Data-driven autonomous driving decision optimization system and method

Abstract

A kind of data-driven automatic driving decision optimization system and method, the system includes: data production module, data screening module, model encapsulation module and parameter tuning module, wherein: data production module is with human driving data as input, after being labeled, preprocessed and format conversion, extract key features and carry out standardization, normalization and coding, generate the original data in data-driven process in line with algorithm input requirement;Data screening module is according to each scene divided by decision maker from training data corresponding data is screened out and is effectively classified;Model encapsulation module encapsulates the decision code of C++ and constructs the trajectory pair evaluation cost map required for training using the true value evaluation method based on trajectory pair;Parameter tuning module is according to the data under different scenes after screening, encapsulated decision algorithm model and trajectory pair evaluation cost map, using the method of black box optimization, obtain the decision parameter under the current decision algorithm in corresponding scene.This application combines rule-based and learning-based methods, and uses the large amount of data generated by autonomous vehicles during operation to improve the upper limit of the decision system, which is more beneficial to the performance of the decision maker in the actual scene and belongs to the data-driven decision parameter adjustment technology of white box nature.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of automatic driving, and particularly relates to a data-driven automatic driving decision optimization system and method. BACKGROUND

[0002] Decision technology is the core of an intelligent driving system and influences safety and comfort. Existing technologies include rule-based and learning-based methods. Rule-based methods rely on experience to design an objective function and seek an optimal solution, but the result is far from that of a skilled driver. Learning-based methods can improve the output effect of decision through model training, but the problem of insufficient explainability exists. SUMMARY

[0003] The application proposes a data-driven automatic driving decision optimization system and method to solve the problem that existing technologies are trained for specific scenes or tasks and are difficult to cope with dynamic and complex environmental changes. The system and method combine rule-based and learning-based methods, utilize a large amount of data generated by an automatic driving vehicle during operation, improve the upper limit of a decision system, and are more beneficial to the performance of a decision maker in an actual scene and belong to a data-driven decision parameter adjustment technology with white-box nature.

[0004] The application is implemented by the following technical scheme:

[0005] The application relates to a data-driven automatic driving decision optimization system, which comprises a data production module, a data screening module, a model encapsulation module and a parameter adjustment module. The data production module takes human driving data as input, extracts key features after labeling, preprocessing and format conversion, and performs standardization, normalization and encoding to generate original data in a data-driven process that meets the input requirements of an algorithm. The data screening module screens corresponding data from training data according to various scenes divided by a decision maker and effectively classifies the data. The model encapsulation module encapsulates decision code in C++ and constructs a trajectory pair evaluation cost map required for training by using a true value evaluation method based on a trajectory pair. The parameter adjustment module obtains decision parameters in a corresponding scene under a current decision algorithm by using a black-box optimization method according to data in different scenes after screening, an encapsulated decision algorithm model and a trajectory pair evaluation cost map.

[0006] The human driving data is generated by experienced drivers driving vehicles and is initially labeled, each small data set is composed of a ten-minute driving segment, and each frame of data needs to include comprehensive self-vehicle state information and surrounding environment information, including self-vehicle position, vehicle speed, acceleration, current lane ID, lane line information, lane association information, self-vehicle current behavior information such as left turn, straight, passing through an intersection, surrounding vehicle information such as vehicle ID, vehicle type, vehicle size, position, vehicle speed, vehicle future trajectory predicted by a prediction module, and the like.

[0007] The original data in the data-driven process includes: sampled trajectory information of the ego vehicle, sampled trajectory information of the other vehicle, a list of vehicle IDs required for game playing, a list of state information of the ego vehicle, a list of state information of the other vehicle, conflict point information, posterior trajectory of the ego vehicle within 6s in the future, posterior trajectory of the other vehicle within 6s in the future, and a list of environment information.

[0008] The decision-making scene includes straight, intersection, merging / merging, and lane changing scenes.

[0009] The corresponding data correspond to the division of the decision-making scene and obtain key data for training through data mining.

[0010] The application relates to an automatic driving decision optimization method based on the above system, which reads original data, calls a decision algorithm and sets rules to generate final data; data cleaning and data mining are performed according to the obtained data; a decision algorithm framework of sampling-target function-evaluation-screening is used to encapsulate the decision algorithm and construct a true value; and a black box optimization method is used to iteratively find the optimal parameter combination value that satisfies the target function of the training data in the corresponding scene, so that the optimized automatic driving decision is obtained.

