Data-driven intelligent car anthropomorphized decision planning system and method

Abstract

The application provides a system and a method, including natural driving data collection, multi-source heterogeneous data preprocessing, driving behavior scene mining, behavior feature screening and extraction, behavior style classification and advantage and disadvantage evaluation, personification model parameter identification and personification network model training. Through analysis of natural driving behavior of a driver, a personification decision planning model can be constructed, the personification degree of common ADAS functions associated with various behaviors is improved, the user experience is improved, and a sufficient trust relationship between a driver and an intelligent driving vehicle is established. The personification decision planning system established through analysis of the driving behavior of the driver can also improve the driving similarity between the intelligent driving system and the traditional driver, promote the harmonious interaction between the vehicle and the car in the heterogeneous traffic flow, reduce the accident rate, and improve the overall stability, so that the key technical bottleneck of intelligent driving is broken through, the high-level and personification intelligent driving system is quickly applied, and the development of intelligent networked vehicles is empowered.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving decision-making and planning, and in particular relates to a data-driven intelligent vehicle anthropomorphic decision-making and planning system and method. Background Technology

[0002] As the deployment rate of intelligent driving assistance systems increases year by year, heterogeneous traffic flows composed of intelligent vehicles and driver-driven vehicles will persist for a long time. Current intelligent driving systems mostly rely on single logic or rule constraints to output decision-making routes and control vehicles, without considering differences in driving styles. However, drivers have diverse driving styles and varying skill levels, resulting in significant differences in driving behavior in specific scenarios such as following other vehicles, changing lanes, and turning. In traditional traffic flows, drivers can reach a certain understanding through the expression and recognition of their subjective intentions, allowing for appropriate yielding and priority. However, for heterogeneous traffic flows, a single-style intelligent driving decision-making and planning system struggles to achieve consistency with drivers of various driving styles, leading to issues related to maneuverability, stability, and safety.

[0003] Decision-making and planning are central to intelligent driving. While many key technologies have been successfully implemented, bottlenecks and problems remain in certain areas requiring breakthroughs. These manifest as delayed decision switching, unsatisfactory planned trajectory quality, vehicle speeds exceeding driver and passenger expectations, and poor lane-changing and obstacle avoidance quality. This indicates that current intelligent driving decision-making and planning systems cannot yet meet the personalized needs of drivers and passengers. Furthermore, while current intelligent driving assistance systems, such as ACC (Adaptive Cruise Control), AEB (Automatic Emergency Braking), and LKA (Lane Keeping Assist), are rapidly being mass-produced and widely adopted, these systems or functions are largely geared towards marketing purposes. For drivers, there are still situations where they are unsure how to use them, hesitant to use them, or find them difficult to use. The main reasons for these problems are the poor applicability of current autonomous driving functions, their limited applicable scenarios, and the significant gap between current autonomous driving decision-making and control and their ability to meet driver and passenger expectations, failing to achieve human-like driving.

[0004] Currently, the lack of human-like decision-making and planning technologies is due to two main reasons: firstly, the limited accumulation of real-world driving behavior data for different drivers hinders the implementation of effective data-driven methods; secondly, the absence of a comprehensive closed-loop data-driven human-like decision-making and planning system and methodology. Therefore, there is an urgent need to construct a data-driven intelligent vehicle human-like decision-making and planning system and methodology to improve the human-like nature of intelligent driving decision-making and planning systems and overcome the technological bottlenecks in intelligent driving decision-making and planning. Summary of the Invention

[0005] In view of this, the present invention aims to propose a data-driven intelligent vehicle anthropomorphic decision-making and planning system and method. Using a data acquisition vehicle as the carrier, it achieves open road data acquisition and storage based on vehicle-side multi-source sensors and controllers, vehicle-side power supply system, vehicle-side data acquisition system, and vehicle-side large-capacity storage system. Based on multi-source heterogeneous data preprocessing modules, behavior scene mining modules, correlation feature filtering and extraction modules, behavior stylization classification and evaluation modules, anthropomorphic model parameter identification modules, and anthropomorphic network training modules, it achieves raw data preprocessing, behavior scene data mining, scene feature data extraction, behavior stylization and merits evaluation labeling, decision-making and planning model parameter identification, and anthropomorphic network model data training. This enables the construction of an anthropomorphic decision-making and planning system oriented towards driving behavior based on a data-driven approach, breaking through the limitations of intelligent driving decision-making and planning technology, and truly starting from human driving behavior habits and personalized needs to develop intelligent driving vehicles that better meet user driving requirements.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0007] The first aspect of this solution discloses a data-driven anthropomorphic decision-making and planning method for intelligent vehicles, including natural driving data collection, multi-source heterogeneous data preprocessing, driving behavior scenario mining, behavior feature screening and extraction, behavior style classification and evaluation of advantages and disadvantages, anthropomorphic model parameter identification and anthropomorphic network model training.

