Multi-vehicle interaction intention prediction method and device based on social value orientation

Abstract

The invention discloses a social value orientation-based multi-vehicle interaction intention prediction method and apparatus. The method comprises the steps of obtaining multi-source input data; preprocessing to obtain a standardized historical interaction behavior data set; learning an interpretable monomer reward function from historical data based on maximum entropy inverse reinforcement learning to quantify the driving behavior utility; vehicle social value priori distribution is initialized, and social value posteriori estimation is obtained through updating of a probability filter in combination with multi-vehicle historical observation data and a single reward function; and inputting the single reward function and the social value posteriori estimation as prior information into a multi-vehicle interaction model to realize probability prediction of future interaction intentions of surrounding vehicles. The method effectively overcomes the dependence of the prior art on a manual reward project, improves the accuracy, interpretability and real-time performance of intention prediction in a complex interaction scene, and provides a reliable basis for automatic driving decision planning.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic driving, and particularly relates to a multi-vehicle interaction intention prediction method and device based on social value orientation. BACKGROUND

[0002] In the research and engineering practice of automatic driving and intelligent transportation, multi-vehicle interaction intention prediction is a core task, which aims to infer the intentions and decision tendencies of surrounding vehicles in the future steps from historical trajectories and real-time perception. Traditional methods include heuristic models based on dynamics, behavior cloning or end-to-end deep learning models, etc., but these methods often have difficulty in simultaneously considering interpretability and robustness, especially when encountering drivers with different social preferences or risk preferences.

[0003] In current research, social value orientation (SVO) is proposed as a social preference representation that synthesizes self-utility and other-utility in proportion, which is used to capture the "selfish-altruistic" tendency in driving behavior. This kind of method usually learns single-vehicle rewards through certain reward assumptions or inverse reinforcement learning methods, and then obtains the social value parameters of individuals through probability filtering or posterior estimation to participate in interaction modeling. This direction of research can theoretically explain various game behaviors and improve prediction interpretability, but in the engineering implementation level, it faces problems such as reward engineering complexity, inverse reinforcement learning sensitivity to samples and features, and high difficulty in reproduction.

[0004] The SVO modeling, histogram filtering update process and decision design based on iterative optimal feedback provided by DiDi American Research Institute provide important inspiration, but its road test results also expose the shortcomings in adapting to large-scale log data and online computing efficiency. Similar methods such as Policy Bayesian Inference, CPT, etc. have advantages in theory and experiment, but in the process of engineering landing to automatic driving systems, they generally face three major problems: first, the definition of single-vehicle reward and the inverse reinforcement learning step are highly dependent on data and feature engineering, making the method difficult to migrate or reproduce; second, real-time estimation of multi-vehicle interaction needs to balance computing efficiency and estimation accuracy, and it is difficult to achieve robust online estimation under limited resources on the vehicle side; third, existing methods provide few complete modular processes and visualization means from logs to rewards to intentions, making it difficult to support systematic troubleshooting and iterative improvement.

[0005] Two representative methods of the prior art are introduced as follows:

[0006] (1) Method based on Policy Bayesian Inference

[0007] This method attempts to estimate the driver's tendency in a lower-dimensional meta-reward space, designs several meta-reward factors (such as egoism, courtesy, confidence, etc.) according to experience, and performs Bayesian inference on the weights of these meta-rewards based on the idea of maximum entropy posterior estimation.

[0008] However, this approach is more restrictive than the SVO method in terms of learning efficiency and engineering implementation, but it also relies on the clear definition of the single reward function, which is often difficult to guarantee in real traffic data. In addition, the strategy Bayesian method often uses Top-K strategy to select candidate actions due to the discretization of action distribution in decision output, which may lead to roughness and underfitting in intention prediction in complex continuous interaction scenarios.

[0009] (2) Cumulative Prospect Theory (CPT) based method

[0010] This method explains the trade-off between safety and aggressiveness of humans through nonlinear transformation of gains and losses and probability weighting. In multi-vehicle interaction, embedding the CPT idea into the overall utility evaluation can capture the asymmetric perception of individuals to potential benefits and risks in different situations.