[0011] The data mining processing is screened according to the difference between TTC (Time To Collision) of the ego vehicle and TTC of the other vehicle, and specifically: DTTC = TTC ego -TTC agent Wherein: DTTC represents the time difference between the ego vehicle and the other vehicle to the conflict point, TTC ego and TTC agent respectively represent the time of the ego vehicle and the other vehicle to the conflict point.

[0012] The conflict point refers to the intersection of the posterior trajectory of the ego vehicle and the posterior trajectory of the other vehicle in space, which is obtained by calculating the collision detection of the ego vehicle and the other vehicle.

[0013] The decision algorithm package refers to: firstly sampling the future driving trajectories of the ego vehicle and the other vehicle, then comprehensively evaluating the ego vehicle sampling and the other vehicle sampling results according to a pre-target function, so as to obtain a trajectory pair satisfying the minimum target function, and thus obtaining the yield strategy between the ego vehicle and the other vehicle.

[0014] The target function func=w s ·F safe +w c ·F comfort +w p ·F pass +w r ·F right +w safe F comfort F pass F right respectively represent the target function expressions of safety, comfort, trafficability and right-of-way, w s w c w p w r represent the weights of each in the target function.

[0015] The true value is constructed, that is, the decision true value required in the training data construction process, which determines the more similar trajectory pair and the decision result of the driver by calculating the similarity of the trajectory pair and the similarity of the mode.

[0016] The similarity of the trajectory pair includes the final point error (fde), the average error (ade) and the lateral error (lde) between the sampling trajectory and the posterior trajectory to measure, specifically: Wherein: Δd is the Euclidean distance, and Δl is the lateral distance in the Frenet coordinate system.

[0017] The similarity of the mode is obtained according to the DTTC of the ego vehicle and the other vehicle, the solution inconsistent with the posterior yield relationship of the sampling result is punished, and the true value evaluation calculation method based on the data and the decision algorithm framework is obtained, specifically: label__cost=w1×ade+w2×fde+w3×lde+w4×bde, wherein: the difference value of the mode is bde, w1, w2, w3 and w4 are hyperparameters, which can be designed artificially.

[0018] The parameter optimization adopts KL divergence loss function in the iteration process, specifically: Wherein: p(x i ) is the true value, and q(x i ) is the value output by the decision algorithm. In order to improve the efficiency, the whole training process can adopt the multi-thread training method.

[0019] Technical effects

[0020] The present application uses the driving style characteristics implied in a large amount of human driving data, extracts parameters reflecting the driving behavior rules through offline calculation and analysis, and uses them to drive the dynamic parameter adjustment in the automatic driving decision system. This method combines the characteristics of white box technology to internalize the driver's operation style into interpretable algorithm logic. The method based on TTC (Time to Collision) difference and conflict point analysis is used to accurately screen the key moment data for decision-making, and the trajectory prediction of the ego vehicle and other vehicles is generated, and through the comprehensive evaluation of the trajectory modal similarity and trajectory error, high-quality true value construction is completed. On the basis of data-driven, combined with the regularization framework, a multi-objective function of safety, comfort, traffic and right-of-way is established. Compared with the prior art, the present application effectively integrates data and logic by combining white-box data-driven methods with rule frameworks, making parameter decision-making more human-like and interpretable, rather than relying solely on simple black-box model optimization; using multi-threading acceleration, conflict point analysis and true value construction method based on modal, the efficiency of data processing and model training is improved, which is a significant improvement over the data cleaning and screening process in the prior art; the offline parameter generation method avoids the instability in online real-time optimization, further ensuring the safety and reliability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The figure is a schematic diagram of the system of the present application. DETAILED DESCRIPTION

[0022] As shown in Figure 1 , the present embodiment relates to a data-driven automatic driving decision parameter adjustment system, which includes a data production module, a data screening module, a model packaging module and a parameter tuning module.

