short timeframe. Examples include stable following, accelerated following, deceleration avoidance, free cruise, lane change preparation, lane change execution, and intersection passage. Let the set of behavioral primitive types be... for:
[0051] in, Indicates the total number of behavior primitive types. Each behavior fragment... Classification mapping based on behavioral primitives Assigned a behavior primitive type index Optionally, the classification can be implemented using a "rule threshold + state machine" approach or a "clustering / latent variable model + semantic naming" approach. Regardless of the method used, a reproducible behavioral primitive type index must be output. To support subsequent aggregation calculations.
[0052] Total number of behavioral primitive types The total number of behavioral primitive types can be determined based on the classification method used. When using predefined rules or state machines, the total number of behavioral primitive types is... The number of predefined behavioral primitive types; the total number of behavioral primitive types when using clustering, latent variable models, or semi-supervised classification methods. The clustering profile coefficient, elbow method, BIC / AIC, clustering stability, minimum class size, scene coverage, and behavioral semantic separability can be comprehensively determined. Low-frequency types with insufficient sample size, semantic instability, or instability across repeated sampling can be merged into higher-level behavioral primitive types or "other" types. Total number of behavioral primitive types. Once determined, the distribution vector of subsequent behavioral primitives, the distribution vector of driving gene prototype composition, and the JSD calculation all use the same total number of behavioral primitive types. This is to ensure the stability of distribution comparison and matching weight calculation.
[0053] Supplementary explanation regarding behavior segment boundaries: The determination of the above segment boundaries can be triggered not only by changes in local control modes, but also by external factors such as changes in the main interaction object, sudden changes in traffic risk status, and changes in scene constraints. For situations where the attribution of behavior within the boundary neighborhood is ambiguous, one or more of the following optional processing methods can be adopted: (1) Soft assignment allows several sampling points or short windows in the neighborhood to be assigned to two adjacent behavioral primitive types at the same time, and outputs confidence or membership degree; (2) Overlapping window processing: Overlapping sliding windows are taken on both sides of the boundary for separate calculation to smooth the boundary truncation effect; (3) Neighborhood smoothing: Apply local smoothing post-processing to the behavioral primitive types of the boundary neighborhood.
[0054] It should be clarified that the aforementioned soft assignment or confidence mechanism is only an optional extended implementation method for this classification step and does not change the main formula chain based on a single reproducible label in subsequent steps. The default implementation method of this application is still to output a single, definite behavioral primitive type index for each behavioral fragment to directly support subsequent aggregation, distribution, projection, and matching calculations.
[0055] Optionally, model-based classification closed-loop verification. In an optional implementation, reliable classification of behavioral primitives can be achieved through the following closed-loop verification process: (1) First, a mapping relationship between behavioral primitives and kinematic feature profiles is constructed based on a small number of labeled samples. The feature profiles may include velocity-acceleration joint distribution profiles, workshop time-distance profiles, lateral offset profiles, etc.; (2) The unlabeled behavioral segments are automatically classified using the mapping relationship, and an initial behavioral primitive type index is output; (3) A verification closed loop is constructed through classification consistency test, temporal continuity test, and expert sampling verification. For segments with inconsistent or discontinuous classifications, backtracking correction is performed; (4) The above steps are iterated until the classification converges above the preset consistency threshold. This closed-loop verification implementation does not change the parallel relationship between the two main classification paths mentioned above, namely "rule threshold + state machine" and "clustering / latent variable model + semantic naming". It is only an optional supplement to enhance the reliability and reproducibility of classification.
[0056] Optionally, semi-supervised classification mapping can be used as another alternative implementation. In another alternative implementation, a behavior primitive classifier can be constructed based on a small number of labeled behavior fragments and a large number of unlabeled behavior fragments using semi-supervised learning methods (such as self-training, co-training, and label propagation) to reduce the dependence on full labeling and enhance the applicability of the method under different data conditions.
[0057] The input-output interfaces of the three methods mentioned above—"rule threshold + state machine", "clustering / latent variable model + semantic naming", and "semi-supervised classification mapping"—are consistent: the input is always a behavior fragment. The output, including the extracted kinematic features, interaction features, and road or scene constraint features, must at least include the scene category. and behavior primitive type index It can optionally output confidence or membership. The differences between the different methods lie in how the model parameters, thresholds, cluster centers, or classifiers are obtained, without changing the data interface required for subsequent aggregation, distribution construction, gene projection, and matching calculations.
[0058] In this step, kinematic features, interaction features, and road constraint features are extracted for each behavior segment. Each behavior segment is then mapped to a scene category and a behavior primitive type index using either rule thresholding and state machine methods or clustering and semantic naming methods.
[0059] Step S202: Based on the evaluation indicators included in the preset multiple evaluation dimensions, the indicators of each behavior segment are calculated and normalized to obtain the segment-level evaluation results of each behavior segment under the evaluation dimensions. In this embodiment, multiple evaluation dimensions are preset and can be set according to actual needs. For example, the evaluation system in this embodiment includes three individual attributable dimensions: safety, comfort, and energy consumption, as well as a system-level parallel output dimension of traffic flow efficiency, totaling twelve core indicators.
[0060] Safety dimensions include: (1) Minimum Collision Time (TTC) / Minimum Corrected Collision Time (MTTC), which reflects the longitudinal collision proximity in following, merging, and lane changing scenarios. The calculation source is the vehicle spacing, relative speed, and optional relative acceleration. The lower this index, the worse the situation. (2) Minimum Post-Encroachment Time (PET), which reflects the remaining time interval of the collision zone. The calculation source is the trajectory collision zone crossing time or simulation logs. The lower this index, the worse the situation. (3) Maximum Deceleration Rate to Avoid a Crash (DRAC), which reflects the severity of the scenario collision. The calculation source is the relative speed and distance in key interactions. The higher this index, the more severe the scenario.
[0061] The comfort dimension includes: (4) Root mean square of longitudinal acceleration, which reflects longitudinal driving smoothness. The calculation source is the speed and acceleration sequence. The higher the index, the worse the ride quality. (5) Root mean square of lateral acceleration, which reflects lateral driving smoothness and steering load. The calculation source is the lateral trajectory or lane change kinematic data. The higher the index, the worse the ride quality. (6) Root mean square of jerk, which reflects transient abruptness and impact. The calculation source is the time derivative of acceleration. The higher the index, the worse the ride quality.
[0062] The energy consumption dimension includes: (7) Energy intensity per unit distance, which reflects the direct energy demand per unit distance. The calculation source is the energy consumption model, fuel model or calibration simulation output. The higher this index, the worse it is; (8) Positive acceleration intensity, which reflects the propellant demand. The calculation source is the integral or mean of positive acceleration. The higher this index, the worse it is; (9) Frequency of intense maneuvers, which reflects the frequency of rapid acceleration and deceleration events. The calculation source is the event count based on the threshold. The higher this index, the worse it is.
[0063] System-level traffic flow efficiency indicators include: (10) Traffic volume, which reflects the number of vehicles passing through per unit time. The calculation source is the cross-section or entrance count. The higher the indicator, the better; (11) Average delay, which reflects the time loss relative to the ideal operating state. The calculation source is the simulation output or trajectory reconstruction. The higher the indicator, the worse; (12) Average queue length, which reflects the difficulty of congestion accumulation and dissipation. The calculation source is the detector, trajectory or simulation state. The higher the indicator, the worse.
[0064] For a more refined interpretation of comfort, indicators such as yaw rate variation or lateral deviation fluctuation can be incorporated. When a calibrated emissions or energy consumption model is available, CO2 emission intensity can be reported as an auxiliary environmental validation variable.
[0065] In this embodiment of the application, the segment-level evaluation result is a three-dimensional vector composed of the dimension scores obtained by each behavioral segment in the three dimensions of safety, comfort and energy consumption.
[0066] Behavioral fragments Upper The first evaluation dimension The original indicator values for each indicator are calculated as follows:
[0067] in, This represents the evaluation dimension index, with the individual level including safety, comfort, and energy consumption dimensions. Indicates the index of evaluation dimensions The index of indicators below, Indicates the first Evaluation Dimension 1 The calculation function for each indicator, For the number to be calculated A behavioral segment.