[0008] It is represented in the form of data streams, corresponding to raw data, preprocessed data, behavioral scene data, behavioral feature data, standardized labeled data, anthropomorphic rule data, and anthropomorphic model data, respectively.

[0009] The corresponding functional modules are: multi-source heterogeneous data preprocessing module, behavior scene mining module, association feature filtering and extraction module, behavior style classification and evaluation module, anthropomorphic model parameter identification module, and anthropomorphic network training module.

[0010] Furthermore, the multi-source heterogeneous data preprocessing module is used to preprocess the raw data collected from the vehicle, which includes whole vehicle data, target object data, lane line data, traffic sign data, vehicle posture data, text data, and high-definition video data from different perspectives.

[0011] Furthermore, the behavior scenario mining module is used to perform scenario mining on the generated preprocessed data, including common behavior scenarios such as following other cars, changing lanes, turning, making U-turns, passing through intersections, entering and exiting ramps, and passing through tunnels, as well as dangerous behavior scenarios such as descending steep slopes, avoiding obstacles at high speeds, turning at high speeds, and avoiding emergencies.

[0012] Furthermore, the associated feature filtering and extraction module is used to filter and extract the behavioral scene associated feature data obtained through mining.

[0013] Furthermore, the behavior stylistic classification and evaluation module is used to perform stylistic classification and quality assessment on the extracted behavioral feature data. Stylistic classification refers to clustering and labeling behavioral styles based on various unsupervised learning methods. Quality assessment is used to formulate evaluation criteria for various driving behaviors and label the quality of the behaviors.

[0014] Furthermore, the anthropomorphic model parameter identification module is used to identify the proxy parameters of the behavior rule model through various optimization algorithms. The rule model refers to a rule-based model containing a large number of proxy parameters. The anthropomorphic model parameter identification module can convert the standardized labeled data into anthropomorphic rule data.

[0015] Furthermore, the anthropomorphic network training module is used to input various standardized labeled data into the constructed neural network model, and to build an anthropomorphic network through model training, which can be used to directly predict specific behaviors.

[0016] Secondly, this solution discloses a data-driven intelligent vehicle anthropomorphic decision-making and planning system, including a data acquisition vehicle, vehicle-side multi-source sensors and sensor controllers, vehicle-side data acquisition system, vehicle-side large-capacity storage system, vehicle-side power supply system, multi-source heterogeneous data preprocessing module, behavior scene mining module, correlation feature screening and extraction module, behavior style classification and evaluation module, anthropomorphic model parameter identification module, anthropomorphic network training module, and various supporting software and hardware devices.

[0017] Furthermore, the data acquisition vehicle serves as the foundation, and is equipped with multi-source sensors and controllers, a vehicle-side data acquisition system, a vehicle-side large-capacity storage system, and a vehicle-side power supply system.

[0018] Sensor controllers include controllers for various types of sensors.

[0019] Furthermore, the vehicle-mounted multi-source sensors include functional cameras, LiDAR, millimeter-wave radar, GNSS inertial navigation equipment, and high-definition cameras;

[0020] The functional cameras include front-view and rear-view cameras, which can collect information on objects, traffic signs, and lane lines in front of and behind the test vehicle. Object information includes the type of object, the relative lateral and longitudinal distances between the object and the vehicle, and the relative lateral and longitudinal speeds. Traffic sign information includes the relative distances of speed limit signs, traffic lights, zebra crossings, and stop lines to the vehicle. Lane line information includes the type of lane line, the color of the lane line, and the distance between the vehicle and the lane line.

[0021] The lidar includes front left, front right, rear left, and rear right lidars, used to collect information on targets to the front and rear of the vehicle, including the relative lateral and longitudinal distances, relative lateral and longitudinal velocities, and relative lateral and longitudinal accelerations between the target and the vehicle.

[0022] Millimeter-wave radar includes left front, left rear, right front, and right rear millimeter-wave radars, which are used to collect information on targets in the front and rear sides, including the relative lateral and longitudinal distances, relative lateral and longitudinal speeds, and relative lateral and longitudinal accelerations between the target and the vehicle. The target information is fused and characterized by millimeter-wave radar, lidar, and functional cameras.

[0023] GNSS inertial navigation equipment is used to collect information on the test vehicle's positioning, heading angle, and road curvature.

[0024] The high-definition cameras include front-facing, rear-facing, side-facing, and in-vehicle driver cameras, which respectively capture synchronous high-definition video of the driver from the front, rear, side, and inside the vehicle.

[0025] The vehicle-side data acquisition system includes hardware and software. The hardware consists of a multi-interface high-performance industrial control computer and a high-definition display screen inside the vehicle for real-time monitoring. The software includes data acquisition software that can access various sensor signals. The acquisition software is used to calibrate various sensors, configure cameras, monitor and display signals, fuse multi-source data, and automatically save data.

[0026] The vehicle-side high-capacity data storage system includes NAS devices and switches deployed at the vehicle end. The NAS devices are connected to a multi-interface high-performance industrial control computer via the switches to store the collected data.