[0023] The data production module comprises a human driving data unit, a main line sampling code packaging unit, a sampling trajectory feature generation unit and an optimal trajectory calibration unit, wherein: the human driving data unit collects and integrates a large amount of data generated by human drivers according to human driving data, including self-vehicle state (position, speed, acceleration, lane information) and surrounding vehicle and environment information (such as vehicle ID, type, trajectory, conflict point, etc.), for reflecting driving style and behavior rules; the main line sampling code packaging unit packages main line sampling code and automatically generates a trajectory sampling scheme meeting the rules according to data sampling rules and target function constraint conditions, to obtain a preliminary sampling trajectory set constrained by rules; the sampling trajectory feature generation unit extracts trajectory sampling features (such as speed, acceleration, TTC difference, etc.) according to trajectory sampling results, including trajectory points of self-vehicle and other vehicles and interaction information, generates trajectory feature vectors for analysis, and obtains sampling trajectory features for subsequent model packaging and evaluation; the optimal trajectory calibration unit performs multi-objective comprehensive evaluation on the sampling trajectory according to the trajectory feature and the target function value (such as safety, comfort, etc.), calibrates the optimal trajectory meeting the target function condition, and obtains the calibrated optimal trajectory for subsequent training and verification.

[0024] The data screening module comprises a scene screening unit, an effective frame screening unit, a key obstacle screening unit and a problem category screening unit, wherein: the scene screening unit classifies the trajectory data according to the trajectory data and the scene classification rule, coarsely classifies the trajectory data according to scenes such as straight driving, intersection, merging and outflow, provides labels for specific scenes, and obtains a classified scene data set to provide a basis for detailed analysis; the effective frame screening unit screens out key frames (such as frames with DTTC in the range of [-1.0, 3.0]) according to time-related indicators (such as TTC, DTTC) in an effective time window, eliminates redundant information, and obtains an effective frame set in a high-risk scene; the key obstacle screening unit screens out key obstacle information (such as objects close to or about to interact with the ego vehicle) that has an important influence on decision-making according to the trajectory data and the environment information (such as obstacle position and moving trajectory), and obtains a scene data subset containing key obstacle information; the problem category screening unit screens scene data according to scene features and potential problem categories (such as collision risk and overtaking behavior) in the data, focuses on solving specific driving problems, and obtains a high-quality training data set optimized for problem categories.

[0025] The model packaging module comprises a model input unit, a trajectory pair white box evaluation unit and a sampling space evaluation distribution unit, wherein: the model input unit inputs the trajectory pair features according to the filtered trajectory feature vectors and the related parameter recommendation rules, generates the recommended parameter set adapted to the model, obtains the recommended parameter set, and provides support for the accuracy and applicability of the model; the trajectory pair white box evaluation parameter unit calculates the evaluation parameters of the trajectory pair using the white box model according to the trajectory pair features and the objective function (such as trajectory similarity, modal similarity, etc.), obtains the trajectory pair white box evaluation parameters, and is used for accurate analysis of the decision effect; and the model output unit outputs the comprehensive evaluation distribution diagram of the sampling trajectory space according to the white box evaluation parameters and the sampling trajectory distribution, is used for describing the advantages and disadvantages of the trajectory under the objective function, and obtains the evaluation distribution of the sampling space, which provides a basis for parameter tuning.

[0026] The parameter tuning module comprises a parameter black box optimization input unit, a black box optimization unit and an iterative parameter recommendation unit, wherein: the parameter black box optimization input unit provides the initial parameter black box optimization input according to the model input parameters and the training data feature vectors, as the starting condition for optimization, obtains the initial parameter configuration for starting the optimization process; the black box optimization unit adjusts the parameters using the black box optimization algorithm (such as genetic algorithm, Bayesian optimization, etc.) according to the parameter configuration and the optimization objective function (KL divergence), gradually approaches the optimal value, and obtains the optimized parameter configuration result; and the iterative parameter recommendation unit iteratively recommends the parameter configuration according to the intermediate result output by the black box optimization and the evaluation index, adjusts the parameter combination to improve the optimization efficiency, and obtains a set of recommended optimization parameters.