[0068] For mean or root mean square indicators, time-weighted calculation is used: Indicates the sampling index, if Represents the given first action within the action fragment. The observations at each sampling time point, , The interval between adjacent sampling times is the behavior segment. Time-weighted mean of internal observations and behavioral fragments Time-weighted root mean square value of internal observations Represented as:
[0069] in, For behavioral fragments Duration; This represents the number of sampling points within the behavior segment; The sampling index takes values from 1 to... .
[0070] Behavioral fragments Evaluation Dimensions The Middle Normalized index values of each indicator (The value is restricted to) ) is denoted as:
[0071] in, Indicates the first Evaluation Dimension 1 The normalized mapping function corresponding to each indicator, Behavioral fragments of input normalized mapping Upper The first evaluation dimension The original indicator values of each indicator.
[0072] Step S203: Aggregate the fragment-level evaluation results under each evaluation dimension according to the scene category and the behavior primitive type to obtain the behavior primitive layer evaluation results.
[0073] In this embodiment of the application, the evaluation result of the behavior primitive layer is an evaluation matrix obtained by aggregating the segment-level scores of each dimension according to the scene and primitive type.
[0074] In this step, behavioral fragments with the same scenario category and the same behavioral primitive type index are grouped into the same aggregation group. Within the same aggregation group, the safety dimension is weighted by an event-weighted strategy, and the comfort and energy consumption dimensions are weighted by the duration of the fragments to obtain the aggregation score matrix of each driver on each type of behavioral primitive in each scenario.
[0075] This application addresses the lack of semantic annotation in original trajectory data by classifying behavioral segments into scenarios and behavioral primitives, assigning semantically meaningful scenario categories and behavioral primitive type indexes to disordered trajectory segments. It calculates and normalizes indicators for each behavioral segment based on multi-dimensional evaluation metrics of safety, comfort, and energy consumption, transforming original indicators with varying physical dimensions into segment-level evaluation results at a unified scale, thus resolving the difficulty of directly fusing and comparing multi-dimensional indicators. Furthermore, it aggregates segment-level evaluation results by scenario category and behavioral primitive type to obtain behavioral primitive layer evaluation results, preserving both the differences in scenario environment and behavioral patterns. This provides a structured intermediate data interface for subsequent style matching and projection evaluation at the driving gene prototype layer. Therefore, this method inserts a behavioral primitive layer as a stable intermediate representation between the trajectory layer and the driving gene prototype layer, solving the problem of the lack of a unified connection mechanism between locally interpretable behavioral units and high-level driver difference structures in existing technologies.
[0076] In another embodiment of this application, step S202 normalizes each behavioral segment based on the evaluation indicators included in a set of preset evaluation dimensions to obtain segment-level evaluation results for each behavioral segment, such as... Figure 3 As shown, it includes: Step S301: For evaluation indicators with preset physical thresholds in multiple evaluation dimensions, the calculation results of each evaluation indicator are normalized by using monotonic mapping based on the Sigmoid function or piecewise function mapping to obtain the first normalized indicator value. In this application embodiment, the evaluation index with preset physical threshold refers to an index with a clear physical safety threshold or regulatory limit, including the minimum collision time TTC, minimum rear intrusion time PET, maximum deceleration required to avoid collision DRAC in the safety dimension, and the root mean square of acceleration Jerk in the comfort dimension.
[0077] Strategy A: Threshold-based normalization based on the Sigmoid function. For safety and comfort dimension indices with explicit physical safety thresholds or regulatory limits (such as TTC, PET, DRAC, Jerk, etc.), monotonic Sigmoid mapping or piecewise function mapping can be used.
[0078] Strategy B (Optional): Transform Sigmoid Normalization. For benefit-type indicators with large dynamic ranges or theoretically approaching infinity, such as TTC and PET, directly using a single Sigmoid mapping may result in insufficient resolution in the high-safety range. In an optional implementation, the original indicator value is first pre-transformed to fall into a numerical range suitable for Sigmoid mapping, and then Sigmoid normalization is performed. The pre-transformation can be one or more of the following: (1) truncation to preset upper and lower bounds followed by linear scaling; (2) monotonically bounded transformation, such as tanh mapping or arctan mapping; (3) reciprocal transformation or logarithmic transformation to map a large range of positive real numbers to a more compact numerical range. Taking TTC as an example, its reciprocal 1 / TTC can be calculated first as a conflict urgency proxy, and then Sigmoid normalization can be performed.
[0079] When reciprocal transformation, logarithmic transformation, or other pre-transformation methods are used, the subsequent segment-level dimension scores, behavioral primitive layer evaluation results, and the safety dimension driving quality score in formula (17) are all based on the safety quality score obtained after pre-transformation and directional normalization. In this case, the physical meaning of the safety dimension driving quality score in formula (17) is: under the weighted behavioral primitive composition of the driving gene prototype, the score for the driver... In scene category The evaluation is based on a weighted average of the conflict urgency suppression capacity or safety margin performance of each behavioral primitive, rather than a linear weighting of the original time quantities such as TTC and PET.
[0080] In another alternative implementation, a piecewise function mapping can be used to divide the original value range of the index into a danger zone, an intermediate zone, and a high-safety zone, with each zone using a different mapping function (e.g., a steep linear mapping for the danger zone and a flat saturation mapping for the high-safety zone) to maintain appropriate resolution within different safety level ranges.
[0081] It should be clarified that the above-described Sigmoid normalization and piecewise mapping can both be used as optional implementations of the normalization process; when using the direct Sigmoid normalization shown in formulas (8) and (9) below, it can be used in combination with the pre-transformation process.
[0082] For benefit-oriented indicators where higher values are preferred (such as TTC and PET), the normalized indicator value... for:
[0083] in, Evaluation Dimensions The Middle Reference thresholds for each indicator The slope parameter of the Sigmoid function. This is used to control the steepness of the gradient near the threshold. hour This indicates that the safety margin is sufficient; when hour This indicates a critical danger state.
[0084] For higher-valued but worse-than-average cost-related metrics (such as the deceleration required to avoid a collision (DRAC) and the root mean square of jerk, which reflects transient abruptness and impact), the normalized metric value... for:
[0085] It is a natural exponential function. Indicates behavior fragment index, Indicates the evaluation dimension index. Indicates the index of the indicator, when hour It indicates smoothness / mildness; when hour This indicates that the conflict is serious.
[0086] This mapping is naturally bounded to ,exist At maximum sensitivity , For the normalization function of the original index value The derivative of; the maximum sensitivity at the reference threshold is / 4, naturally saturates when far from the threshold, requiring no additional truncation operation.
[0087] The reference thresholds are sourced as follows: TTC can use 1.5s (critical collision hazard) or 3.0s (normal safety margin); PET can use 1.0s or 1.5s; DRAC can use a comfort deceleration threshold of 2.0s. Or emergency deceleration threshold 4.0 Jerk can refer to ISO 2631-1 comfort reference limits. Slope parameter. The setting satisfies: the score is 0.5 at the threshold, and , This is the index offset relative to the reference threshold; When, the deviation from the threshold ± Achieve the specified score at the designated location. Let be the natural logarithm function, such that when deviating from the threshold The scores were 0.25 and 0.75 respectively, among which, Subtraction before addition, i.e. ( →0.25、+ →0.75), indicating that it is below the reference threshold. The normalized score is 0.25, which is higher than the reference threshold. The normalized score is 0.75. For cost-related indicators, ± represents addition followed by subtraction, with opposite directions (+). →0.25、 →0.75).
[0088] For cost-type indicators (such as DRAC) with long-tailed distribution or occasional extreme pulses, it is necessary to first perform truncation or robust transformation (such as Winsorized truncation or logarithmic transformation) and then input formula (9) for Sigmoid normalization to reduce the excessive influence of outliers on the normalized indicator value.