[0027] The vehicle-side power supply system consists of a power conversion module and an inverter, which is used to power various sensors and controllers in the vehicle, vehicle-side data acquisition system hardware, and vehicle-side large-capacity storage system. The inverter is connected to the vehicle battery and is used to power multi-interface high-performance industrial control computer, NAS equipment, and switch.

[0028] The power conversion module is connected to the vehicle's battery and is used to power various sensors and controllers.

[0029] Compared with existing technologies, the data-driven intelligent vehicle anthropomorphic decision-making and planning system and method described in this invention have the following advantages:

[0030] The data-driven intelligent vehicle anthropomorphic decision-making and planning system and method can analyze the driver's natural driving behavior to construct anthropomorphic decision-making and planning models, improve the anthropomorphism of commonly used ADAS functions related to various behaviors, enhance the user experience, and help drivers and passengers establish sufficient trust with intelligent driving vehicles. The anthropomorphic decision-making and planning system built through driver behavior analysis can also improve the similarity between intelligent driving systems and traditional drivers, promote harmonious vehicle-to-vehicle interaction under heterogeneous traffic flow, reduce the accident rate, improve overall stability, and ultimately break through the key technical bottlenecks of intelligent driving, promote the rapid deployment and application of high-level, anthropomorphic intelligent driving systems, and empower the development of intelligent connected vehicles. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0032] Figure 1 This is a diagram showing the positional relationships of various components on the vehicle side.

[0033] Figure 2 Data closed-loop driven anthropomorphic processing flowchart;

[0034] Figure 3 Here is a flowchart of the data processing for each module;

[0035] Figure 4 This is a schematic diagram illustrating a car-following behavior scenario.

[0036] Figure 5 This is a cluster distribution diagram of some typical indicator parameters.

[0037] Explanation of reference numerals in the attached figures:

[0038] 1-NAS device; 2-Switch; 3-Display screen; 4-Multi-interface high-performance industrial computer; 5-GNSS inertial navigation equipment; 6-Functional camera; 7-LiDAR; 8-High-definition camera; 9-Inverter; 10-Power conversion module; 11-Millimeter-wave radar. Detailed Implementation

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0040] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0041] A data-driven intelligent vehicle anthropomorphic decision-making and planning system mainly consists of a data acquisition vehicle, vehicle-side multi-source sensors and controllers, vehicle-side data acquisition system, vehicle-side large-capacity storage system, vehicle-side power supply system, multi-source heterogeneous data preprocessing module, behavior scene mining module, correlation feature screening and extraction module, behavior style classification and evaluation module, anthropomorphic model parameter identification module, anthropomorphic network training module, and various supporting software and hardware devices.

[0042] Among them, such as Figure 1 As shown, a data acquisition vehicle is used as the basis, and the vehicle is equipped with multi-source sensors and controllers, a vehicle-side data acquisition system, a vehicle-side large-capacity storage system, and a vehicle-side power supply system.

[0043] The vehicle-mounted multi-source sensors include a functional camera (6), a lidar (7), a millimeter-wave radar (11), a GNSS inertial navigation system (5), and a high-definition camera (8). The sensor controllers include controllers for each type of sensor.

[0044] Among them, the six functional cameras are divided into front-view and rear-view cameras, which can collect information on target objects, traffic signs, and lane lines in front of and behind the test vehicle. Target object information includes target type, relative lateral and longitudinal distances between the target and the vehicle, and relative lateral and longitudinal speeds; traffic sign information includes speed limit signs, traffic lights, zebra crossings, stop lines, and their relative distances from the vehicle; lane line information includes lane line type, lane line color, and the distance between the vehicle and the lane line.

[0045] The LiDAR 7 is divided into left front, right front, left rear, and right rear LiDARs, which are mainly used to collect information on targets in front of and behind the vehicle, including the relative lateral and longitudinal distances, relative lateral and longitudinal velocities, and relative lateral and longitudinal accelerations between the target and the vehicle.

[0046] The millimeter-wave radar 11 is divided into left front, left rear, right front, and right rear millimeter-wave radars. It mainly collects information on targets in the front and rear sides, including the relative lateral and longitudinal distances, relative lateral and longitudinal velocities, and relative lateral and longitudinal accelerations between the target and the vehicle. The target information is fused and characterized by the millimeter-wave radar 11, the lidar 7, and the functional camera 6.

[0047] The GNSS inertial navigation equipment 5 mainly collects information such as the test vehicle's positioning, heading angle, and road curvature.

[0048] The high-definition camera consists of eight cameras: a front-facing camera, a rear-facing camera, a side-facing camera, and a driver's camera inside the vehicle. These cameras simultaneously capture high-definition video from the front, rear, side, and driver's camera inside the vehicle.

[0049] The vehicle-side data acquisition system consists of hardware and software. The hardware comprises a multi-interface high-performance industrial control computer 4 and a high-definition display screen 3 inside the vehicle for real-time monitoring. The software is a data acquisition software capable of accessing various sensor signals. The software can perform functions such as sensor calibration, camera configuration, signal monitoring and display, multi-source data fusion, and automatic saving.

