Automatic driving personification risk control system and method

Abstract

The invention provides an automatic driving personification risk control system and method, and the system comprises an anti-fact track generator which is configured to receive environment perception information, and generate at least one anti-fact track based on the environment perception information, the anti-fact trajectory is used for representing an expected trajectory conforming to the cautious driving behavior mode under the environment perception information; the anthropomorphic risk comparator is configured to obtain a current planning track output by the vehicle main planning decision-making system and perform quantitative comparison on the current planning track and the anti-fact track to obtain an anthropomorphic risk level; and the grading intervention device is configured to output an intervention instruction to a man-machine interaction system or a vehicle main planning decision system of the vehicle according to the personification risk grade in order to reduce or eliminate the personification risk. According to the method, the computable measurement and the flexible intervention of the non-personification degree of the driving behavior are realized, and the social acceptability and the riding comfort of the automatic driving behavior are remarkably improved while the basic safety is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to an anthropomorphic risk control system and method for autonomous driving. Background Technology

[0002] With the rapid development of autonomous driving technology, especially Level 2 and above advanced driver assistance systems (ADAS) and Level 4 and above fully autonomous driving systems, the core objective is to improve traffic efficiency while ensuring driving safety. Existing autonomous driving decision-making and planning systems, by integrating data from multiple sensors such as LiDAR, cameras, and millimeter-wave radar, and leveraging advanced prediction and planning algorithms, have achieved significant results in avoiding direct collisions with obstacles. However, when widely applied in complex urban road environments, autonomous driving systems face a typical "Safety of Intended Function" (SOTIF) challenge. Specifically, even when strictly adhering to traffic regulations and without direct collision risk, the system's driving behavior may significantly deviate from the general expectations and psychological comfort zones of surrounding human drivers, exhibiting "non-human" or "aggressive" characteristics.

[0003] Currently, mainstream technical solutions mainly focus on addressing hard, imminent collision risks, such as triggering warnings or emergency braking by setting a fixed Time-to-Collision (TTC) threshold. While these methods are effective for clear physical hazards, they lack a dynamic assessment benchmark that aligns with human driving habits and social norms. Consequently, they cannot quantify and manage the aforementioned soft SOTIF (Safety-Oriented Impact Factor) issues—those without collision risk but exhibiting inhuman or uncomfortable behavior.

[0004] Therefore, there is an urgent need to provide an anthropomorphic risk control system and method for autonomous driving that can identify, assess and smoothly intervene in anthropomorphic risks in order to improve the social acceptability and ride comfort of autonomous driving systems. Summary of the Invention

[0005] In view of this, it is necessary to provide an anthropomorphic risk control system and method for autonomous driving, in order to solve the technical problem that the existing technology cannot quantify and intervene in anthropomorphic risks, resulting in poor social acceptability and ride comfort of autonomous driving systems.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides an anthropomorphic risk control system for autonomous driving, comprising: A counterfactual trajectory generator is configured to receive environmental perception information and generate at least one counterfactual trajectory based on the environmental perception information, the counterfactual trajectory being used to characterize an expected trajectory that conforms to a cautious driving behavior pattern under the environmental perception information. An anthropomorphic risk comparator is configured to acquire the current planning trajectory output by the vehicle's main planning decision system, and to quantitatively compare the current planning trajectory with the counterfactual trajectory to obtain an anthropomorphic risk level; A graded intervention device is configured to output intervention commands to the vehicle's human-machine interaction system or the vehicle's main planning and decision-making system based on the level of anthropomorphic risk, with the aim of reducing or eliminating anthropomorphic risk.

[0007] In one possible implementation, the counterfactual trajectory generator includes a trajectory generation model built around a conditional variational autoencoder, which is trained on a dataset labeled with driving behavior tags. The counterfactual trajectory generator is specifically configured to: receive conditional variables representing a cautious driving behavior pattern, and generate at least one counterfactual trajectory based on the environmental perception information and the conditional variables.

[0008] In one possible implementation, the loss function during training of the trajectory generation model includes a standard loss term and a weighted cross-entropy loss term based on driving behavior labels, wherein the standard loss term includes a reconstruction loss term and a KL divergence term.