[0089] Step S302: For evaluation indicators without preset physical thresholds in multiple evaluation dimensions, normalization is performed by scaling to a preset range to obtain the second normalized indicator value. In this application embodiment, the evaluation index without a preset physical threshold refers to a relative index that does not have a uniform physical limit value, including the energy consumption intensity per unit distance, positive acceleration intensity, and frequency of intense maneuvers in the energy consumption dimension.
[0090] For relative indicators without a clear physical threshold (such as energy consumption proxy indicators), Min-Max normalization is used; for traffic flow efficiency indicators, they can be normalized or summarized separately according to the calculation method of system-level traffic flow efficiency results.
[0091] For benefit-type indicators where higher values are preferred:
[0092] For cost-related indicators where higher values are worse:
[0093] in, For the first The first evaluation dimension The normalized lower and upper bounds of each indicator. and It can be determined by the full sample statistical range, quantile cutoff interval, or literature threshold. These are the original index value and the normalized index value after Min-Max normalization, respectively. The minimum and maximum value operators are used to truncate the normalized index value to... In practice, a truncation operation is applied to the normalized index value. To prevent outliers from exceeding the limit.
[0094] Step S303: The first normalized index value and the second normalized index value are weighted and summed according to the preset index weights within the evaluation dimension to which the evaluation index belongs, so as to obtain the segment-level dimension score of each behavior segment under each evaluation dimension. In practical engineering deployments, the calculation of safety dimension indicators such as TTC, PET, and DRAC may be affected by factors such as sensing noise, target loss, target ID switching, and intermittent sensor failure. To improve the reliability and robustness of indicator calculations, one or more of the following safeguards can be added in steps S301 to S304: (1) Indicator calculation feasibility check module: Before calculating the safety dimension indicators, first determine whether the necessary input quantities (such as vehicle spacing, relative speed, relative acceleration) are within the effective range. If the input quantities are missing or obviously abnormal, mark the corresponding indicator of that moment or segment as unusable. (2) Outlier removal and smoothing: Apply a sliding window median filter or Hampel filter to the index sequence to remove outliers that exceed the median absolute deviation by a preset multiple; (3) Exclusion of low-confidence samples: For sampling points whose target loss time exceeds the preset threshold or within the preset time window before and after the ID-Switch event, reduce their weight in the index aggregation calculation or exclude them directly; (4) Robustness statistics substitution: When the kurtosis or skewness of the original index sequence exceeds the preset threshold, the median can be used to replace the mean and the median absolute deviation (MAD) can be used to replace the standard deviation as aggregate statistics.
[0095] The above safeguards are optional supplements to enhance the feasibility of the project and do not change the main calculation logic of steps S301 to S304 and the subsequent aggregation formula chain.
[0096] In this embodiment of the application, the segment-level dimension score is the score obtained by weighting and summing multiple normalized index values within the same dimension according to the weights within the dimension, and the sum of the weights within each dimension is 1.
[0097] Behavioral fragments In evaluation dimensions The following segment-level dimension scores are:
[0098] in, Evaluation Dimensions The number of indicators below Evaluation Dimensions Next The weights within each dimension of the indicator, with the sum of all weights within the same dimension being 1. For indexing the sum of indicators, values range from 1 to... .
[0099] Weight The determination is achieved through a dual-path process: a structured expert judgment path and a reference equal-weighted path. In the structured path, candidate indicators are scored based on theoretical relevance, practical measurability, and safety-critical impact, and after normalization, they meet the following requirements. In the reference path, indicators within each dimension are set to equal weight. The final report must specify the methods used. And a robustness comparison with the equal-weighted path to transparently examine ranking stability.
[0100] Step S304: Determine the segment-level evaluation result of each behavioral segment based on the segment-level dimension score of each behavioral segment under each evaluation dimension.
[0101] In this embodiment of the application, the segment-level evaluation result is a comprehensive evaluation vector composed of the segment-level dimension scores obtained by each behavioral segment under the safety dimension, comfort dimension and energy consumption dimension.
[0102] In this step, the safety, comfort, and energy consumption scores of the same behavioral segment are combined to form a segment-level evaluation result for that segment, with each segment having a three-dimensional score vector.
[0103] For example, if a certain behavioral segment has a safety dimension score of 0.77, a comfort dimension score of 0.82, and an energy consumption dimension score of 0.65, then the segment-level evaluation result of this segment is (0.77, 0.82, 0.65).
[0104] This application's embodiments employ a dual-track strategy: using Sigmoid threshold normalization for indicators with physical thresholds and Min-Max interval normalization for indicators without physical thresholds. This allows evaluation indicators of different properties to be fused under a unified scale, preserving sensitivity near the safety threshold while avoiding excessive influence of outliers on relative indicators, thus providing standardized quantitative input for subsequent hierarchical aggregation.
[0105] In another embodiment of this application, step S203 aggregates the fragment-level evaluation results under each of the evaluation dimensions according to the scene category and the behavior primitive type to obtain the behavior primitive layer evaluation results, such as... Figure 4 As shown, it includes: Step S401: Group fragments with the same scene category and the same behavior primitive type into the same aggregation group; In this embodiment of the application, an aggregation group is a collection of fragments defined by two dimensions: scene category and behavior primitive type. All behavior fragments within the same aggregation group have the same scene category and the same behavior primitive type index.
[0106] For the driver In scene category The next line is a primitive type. The corresponding set of behavior fragment indices for:
[0107] in, Indicates driver The complete set of indexes of all behavioral fragments, Classify and map behavior primitives, and output behavior fragments. Behavioral primitive type index , Classify and map scenes, and output behavior fragments. Scene categories . Index for behavior fragments.
[0108] driver In scene category The next line is the primitive type index. Corresponding evaluation dimensions aggregate score for:
[0109] in, For the driver In scene category Behavioral primitive type index The following is a list of fragments. The aggregate weights are such that the sum of the weights within the same aggregate group is 1. For behavioral fragments In evaluation dimensions The segment-level dimension score below.
[0110] In this step, all behavior fragments are filtered based on both their scene category and behavior primitive type, and those that meet the scene category criteria are selected. Same and behavior primitive type index All identical fragments are grouped into the same aggregation group.
[0111] Step S402: Within any aggregation group, according to the duration of the behavior segment or by adopting an event equal weighting strategy, the segment-level dimension scores in the segment-level evaluation results of each behavior segment are weighted and aggregated to obtain the aggregation score. Specifically, within any aggregation group, based on the segment-level evaluation results of each behavioral segment, the evaluation indicators for the comfort and energy consumption dimensions are weighted and aggregated according to the duration of the behavioral segment to obtain the first aggregation score matrix; within any aggregation group, based on the segment-level evaluation results of each behavioral segment, the evaluation indicators for the safety dimension are aggregated using an event-weighted strategy to obtain the second aggregation score matrix. In this embodiment, the first aggregated score matrix is a matrix composed of the aggregated scores of the comfort dimension and energy consumption dimension on various behavioral primitives in various scenarios. The aggregated score within each aggregate group is equal to the average of the segment dimension scores within that group, weighted by segment duration. In this embodiment, the second aggregated score matrix is a matrix composed of the aggregated scores of the safety dimension on various behavioral primitives in various scenarios. The safety dimension aggregated score within each aggregate group is equal to the arithmetic mean of the safety dimension scores of each segment within that group.
[0112] Comfort and energy consumption dimensions: weighted by segment duration, aggregated weights. , For behavioral fragments Duration; The index for summing fragments within an aggregation group. For aggregation group Behavioral fragments for indexing Duration.
[0113] Safety incidents are characterized by "low frequency and high severity" (e.g., 1 second of emergency braking). Weighting them by duration would dilute these critical events with a large number of low-risk segments. Therefore, safety maintenance adopts an event-weighted approach. This ensures that each segment contributes equally to the safety score. A mixed weighting approach can also be used. ,in, It is a hybrid coefficient between duration weighting and event equal weighting. Security is the default setting. (Pure event equal weight). The number of behavioral fragments in the aggregate group, i.e., the set The base number, in its physical sense, means that segments with longer durations are more representative of the driver's comfort and energy consumption performance in that behavioral primitive type.