[0027] The embodiment is based on the automatic driving decision optimization method of the above system, which comprises:

[0028] Step 1, data production: reading original data, calling decision algorithm, setting rules to generate final data, specifically including:

[0029] 1.1 driven by experienced drivers, each piece of data is divided into 10-minute segments, and each frame of data contains the following information: vehicle state (position, speed, acceleration, lane information, current behavior, etc.); surrounding vehicle information (vehicle ID, type, size, speed, predicted trajectory, etc.); conflict point, future 6-second trajectory information of the ego vehicle and other vehicles, environmental information, etc.

[0030] 1.2 collate and output data for training, including: sampling trajectory, key vehicle ID, state information of the ego vehicle and other vehicles, conflict point information, etc.

[0031] 1.3 improve data production efficiency through multi-threaded operation, call decision algorithm and rules to generate final data.

[0032] Step 2, data screening: data cleaning, data mining process, specifically including:

[0033] 2.1 Scene division: according to the straight, intersection, and other scenes such as in-out, the data is preliminarily classified, and further subdivided, such as for the large and small car scene of the other car cutting in.

[0034] 2.2 Key frame extraction: use TTC (time to collision) difference to screen key frames, screen frames with DTTC between [-1.0, 3.0].

[0035] 2.3 Conflict point identification: calculate the intersection of the ego car and the other car's posterior trajectory.

[0036] Step 3, model packaging: adopt the decision algorithm framework of sampling-target function-evaluation-screening to package the decision algorithm and build the true value, specifically including:

[0037] 3.1 Core framework: the decision algorithm is based on the process of sampling→target function→evaluation→screening, the target function considers safety, comfort, trafficability, and road right factors, and evaluates the trajectory pair to get the optimal yield strategy.

[0038] 3.2 True value construction method: measure the similarity of the trajectory pair (final point error FDE, average error ADE, lateral error LDE). Modal similarity: punish the deviation (BDE) of the sampling result and the posterior yield relationship. Build the final true value evaluation function for the training process.

[0039] Step 4, use the method of black box optimization to iteratively find the optimal parameter combination value that meets the target function of the training data in the corresponding scene, specifically including:

[0040] 4.1 The screened training data and calibrated true value are used as input to iteratively find the optimal parameter combination through black box optimization.

[0041] 4.2 The loss function used in the training process is the KL divergence loss function Where: p(x i ) is the true value, q(x i ) is the value output by the decision algorithm. The whole training process uses a multi-thread training method to improve efficiency.

[0042] Through specific experiments, under the Simulink-Carsim-Prescan joint simulation platform environment, the experimental scene is built by Prescan, the vehicle dynamics model is provided by Carsim, and the algorithm is deployed by Simulink. The high-speed dynamic scene of multiple lanes and multiple vehicles, the other vehicle cut-in scene and the ego vehicle lane-changing scene are selected for simulation. The PHP algorithm using the method is compared with the existing fixed strategy modeling uncertainty method (PHP*) in the above three types of simulation scenes.

[0043] In the high-speed dynamic scene of multiple lanes and multiple vehicles, the speed of the PHP method is significantly higher than that of PHP*, indicating that the PHP algorithm using the decision optimization method of the present application has higher passing efficiency. For the other vehicle cut-in scene, the PHP method performs better in comfort, with a 86.5% improvement in the absolute value of the maximum deceleration compared to PHP*; for the ego vehicle lane-changing scene, the PHP method performs better in passing, with a 28.5% improvement in average speed compared to PHP*. As can be seen, the PHP method exhibits higher comfort and passing in real vehicle experiments, especially in handling dynamic vehicle cut-in and lane-changing scenes, which reflects the effectiveness of the decision optimization method of the present application.

[0044] Compared with the prior art, the present method comprehensively considers the ego vehicle, other vehicles and environmental information, builds more realistic and detailed data input, and has more comprehensive data dimensions; through conflict point analysis and posterior trajectory generation, the data generation is improved in terms of pertinence and scene applicability; based on TTC, DTTC and scene classification, key frames are accurately extracted, invalid data is removed, and cleaning and screening are more efficient; the true value data is more consistent with the actual demand by comprehensively considering the trajectory and modal similarity; the parameter tuning efficiency is improved by using black box optimization and KL divergence, and the optimal configuration is quickly obtained. The present method has significant improvements in efficiency, accuracy and applicability, and can achieve higher performance in the fields of training data generation and algorithm optimization.