[0009] In one possible implementation, the anthropomorphic risk comparator includes a dynamic threshold determination unit, an indicator value determination unit, and a risk judgment unit. The dynamic threshold determination unit is used to determine the dynamic threshold of the evaluation index based on historical driving data within a preset historical time period. The indicator value determination unit is used to determine multiple first indicator values ​​based on the counterfactual trajectory and multiple second indicator values ​​based on the current planned trajectory; The risk determination unit is used to determine the number of indicators whose multiple second indicator values ​​are worse than the multiple first indicator values. When the number of indicators is less than a preset number, the anthropomorphic risk level is determined to be low. When the number of indicators is greater than or equal to the preset number, and each of the second indicator values ​​is less than or equal to the dynamic threshold, the anthropomorphic risk level is determined to be medium risk. When the number of indicators is greater than or equal to the preset number, and each of the second indicator values ​​is greater than the dynamic threshold, the anthropomorphic risk level is determined to be high risk.

[0010] In one possible implementation, the plurality of first indicators and the plurality of second indicator values ​​all include the time difference to reach the conflict point, the maximum deceleration required for the entire journey, and the minimum lateral distance from the predicted path of pedestrians or non-motorized vehicles.

[0011] In one possible implementation, the dynamic threshold determination unit includes a data standardization subunit, a clustering subunit, and a threshold determination subunit; The data standardization subunit is used to standardize the historical driving data to obtain standardized data. The clustering subunit is used to cluster the standardized data to obtain multiple driving mode clusters; The threshold determination subunit is used to determine the cautious driving mode cluster among the multiple driving mode clusters, and to determine the statistical distribution of each evaluation index in the cautious driving mode cluster, and to determine the dynamic threshold based on the statistical distribution.

[0012] In one possible implementation, the hierarchical interventioner is specifically configured as follows: When the anthropomorphic risk level is low, informational prompts are provided through the human-computer interaction interface; When the anthropomorphic risk level is medium, trajectory fine-tuning parameters are sent to the vehicle's main planning and decision-making system to optimize the comfort of the current planned trajectory; When the anthropomorphic risk level is high, a negotiated takeover request is initiated, and an alarm is displayed on the human-computer interaction interface, providing counterfactual trajectories as suggested actions.

[0013] In one possible implementation, the hierarchical interventioner is further configured as follows: If the suggested action is not responded to within a preset time period, one of the at least one counterfactual trajectory is executed.

[0014] In one possible implementation, the environmental perception information includes high-precision map data, dynamic target information, traffic signal information, and vehicle status information.

[0015] Secondly, the present invention also provides a method for anthropomorphic risk control in autonomous driving, comprising: Receive environmental perception information and generate at least one counterfactual trajectory based on the environmental perception information, wherein the counterfactual trajectory is used to characterize the expected trajectory that conforms to the cautious driving behavior pattern under the environmental perception information; Obtain the current planned trajectory output by the vehicle's main planning decision system, and quantitatively compare the current planned trajectory with the counterfactual trajectory to obtain the anthropomorphic risk level; With the aim of reducing or eliminating anthropomorphic risks, intervention commands are output to the vehicle's human-machine interaction system or the vehicle's main planning and decision-making system based on the anthropomorphic risk level.

[0016] The beneficial effects of this invention are as follows: The autonomous driving anthropomorphic risk control system provided by this invention firstly generates counterfactual trajectories that conform to cautious driving behavior patterns through a counterfactual trajectory generator, establishing an objective and dynamic quantitative evaluation benchmark for the subjective concept of anthropomorphism, thus solving the problem in existing technologies where the lack of such benchmarks makes it impossible to assess soft SOTIF risks. Secondly, an anthropomorphic risk comparator quantitatively compares the vehicle's actual current planned trajectory with the counterfactual trajectory to obtain a risk level, achieving real-time and calculable measurement of the degree of non-anthropomorphism in driving behavior. Finally, a graded intervention device outputs intervention commands to the vehicle system according to the risk level. The system can achieve a complete closed loop from risk identification to flexible intervention without directly overturning the main planning decision, thereby significantly improving the social acceptability and ride comfort of autonomous driving behavior while ensuring basic safety.