[0114] In this step, for the aggregation groups of comfort and energy consumption dimensions, the comfort dimension scores or energy consumption dimension scores of each segment within the same aggregation group are weighted and summed according to the proportion of each segment's duration to the total duration, to obtain the driver's comfort dimension aggregation score and energy consumption dimension aggregation score for this behavior primitive type in this scenario. For the aggregation groups of safety dimensions, the safety dimension scores of each segment within the same aggregation group are averaged in an equally weighted manner, meaning that each segment contributes equally to the safety score, to obtain the driver's safety dimension aggregation score for this behavior primitive type in this scenario.
[0115] Step S403: Determine the evaluation result of the behavioral primitive layer based on the aggregation scores corresponding to multiple aggregation groups. Specifically, the evaluation result of the behavioral primitive layer can be determined based on the first aggregation score matrix and the second aggregation score matrix.
[0116] In this embodiment of the application, the evaluation result of the behavioral primitive layer is jointly composed of the first aggregated score matrix of the comfort dimension and the energy consumption dimension and the second aggregated score matrix of the safety dimension.
[0117] In this step, the aggregated scores of safety, comfort and energy consumption dimensions on various behavioral primitives in each scenario are combined into a complete aggregated score matrix, which is output as the evaluation result of the behavioral primitive layer to the subsequent driving gene prototype layer processing.
[0118] when At that time, aggregate score Missing values should not be forced to zero; they should be treated as missing values. The missing data handling strategy is as follows: If a certain scene category If all behavioral primitive types are missing (no observations), then this scenario will not participate in the driver aggregation calculation.
[0119] If scene category If the lower part of the behavioral primitive types are missing, then the weight of the behavioral primitive composition in that scenario will be determined when aggregating the driving gene prototype layer. Renormalization is performed only for non-missing types:
[0120] in, This represents the weight of the behavioral primitive composition of the driving gene prototype after renormalization of the set of non-missing behavioral primitive types. Indicates scene category There exists a set of behavioral primitive types for the observed fragment. This represents the empty set.
[0121] Alternatively, interpolation can be performed using neighboring behavior primitive types within the same scene.
[0122] This application embodiment groups fragments with the same scenario and primitive type into aggregation groups and adopts a differentiated aggregation strategy within the groups, which uses equal weighting for safety dimension events and duration weighting for comfort energy consumption dimension. This ensures that the safety dimension evaluation is not diluted by low-risk long-term fragments, while the comfort energy consumption dimension evaluation can reflect the dominant contribution of long-term fragments. This provides an aggregation score matrix for subsequent driving gene prototype layer projection that retains the sensitivity to safety-critical events and reflects long-term energy consumption comfort characteristics.
[0123] In another embodiment of this application, step S103 projects the evaluation results of the behavioral primitive layer onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype, such as... Figure 5 As shown, it includes: Step S501: Based on the group driving data under multiple scenario categories, obtain the driving gene prototype corresponding to each scenario category in advance. The driving gene prototype includes the weights of behavioral primitives corresponding to multiple behavioral primitive types. In this step, group driving data covering multiple driver types and various traffic scenarios are collected. Distribution vectors are constructed for each driver in the group data according to the frequency of occurrence of behavioral primitives under the same scenario category. Clustering or factor analysis is performed on all distribution vectors to extract Z shared behavioral patterns. Each pattern corresponds to a driving gene prototype and its weights as composed of behavioral primitives in each scenario.
[0124] The driving gene prototype was learned from a population dataset containing multiple drivers. Specifically, it was extracted by clustering or factor analysis of the frequency (rather than quality scores) of behavioral primitives of all drivers in scene categories. Each shared behavioral pattern. Each driving gene prototype consists of a set of behavioral primitives across various scenario categories, described by weights:
[0125] in, Indicates the primitive type index of the behavior. This represents the prototype index of the driving gene. Indicates the driving gene prototype in the scenario category Index of Behavioral Primitive Types The behavioral primitives are composed of weights, for The sum is 1. The total number of behavioral primitive types; This represents the total number of driving gene prototypes. Values range from 1 to This weight describes "which behavioral primitives this driving gene prototype prefers and their relative proportions"—a behavioral style description independent of quality assessment dimensions. In other words, the same driving gene prototype The calculations for safety, comfort, and energy consumption remain consistent, with the only difference being the aggregated scores of each behavioral primitive across different dimensions. The traffic flow efficiency results are not the weights themselves, but rather the system-level parallel outputs; they are not included in the calculation of the driving gene prototype projection.
[0126] Once these weights are determined through group learning, they become a fixed structure. Subsequent evaluations of any individual driver simply require mapping their behavioral scores to this structure. Based on a known driving gene prototype, there is no need to relearn the gene structure.
[0127] For example, a gene might assign higher weights to behavioral motifs like "accelerating and following" and "frequent lane changes" in high-speed scenarios, and "running a yellow light" and "emergency deceleration and avoidance" in intersection scenarios, thus obtaining a semantic label of "aggressive." This behavioral motif preference is consistently applied to evaluations across various dimensions such as safety and comfort, with the same behavioral motif exhibiting different quality performance across different dimensions.
[0128] A reference implementation process for learning driving gene prototypes: In one specific implementation, the population learning of the driving gene prototype can be performed according to the following process: (1) Group data coverage requirements: The collected group driving dataset covers multiple driver types, typical regional / road environments (urban expressways, main roads, secondary roads, highways, etc.) and typical traffic scenarios (free flow, synchronous flow, congested flow) to ensure that the extracted driving gene prototype has the ability to generalize scenarios.
[0129] (2) Preprocessing of the vector of behavioral primitive distribution: For the vector of behavioral primitive distribution of each driver in the same scene category, feature filtering (such as removing behavioral primitive types with near zero variance), dimensionality reduction (such as PCA or t-SNE), sparse representation (such as retaining only behavioral primitives with a proportion higher than a preset threshold) or group statistics (such as merging low-frequency behavioral primitives into higher categories according to semantic similarity) can be used to preprocess the vector to reduce the interference of high-dimensional sparsity on the clustering effect.
[0130] (3) Processing of sparse scene / low-frequency behavior primitives: When a scene category has sparse samples in the group data, the following methods can be used: internal reweighting of behavior primitives in the scene (increasing the weight of low-frequency but safety-critical behavior primitives), merging of neighboring scenes (merging semantically similar scene categories and then clustering them), or merging of low-frequency categories (merging behavior primitives with a frequency lower than a preset threshold into the "other" category).
[0131] (4) Determination of the number of clusters: Total number of driving gene prototypes It can be determined comprehensively through methods such as Silhouette Score, Elbow Method, Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC), or cluster stability analysis (repeatedly clustering random subsamples and calculating the consistency of cluster assignments), rather than a single fixed value.
[0132] (5) Validation of prototype learning: After the prototype learning of driving genes is completed, it can be validated by resampling consistency test (independent clustering is performed on different random subsets of the population data to compare the stability of cluster centers and the weights of behavioral primitives), leave-one-driver cross-validation (re-clustering after removing drivers one by one to check whether the prototype structure remains stable) or a small amount of manual semantic verification (checking whether the behavioral primitive composition of typical driving gene prototypes is consistent with domain knowledge).
[0133] It should be noted that the above process is a reference implementation method used to illustrate the engineering feasibility of learning the driving gene prototype. The learning results of the driving gene prototype are not necessarily subject to the submission of complete population learning experimental data.
[0134] Step S502: For each evaluation dimension, based on the weights of the behavioral primitives corresponding to each behavioral primitive type in the driving gene prototype, the evaluation results of the behavioral primitive layer are projected onto multiple driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype.
[0135] In this embodiment of the application, the driving quality score is the result of weighted summation of the behavioral primitive layer evaluation results according to the weights of the behavioral primitive composition corresponding to multiple behavioral primitive types in the driving gene prototype.