[0011] However, as a psychological characterization, direct application of CPT to decision-making or inference often requires special calibration of parameters, and numerical stability and real-time performance in combinatorial games or continuous control scenarios are still challenges for engineering implementation. SUMMARY

[0012] To solve the above technical problems, the present application proposes a multi-vehicle interaction intention prediction method and device based on social value orientation.

[0013] To achieve the above purpose, the technical scheme of the present application is as follows:

[0014] In a first aspect, the present application discloses a multi-vehicle interaction intention prediction method based on social value orientation, comprising:

[0015] Step S1: obtaining multi-source input data in an autonomous driving scenario, the multi-source input data including sensor logs generated by real vehicles or simulation, vehicle historical trajectory labels, and road map information;

[0016] Step S2: preprocessing the multi-source input data to obtain a standardized historical interaction behavior dataset;

[0017] Step S3: based on a maximum entropy inverse reinforcement learning framework, learning an interpretable single reward function from the historical interaction behavior data, wherein the single reward function is used to quantify the utility of driving behavior;

[0018] Step S4: Initialize the social value prior distribution for each vehicle with the intention to be predicted, obtain multi-vehicle multi-step historical observation data, use the individual reward function learned in step S3 as the observation weight, update the social value prior distribution online through a probability filter, and obtain the social value posterior estimate of each vehicle in the current interaction scenario.

[0019] Step S5: Using the individual reward function learned in step S3 and the posterior estimate of social value obtained in step S4 as prior information, input them into the multi-vehicle interaction model for intention prediction to predict the probability of future interaction intentions of surrounding vehicles.

[0020] Based on the above technical solution, the following improvements can be made:

[0021] As a preferred embodiment, step S3 includes:

[0022] Step S3.1: Input the historical interaction behavior dataset into the maximum entropy inverse reinforcement learning framework;

[0023] Step S3.2: Based on the maximum entropy inverse reinforcement learning framework, sample the historical vehicle trajectories in the historical interaction behavior dataset to generate a candidate trajectory set;

[0024] Step S3.3: From the generated candidate trajectory set, select trajectories that conform to the driving patterns of real roads as expert trajectory samples;

[0025] Step S3.4: Construct a feature system for inverse learning, which includes one or more of the following features: relative vehicle position, speed and acceleration difference, and historical behavior semantics;

[0026] Step S3.5: Based on the constructed feature system and the selected expert trajectory samples, execute the maximum entropy inverse reinforcement learning process, and finally output an interpretable single reward function through inverse learning.

[0027] As a preferred embodiment, step S3 further includes: verification of the individual reward function, which includes the following:

[0028] The obtained individual reward function is applied to the candidate trajectory set to calculate the score difference between expert trajectories and non-expert trajectories. Based on this score difference, the effectiveness and interpretability of the individual reward function are verified.

[0029] As a preferred option, the following also include:

[0030] Step S6: Evaluate and visualize the prediction process and results, including corner case sampling, keyframe visualization, indexed accuracy alignment testing, and system resource consumption evaluation.

[0031] Secondly, this invention discloses a multi-vehicle interaction intent prediction device based on social value orientation, comprising:

[0032] The data acquisition module is used to acquire multi-source input data in autonomous driving scenarios. The multi-source input data includes sensor logs generated from real vehicles or simulations, vehicle historical trajectory labels, and road map information.

[0033] The preprocessing module is used to preprocess multi-source input data to obtain a standardized historical interaction behavior dataset;

[0034] The reward learning module is used to learn an interpretable individual reward function from historical interaction behavior data based on the maximum entropy inverse reinforcement learning framework. The individual reward function is used to quantify the utility of driving behavior.

[0035] The social value estimation module is used to initialize the social value prior distribution for each vehicle with the intention to be predicted, acquire multi-vehicle multi-step historical observation data, use the individual reward function learned by the reward learning module as the observation weight, and update the social value prior distribution online through a probability filter to obtain the social value posterior estimate of each vehicle in the current interaction scenario.

[0036] The intent prediction module uses the individual reward function learned by the reward learning module and the posterior estimate of social value obtained by the social value estimation module as prior information, and inputs them into the multi-vehicle interaction model of intent prediction to predict the probability of future interaction intentions of surrounding vehicles.