[0017] Furthermore, the system proposed in this invention operates as an independent bypass system in parallel with the vehicle's main planning and decision-making system, without affecting the original real-time decision-making and core safety functions of the vehicle's main planning module. By employing counterfactual trajectories as a dynamic external evaluation benchmark, it is possible to independently assess and externally intervene in the anthropomorphic risks of the decision results without intruding into or interfering with the internal logic of the main planning algorithm. This ensures that the vehicle's main planning and decision-making system always operates efficiently based on defined physical safety rules, while the autonomous driving anthropomorphic risk control system focuses on the social optimization and comfort adjustment of driving behavior. The two are functionally decoupled but synergistic in terms of results, significantly improving the functional safety and deployment flexibility of the autonomous driving system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of an embodiment of the anthropomorphic risk control system for autonomous driving provided by the present invention; Figure 2 A schematic diagram of an embodiment of the anthropomorphic risk comparator provided by the present invention; Figure 3 A schematic diagram of an embodiment of the dynamic threshold determination unit provided by the present invention; Figure 4 A schematic diagram illustrating a specific embodiment of the control method provided by the present invention in the scenario of the yellow light ending at an intersection; Figure 5This is a schematic flowchart of an embodiment of the anthropomorphic risk control method for autonomous driving provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] It should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this invention illustrate operations implemented according to some embodiments of the invention. It should be understood that the operations in the flowcharts may be implemented out of order, and steps without logical contextual relationships may be reversed or performed simultaneously. Furthermore, those skilled in the art, guided by the content of this invention, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] To illustrate what constitutes non-anthropomorphism, consider this example: At the end of the yellow light at an intersection, a vehicle's primary road planning system might, based on precise calculations, determine that it is safe to proceed and therefore choose to accelerate through, rather than slowing down and stopping as most cautious human drivers would. While this behavior doesn't directly cause an accident, it can confuse, stress, or lead to misjudgments by other road users, increasing the potential risk of indirect accidents. This falls under the category of "insufficient expected performance" in SOTIF (Situation of Expected Performance), and this problem stems from non-anthropomorphic driving behavior.

[0024] To address the aforementioned technical problems, this invention proposes an anthropomorphic risk control system and method for autonomous driving, which is deployed on the computing platform of an autonomous vehicle. These will be described in detail below.

[0025] Figure 1A schematic diagram of an embodiment of the anthropomorphic risk control system for autonomous driving provided by the present invention is shown below. Figure 1 As shown, the anthropomorphic risk control system 10 for autonomous driving includes: The counterfactual trajectory generator 100 is configured to receive environmental perception information and generate at least one counterfactual trajectory based on the environmental perception information. The counterfactual trajectory is used to characterize the expected trajectory that conforms to the cautious driving behavior pattern under the environmental perception information.

[0026] In a specific embodiment of the present invention, the environmental perception information includes high-precision map data, dynamic target information, traffic signal information, and vehicle status information. The high-precision map data includes lane lines, road boundaries, stop lines, and intersection topology. The dynamic target information includes the position, speed, acceleration, and orientation of surrounding vehicles, pedestrians, and non-motorized vehicles. The traffic signal information includes the traffic light status (red, yellow, green) and remaining time. The vehicle's own status includes its current position, speed, and acceleration.

[0027] In real-world scenarios, to reduce computational load and improve the real-time performance of the autonomous driving anthropomorphic risk control system 10, the number of counterfactual trajectories is 1-3.

[0028] The anthropomorphic risk comparator 200 is configured to acquire the current planned trajectory output by the vehicle's main planning decision system, and to quantitatively compare the current planned trajectory with the counterfactual trajectory to obtain the anthropomorphic risk level.

[0029] Specifically, the goal of the current planned trajectory is to efficiently and stably complete the navigation task from start to finish while meeting traffic regulations and strict safety constraints. Its generation process comprehensively optimizes multiple objectives, including driving efficiency, ride comfort, and dynamic feasibility. In contrast, the counterfactual trajectory is not generated for direct execution. Its core objective is to simulate the more conservative, defensive, or socially expected driving choices that a typical human driver might make in a given scenario. In other words, the current planned trajectory reflects the decision-making results of the vehicle's main planning system, while the counterfactual trajectory is used to quantify the difference between the decision-making results and cautious human driving habits, thereby revealing potential anthropomorphic risks.

[0030] The graded intervention device 300 is configured to output intervention commands to the vehicle's human-machine interaction system or the vehicle's main planning and decision-making system based on the level of anthropomorphic risk, with the aim of reducing or eliminating anthropomorphic risk.

[0031] Compared with existing technologies, the autonomous driving anthropomorphic risk control system 10 provided in this embodiment of the invention firstly generates a counterfactual trajectory that conforms to the cautious driving behavior pattern through a counterfactual trajectory generator 100, establishing an objective and dynamic quantitative evaluation benchmark for the subjective concept of anthropomorphism, thus solving the problem in existing technologies where the lack of such a benchmark makes it impossible to assess soft SOTIF risks. Secondly, an anthropomorphic risk comparator 200 quantitatively compares the vehicle's actual current planned trajectory with the counterfactual trajectory to obtain a risk level, achieving real-time and calculable measurement of the degree of non-anthropomorphism in driving behavior. Finally, a graded intervention device 300 outputs intervention commands to the vehicle system according to the risk level. The system can achieve a complete closed loop from risk identification to flexible intervention without directly overturning the main planning decision, thereby significantly improving the social acceptability and ride comfort of autonomous driving behavior while ensuring basic safety.