[0136] driver In scene category The following are evaluation dimensions of the driving gene prototype. Driving quality score Defined as:
[0137] in, This represents the weight of the driving gene prototype in relation to the behavioral primitive composition of each behavioral primitive. To aggregate the scores, This is the index for the behavior primitive type, with values ranging from 1 to the total number of behavior primitive types. The driving quality score answers the question: "If behavioral primitive preferences based on the driving gene prototype are applied, what are the driver's preferences?" What is the quality score? For example, an "aggressive" gene in a high-speed scenario assigns high weights to "acceleration and following" and "lane change execution". Its driving quality score measures the driver's overall quality on these key behavioral primitives, rather than a simple average of all behavioral primitives.
[0138] It needs to be clearly pointed out that the driving quality score Based solely on the driver In scene category The aggregate score formed by the actual observed behavioral segments calculate. This reflects a quality review of the driver's existing behavioral records under a specific gene template and does not constitute an inference of the driver's potential executive or counterfactual ability on behavioral primitives that have not occurred. For behavioral primitive types with insufficient sample size or that have not been observed, their corresponding aggregate scores should not be interpreted as evidence of driving ability.
[0139] It should be noted that, for fixed , and ,exist Each driver has a driving quality score (one for each scenario), which ensures that even when evaluating a single driver, the driving gene prototype layer does not degenerate into a single scalar.
[0140] In this step, for the evaluation results of the target driving object's behavior primitive layer in a specific scenario, the aggregate score of each behavior primitive type in the scenario is multiplied by the composition weight of the corresponding behavior primitive type of the driving gene prototype in that scenario and then summed to obtain the driving quality score of the driver corresponding to the driving gene prototype in that scenario. Each driver generates a driving quality score for each driving gene prototype in each scenario.
[0141] This application embodiment obtains a fixed number of driving gene prototypes by clustering from group driving data, so that subsequent target driving objects only need to be projected onto the known prototype space without relearning the gene structure, which reduces the computational complexity of online evaluation. At the same time, the driving quality score can measure the comprehensive quality performance of drivers under a specific gene preference template, providing a quantitative basis for style perception evaluation.
[0142] In another embodiment of this application, step S104 determines the matching weight between the target driving object and each driving gene prototype based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype under each scenario category. Figure 6 As shown, it includes: Step S601: Calculate the distribution difference measure between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype; Relying solely on driving quality scores It is impossible to distinguish between "a driver performing well under this gene template" and "a driver's actual driving habits matching this gene pattern." A driver may perform well on all behavioral motifs (i.e., all...). The scores are all relatively high, resulting in high scores under any gene projection, but this does not mean that the driver's behavior style is consistent with that gene.
[0143] To separate "performance" from "style," a matching mechanism is introduced. The matching degree measures the driver's performance. In scene category Distribution of actual behavioral primitives The behavioral primitive composition pattern of the driving gene prototype in this scenario The degree of proximity between them.
[0144] The aforementioned matching weights are also constructed solely based on the actual experience distribution of observed behavioral primitives, without involving implicit inferences about unobserved behavioral primitives. When the number of observed samples of behavioral primitives in a certain scenario is insufficient, a low-confidence label can be added to the matching weights for that scenario at the reporting layer. These labels are generally not used in the comparative interpretation of cross-driver comprehensive evaluations. The low-confidence label is an auxiliary constraint at the reporting or interpretation layer and is not directly written back into the core calculation formulas (18)-(20) of the matching weights.
[0145] In the main implementation, this embodiment uses Jensen-Shannon divergence (JSD) as the difference measure. The Jensen-Shannon divergence is defined as:
[0146] in, and Let represent two probability distribution vectors to be compared. This represents the mixed distribution obtained by averaging two probability distribution vectors to be compared. is the Jensen-Shannon divergence, used to measure the difference between two probability distributions, and its value ranges from [0, ln2].
[0147] in, This is the Kullback-Leibler divergence. This definition ensures... ,when hour and (Perfect match) When the two distributions do not overlap at all and Here The unnormalized matching degree is defined by formula (19) below (obtained by JSD through negative exponential kernel transformation). >0 represents the temperature parameter in this formula.
[0148] In the KL divergence expansion calculation, for terms with zero probability components, the limiting convention is used. The positive probability component ,Right now , The probability components that approach 0 are dummy variables in local limit operations, not the original values of the index. , log is the logarithmic function; lim is the limit operator. If the probability of a certain row primitive type is 0 in both distributions, then that dimension is also 0 in the mixed distribution, and its contribution to both KL terms is recorded as 0; if the probability of one of the distributions is positive, then the corresponding component of the mixed distribution is positive, and there will be no case where the positive probability term is divided by zero.
[0149] The reason for choosing JSD instead of other distribution distance metrics such as cosine similarity is: (1) behavioral primitive distribution and gene composition distribution Both are probability distribution vectors. JSD is a standard difference measure between probability distributions, which is suitable for this scenario; (2) When both distributions are close to uniform distributions, the cosine similarity approaches 1.0 and loses its ability to distinguish, while JSD can sensitively capture subtle structural differences; (3) JSD is symmetric and bounded to The square root of this distance constitutes an effective distance metric that satisfies the triangle inequality. When using the Jensen-Shannon divergence as a measure of distributional dissimilarity, the matching degree is defined as:
[0150] in, For the driver In scene category The unnormalized match degree with the driving gene prototype was obtained by JSD through a negative exponential kernel transformation. The temperature parameter controls the sensitivity of the matching degree to distribution differences. Drivers In scene category The distribution vector of behavioral primitives and the index of driving gene prototypes. In scene category The distribution vector is composed of the basic elements of the behavior. exp is the natural exponential function; JSD is the Jensen-Shannon divergence. Since... Therefore The range of values is .
[0151] When alternative difference measures such as Bhattacharyya distance, Hellinger distance, or EarthMover's Distance are used, only the distribution difference measure term in formula (19) is replaced, and the output interface of subsequent exponential kernel transformation, matching degree normalization, and normalized matching weight remains unchanged; the above reasons for choosing JSD apply to the main implementation method and do not constitute an exclusion of alternative measure implementation methods.
[0152] In addition to using Jensen-Shannon divergence (JSD) as a measure of distribution difference, in an optional implementation, distance measures between probability distributions such as Bhattacharyya distance, Hellinger distance, or EarthMover's Distance can also be used to calculate the matching weight between the driver behavior primitive distribution and the gene composition distribution. The specific measure to be used can be determined according to the application scenario's requirements for the sensitivity of distribution difference.
[0153] Step S602: Determine the matching weight between the target driving object and each of the driving gene prototypes based on the multiple distribution difference measures.
[0154] In this step, each of the distribution difference measures can be used as the independent variable of the exponential kernel function. After undergoing a negative exponential transformation using the exponential kernel function, the matching degree is obtained. Multiple matching degrees are then normalized to obtain the matching weights between the target driver and each of the driver gene prototypes. By converting the distribution difference measure into a matching degree through a negative exponential transformation and then normalizing it into a matching weight, the strict positive property of the exponential kernel function ensures that the normalization denominator is never zero, making the matching weight calculation mathematically stable and reliable. Simultaneously, the temperature parameter can flexibly adjust the smoothness of the matching distribution to accommodate different numbers of driver gene prototypes.
[0155] In this embodiment of the application, the normalized matching weight is:
[0156] To normalize the matching weights obtained from the matching degrees of different driving gene prototypes under the same driver and scenario, the driving gene prototype index is... The sum is 1. This represents the unnormalized matching degree. Summation index for the driving gene prototype; This specification represents the total number of driving gene prototypes. , , Symbols marked with an apostrophe are summation indices for the corresponding set or number range.
[0157] because It always holds true (the exponential function is strictly positive), the denominator is never zero, and normalization is always defined.
[0158] In this embodiment, the transitions between the four levels answer complementary causal questions: Excerpt Behavioral primitive: "How would this type of behavior typically score in this scenario?" — Answered by aggregating the quality of similar segments (weighted by duration).
[0159] Behavioral primitives Gene (projection): "If the behavioral primitive preferences of the driving gene prototype are applied, the driver..." What is the quality score? — Weights are composed of behavioral primitives under each scenario category. right Linearly weighted answer.