[0037] As a preferred option, the reward-based learning module includes:

[0038] The input unit is used to input the historical interaction behavior dataset into the maximum entropy inverse reinforcement learning framework.

[0039] The candidate trajectory generation unit is used to sample the historical vehicle trajectories in the historical interaction behavior dataset based on the maximum entropy inverse reinforcement learning framework to generate a set of candidate trajectories.

[0040] The filtering unit is used to select trajectories that conform to real-world driving patterns from the generated candidate trajectory set, and use them as expert trajectory samples.

[0041] The building unit is used to construct a feature system for inverse learning, which includes one or more of the following features: relative vehicle position, speed and acceleration difference, and historical behavior semantics.

[0042] The reward learning unit is used to execute the maximum entropy inverse reinforcement learning process based on the constructed feature system and the selected expert trajectory samples, and finally outputs an interpretable single reward function through inverse learning.

[0043] As a preferred option, the reward-based learning module also includes:

[0044] The verification unit is used to verify the individual reward function;

[0045] The obtained individual reward function is applied to the candidate trajectory set to calculate the score difference between expert trajectories and non-expert trajectories. Based on this score difference, the effectiveness and interpretability of the individual reward function are verified.

[0046] As a preferred option, the following also include:

[0047] The evaluation and display module is used to evaluate and visualize the prediction process and results, including corner case sampling, keyframe visualization, indexed accuracy alignment testing, and system resource consumption evaluation.

[0048] Thirdly, the present invention discloses a computing device, comprising:

[0049] One or more processors;

[0050] Memory;

[0051] And one or more programs, wherein the one or more programs are stored in memory and configured to be executed by one or more processors, and the one or more programs include instructions for any of the above-described social value-based multi-vehicle interaction intent prediction methods.

[0052] Fourthly, the present invention discloses a storage medium storing one or more computer-readable programs, the one or more programs including instructions adapted to be loaded by a memory and executed by any of the above-described multi-vehicle interaction intention prediction methods based on social value orientation.

[0053] To address the problems in existing technologies such as the difficulty in constructing individual rewards, limitations in estimating social preferences, and the difficulty in balancing accuracy and real-time performance in online inference of risk preferences and multi-vehicle interactions, this invention discloses a method and apparatus for predicting multi-vehicle interaction intentions based on social value orientation, which has the following beneficial effects:

[0054] First, this invention employs maximum entropy inverse reinforcement learning to learn an interpretable individual reward function from historical interaction behavior data, avoiding the bias and weak generalization problems caused by relying on manually designed rewards in traditional methods. This data-driven reward function can more realistically reflect the utility of driving behavior, thus providing a more reliable and accurate numerical basis for subsequent social value orientation estimation and intention prediction, significantly reducing the reliance on human reward shaping and improving the interpretability of the behavioral model.

[0055] Second, by combining social value orientation estimation with commonly used filtering methods for posterior estimation, this invention can more accurately characterize drivers' social preferences and personalized decisions in different situations, thereby improving the accuracy and robustness of short- and medium-term intention prediction.

[0056] Third, this invention not only outputs visualized prediction results, but also provides a clear intermediate inference process, which helps R&D and testing personnel understand the system's decision-making logic, increases trust in the autonomous driving system's decisions, facilitates alignment and problem localization by the engineering side, thereby accelerating the model's deployment process from training to deployment and reducing reproduction and maintenance costs.

[0057] In summary, this invention obtains interpretable individual reward functions from data through maximum entropy inverse reinforcement learning, and uses these functions to drive online estimation of social value orientation and inference of multi-vehicle game theory. This significantly improves the accuracy, interpretability, and engineering implementation efficiency of intent prediction, providing a reliable technical guarantee for the safe and smooth operation of autonomous vehicles in complex interactive environments. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart of a multi-vehicle interaction intent prediction method provided in an embodiment of the present invention.

[0060] Figure 2 This is an example diagram illustrating data acquisition and preprocessing operations provided in an embodiment of the present invention.

[0061] Figure 3 An example diagram illustrating reward-based learning provided in an embodiment of the present invention.

[0062] Figure 4 An example diagram illustrating the estimation of social value orientation provided for embodiments of the present invention.