[0032] Furthermore, the system proposed in this embodiment operates as an independent bypass system in parallel with the vehicle's main planning and decision-making system, without affecting the original real-time decision-making and core safety functions of the vehicle's main planning module. By employing counterfactual trajectories as a dynamic external evaluation benchmark, it is possible to independently assess and externally intervene in the anthropomorphic risks of the decision results without intruding into or interfering with the internal logic of the main planning algorithm. This ensures that the vehicle's main planning and decision-making system always operates efficiently based on defined physical safety rules, while the autonomous driving anthropomorphic risk control system focuses on the social optimization and comfort adjustment of driving behavior. The two are functionally decoupled but synergistic in terms of results, significantly improving the functional safety and deployment flexibility of the autonomous driving system.

[0033] In some embodiments of the present invention, the counterfactual trajectory generator 100 includes a trajectory generation model built around a conditional variational autoencoder, which is trained on a dataset labeled with driving behavior tags.

[0034] It should be noted that driving behavior labels include cautious, conservative, and typical, meaning the trajectory generation model can determine counterfactual trajectories for multiple driving modes (different driving behavior labels). However, since the purpose of this embodiment is to assess the anthropomorphic risk between the currently planned trajectory and the expected trajectory under the cautious driving behavior mode, in a specific embodiment of this invention, the counterfactual trajectory generator is specifically configured to: receive conditional variables characterizing the cautious driving behavior mode, and generate at least one counterfactual trajectory based on environmental perception information and the conditional variables.

[0035] Specifically, after receiving environmental perception information and condition variables, the encoder part of the conditional variational autoencoder outputs a probability distribution in the latent space. When generating the trajectory, the decoder randomly samples multiple times from this probability distribution, obtaining a different latent variable vector each time. Instead of using a fixed value, the decoder randomly samples multiple times from the distribution, resulting in a different latent variable vector each time. The decoder combines each sampled latent vector with the same condition variable to decode a reasonable trajectory that meets the conditions. Due to the differences in the latent vectors, the generated trajectory will exhibit reasonable and diverse possibilities in details (such as velocity change curves, minor deviations in the path, and deceleration timing). These trajectories all meet the overall requirement of caution but represent different specific operational choices.

[0036] In other words, the mechanism of multiple outputs for the same input simulates a variety of different, but all cautious, driving strategies that a human driver might adopt in the same situation (e.g., smoothly decelerating to a stop, or braking lightly in advance with a larger margin). This provides a richer and more statistically significant benchmark for subsequent risk comparisons, making the risk assessment results more robust and reliable.

[0037] The training objectives of conventional conditional variational autoencoders primarily ensure that the model achieves two things: 1. Accurate reconstruction: given a real trajectory, after encoding and decoding, the output trajectory should closely resemble the input. 2. Regular latent space: when encoding different trajectories into the latent space, their distribution should be smooth, continuous, and conform to a pre-defined prior distribution (such as a standard normal distribution). Based solely on this objective, the generated counterfactual trajectory may not be the trajectory a cautious human driver would collect. Using this benchmark to evaluate the currently planned trajectory will lead to errors in the entire evaluation result.

[0038] To address this technical problem, in some embodiments of the present invention, the loss function during trajectory generation model training includes a standard loss term and a weighted cross-entropy loss term based on driving behavior labels. The standard loss term includes a reconstruction loss term and a KL divergence term.

[0039] Specifically, the reconstruction loss term achieves accurate reconstruction, the KL divergence term achieves latent space regularization, and the weighted cross-entropy loss term based on driving behavior labels forces the model to simultaneously consider the physical realism of trajectory reconstruction when encoding and reconstructing trajectories in the latent space. Therefore, the trained model exhibits higher sensitivity and specificity to the caution condition, generating counterfactual trajectories that are statistically more faithful to human cautious driving patterns, thus significantly improving the benchmark quality and the credibility of the final judgment in the entire anthropomorphic risk assessment process.