[0050] The vehicle-side high-capacity data storage system includes a NAS (Network Attached Storage) device 1 and a switch 2 deployed on the vehicle side. The NAS device 1 is connected to a multi-interface high-performance industrial control computer 4 through the switch 2 to store the collected data.

[0051] The vehicle-side power supply system consists of a power conversion module 10 and an inverter 9, used to power various sensors and controllers, vehicle-side data acquisition system hardware, and vehicle-side large-capacity storage system inside the vehicle. The inverter 9 is connected to the vehicle battery and is used to power the multi-interface high-performance industrial control computer 4, NAS device 1, and switch 2; the power conversion module 10 is connected to the vehicle battery and powers various sensors and controllers.

[0052] A data-driven, anthropomorphic decision-making and planning method for intelligent vehicles mainly includes natural driving data collection, multi-source heterogeneous data preprocessing, driving behavior scenario mining, behavioral feature selection and extraction, behavioral style classification and merit evaluation, anthropomorphic model parameter identification, and anthropomorphic network model training. This is represented in the form of data streams, such as... Figure 2 As shown, these correspond to raw data, preprocessed data, behavioral scene data, behavioral feature data, standardized labeled data, anthropomorphic rule data, and anthropomorphic model data, respectively. The corresponding functional modules are: multi-source heterogeneous data preprocessing module, behavioral scene mining module, correlation feature filtering and extraction module, behavioral stylistic classification and evaluation module, anthropomorphic model parameter identification module, and anthropomorphic network training module. The data processing flow for each module is as follows: Figure 3 As shown.

[0053] The multi-source heterogeneous data preprocessing module is used to preprocess the raw data collected from the vehicle. This raw data includes textual data such as vehicle data, target object data, lane line data, traffic sign data, and vehicle posture data, as well as high-definition video data from different perspectives. First, the textual data is preprocessed. Since the data acquisition frequencies and initial timestamps of different sensors vary, all raw textual data needs to be synchronized. After synchronization, the textual data needs to be cleaned, including outlier removal, missing data insertion, null value filling, and data filtering. The cleaned data is then merged into a single data table, completing the transformation from raw data to preprocessed data. For video data, timestamp synchronization with the textual data is required, and multiple video streams need to be stitched and merged for synchronized viewing with the textual data. This module primarily converts raw data into preprocessed data.

[0054] The behavior scene mining module is used to mine scenes from the generated preprocessed data, including but not limited to common behavior scenes such as following other cars, changing lanes, turning, making a U-turn, passing through intersections, entering and exiting ramps, and passing through tunnels, as well as dangerous behavior scenes such as descending steep slopes, high-speed obstacle avoidance, high-speed cornering, and emergency avoidance. This module mainly realizes the generation of behavior scene data from preprocessed data. The main steps include (1) defining the start and end points of the behavior; (2) defining process constraints; (3) writing an extraction program for batch mining; (4) manually labeling and verifying the results; and (5) optimizing the batch scene mining program.

[0055] (1) First, define the start and end points of the behavior. The start point refers to the moment when the vehicle information, target object information, and traffic environment information meet the triggering conditions of the behavior scenario. Among them, the vehicle information mainly focuses on the vehicle's motion state, including lateral and longitudinal vehicle speeds, yaw rate, and acceleration / deceleration. The constraints are as follows:

[0056]

[0057] In the formula, v min_x v ego_x v max_x These represent the minimum longitudinal speed of the vehicle, the actual longitudinal speed of the vehicle, and the maximum longitudinal speed of the vehicle that meet the trigger conditions of the behavior scenario, respectively; v min_y v ego_y v max_y ω represents the minimum lateral speed of the vehicle, the actual lateral speed of the vehicle, and the maximum lateral speed of the vehicle that meet the trigger conditions of the behavior scenario, respectively; min ω ego ω maxThese are, respectively, the minimum yaw rate of the vehicle, the actual yaw rate of the vehicle, and the maximum yaw rate of the vehicle that meet the triggering conditions of the behavior scenario; a min_x a ego_x a max_x These are, respectively, the minimum longitudinal acceleration of the vehicle that meets the triggering conditions of the behavior scenario, the actual longitudinal acceleration of the vehicle, and the maximum longitudinal acceleration of the vehicle that meets the triggering conditions of the behavior scenario; a min_y a ego_y a max_y These represent the minimum lateral acceleration of the vehicle, the actual lateral acceleration of the vehicle, and the maximum lateral acceleration of the vehicle when the behavior scenario trigger conditions are met, respectively.

[0058] Target information mainly focuses on the six primary targets (directly in front, directly behind, left front, left rear, right front, and right rear) after target tracking is completed. It is determined by the target type, the relative lateral and longitudinal velocities between the vehicle and each target, and the relative lateral and longitudinal distances.