[0063] Figure 5 Example diagrams for effect evaluation and visualization provided in embodiments of the present invention. Detailed Implementation

[0064] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0065] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] The expression “includes” is an “open-ended” expression, which means that there is a corresponding component or step, and should not be interpreted as excluding additional components or steps.

[0067] To achieve the objectives of this invention, some embodiments of a multi-vehicle interaction intent prediction method and apparatus based on social value orientation are provided, such as... Figure 1 As shown, the multi-vehicle interaction intent prediction method includes:

[0068] Step S101: Obtain multi-source input data in the autonomous driving scenario. The multi-source input data includes sensor logs generated from the real vehicle or simulation, vehicle historical trajectory labels, and road map information.

[0069] Step S102: Preprocess the multi-source input data to obtain a standardized historical interaction behavior dataset;

[0070] Step S103: Based on the maximum entropy inverse reinforcement learning framework, learn an interpretable individual reward function from historical interaction behavior data, where the individual reward function is used to quantify the utility of driving behavior;

[0071] Step S104: Initialize the social value prior distribution for each vehicle with the intention to be predicted, obtain multi-vehicle multi-step historical observation data, use the individual reward function learned in step S103 as the observation weight, update the social value prior distribution online through a probability filter, and obtain the social value posterior estimate of each vehicle in the current interaction scenario.

[0072] Step S105: Using the individual reward function learned in step S103 and the posterior estimate of social value obtained in step S104 as prior information, input them into the multi-vehicle interaction model for intention prediction to predict the probability of future interaction intentions of surrounding vehicles.

[0073] This invention focuses on addressing the problem of unobservable intent caused by the diversity of human driving behavior, differences in social value, and differences in risk preferences. By combining log-driven reward learning, social value estimation, and probabilistic inference, it improves the accuracy and interpretability of multi-vehicle interaction intent prediction, thereby providing reliable priors for vehicle decision-making and planning.

[0074] The steps of this invention will be described in detail below.

[0075] like Figure 2 As shown, step S101 is the data acquisition step.

[0076] The preprocessing operations in step S102 include one or more of the following: timestamp alignment, coordinate system I, map information matching, and feature normalization: performing timestamp alignment, coordinate system I, map information matching, and normalization on multi-source input data, and extracting one or more features including relative position, velocity, acceleration, heading angle difference, and behavioral semantic labels.

[0077] This invention, after unified time and coordinate alignment, feature normalization, and map information cropping, outputs an aligned historical interaction behavior dataset for subsequent learning and estimation. Notably, the preprocessing fully considers the computational and latency constraints of the deployment end, providing a "global / local" map search strategy selection and offering engineering implementation suggestions for multi-threaded parallel processing of real-time sensitive data paths to ensure consistency between the online processing efficiency on the vehicle end and the training data on the model end.

[0078] like Figure 3 As shown, step S103 is the reward learning step, which includes:

[0079] Step S103.1: Input the historical interaction behavior dataset into the maximum entropy inverse reinforcement learning framework;

[0080] Step S103.2: Based on the maximum entropy inverse reinforcement learning framework, sample the historical vehicle trajectories in the historical interaction behavior dataset to generate a candidate trajectory set;

[0081] Step S103.3: Select trajectories that conform to real-world driving patterns from the generated candidate trajectory set and use them as expert trajectory samples;

[0082] Step S103.4: Construct a feature system for inverse learning, which includes one or more of the following features: relative vehicle position, speed and acceleration difference, historical behavior semantics, etc.

[0083] Step S103.5: Based on the constructed feature system and the selected expert trajectory samples, execute the maximum entropy inverse reinforcement learning process, and finally output an interpretable single reward function through inverse learning.

[0084] This single reward function not only serves as the numerical basis for subsequent estimation and decision-making, but also incorporates an automated verification process designed in engineering to verify whether the learned reward can significantly distinguish expert trajectories on the candidate trajectory set, thereby reducing the engineering burden of human reward shaping.

[0085] The verification of the individual reward function includes the following:

[0086] The obtained individual reward function is applied to the candidate trajectory set to calculate the score difference between expert trajectories and non-expert trajectories. Based on this score difference, the effectiveness and interpretability of the individual reward function are verified.