[0040] In a specific embodiment of the present invention, the complete loss function is: ; In the formula, This is the total loss function; For standard loss items; The weighted cross-entropy loss term is based on driving behavior labels; is the weight of the weighted cross-entropy loss term based on driving behavior labels, where the driving behavior label is 0 or 1, where 0 represents aggressive and 1 represents cautious.

[0041] in, To balance trajectory generation quality and behavioral bias, by setting the conditional variable to caution during training and optimizing this joint loss function, the model can be guided to generate trajectories that conform to the statistical characteristics of cautious human drivers during inference.

[0042] The trajectory generation model can be obtained through the above training process. Based on this, the counterfactual trajectory generator is specifically configured to: receive conditional variables representing cautious driving behavior patterns, and generate at least one counterfactual trajectory based on environmental perception information and conditional variables.

[0043] Specifically, the counterfactual trajectory is a four-dimensional spatiotemporal trajectory sequence (x,y,v,t), representing the vehicle's location, speed, and time.

[0044] To improve the accuracy of anthropomorphic risk level determination, in some embodiments of the present invention, such as Figure 2 As shown, the anthropomorphic risk comparator 200 includes a dynamic threshold determination unit 210, an indicator value determination unit 220, and a risk judgment unit 230. The dynamic threshold determination unit 210 is used to determine the dynamic threshold of the evaluation index based on historical driving data within a preset historical time period.

[0045] The preset historical time period can be set or adjusted according to the actual application scenario. In a specific embodiment of the present invention, the preset historical time period is 5 minutes from the current time.

[0046] The indicator value determination unit 220 is used to determine multiple first indicator values ​​based on the counterfactual trajectory and multiple second indicator values ​​based on the current planned trajectory; The risk assessment unit 230 is used to determine the number of indicators whose second indicator values ​​are worse than the number of first indicator values. When the number of indicators is less than a preset number, the anthropomorphic risk level is determined to be low. When the number of indicators is greater than or equal to the preset number and each second indicator value is less than or equal to the dynamic threshold, the anthropomorphic risk level is determined to be medium risk. When the number of indicators is greater than or equal to the preset number and each second indicator value is greater than the dynamic threshold, the anthropomorphic risk level is determined to be high risk.

[0047] The second indicator value being inferior to the first indicator value means that the driving mode represented by the second indicator value is more aggressive, that is, the aggressiveness of the current planned trajectory is greater than that of the counterfactual trajectory.

[0048] This invention determines dynamic thresholds for evaluation indicators based on historical driving data within a preset historical time period, avoiding the rigidity of adaptation caused by fixed thresholds and improving the adaptability of thresholds to changing traffic scenarios. Simultaneously, it uses two dimensions—the number of indicators and their values—to determine the anthropomorphic risk level, comprehensively reflecting the aggressiveness of driving behavior and improving the accuracy of the anthropomorphic risk level assessment.

[0049] It should be understood that the preset quantity can be set or adjusted according to the actual application scenario, and no specific limit is made here.

[0050] In a specific embodiment of the present invention, the multiple first indicators and multiple second indicator values ​​all include the time difference to reach the conflict point, the maximum deceleration required for the entire journey, and the minimum lateral distance to the predicted path of pedestrians or non-motorized vehicles.

[0051] This invention constructs a multi-dimensional, quantifiable evaluation standard using parameters across three dimensions. Specifically: the time difference to the point of conflict measures the aggressiveness of the behavior in the time dimension; earlier arrival implies a more urgent strategy that compresses others' reaction time. The maximum deceleration required throughout the journey is evaluated from the perspective of longitudinal dynamic comfort and safety margins; a greater deceleration requirement indicates that the current plan relies more on emergency braking, revealing insufficient margins for scenario uncertainties, significantly reducing ride comfort and increasing the risk of rear-end collisions. The minimum lateral distance to pedestrians / non-motorized vehicles focuses on the sense of spatial oppression and social etiquette; excessively close distances, even without a collision, pose a psychological threat to vulnerable road users, causing unease and misunderstanding. These three parameters, from the perspectives most sensitive to human drivers—time pressure, abruptness of action, and spatial aggression—comprehensively quantify the degree of anthropomorphism in autonomous driving behavior, making the risk assessment results comprehensive, objective, and consistent with the intuitive feelings of human drivers.

[0052] In specific embodiments of the present invention, such as Figure 3 As shown, the dynamic threshold determination unit 210 includes a data standardization subunit 211, a clustering subunit 212, and a threshold determination subunit 213; The data standardization subunit 211 is used to standardize historical driving data to obtain standardized data.

[0053] Specifically, historical driving data is Z-score standardized to eliminate the influence of dimensions.