[0059]

[0060] In the formula, obj type_i The type of the target object; v ri_x min v ri_x v ri_x max These represent the minimum, the actual measured, and the maximum relative longitudinal speeds between the vehicle and the target object, respectively, when the behavior scene trigger conditions are met; v ri_y min v ri_y v ri_y max These represent the minimum, measured, and maximum relative lateral velocities between the vehicle and the target object, respectively, when the behavior scene trigger conditions are met; s ri_x min s ri_x s ri_x max These represent the minimum, measured, and maximum relative longitudinal distances between the vehicle and the target object, respectively, when the behavioral scenario trigger conditions are met; s ri_y min s ri_y s ri_y max These represent the minimum relative lateral distance between the vehicle and the target object, the actual measured relative lateral distance between the vehicle and the target object, and the maximum relative lateral distance between the vehicle and the target object, respectively, when the behavior scene trigger conditions are met.

[0061] Environmental information primarily focuses on road traffic signs and lane markings. Constraints on road traffic signs include determining the type of sign (speed limit, traffic lights, or signs, etc.) and the relative lateral and longitudinal distances between the vehicle and the traffic sign. Constraints on lane markings mainly refer to lane marking types and the relative distance between the vehicle and the lane markings.

[0062]

[0063] In the formula, TS type Indicates the type of traffic sign; s ts_x min s ts_x s ts_x max These represent the minimum relative longitudinal distance between the vehicle and the traffic sign, the actual measured relative longitudinal distance between the vehicle and the traffic sign, and the maximum relative longitudinal distance between the vehicle and the traffic sign, respectively, when the behavior scenario trigger conditions are met; s ts_y min s ts_y s ts_y max These represent the minimum, measured, and maximum relative lateral distances between the vehicle and traffic signs, respectively, when the behavior scenario trigger conditions are met. type Indicates the type of lane markings; s ego_llmin s ego_ll s ego_llmax These represent the minimum distance between the vehicle and the left lane line, the actual measured distance between the vehicle and the left lane line, and the maximum distance between the vehicle and the left lane line, respectively, all meeting the trigger conditions for the behavior scenario. ego_rlmin s ego_rl s ego_rlmax These represent the minimum distance between the vehicle and the right lane line that meets the triggering conditions of the behavior scenario, the actual measured distance between the vehicle and the right lane line, and the maximum distance between the vehicle and the right lane line that meets the triggering conditions of the behavior scenario.

[0064] When the above conditions are met, it is defined as the starting point of the behavior scene. When the ending point of the behavior scene is defined in combination with the scene characteristics, it is defined as the end of the scene when one or more of the above conditions are not met.

[0065] (2) Define process constraints, which mainly include vehicle information constraints, target object information constraints, traffic environment information constraints, and duration constraints. Among them, the vehicle information constraints, target object information constraints, and traffic environment information constraints are consistent with the above constraints, while the duration constraint is for the duration of the behavior, that is, the total time from the start point to the end point of the behavior, which must meet the minimum time requirement.

[0066] t min ≤T

[0067] In the formula, t minT represents the minimum duration required for the behavior scenario to meet the requirements; T represents the actual duration of the behavior scenario.

[0068] (3) After completing the behavioral scenario constraints, program the behavior extraction algorithm. Programming tools include, but are not limited to, VS, Matlab, PyCharm, etc., and programming languages ​​include, but are not limited to, C, Python, etc. Batch mining is performed on the collected preprocessed data, and naming rules are defined. Each behavior segment has a timestamp and a name indicating the behavior's category.

[0069] (4) After extraction is completed, the mining results are verified by manual annotation and the accuracy index is output.

[0070] (5) Optimize the boundary constraints and thresholds of the above behavior extraction definition through the annotation results to improve the accuracy of behavior scene mining.

[0071] The associated feature filtering and extraction module is used to filter and extract the associated feature data of the behavior scene obtained through mining. After obtaining the behavior scene fragment data, it is necessary to perform detailed characterization of the vehicle information, relative positional relationship with surrounding objects, lane lines, traffic signs and motion state information under the behavior scene. Therefore, it is necessary to filter and extract the key feature vectors and parameters in the behavior scene data. The associated feature filtering and extraction module can be used to convert behavior scene data into behavior feature data. The main steps include (1) behavior scene classification; (2) feature vector filtering; (3) feature parameter acquisition; (4) parameter feature selection; (5) parameter feature extraction.

[0072] (1) Considering that vehicle motion is mainly divided into three categories, including lateral behavior, longitudinal movement and coupled movement, the specific driving behavior is first classified. If the focus of this type of behavior is lateral movement, then the lateral movement characteristics of the vehicle are mainly considered. If the focus of this type of behavior is longitudinal movement, then the longitudinal movement characteristics of the vehicle are mainly considered. If this type of behavior is classified as coupled movement, then the lateral and longitudinal coupled movement characteristics need to be comprehensively considered.