[0160] Behavioral primitives Gene (match): "driver" How closely does the actual behavioral distribution resemble the composition of the driving gene prototype? — Through and The JSD kernel-based matching answer between them.
[0161] Gene Driver: "What is the complete profile of this driver across different scenarios and genes?" — Through scenarios The answer is the product matrix of the gene's driving quality score and the matching weight.
[0162] This forms the specific chain of a hierarchical causal chain:
[0163] The core function of the driving gene prototype layer is not to reduce the dimensionality or compress information of behavioral primitive scores, but to transform the scenario-level driver evaluation results... The complete formula can be expanded to obtain:
[0164] Among them, effective weights .
[0165] That is, the driver's final score is still a linear combination of behavioral primitive scores, but the effective weights... It is driver-adaptive. Two drivers obtain exactly the same aggregate score on the same behavioral primitives. They will still differ due to their different behavioral styles. Different results (from different values) lead to different overall evaluations, thus resulting in different effective weights. .
[0166] Therefore, the role of the driving gene prototype layer is adaptive weighting: it forms corresponding evaluation weight templates based on the driver's actual behavioral style. Safety-oriented drivers are evaluated using the safety-oriented template, and aggressive drivers are evaluated using the aggressive template, thus making the comprehensive evaluations between drivers of different styles style-aware and comparable.
[0167] It is important to emphasize that the adaptive weighting mechanism of the driving gene prototype layer is a weighted reconstruction performed within the scope of observed behavioral primitives, aiming to improve style comparability rather than speculating on the driver's potential performance on unobserved behavioral primitives. Effective weights The physical meaning of this is "the relative importance of the quality of each behavioral element to the overall evaluation given the actual behavioral style of the driver", and it should not be interpreted as a counterfactual evaluation of behaviors that did not occur.
[0168] Specifically, when the number of driving gene prototypes is less than or equal to a preset threshold, the temperature parameter in the exponential kernel function is a preset temperature parameter; when the number of driving gene prototypes is greater than the preset threshold, the temperature parameter in the exponential kernel function is determined based on the number of driving gene prototypes. By using a fixed temperature parameter when the number of driving gene prototypes is small to maintain a smooth matching transition, and using an adaptive temperature parameter when the number of driving gene prototypes is large to enhance the matching discriminativeness, the matching weight calculation can adapt to scenarios with different numbers of driving gene prototypes, ensuring the stability and interpretability of the matching results under different implementation conditions.
[0169] In practical applications, temperature parameters The settings provide two heuristic starting points: Reference 1: Uniform discrimination force design At this point, given Minimum Unnormalized Matching Degree This means that genes with no matching retain approximately 36.7% of the weights of genes with a perfect match. This design is derived from the natural decay property of the exponential function when the parameter is 1. The weight ratio between completely mismatched and perfectly matched cases is... This provides a moderate baseline that is "discriminatory but not overly exclusive." When Furthermore, it can serve as a reference value when a smooth transition between genes is desired. ln is the base of the natural logarithm; ln is the natural logarithm. This represents the number of candidate driving gene prototypes.
[0170] Reference 2: Adaptive design for the number of prototypes. .at this time That is, the gene weight ratio in the case of perfect matching and perfect mismatch is Its intuition is: when it exists When considering candidate gene prototypes, the weights in the perfect match case can be maintained relative to the perfect mismatch case. The difference in magnitude allows soft allocation to behave similarly to hard allocation while retaining smoothness. It can also be used as a reference value when a clearer distinction between genes is needed.
[0171] The above The values are heuristic reference points, not rigorous theoretical derivations within this framework. The applicable values for the exponential core temperature parameter depend on the actual distribution characteristics of the data; therefore, specific values need to be considered for different... Sensitivity analysis was performed on the value to examine its impact on the final evaluation ranking.
[0172] This application embodiment quantifies the degree of similarity between the actual distribution of driver behavior and the composition distribution of each driving gene prototype by measuring the distribution difference. Then, through exponential kernel transformation and normalization, the difference measure is transformed into normalized matching weights, realizing a quantitative mapping of behavioral style from the distribution space to the weight space, and providing a calculable matching basis for style adaptive evaluation weighting.
[0173] While traditional micro-models offer good semantic interpretability, they struggle to fully represent driver heterogeneity; natural driving data-driven methods, while exhibiting high behavioral realism, lack stable and transparent intermediate layer representation structures; and existing behavior evaluation methods fail to differentiate system-level traffic flow efficiency. Therefore, in another embodiment of this application, as... Figure 7 As shown, the method further includes: Step S701: Based on the scenario-level driver evaluation results, aggregate the scenario-level driver evaluation results according to the preset scenario exposure weight to obtain individual behavior quality scores. In this embodiment of the application, the individual behavior quality score is a single summary value obtained by weighting and aggregating the scene-level evaluation results of the three individual attributable dimensions—safety, comfort, and energy consumption—according to scene exposure weights, and then combining them according to dimension weights.
[0174] driver In scene category Evaluation Dimensions Scenario-level driver evaluation results Obtained through weighted aggregation based on matching weights:
[0175] These are the normalized matching weights and the driving quality score, respectively. For the driving gene prototype index, the value ranges from 1 to ; This represents the total number of driving gene prototypes. The resulting scenario-level driver evaluation results constitute a scenario-level evaluation result matrix across the safety, comfort, and energy consumption dimensions.
[0176] To further preserve the interpretable decomposition of gene contributions, define the driver. In scene category Lower driving gene prototype on evaluation dimensions The contribution is as follows:
[0177] thereby The contribution components of each driving gene prototype can be directly used in radar charts or stacked bar charts to reveal "which behavioral styles dominate the driver's quality performance in a specific scenario".
[0178] If a single aggregated value is required, it can be further aggregated by scenario-based exposure weights:
[0179] in, Indicates driver In evaluation dimensions The comprehensive dimensional score is obtained after aggregating the weights of the scenarios exposed. This represents the total number of scene categories. Indicates driver In scene category The scene exposure weight, Values range from 1 to This embodiment defaults to using scene duration exposure to ensure a single, reproducible aggregate definition. The scene exposure weight is determined according to the proportion of the total duration of behavioral segments under each scene category to the total duration of all behavioral segments of the target driving object.
[0180] in, For the driver In scene category The complete set of indexes for all behavior fragments below. For fragments Duration; Index for behavior fragments; The index for summing the scenarios in the denominator. For the driver In the context of summation index The set of indexes for all behavioral fragments. It can also be determined by the proportion of fragments or the reliability of the sample. Attention should be paid to the comprehensive dimension score. Exposure weights based on scene Weighted, therefore, the overall score is calculated. Direct comparison is only possible when there are similarities. The distribution of drivers is meaningful. For drivers with significantly different exposure patterns, it is advisable to combine scenario-level driver evaluation results. .
[0181] In this step, the driver's scenario-level driver evaluation results in each scenario are multiplied by the scenario exposure weight of each scenario and then summed to obtain the comprehensive dimension score for each dimension. Then, the comprehensive dimension scores of safety, comfort and energy consumption are multiplied by their respective first-level dimension weights and summed to obtain the individual behavior quality score.
[0182] Step S702: Determine the traffic flow efficiency result separately based on the system-level traffic flow efficiency evaluation index; In this embodiment of the application, the system-level traffic flow efficiency evaluation indicators include traffic volume, average delay and average queue length. These indicators reflect the operational status of the entire traffic flow or road segment, including the target driver.
[0183] The indicators for safety, comfort, and energy consumption can all be calculated from individual vehicle trajectories. However, the three indicators for traffic flow efficiency (traffic volume, average delay, and average queue length) are system-level indicators, reflecting the operational status of the entire traffic flow or road segment, including the target driver. Therefore, traffic flow efficiency results... As a separate, parallel output report (rather than incorporating individual behavioral quality scores) Individual behavior quality score The calculation method is shown in formula (27). For the driver The system-level traffic flow efficiency results of the traffic environment are reported independently from the individual behavior quality scores: Single-vehicle simulation / isolated data: If the simulation environment only contains the target driver, the system metrics are entirely attributed to that driver, with no attribution ambiguity.