[0054] It should be noted that, in order to reduce the amount of data processing and further improve the timeliness of system control, before standardizing historical driving data, key data that affects dynamic thresholds, such as the time difference to the conflict point and deceleration, can be extracted from the historical driving data to reduce the amount of computation.

[0055] Clustering subunit 212 is used to cluster standardized data to obtain multiple driving mode clusters.

[0056] In a specific embodiment of the present invention, the clustering algorithm used is the density-based clustering algorithm DBSCAN. DBSCAN can automatically identify the main patterns in the data and identify abnormal behavior as noise.

[0057] The driving mode clusters include the cautious driving mode cluster, the conservative driving mode cluster, and the typical driving mode cluster. In terms of safety margin: conservative driving mode cluster > cautious driving mode cluster > typical driving mode cluster; in terms of traffic efficiency: typical driving mode cluster > cautious driving mode cluster > conservative driving mode cluster.

[0058] The threshold determination subunit 213 is used to determine the cautious driving mode cluster among multiple driving mode clusters, and to determine the statistical distribution of each evaluation index in the cautious driving mode cluster, and to determine the dynamic threshold based on the statistical distribution.

[0059] The specific method for determining the dynamic threshold based on statistical distribution is as follows: the dynamic threshold is determined based on preset percentiles and statistical distribution. For example, the 15th percentile of the time difference between arriving at the conflict point (the smaller the time difference between arriving at the conflict point, the more dangerous it is, so a smaller percentile is used) and the 85th percentile of the maximum deceleration are calculated in the cautious driving mode cluster.

[0060] In some embodiments of the present invention, the graded interventioner 300 is specifically configured as follows: When the anthropomorphic risk level is low, informative prompts are provided through the human-computer interaction interface (HMI).

[0061] Specifically, an informative prompt could be: "The current approach to the intersection may be slightly aggressive." This helps the driver understand the vehicle's driving style without interfering.

[0062] When the anthropomorphic risk level is medium, trajectory fine-tuning parameters are sent to the vehicle's main planning and decision-making system to optimize the comfort of the current planned trajectory.

[0063] Specifically, the maximum deceleration of the current planned trajectory can be smoothly adjusted from -3.5m / s² to a more comfortable -3.0m / s², or the lateral distance to adjacent vehicles can be slightly increased.

[0064] When the anthropomorphic risk level is high, a negotiated takeover request is initiated, and an alarm is displayed on the human-computer interaction interface, providing counterfactual trajectories as suggested actions.

[0065] It should be understood that when the anthropomorphic risk level is high, it indicates that the current driving behavior is highly inconsistent with human expectations and may trigger panic or misjudgment. The system will initiate negotiated takeover, issuing clear visual and auditory alerts on the HMI, such as "Immediate takeover recommended! The current plan has a high anthropomorphic risk," and clearly displaying a counterfactual cautious trajectory as a suggested action, awaiting driver confirmation or voluntary takeover.

[0066] It should be noted that the counterfactual trajectory provided on the human-computer interaction interface should be the most cautious of all counterfactual trajectories, that is, the counterfactual trajectory with the largest time difference to the conflict point, the smallest maximum deceleration, and the largest minimum lateral distance to pedestrians / non-motorized vehicles, in order to improve driving safety.

[0067] To further ensure driving safety, the graded intervention device 300 is also specifically configured as follows: If the suggested action is not responded to within a preset time period, execute at least one of the counterfactual trajectories.

[0068] The preset duration can be set or adjusted according to the actual application scenario, and no specific limit is set here.

[0069] This invention establishes three progressively higher levels of intervention: information prompts for low-risk situations, trajectory fine-tuning for medium-risk situations, and negotiated takeover for high-risk situations. This tiered response ensures the precision and humanization of system intervention. In other words, during the initial stages of risk, prompts are used to raise driver awareness and avoid unnecessary interference; when risks become apparent, flexible trajectory optimization proactively corrects driving style, balancing safety and comfort; and a clear takeover process is initiated only when the risk is significant, prioritizing human-machine collaboration through suggestions. This achieves a smooth, predictable, and explainable intervention experience, significantly reducing the risk of misoperation or user resistance caused by abrupt system intervention, and enhancing the trustworthiness and acceptability of the autonomous driving system.