[0073] (2) If the behavior scenario is classified as lateral, then lateral motion features need to be selected, including the vehicle's lateral speed, lateral acceleration, relative distance to the lane line, lateral distance to the target object, and lateral relative speed to the target object. If it is longitudinal motion, then the longitudinal motion features of the vehicle should be mainly considered, including the vehicle's longitudinal speed, longitudinal acceleration, longitudinal distance to the target object, and longitudinal relative speed to the target object. If it is coupled motion, then the lateral and longitudinal correlation features should be comprehensively considered, including the vehicle's lateral and longitudinal speeds, lateral and longitudinal accelerations, lateral and longitudinal distances to the target object, lateral and longitudinal speeds to the target object, yaw rate, and distance to the lane line.

[0074] (3) After extracting the feature vector, obtain the feature parameters. The feature parameters include various statistics for the feature vector, including but not limited to minimum value, maximum value, standard deviation, variance, mean, median, starting value, ending value, values ​​at key time points of the behavior, and summation values ​​at each time point, etc. The main parameters are extracted as follows:

[0075]

[0076] (4) Perform feature selection on the acquired parameters. The selection methods include, but are not limited to, correlation analysis, univariate feature selection, variance method, etc. By using feature selection methods, various parameters with high correlation and small variance are eliminated to achieve feature dimensionality reduction.

[0077] (5) Perform feature extraction on the acquired parameters. The extraction methods include, but are not limited to, principal component analysis, independent component analysis, and linear discriminant analysis. Obtain the principal component factors associated with the behavior through various feature extraction methods to achieve feature dimensionality reduction.

[0078] The behavior stylization classification and evaluation module is used to perform stylization classification and quality assessment on the extracted behavior feature data. Stylization classification refers to clustering and labeling behavior styles based on various unsupervised learning methods. Quality assessment is used to formulate evaluation criteria for various driving behaviors and label the quality of behaviors. The behavior stylization classification and evaluation module can be used to convert behavior feature data into standardized labeled data. The main steps include (1) unsupervised clustering of behaviors; (2) stylization labeling of behaviors; (3) construction of behavior evaluation system; (4) determination of optimal parameter intervals; and (5) labeling of behavior quality.

[0079] (1) Since there are no prior stylistic labels for each behavior segment, driving style annotation is performed based on unsupervised learning clustering. General clustering methods include, but are not limited to, mean clustering, hierarchical clustering, and fuzzy clustering. To evaluate the behavior clustering results, clustering evaluation metrics are introduced. Since there are no mature external annotations for this sample set, an internal metric method is selected, with the minimum intra-class distance and the maximum inter-class distance as the evaluation objectives. Internal metrics are used as the main evaluation metrics, including but not limited to the silhouette coefficient.

[0080] (2) First, define the style categories. Generally, three gradients can be used: aggressive, conservative, and calm. Since the clustering method used is unsupervised learning, the generated clusters do not have clear label attributes. Therefore, it is necessary to analyze the specific parameters to clarify which driving style each cluster corresponds to. By drawing the statistical distribution of various feature parameters, the style labels can be mapped one-to-one with each cluster.

[0081] (3) Performance Evaluation: For various typical driving behaviors, safety, passability, comfort, and fuel efficiency are the main criteria. Safety indicators include, but are not limited to, time to collision (TTC) and headway (THW). Passability indicators include, but are not limited to, average speed, acceleration, and traffic flow. Comfort indicators include acceleration and dynamism. Fuel efficiency indicators include power and braking performance, such as acceleration, deceleration, number of rapid accelerations, and number of rapid decelerations. The weights of various criteria in the behavior are determined, and driving behaviors are divided into conventional driving behaviors and dangerous driving behaviors. Conventional driving behaviors include, but are not limited to, following other vehicles, changing lanes, turning, U-turns, and stop-and-go driving. The main consideration is driving style factors, with a slight reduction in the weight of safety criteria and a slight increase in the weight of comfort, passability, or fuel efficiency criteria. Dangerous driving behaviors include, but are not limited to, descending steep slopes, high-speed obstacle avoidance, emergency braking, and navigating sharp curves. The main consideration is safety factors, with a slight increase in the weight of safety criteria.

[0082] (4) Determine the optimal range for each type of indicator. After accumulating data over a period of time, draw statistical distribution charts of various indicator parameters and define the parameter range that conforms to traffic regulations, does not cause traffic accidents, and conforms to the behavior patterns of most drivers as the optimal range.

[0083] (5) By setting the indicator weights and the optimal range of the corresponding indicators in the evaluation system, the merits and demerits of each behavioral scenario segment are evaluated, and the merits and demerits labels are added to the specific behaviors based on the evaluation results.

[0084] The anthropomorphic model parameter identification module is used to identify the parameters of the behavioral rule model through various optimization algorithms. The rule model refers to a rule-based model containing a large number of parameters. The anthropomorphic model parameter identification module can convert the standard labeled data into anthropomorphic rule data. The main steps include (1) constructing the rule model; (2) parameter identification and optimization; (3) rule model evaluation; and (4) generating the anthropomorphic rule model.