[0087] like Figure 4 As shown, step S104 is the social value estimation and Bayesian inference step. This invention draws on and extends the ideas of Social Value Orientation (SVO) and Policy Bayesian Inference. Specifically, a prior social value distribution is maintained for each vehicle. This prior distribution is updated online using a probability filter (such as a general framework of histogram filtering or Bayesian posterior update) based on multi-step historical observations and the learned individual reward function, to obtain the posterior social value estimate of each vehicle in the current scenario.

[0088] Unlike traditional SVO methods, this invention introduces the reward obtained from maximum entropy IRL as the weight of the observation during the estimation process. This approach retains the explanatory power of SVO while taking into account the impact of the driver's nonlinear perception of risk. This combination helps to more realistically reflect human decision-making preferences in critical interaction scenarios.

[0089] Step S105 is the intent prediction and interaction inference step. This invention uses the obtained individual reward function and social value posterior as prior information to construct a probabilistic inference process for multi-vehicle interactions. This process includes candidate policy generation, policy scoring and probability normalization based on the current posterior, and approximate solution of the multi-vehicle joint policy space. When solving the approximation, engineering-feasible approximation methods are preferentially adopted, such as probability rearrangement based on the Top-K candidate set and iterative optimal response, so that the system maintains the computability of online inference while ensuring prediction accuracy.

[0090] For offline analysis scenarios requiring higher precision, this invention also supports the introduction of more complex game solvers or hierarchical game models to further improve the rationality of decision-making.

[0091] Furthermore, the multi-vehicle interaction intent prediction method of the present invention also includes step S106, which is a step of effect evaluation and visualization display.

[0092] like Figure 5 As shown, step S106: Evaluate and visualize the prediction process and results, including corner case sampling, keyframe visualization, indexed accuracy alignment test (based on trajectory error, confusion matrix of intent category, etc.) and system resource consumption evaluation.

[0093] It is worth noting that, in order to facilitate engineering troubleshooting, this invention proposes a troubleshooting strategy for control variables, that is, during the debugging phase, only a single variable (such as map clipping radius, model quantization level, or Bayesian prior parameters) is allowed to be changed at a time, so as to quickly locate the problem that causes the sudden change in performance.

[0094] The above-mentioned prediction of multi-vehicle interaction intentions based on social value orientation has the following effects:

[0095] 1) Reduce reliance on manual intervention and improve the reliability and interpretability of the reward function.

[0096] This invention utilizes a maximum entropy inverse reinforcement learning framework to automatically learn individual reward functions from standardized historical interaction behavior data. This eliminates the need for manually pre-setting reward factors (such as the meta-rewards like "selfishness" and "politeness" relied upon in traditional Policy Bayesian Inference methods). Furthermore, the process of candidate trajectory generation, expert trajectory screening, and feature system construction ensures that the reward function closely reflects real-world driving patterns. Simultaneously, the "individual reward verification step" verifies the effectiveness of the reward by calculating the score difference between expert and non-expert trajectories, further avoiding biases caused by manually designed rewards. This invention significantly reduces reliance on manual reward shaping, solving the problems of difficult individual reward construction and poor reproducibility in existing technologies. Moreover, because the reward function is learned inversely from real trajectory data, the utility quantification logic of driving behavior can be clearly explained.

[0097] 2) Accurately characterize driving preferences to improve the accuracy and robustness of intention prediction.

[0098] This invention uses the individual reward function learned in step S103 as the observation weight, and updates the prior distribution of vehicle social value online through a probability filter to obtain the posterior estimate of social value. This not only retains the ability of Social Value Orientation (SVO) to characterize the "self-interest-altruism" tendency, but also incorporates the driver's nonlinear perception of risk (such as lower reward weights for high-risk behaviors) through individual rewards, thus overcoming the shortcomings of traditional SVO methods that only focus on social preferences and ignore risk perception. At the same time, by using individual rewards and the posterior of social value as prior information for multi-vehicle interaction models, it enables probabilistic prediction of the future intentions of surrounding vehicles. This can accurately adapt to drivers with different social preferences and risk preferences, solving the problems of rough prediction and underfitting in complex interaction scenarios of existing technologies, and significantly improving the accuracy and robustness of short-term and medium-term intention prediction.