[0070] To verify the effectiveness of the anthropomorphic risk control system for autonomous driving proposed in this embodiment of the invention, the scenario of the end of the yellow light at an intersection will also be used as an example for illustration. Figure 4 As shown, specifically: The vehicle's main planning and decision-making system, based on sensor data and algorithms, makes the decision to accelerate through a yellow light (T_ego). This decision may technically comply with traffic regulations and pose no immediate collision risk. However, a parallel counterfactual trajectory generator, based on the exact same environmental conditions, generates an alternative solution—deceleration and stopping (T_cf)—that aligns with human cautious driving habits. A human-like risk comparator quantitatively compares the two decisions using key indicators, objectively determining that the T_ego behavior is aggressive and classifying it as high-risk. This process demonstrates that the present invention can effectively detect soft risks based on behavioral comfort and social expectations that traditional safety systems cannot identify. Furthermore, after determining a high-risk scenario, the graded intervention does not simply trigger emergency braking or abrupt takeover, but instead outputs the suggested T_cf as a new planning instruction. This instruction is presented to the driver through a human-machine interface or directly sent to the planning system for trajectory replanning.

[0071] This example illustrates that the embodiments of the present invention can dynamically generate humanized benchmarks, quantitatively assess behavioral risks, and trigger smooth closed-loop interventions, thereby significantly improving the acceptability and safety of autonomous driving systems in complex social and traffic scenarios. This fully demonstrates the effectiveness and practical value of the present invention in solving the "insufficient expected performance" problem in SOTIF.

[0072] In summary, the anthropomorphic risk control system for autonomous driving proposed in this invention is the first to use algorithmically generated counterfactual trajectories that conform to cautious human expectations as a dynamic evaluation benchmark. It quantifies subjective and vague concepts such as anthropomorphism and acceptability through objective and calculable indicator differences, solving a long-standing evaluation challenge in this field. It enhances the social acceptability of driving behavior by identifying and proactively avoiding aggressive behaviors that are technically legal but have poor social perception. This system enables autonomous vehicles to behave more like experienced and courteous drivers, significantly improving passenger comfort and trust, as well as the sense of security for other road users, effectively reducing SOTIF risks. Simultaneously, this invention specifically addresses the "insufficient expected performance" problem within "SOTIF" by providing early warnings and interventions for atypical behaviors, reducing the potential risk of accidents caused by human drivers misunderstanding the intentions of autonomous vehicles and taking erroneous actions. It achieves flexible and gradual intervention: compared to traditional hard interventions such as "sudden braking" or "forced takeover" based on fixed thresholds, the hierarchical warning and strategy fine-tuning mechanism of this invention is smoother and more human-centered. Prioritizing comfort-oriented fine-tuning and information prompts minimizes interference with main planning decisions and impacts on the riding experience, achieving a better balance between safety and efficiency. Furthermore, because the risk assessment threshold is dynamically calculated based on recent natural driving data, this system can adapt to different traffic environments, time periods, and regional driving cultures, exhibiting stronger robustness and scenario adaptability.

[0073] On the other hand, embodiments of the present invention also provide a method for anthropomorphic risk control in autonomous driving, such as... Figure 5 As shown, the methods for controlling risks associated with the anthropomorphic design of autonomous driving include: S501. Receive environmental perception information and generate at least one counterfactual trajectory based on the environmental perception information. The counterfactual trajectory is used to characterize the expected trajectory that conforms to the cautious driving behavior pattern under the environmental perception information. S502. Obtain the current planned trajectory output by the vehicle's main planning decision system, and quantitatively compare the current planned trajectory with the counterfactual trajectory to obtain the anthropomorphic risk level; S503. With the aim of reducing or eliminating anthropomorphic risks, intervention commands are output to the vehicle's human-machine interaction system or the vehicle's main planning and decision-making system according to the level of anthropomorphic risk.

[0074] It should be noted that the autonomous driving anthropomorphic risk control method provided in the above embodiments can realize the technical solutions described in the above autonomous driving anthropomorphic risk control system embodiments. The specific implementation principles of each step can be found in the corresponding content in the above autonomous driving anthropomorphic risk control system embodiments, and will not be repeated here.

[0075] The above provides a detailed description of an anthropomorphic risk control system and method for autonomous driving provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A human-like risk control system for autonomous driving, characterized in that, include: A counterfactual trajectory generator is configured to receive environmental perception information and generate at least one counterfactual trajectory based on the environmental perception information, the counterfactual trajectory being used to characterize an expected trajectory that conforms to a cautious driving behavior pattern under the environmental perception information. An anthropomorphic risk comparator is configured to acquire the current planning trajectory output by the vehicle's main planning decision system, and to quantitatively compare the current planning trajectory with the counterfactual trajectory to obtain an anthropomorphic risk level; A graded intervention device is configured to output intervention commands to the vehicle's human-machine interaction system or the vehicle's main planning and decision-making system based on the level of anthropomorphic risk, with the aim of reducing or eliminating anthropomorphic risk.