[0085] (1) Combine specific behavioral characteristics to cite or construct rule models, including but not limited to semi-empirical physical models based on dynamic constraints or various mathematical models based on polynomials, spline curves, etc. The constructed rule model can characterize behavioral features, and the model contains some identification parameters, which need to be identified through data input.

[0086] (2) Considering that there are a large number of identification parameters in the model, the data with excellent behavior labels in the aforementioned standardized labeled data are input into the rule model constructed in (1) according to the corresponding stylistic classification. Various optimization identification algorithms are used to perform anthropomorphic optimization identification of the parameters. The optimization identification algorithms include, but are not limited to, genetic algorithm, particle swarm optimization algorithm, simulated annealing algorithm, etc.

[0087] (3) After completing the stylistic parameterization identification, anthropomorphic rule models with style classification attributes will be generated. For example, if the aforementioned categories are radical, conservative, and calm, then radical, conservative, and calm anthropomorphic rule models will be generated here. The identified anthropomorphic rule models are evaluated using various error functions as evaluation criteria. These error functions include, but are not limited to, mean absolute error (MAE) and mean squared error (MSE). After evaluation, the identification parameters are optimized based on the evaluation results.

[0088] (4) Define the accuracy threshold. When the identification result meets the defined accuracy threshold through evaluation, the anthropomorphic rule model is considered to be effective. This set of identification parameters can be used to construct and generate the corresponding anthropomorphic rule model.

[0089] The anthropomorphic network training module is used to input various types of standardized labeled data into the constructed neural network model. Through model training, an anthropomorphic network is built, which can be used to directly predict specific behaviors. The anthropomorphic network training module can convert standardized labeled data into anthropomorphic model data. The main steps include (1) building a network model; (2) model training; (3) model evaluation; and (4) generating an anthropomorphic network model.

[0090] (1) Construct a neural network model. Based on the behavioral characteristics, construct a neural network model framework to train the behavior. The network framework includes, but is not limited to, single-layer feedforward, multi-layer feedforward, RNN, CNN and their corresponding variant network structures. Initially define the network model parameters, including but not limited to the number of hidden layers, the number of neurons, batch size, and time step.

[0091] (2) Use high-quality standard data with stylized annotations as input to train the neural network model. Stop when the iteration reaches a certain number of times. According to the aforementioned definition of three types of driving styles, namely aggressive, conservative and calm, we can obtain an aggressive anthropomorphic network model, a conservative anthropomorphic network model and a calm anthropomorphic network model.

[0092] (3) Evaluate and optimize the model prediction results, input test set data of the corresponding style to test the model, evaluate it through various error functions, and further optimize and iterate the model based on the results.

[0093] (4) Define a certain accuracy threshold. When the prediction results of the network model can meet the defined accuracy threshold, it is considered that the anthropomorphic network model has a good effect and meets the final application requirements.

[0094] The supporting hardware and software equipment includes two parts: hardware and software. The software includes software that can realize the above-mentioned data analysis, processing and model training functions; the hardware includes various hardware devices that can support data processing and model training, including but not limited to high-performance computers, GPUs, cables, connectors, network cables, power adapters, etc.

[0095] The following section uses a car-following behavior scenario as an example to provide a detailed description of the data-driven intelligent vehicle anthropomorphic decision-making and planning system and method described in this invention:

[0096] After large-scale data collection from the vehicle is completed, the collected data is preprocessed uniformly through a multi-source heterogeneous data preprocessing module. This includes frequency synchronization of text and video data, outlier filtering, and missing data supplementation. Following data preprocessing, a behavior scene mining module extracts car-following behavior from the preprocessed data. The car-following behavior scene description is as follows: Figure 4 As shown, the constraints are defined as follows:

[0097]

[0098] In the formula: D ego W represents the width of the vehicle; L represents the width of the lane; l L r v represents the distance from the origin of the coordinate system to the left and right lane lines. ego X represents the vehicle's speed. n X represents the relative longitudinal distance between the vehicle and an object within the same lane. obj ,Y obj These represent the relative longitudinal distance and relative lateral distance to the target object being followed; D obj v is the width of the target object. obj X represents the speed of the target vehicle. max ,X min These represent the relative longitudinal distance thresholds between the vehicle and the target object; THW represents the headway; THW min The time distance threshold is defined as the distance to the front of the vehicle. When the above constraints are met, the start of car-following is defined, and segments with a duration greater than 20 seconds are defined as the desired car-following segments. After defining the constraints and start / end points, a batch processing program for car-following behavior scene mining was written in PyCharm using Python to extract data from the preprocessed data. The extracted data was then manually verified, and the constraints and defined thresholds were optimized.