[0045] The above specific embodiments can be adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present application, the protection scope of the present application is subject to the claims and is not limited by the above specific embodiments, and each implementation scheme within the scope is subject to the constraints of the present application.

Claims

1. A data-driven autonomous driving decision optimization system, characterized in that, include: The system comprises a data production module, a data filtering module, a model encapsulation module, and a parameter tuning module. Specifically: the data production module takes driver and passenger data as input, and after annotation, preprocessing, and format conversion, extracts key features and performs standardization, normalization, and encoding to generate raw data that meets the algorithm's input requirements for the data-driven process; the data filtering module filters relevant data from the training data according to the scenarios defined by the decision-maker and performs effective classification; the model encapsulation module encapsulates the C++ decision code and uses a trajectory pair-based truth evaluation method to construct the trajectory pair evaluation cost map required for training; and the parameter tuning module, based on the filtered data from different scenarios, the encapsulated decision algorithm model, and the trajectory pair evaluation cost map, uses black-box optimization methods to obtain the decision parameters for the corresponding scenarios under the current decision algorithm. The data production module includes: a human-vehicle data unit, a mainline sampling code encapsulation unit, a sampling trajectory feature generation unit, and an optimal trajectory calibration unit; The data filtering module includes: a scene filtering unit, a valid frame filtering unit, a key obstacle filtering unit, and a problem category filtering unit; The model encapsulation module includes: a model input unit, a trajectory-to-white-box evaluation unit, and a sampling space evaluation distribution unit.

2. The data-driven autonomous driving decision optimization system according to claim 1, characterized in that, The aforementioned driver data is generated by experienced drivers driving vehicles and has undergone initial annotation. Each small dataset consists of a ten-minute driving segment, and each frame of data needs to contain comprehensive information about the vehicle's status and the surrounding environment. The vehicle status information includes: vehicle position, speed, acceleration, current lane ID, lane line information, lane association information, and current vehicle behavior information; The vehicle's current behavior information includes turning left, going straight, and passing through intersections; The surrounding environment information includes: vehicle ID, vehicle type, vehicle size, location, vehicle speed, and the vehicle's future trajectory predicted by the prediction module; The raw data in the data-driven process includes: the sampling trajectory information of the vehicle, the sampling trajectory information of other vehicles, the list of vehicle IDs to be played, the list of the status information of the vehicle, the list of the status information of other vehicles, the conflict point information, the posterior trajectory of the vehicle in the next 6 seconds, the posterior trajectory of other vehicles in the next 6 seconds, and the list of environmental information. The decision-making scenarios include straight-ahead, intersection, merging / merging out, and lane-changing scenarios.

3. The data-driven autonomous driving decision optimization system according to claim 1, characterized in that, The human-driven data unit collects and integrates a large amount of data generated by human drivers based on human driving data, including the vehicle's status and information about surrounding vehicles and the environment, in order to reflect driving style and behavioral patterns. The mainline sampling code encapsulation unit encapsulates the mainline sampling code and automatically generates trajectory sampling schemes that satisfy the rules based on the data sampling rules and objective function constraints, resulting in a preliminary set of sampled trajectories constrained by the rules. The sampled trajectory feature generation unit extracts trajectory sampling features based on the trajectory sampling results, including the trajectory points and interaction information of the vehicle and other vehicles, and generates trajectory feature vectors for analysis, obtaining sampled trajectory features for subsequent model encapsulation and evaluation. The optimal trajectory calibration unit performs multi-objective comprehensive evaluation on the sampled trajectories based on the trajectory features and objective function values, calibrates the optimal trajectory that satisfies the objective function conditions, and obtains the calibrated optimal trajectory for subsequent training and validation.

4. The data-driven autonomous driving decision optimization system according to claim 1, characterized in that, The scene filtering unit classifies the trajectory data into scenes based on the trajectory data and scene classification rules. It performs coarse classification into straight-through, intersection, and merging / exit scenes, and provides labels for specific scenes to obtain a classified scene dataset, which provides a basis for detailed analysis. The effective frame filtering unit filters key frames according to time-related indicators and removes redundant information to obtain a set of effective frames in high-risk scenes. The key obstacle screening unit filters out key obstacle information that has an important impact on decision-making based on trajectory data and environmental information, and obtains a subset of scene data containing key obstacle information; The problem category filtering unit filters scenario data according to the scenario characteristics and potential problem categories in the data, focusing on the solution of specific driving problems, and obtains a high-quality training dataset optimized for problem categories.