2. The anthropomorphic risk control system for autonomous driving according to claim 1, characterized in that, The counterfactual trajectory generator includes a trajectory generation model built around a conditional variational autoencoder, which is trained on a dataset labeled with driving behavior tags. The counterfactual trajectory generator is specifically configured to: receive conditional variables representing a cautious driving behavior pattern, and generate at least one counterfactual trajectory based on the environmental perception information and the conditional variables.

3. The anthropomorphic risk control system for autonomous driving according to claim 2, characterized in that, The loss function during the training of the trajectory generation model includes a standard loss term and a weighted cross-entropy loss term based on driving behavior labels. The standard loss term includes a reconstruction loss term and a KL divergence term.

4. The anthropomorphic risk control system for autonomous driving according to claim 1, characterized in that, The anthropomorphic risk comparator includes a dynamic threshold determination unit, an indicator value determination unit, and a risk judgment unit. The dynamic threshold determination unit is used to determine the dynamic threshold of the evaluation index based on historical driving data within a preset historical time period. The indicator value determination unit is used to determine multiple first indicator values ​​based on the counterfactual trajectory and multiple second indicator values ​​based on the current planned trajectory; The risk determination unit is used to determine the number of indicators whose multiple second indicator values ​​are worse than the multiple first indicator values. When the number of indicators is less than a preset number, the anthropomorphic risk level is determined to be low. When the number of indicators is greater than or equal to the preset number, and each of the second indicator values ​​is less than or equal to the dynamic threshold, the anthropomorphic risk level is determined to be medium risk. When the number of indicators is greater than or equal to the preset number, and each of the second indicator values ​​is greater than the dynamic threshold, the anthropomorphic risk level is determined to be high risk.

5. The anthropomorphic risk control system for autonomous driving according to claim 4, characterized in that, The values ​​of the multiple first indicators and the multiple second indicators all include the time difference to reach the conflict point, the maximum deceleration required for the entire journey, and the minimum lateral distance from the predicted path of pedestrians or non-motorized vehicles.

6. The anthropomorphic risk control system for autonomous driving according to claim 4, characterized in that, The dynamic threshold determination unit includes a data standardization subunit, a clustering subunit, and a threshold determination subunit; The data standardization subunit is used to standardize the historical driving data to obtain standardized data. The clustering subunit is used to cluster the standardized data to obtain multiple driving mode clusters; The threshold determination subunit is used to determine the cautious driving mode cluster among the multiple driving mode clusters, and to determine the statistical distribution of each evaluation index in the cautious driving mode cluster, and to determine the dynamic threshold based on the statistical distribution.

7. The anthropomorphic risk control system for autonomous driving according to claim 4, characterized in that, The graded interventioner is specifically configured as follows: When the anthropomorphic risk level is low, informational prompts are provided through the human-computer interaction interface; When the anthropomorphic risk level is medium, trajectory fine-tuning parameters are sent to the vehicle's main planning and decision-making system to optimize the comfort of the current planned trajectory; When the anthropomorphic risk level is high, a negotiated takeover request is initiated, and an alarm is displayed on the human-computer interaction interface, providing counterfactual trajectories as suggested actions.

8. The autonomous driving anthropomorphic risk control system according to claim 7, characterized in that, The graded interventioner is further specifically configured as follows: If the suggested action is not responded to within a preset time period, one of the at least one counterfactual trajectory is executed.

9. The anthropomorphic risk control system for autonomous driving according to claim 1, characterized in that, The environmental perception information includes high-precision map data, dynamic target information, traffic signal information, and vehicle status information.

10. A human-like risk control method for autonomous driving, characterized in that, include: Receive environmental perception information and generate at least one counterfactual trajectory based on the environmental perception information, wherein the counterfactual trajectory is used to characterize the expected trajectory that conforms to the cautious driving behavior pattern under the environmental perception information; Obtain the current planned trajectory output by the vehicle's main planning decision system, and quantitatively compare the current planned trajectory with the counterfactual trajectory to obtain the anthropomorphic risk level; With the aim of reducing or eliminating anthropomorphic risks, intervention commands are output to the vehicle's human-machine interaction system or the vehicle's main planning and decision-making system based on the anthropomorphic risk level.

Images