[0099] 3) Balancing vehicle-side real-time performance with engineering feasibility to accelerate the technology implementation process.

[0100] By preprocessing operations (such as timestamp alignment, coordinate system unification, map information matching, etc.) to create a standardized dataset, it can be directly adapted to large-scale log data from real vehicles or simulations, avoiding processing delays caused by the chaotic formats of multi-source data. A probabilistic filter is used to achieve online updates of social value distribution, eliminating the need for complex game theory calculations and meeting the real-time requirements under vehicle-side resource constraints. Furthermore, step S106, through "corner case sampling, keyframe visualization, and resource consumption evaluation," can intuitively present the intermediate results and performance bottlenecks of each stage of data processing, reward learning, and intent prediction. It also facilitates the location of problems through quantitative indicators (such as trajectory error and intent confusion matrix), and allows for rapid optimization combined with control variable investigation strategies (such as adjusting map clipping radius and prior distribution parameters). This significantly reduces the alignment difficulty and maintenance cost of the model from training to deployment, accelerating the technology engineering implementation process.

[0101] In other embodiments, the present invention discloses a multi-vehicle interaction intent prediction device based on social value orientation, comprising:

[0102] The data acquisition module is used to acquire multi-source input data in autonomous driving scenarios. The multi-source input data includes sensor logs generated from real vehicles or simulations, vehicle historical trajectory labels, and road map information.

[0103] The preprocessing module is used to preprocess multi-source input data to obtain a standardized historical interaction behavior dataset;

[0104] The reward learning module is used to learn an interpretable individual reward function from historical interaction behavior data based on the maximum entropy inverse reinforcement learning framework. The individual reward function is used to quantify the utility of driving behavior.

[0105] The social value estimation module is used to initialize the social value prior distribution for each vehicle with the intention to be predicted, acquire multi-vehicle multi-step historical observation data, use the individual reward function learned by the reward learning module as the observation weight, and update the social value prior distribution online through a probability filter to obtain the social value posterior estimate of each vehicle in the current interaction scenario.

[0106] The intent prediction module uses the individual reward function learned by the reward learning module and the posterior estimate of social value obtained by the social value estimation module as prior information, and inputs them into the multi-vehicle interaction model of intent prediction to predict the probability of future interaction intentions of surrounding vehicles.

[0107] Furthermore, the reward-based learning module includes:

[0108] The input unit is used to input the historical interaction behavior dataset into the maximum entropy inverse reinforcement learning framework.

[0109] The candidate trajectory generation unit is used to sample the historical vehicle trajectories in the historical interaction behavior dataset based on the maximum entropy inverse reinforcement learning framework to generate a set of candidate trajectories.

[0110] The filtering unit is used to select trajectories that conform to real-world driving patterns from the generated candidate trajectory set, and use them as expert trajectory samples.

[0111] The building unit is used to construct a feature system for inverse learning, which includes one or more of the following features: relative vehicle position, speed and acceleration difference, and historical behavior semantics.

[0112] The reward learning unit is used to execute the maximum entropy inverse reinforcement learning process based on the constructed feature system and the selected expert trajectory samples, and finally outputs an interpretable single reward function through inverse learning.

[0113] Furthermore, the reward-based learning module also includes:

[0114] The verification unit is used to verify the individual reward function;

[0115] The obtained individual reward function is applied to the candidate trajectory set to calculate the score difference between expert trajectories and non-expert trajectories. Based on this score difference, the effectiveness and interpretability of the individual reward function are verified.

[0116] Furthermore, it also includes:

[0117] The evaluation and display module is used to evaluate and visualize the prediction process and results, including corner case sampling, keyframe visualization, indexed accuracy alignment testing, and system resource consumption evaluation.

[0118] Furthermore, it should be noted that the multi-vehicle interaction intent prediction device based on social value orientation provided in the above embodiments is only illustrated by the division of the above functional modules when performing multi-vehicle interaction intent prediction. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the multi-vehicle interaction intent prediction device based on social value orientation can be divided into different functional modules to complete all or part of the functions described above.