5. The data-driven autonomous driving decision optimization system according to claim 1, characterized in that, The model input unit, based on the trajectory feature vectors after data filtering and related parameter recommendation rules, inputs trajectory pair features to generate a recommended parameter set adapted to the model, thus providing support for the model's accuracy and applicability. The trajectory pair white-box evaluation parameter unit, based on the trajectory pair features and the objective function, uses a white-box model to calculate the evaluation parameters of the trajectory pair, thus obtaining the trajectory pair white-box evaluation parameters for accurate analysis of decision-making effectiveness. The model output unit, based on the white-box evaluation parameters and the sampled trajectory distribution, outputs a comprehensive evaluation distribution map of the sampled trajectory space, used to describe the performance of the trajectories under the objective function, thus obtaining the evaluation distribution of the sampled space and providing a basis for parameter tuning.

6. The data-driven autonomous driving decision optimization system according to claim 1, characterized in that, The parameter tuning module includes: a parameter black-box optimization input unit, a black-box optimization unit, and an iterative parameter recommendation unit. Specifically: the parameter black-box optimization input unit provides initial parameter black-box optimization input based on the model input parameters and training data feature vectors, serving as the starting condition for optimization, and obtains an initial parameter configuration to initiate the optimization process; the black-box optimization unit adjusts the parameters using a black-box optimization algorithm based on the parameter configuration and the optimization objective function, gradually approaching the optimal value to obtain the optimized parameter configuration result; the iterative parameter recommendation unit iteratively recommends parameter configurations based on the intermediate results and evaluation metrics from the black-box optimization output, adjusting parameter combinations to improve optimization efficiency, and obtains a set of recommended optimized parameters.

7. An autonomous driving decision optimization method based on the system described in any one of claims 1-6, characterized in that, The process involves reading raw data, invoking decision-making algorithms, setting rules to generate final data, and then performing data cleaning and data mining based on the obtained data. The decision algorithm is encapsulated and ground truth is constructed using a sampling-objective function-evaluation-screening decision algorithm framework. The optimal parameter combination values ​​that satisfy the objective function of the training data under the corresponding scenario are iteratively found using black-box optimization methods to obtain optimized autonomous driving decisions. The data mining process described above filters data based on the difference in Time To Call (TTC) between the vehicle and other vehicles reaching the conflict point. Specifically: ,in: This indicates the time difference between your vehicle and the other vehicle reaching the point of conflict. and These represent the time it takes for the vehicle itself and the other vehicle to reach the point of conflict, respectively. The conflict point refers to the spatial intersection of the posterior trajectory of the vehicle and the posterior trajectory of the other vehicle, which is obtained by calculating the collision detection between the vehicle and the other vehicle. The decision-making algorithm is encapsulated as follows: first, the future driving trajectories of the vehicle and other vehicles are sampled; then, the sampling results of the vehicle and other vehicles are comprehensively evaluated according to a pre-defined objective function to obtain the trajectory pair that satisfies the minimum objective function, thereby obtaining the yielding strategy between the vehicle and other vehicles. The objective function ,in: These respectively represent safety, comfort, accessibility, and right-of-way. This indicates the weight each entity occupies in the objective function; The aforementioned construction of ground truth refers to the decision ground truth required in the process of constructing training data. It determines more similar trajectory pairs and driver decision results by calculating the similarity of trajectory pairs and modal similarity. The similarity of the trajectory pairs is measured by the endpoint error fde, mean error ade, and lateral error lde between the sampled trajectory and the posterior trajectory, specifically: , , ,in: For Euclidean distance, The horizontal distance in the Frenet coordinate system; The modal similarity is calculated by obtaining the posterior yielding relationship based on the DTTC of the vehicle and other vehicles, and penalizing solutions whose sampling results are inconsistent with the posterior yielding relationship. This yields a truth evaluation calculation method based on a data and decision algorithm framework, specifically: Where: the modal difference value is bde, For hyperparameters; The parameter tuning process described above uses the KL divergence loss function in its iterative process, specifically as follows: ,in: If it is true, The value output by the decision algorithm.

Images