[0184] Multi-vehicle interaction simulation: If other drivers (human-driven vehicles or other autonomous vehicles) are present in the traffic flow, the efficiency index reflects the overall system performance rather than the contribution of individual vehicles. This should be understood as "the efficiency state of the traffic environment in which the target driver is located" rather than "the driver's efficiency contribution." If it is necessary to estimate the individual net contribution, it can be calculated through comparative simulation with and without the target driver.
[0185] From the perspective of efficiency The core basis for separation lies in: As a driver A comprehensive measure of individual behavioral quality should only include dimensions controllable by the individual (safety, comfort, energy consumption). Incorporating system-level traffic flow efficiency status into individual quality scores introduces irreducible confounding noise, reducing the validity of cross-environment comparisons.
[0186] In this step, traffic flow efficiency results are calculated based on system-level traffic flow efficiency indicators such as traffic volume, average delay, and average queue length, in a manner independent of individual behavior quality scores. The calculation window for these results is determined by scenario, road segment, or simulation time window.
[0187] Step S703: The traffic flow efficiency result is used as a parallel output result independent of the individual behavior quality score.
[0188] In this embodiment of the application, parallel output means that individual behavior quality score and traffic flow efficiency result are output simultaneously as two independent and non-inclusive evaluation variables.
[0189] In this step, the traffic flow efficiency results are not written back into the calculation of individual behavior quality scores, nor are they mixed with safety, comfort, and energy consumption dimensions as a single scalar. Instead, they are presented alongside individual behavior quality scores in the evaluation report.
[0190] Due to the specific individual attribution of the efficiency dimension, two parallel output variables are generated: Individual behavioral quality score (safety, comfort, energy consumption):
[0191] in, This represents the third of the three dimensions attributable to an individual: safety, comfort, and energy consumption. Each dimension corresponds to a first-level combined weight, and the sum of all weights is 1. For the driver In dimensions The comprehensive score across the following dimensions; Values are taken from three dimensions: safety, comfort, and energy consumption, and used as the summation index.
[0192] System traffic flow efficiency results It is obtained by normalizing, weighting or regularizing system-level traffic flow efficiency indicators such as traffic volume, average delay and average queue length. Its calculation window can be determined according to scenario, road segment or simulation time window.
[0193] It reflects the quality of a driver's own behavior and can be directly compared across drivers. It reflects the operational status of the traffic flow environment in which the driver is located, and its results are not written back into the individual behavior quality score; they can only be compared between drivers in the same or similar traffic environments. This dual-output design avoids conflating individual controllable behavior quality with system-level traffic flow status into a single scalar.
[0194] To ensure the traceability and reproducibility of all weightings, this section only specifies the final quality combination weights used in the driver aggregation phase. The first-level dimension weights are initialized as follows:
[0195] These are the first-level combined weights for the safety, comfort, and energy consumption dimensions, respectively; all are initialized to 1 / 3 in the absence of prior preferences, serving as the neutral baseline for the three individual attributability dimensions in the absence of prior preferences. The efficiency dimension remains... In addition, as Independent report.
[0196] This application's embodiments are based on a dual-output architecture that decouples individual quality and system efficiency. It aggregates safety, comfort, and energy consumption evaluation dimensions into an individual behavior quality score, while simultaneously treating traffic flow efficiency as an independent parallel output. This reduces the confounding influence of system-level traffic flow conditions on individual driving behavior quality evaluations and improves the validity of cross-environment comparisons. This effect can be quantitatively evaluated using indicators such as the test-retest correlation coefficient of individual behavior quality scores, the scoring variance under different scenario exposure combinations, the correlation or decoupling degree between system efficiency results and individual behavior quality scores, and cross-environment scoring offset.
[0197] This embodiment provides an optional application method for mapping driving behavior evaluation results to the micro-simulation execution layer.
[0198] Step 1: Obtain the driving quality score, normalized matching weight, and scene-level driver evaluation results of the target driving object in each scene category according to the method of the aforementioned embodiment.
[0199] Step 1a (Parameter Calibration): Before mapping the evaluation results to the micro-simulation execution layer, the parameter calibration step is performed first. The calibration step includes: (1) Based on the historical behavior statistics of the target driving object, extract the distribution of its behavioral primitives in each scenario category and the empirical distribution of key kinematic parameters within each behavioral primitive (such as the following distance distribution, lane change gap acceptance distribution, acceleration and deceleration distribution, etc.); (2) Map the above empirical distributions to the initial values or prior ranges of the parameters to be modulated in the micro-simulation execution layer, for example, mapping the percentile of the following distance to the safe distance parameter of the following model, and mapping the lane change gap acceptance distribution to the gap threshold of the lane change model; (3) Fine-tuning and optimizing the calibration parameters based on minimizing the error between the simulation output trajectory and the real trajectory. The error can be measured by trajectory position error, velocity profile error, or behavioral primitive sequence matching rate, etc. After completing the parameter calibration, the gene matching result is applied as the modulation coefficient to the calibrated parameters.
[0200] Step 2: Determine the object to be modulated in the micro-simulation execution layer. The object to be modulated includes, but is not limited to, at least one of rule preferences, response thresholds, maneuver priors, and activation probabilities of behavioral primitives. The micro-simulation execution layer may adopt control logic forms such as car following, lane changing, yielding, or fuzzy rule-based control.
[0201] Step 3: Based on the normalized matching weights of the target driver for different driving gene prototypes, determine the set of modulation parameters consistent with their behavioral style. The higher the normalized matching weight of a driving gene prototype, the greater its corresponding modulation influence.
[0202] Step 4 involves applying the set of modulation parameters to the micro-simulation execution layer to induce different local responses in different target driving subjects under the same traffic stimuli. Specifically, the driving gene prototype does not replace the micro-rules but modulates the activation or weighting of these rules in a given traffic context.
[0203] Step 5: Output heterogeneous driver behavior results. The output results include, but are not limited to, simulation trajectories, behavioral primitive sequences, scenario-level driver-level evaluation results, or modulated micro-decision execution states.
[0204] In this embodiment, the micro-simulation execution layer is not limited to a specific software platform and can operate on fuzzy rule-based micro-simulation backbones, rule-priority-based decision-making backbones, or other micro-simulation backbones with transparent decision-making structures. This embodiment limits the mapping logic from evaluation results to micro-simulation execution parameters, rather than a specific simulation software platform.
[0205] Furthermore, based on the scenario-level driver evaluation results and the contribution components of each driving gene prototype, This study aims to interpret the local behavioral differences among different driving subjects in the simulation, thereby enhancing the analyzability of the simulation results.
[0206] In another embodiment of this application, a driving behavior evaluation device is also provided, such as... Figure 8 As shown, it includes: The first acquisition module 11 is used to acquire trajectory data of the target driving object and continuously segment the trajectory data to obtain multiple behavior segments; The first determining module 12 is used to determine the behavior primitive layer evaluation result of the target driving object based on multiple behavior fragments for each scene category, and to determine the behavior primitive distribution of the target driving object; Projection module 13 is used to project the evaluation results of the behavior primitive layer onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype. Each driving gene prototype is determined according to the distribution of the behavior primitive composition of each driving object in the group driving data under multiple scene categories. The second determining module 14 is used to determine the matching weight between the target driving object and each driving gene prototype based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype. The third determining module 15 is used to determine the scenario-level driver evaluation result based on the matching weight and the driving quality score.
[0207] In another embodiment of this application, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the driving behavior evaluation method described in any of the foregoing method embodiments.