[0119] Furthermore, the embodiments of the multi-vehicle interaction intent prediction device based on social value orientation and the multi-vehicle interaction intent prediction method based on social value orientation provided in the above embodiments belong to the same concept. For details of its specific implementation process, please refer to the method embodiments, which will not be repeated here.

[0120] In other embodiments, the present invention discloses a computing device, comprising:

[0121] One or more processors;

[0122] Memory;

[0123] And one or more programs, wherein the one or more programs are stored in memory and configured to be executed by one or more processors, and the one or more programs include instructions for any of the above-described social value-based multi-vehicle interaction intent prediction methods.

[0124] In other embodiments, the present invention discloses a storage medium storing one or more computer-readable programs, the programs including instructions adapted to be loaded by a memory and executed in any of the above-described multi-vehicle interaction intent prediction methods based on social value orientations.

[0125] To address the problems in existing technologies such as the difficulty in constructing individual rewards, limitations in estimating social preferences, and the difficulty in balancing accuracy and real-time performance in online inference of risk preferences and multi-vehicle interactions, this invention discloses a method and apparatus for predicting multi-vehicle interaction intentions based on social value orientation, which has the following beneficial effects:

[0126] First, this invention employs maximum entropy inverse reinforcement learning to learn an interpretable individual reward function from historical interaction behavior data, avoiding the bias and weak generalization problems caused by relying on manually designed rewards in traditional methods. This data-driven reward function can more realistically reflect the utility of driving behavior, thus providing a more reliable and accurate numerical basis for subsequent social value orientation estimation and intention prediction, significantly reducing the reliance on human reward shaping and improving the interpretability of the behavioral model.

[0127] Second, by combining social value orientation estimation with commonly used filtering methods for posterior estimation, this invention can more accurately characterize drivers' social preferences and personalized decisions in different situations, thereby improving the accuracy and robustness of short- and medium-term intention prediction.

[0128] Third, this invention not only outputs visualized prediction results, but also provides a clear intermediate inference process, which helps R&D and testing personnel understand the system's decision-making logic, increases trust in the autonomous driving system's decisions, facilitates alignment and problem localization by the engineering side, thereby accelerating the model's deployment process from training to deployment and reducing reproduction and maintenance costs.

[0129] In summary, this invention obtains interpretable individual reward functions from data through maximum entropy inverse reinforcement learning, and uses these functions to drive online estimation of social value orientation and inference of multi-vehicle game theory. This significantly improves the accuracy, interpretability, and engineering implementation efficiency of intent prediction, providing a reliable technical guarantee for the safe and smooth operation of autonomous vehicles in complex interactive environments.

[0130] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope. All such changes and modifications fall within the scope of the present invention as claimed, which is defined by the appended claims and their equivalents.

Claims

1. A multi-vehicle interaction intent prediction method based on social value orientation, characterized in that, include: Step S1: Obtain multi-source input data in the autonomous driving scenario. The multi-source input data includes sensor logs generated from the real vehicle or simulation, vehicle historical trajectory labels, and road map information. Step S2: Preprocess the multi-source input data to obtain a standardized historical interaction behavior dataset; Step S3: Based on the maximum entropy inverse reinforcement learning framework, learn an interpretable individual reward function from the historical interaction behavior data, wherein the individual reward function is used to quantify the utility of driving behavior; Step S4: Initialize the social value prior distribution for each vehicle with the intention to be predicted, obtain multi-vehicle multi-step historical observation data, use the individual reward function learned in step S3 as the observation weight, and update the social value prior distribution online through a probability filter to obtain the social value posterior estimate of each vehicle in the current interaction scenario. Step S5: Using the individual reward function learned in step S3 and the posterior estimate of social value obtained in step S4 as prior information, input them into the multi-vehicle interaction model for intention prediction to predict the probability of future interaction intentions of surrounding vehicles.