[0208] The electronic device provided in this application embodiment has a processor that executes a program stored in the memory to acquire trajectory data of a target driving object and continuously segment it. For each scenario category, it determines the behavioral primitive layer evaluation result and behavioral primitive distribution of the target driving object based on behavioral fragments. By projecting the behavioral primitive layer evaluation result onto multiple driving gene prototypes, it obtains the driving quality score of the target driving object from the perspective of multiple driving gene prototypes. That is, the score corresponding to the driving quality performance of the target driving object when each driving gene prototype is used as an evaluation benchmark. Then, by matching the behavioral primitive distribution of the target driving object with the behavioral primitive composition distribution of each driving gene prototype, it determines the matching weight of the driving style corresponding to each driving gene prototype. Finally, it comprehensively determines the scenario-level driver layer evaluation result based on the driving quality score (performance) and the matching weight (style). This makes the comprehensive evaluation between drivers of different styles style-aware. Even if two drivers obtain the same driving quality score on the same behavioral primitive, they will still obtain different scenario-level driver layer evaluation results due to their different behavioral styles. Drivers of different styles can obtain accurate evaluations under their respective matched driving gene prototypes.
[0209] The communication bus 1140 mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.
[0210] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.
[0211] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0212] The processor 1110 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0213] In another embodiment of this application, a computer-readable storage medium is also provided, on which a program for a driving behavior evaluation method is stored. When the program for the driving behavior evaluation method is executed by a processor, it implements the steps of the driving behavior evaluation method described in any of the foregoing method embodiments.
[0214] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0215] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.
Claims
1. A driving behavior evaluation method characterized by comprising: include: The trajectory data of the target driving object is acquired, and the trajectory data is continuously segmented to obtain multiple behavior segments; For each scenario category, the evaluation result of the target driving object's behavior primitive layer is determined based on multiple behavior fragments, and the distribution of the target driving object's behavior primitives is determined; The evaluation results of the behavioral primitive layer are projected onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype. Each driving gene prototype is determined according to the distribution of behavioral primitive composition of each driving object in the group driving data under multiple scene categories. Based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype, the matching weight between the target driving object and each driving gene prototype is determined. The scenario-level driver evaluation result is determined based on the matching weight and the driving quality score.
2. The driving behavior evaluation method according to claim 1, characterized in that, The behavioral primitive layer evaluation results of the target driving object are determined based on multiple behavioral fragments, including: Each behavior segment is classified into scene category and behavior primitive category to obtain the scene category and behavior primitive type corresponding to each behavior segment. Based on the evaluation indicators included in multiple preset evaluation dimensions, the evaluation indicators of each behavior segment are calculated and normalized to obtain the segment-level evaluation results of each behavior segment under the evaluation dimensions. The fragment-level evaluation results under each evaluation dimension are aggregated according to the scenario category and the behavior primitive type to obtain the behavior primitive layer evaluation results.
3. The driving behavior evaluation method according to claim 2, characterized in that, Based on the evaluation indicators included in multiple preset evaluation dimensions, each behavioral segment is normalized to obtain a segment-level evaluation result for each behavioral segment, including: For evaluation indicators with preset physical thresholds in multiple evaluation dimensions, the calculation results of each evaluation indicator are normalized by using monotonic mapping or piecewise function mapping based on the Sigmoid function to obtain the first normalized indicator value. For evaluation indicators without preset physical thresholds in multiple evaluation dimensions, normalization is performed by scaling to a preset range to obtain the second normalized indicator value. The first normalized index value and the second normalized index value are weighted and summed according to the preset index weights within the evaluation dimension to which the evaluation index belongs, so as to obtain the segment-level dimension score of each behavior segment under each evaluation dimension. The segment-level evaluation result of each behavioral segment is determined based on the segment-level dimension score of each evaluation dimension.
4. The driving behavior evaluation method according to claim 2, characterized in that, The fragment-level evaluation results under each evaluation dimension are aggregated according to the scenario category and the behavior primitive type to obtain the behavior primitive layer evaluation results, including: Fragments with the same scene category and the same behavior primitive type are grouped into the same aggregation group; Within any aggregation group, the segment-level dimension scores in the segment-level evaluation results of each behavioral segment are weighted and aggregated according to the duration of the behavioral segment or by adopting an event-weighted strategy to obtain the aggregation score; The evaluation result of the behavioral primitive layer is determined based on the aggregation scores corresponding to multiple aggregation groups.
5. The driving behavior evaluation method according to claim 1, characterized in that, The evaluation results of the behavioral primitive layer are projected onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype, including: Based on group driving data under multiple scenario categories, a driving gene prototype corresponding to each scenario category is obtained in advance. The driving gene prototype includes the weights of behavioral primitives corresponding to multiple behavioral primitive types. For each evaluation dimension, based on the weights of the behavioral primitives corresponding to each behavioral primitive type in the driving gene prototype, the evaluation results of the behavioral primitive layer are projected onto multiple driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype.
6. The driving behavior evaluation method according to claim 5, characterized in that, The driving quality score is the result of a weighted sum of the evaluation results of the behavioral primitive layer according to the weights of the behavioral primitive composition corresponding to multiple behavioral primitive types in the driving gene prototype.
7. The driving behavior evaluation method according to claim 1, characterized in that, Based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype under each scenario category, the matching weight between the target driving object and each of the driving gene prototypes is determined, including: Calculate the distribution difference measure between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype; The matching weights between the target driving object and each of the driving gene prototypes are determined based on multiple distribution difference measures.
8. The driving behavior evaluation method according to claim 7, characterized in that, The matching weights between the target driving object and each of the driving gene prototypes are determined based on multiple distribution difference measures, including: Each of the aforementioned distribution difference measures is used as the independent variable of the exponential kernel function, and the matching degree is obtained by performing a negative exponential transformation through the exponential kernel function. The matching degrees are normalized to obtain the matching weights between the target driving object and each driving gene prototype.
9. The driving behavior evaluation method according to claim 8, characterized in that, When the number of driving gene prototypes is less than or equal to a preset number threshold, the temperature parameter in the exponential kernel function is the preset temperature parameter; When the number of driving gene prototypes exceeds a preset threshold, the temperature parameter in the exponential kernel function is determined based on the number of driving gene prototypes.
10. The driving behavior evaluation method according to claim 1, characterized in that, The scenario-level driver evaluation result is determined based on the matching weight and the driving quality score, including: For each evaluation dimension, a weighted aggregation is performed based on the matching weights and driving quality scores corresponding to each driving gene prototype to obtain the scenario-level driver-level evaluation results of the target driving object under multiple evaluation dimensions.
11. The driving behavior evaluation method according to claim 1, characterized in that, The method further includes: Based on the scenario-level driver evaluation results, the scenario-level driver evaluation results are aggregated according to the preset scenario exposure weights to obtain individual behavior quality scores. Traffic flow efficiency results are determined separately based on system-level traffic flow efficiency evaluation indicators; The traffic flow efficiency results are used as parallel outputs independent of the individual behavior quality scores.
12. The driving behavior evaluation method according to claim 11, characterized in that, The scene exposure weight is determined according to the proportion of the total duration of the behavioral segments under each scene category to the total duration of all behavioral segments of the target driving object.
13. A driving behavior evaluation device, characterized in that, include: The first acquisition module is used to acquire trajectory data of the target driving object and continuously segment the trajectory data to obtain multiple behavior segments; The first determining module is used to determine the behavioral primitive layer evaluation result of the target driving object based on multiple behavioral fragments for each scene category, and to determine the behavioral primitive distribution of the target driving object; The projection module is used to project the evaluation results of the behavioral primitive layer onto multiple preset driving gene prototypes to obtain the driving quality score of the target driving object under each driving gene prototype. Each driving gene prototype is determined according to the distribution of the behavioral primitive composition of each driving object in the group driving data under multiple scene categories. The second determining module is used to determine the matching weight between the target driving object and each driving gene prototype based on the matching relationship between the distribution of behavioral primitives of the target driving object and the distribution of behavioral primitives of each driving gene prototype. The third determining module is used to determine the scenario-level driver evaluation result based on the matching weight and the driving quality score.
14. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the driving behavior evaluation method according to any one of claims 1-12.
15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for a driving behavior evaluation method, which, when executed by a processor, implements the steps of the driving behavior evaluation method according to any one of claims 1-12.