2. The multi-vehicle interaction intent prediction method according to claim 1, characterized in that, Step S3 includes: Step S3.1: Input the historical interaction behavior dataset into the maximum entropy inverse reinforcement learning framework; Step S3.2: Based on the maximum entropy inverse reinforcement learning framework, sample the historical vehicle trajectories in the historical interaction behavior dataset to generate a candidate trajectory set; Step S3.3: From the generated candidate trajectory set, select trajectories that conform to the driving patterns of real roads as expert trajectory samples; Step S3.4: Construct a feature system for inverse learning, which includes one or more of the following features: relative vehicle position, speed and acceleration difference, and historical behavior semantics; Step S3.5: Based on the constructed feature system and the selected expert trajectory samples, execute the maximum entropy inverse reinforcement learning process, and finally output an interpretable single reward function through inverse learning.

3. The multi-vehicle interaction intent prediction method according to claim 2, characterized in that, Step S3 further includes: verification of the individual reward function, which includes the following: The obtained individual reward function is applied to the candidate trajectory set to calculate the score difference between expert trajectories and non-expert trajectories. Based on this score difference, the effectiveness and interpretability of the individual reward function are verified.

4. The multi-vehicle interaction intent prediction method according to claim 1, characterized in that, Also includes: Step S6: Evaluate and visualize the prediction process and results, including corner case sampling, keyframe visualization, indexed accuracy alignment testing, and system resource consumption evaluation.

5. A multi-vehicle interaction intent prediction device based on social value orientation, characterized in that, include: The acquisition module is used to acquire multi-source input data in autonomous driving scenarios. The multi-source input data includes sensor logs generated from real vehicles or simulations, vehicle historical trajectory labels, and road map information. The preprocessing module is used to preprocess the multi-source input data to obtain a standardized historical interaction behavior dataset; The reward learning module is used to learn an interpretable individual reward function from the historical interaction behavior data based on the maximum entropy inverse reinforcement learning framework, wherein the individual reward function is used to quantify the utility of driving behavior. The social value estimation module is used to initialize the social value prior distribution for each vehicle with the intention to be predicted, acquire multi-vehicle multi-step historical observation data, use the individual reward function learned by the reward learning module as the observation weight, and update the social value prior distribution online through a probability filter to obtain the social value posterior estimate of each vehicle in the current interaction scenario. The intent prediction module is used to input the individual reward function learned by the reward learning module and the posterior estimate of social value obtained by the social value estimation module as prior information into the multi-vehicle interaction model of intent prediction, and to predict the future interaction intent of surrounding vehicles probabilistically.

6. The multi-vehicle interaction intent prediction device according to claim 5, characterized in that, The reward-based learning module includes: The input unit is used to input the historical interaction behavior dataset into the maximum entropy inverse reinforcement learning framework. The candidate trajectory generation unit is used to sample the historical vehicle trajectories in the historical interaction behavior dataset based on the maximum entropy inverse reinforcement learning framework to generate a set of candidate trajectories. The filtering unit is used to select trajectories that conform to real-world driving patterns from the generated candidate trajectory set, and use them as expert trajectory samples. The building unit is used to construct a feature system for inverse learning, which includes one or more of the following features: relative vehicle position, speed and acceleration difference, and historical behavior semantics. The reward learning unit is used to execute the maximum entropy inverse reinforcement learning process based on the constructed feature system and the selected expert trajectory samples, and finally outputs an interpretable single reward function through inverse learning.

7. The multi-vehicle interaction intent prediction device according to claim 6, characterized in that, The reward-based learning module also includes: The verification unit is used to verify the individual reward function; The obtained individual reward function is applied to the candidate trajectory set to calculate the score difference between expert trajectories and non-expert trajectories. Based on this score difference, the effectiveness and interpretability of the individual reward function are verified.

8. The multi-vehicle interaction intent prediction device according to claim 5, characterized in that, Also includes: The evaluation and display module is used to evaluate and visualize the prediction process and results, including corner case sampling, keyframe visualization, indexed accuracy alignment testing, and system resource consumption evaluation.

9. A computing device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, and the one or more programs include instructions for the social value orientation-based multi-vehicle interaction intent prediction method as described in any of claims 1-4.

10. A storage medium, characterized in that, The storage medium stores one or more computer-readable programs, the programs including instructions adapted to be loaded by memory and executed as described in any of claims 1-4 above, for the multi-vehicle interaction intent prediction method based on social value orientation.

Images