[0099] After completing the mining and verification of car-following behavior scenarios, a large number of valid car-following behavior scenario fragments were obtained. Based on the correlation feature filtering and extraction module, key features of car-following behavior scenarios were extracted. Since car-following behavior is classified as longitudinal movement, six time-domain indicators were taken as feature vectors: longitudinal speed of the vehicle, longitudinal acceleration, longitudinal jerk, longitudinal distance from the car-following target, distance from the lane centerline, and relative longitudinal speed to the target. The standard deviation, mean, maximum, and minimum values ​​of each vector were taken as statistical feature parameters. Eight macroscopic mathematical statistical parameters were added: duration of behavior scenario fragments, percentage of acceleration time, percentage of constant speed time, percentage of deceleration time, percentage of time approaching the target, percentage of time moving away from the target, percentage of time relatively stable with the target, and type of car-following target. A total of 32 parameters were used as correlation feature parameters for car-following behavior.

[0100] To address the issues of numerous types and strong correlations among some parameters in 32-dimensional features, feature selection was first performed on all parameters. Using the variance method, the indicator of "minimum distance from lane centerline" was found to be almost entirely zero, indicating low variance, and therefore it was excluded. After feature parameter selection, feature extraction was performed using Principal Component Analysis (PCA). With a target contribution rate of 90%, the contribution rate of each component was calculated. The cumulative contribution of the first eight principal components exceeded 90%, so the first eight principal components were ultimately extracted, reducing the original sample set from 32 dimensions to 8 dimensions.

[0101] After extracting key features, the extracted car-following behavior segments are stylized and labeled for quality based on the behavioral stylization classification and evaluation module. The styles are defined as aggressive, conservative, and calm. Since each car-following segment lacks prior labels for driving style classification, unsupervised learning clustering is used for driving style labeling. Considering generality, the K-means unsupervised clustering method is chosen. K-means clustering is performed on the dimensionality-reduced sample set, initial iteration centers are set, and iteration records are observed. Clustering convergence is achieved, and the clustering results are good, as shown by calculating the silhouette coefficients of the entire sample. To determine which driving style each cluster corresponds to, specific parameters need to be analyzed. The distribution of some typical indicator parameters is obtained using mathematical statistics methods, such as... Figure 5As shown, the average relative speed between cluster 1 and the target object is mostly negative, and the asymptotic time with the target object accounts for more than 0.5% of the total time. The average relative speed between cluster 2 and the target object is mostly positive, and the asymptotic time with the target object accounts for less than 0.5% of the total time. The average relative speed between cluster 3 and the target object is mostly between -2m / s and 2m / s, relatively stable, and shows good following ability. Comprehensive analysis shows that cluster 1 is aggressive, cluster 2 is conservative, and cluster 3 is calm. After completing the stylized labeling, the merits and demerits of each following behavior segment are evaluated based on the following behavior evaluation criteria (safety, comfort, efficiency, and energy saving). Considering that following is a regular driving behavior, the weights of comfort, efficiency, and energy saving are increased, while the weight of safety is appropriately reduced. After defining the weights of each criterion, the optimal range of following-related parameters is determined. By statistically analyzing the collected segments, the optimal range distribution of parameters can be obtained. Based on this, combined with the weights of each criterion index, behavioral scene data with stylized labels and merits and demerits are finally obtained.

[0102] Excellent aggressive, calm, and conservative car-following behavior data obtained from unsupervised clustering were used as input to the theoretical model. The Gipps semi-empirical car-following model was selected, and parameter optimization was performed using a human-like parameter identification module. The Gipps model comprehensively considers acceleration / deceleration and collision avoidance distance constraints, and is a safe distance-based car-following model. Theoretically, it is expressed as follows: When the vehicle in front brakes suddenly, the following vehicle controls its speed to avoid a collision.

[0103]

[0104] In the formula: Δt is the driver's reaction time; a m ,b m These represent the driver's desired acceleration and deceleration; v m The speed expected by the driver; L is the driver's estimated deceleration of the vehicle ahead. n-1 V is the length of the vehicle in front; Δx(t) is the distance between the two vehicles at time t; v n (t) represents the velocity of the vehicle at time t; v n (t+△t) represents the predicted speed at time t+△t. Since there are many parameters to be identified in this rule-based model, a genetic algorithm (GA) is used to calibrate these parameters. The initial population is set to 100, and crossover and mutation operators are selected. The mutation rate and crossover rate are defined, and the root mean square percentage error (RMSPE) is introduced as the objective function. After multiple iterations, the error eventually stabilizes, and the population mean at this point is selected as the identification result. Aggressive, conservative, and calm following segments are taken from the test set, and the vehicle speed at the next moment is predicted based on the anthropomorphic following rule model and compared. When the comparison result meets the initially defined deviation threshold, the corresponding anthropomorphic following rule model is generated.

[0105] Considering car-following behavior as a time-series-based continuous behavior, an anthropomorphic data-driven model is constructed based on an anthropomorphic network training module. The model takes feature vectors corresponding to car-following segments with stylized labels and quality ratings as input, and the predicted motion state of the vehicle at the next moment as output. A learning model is built based on a Long Short-Term Memory (LSTM) network with a learning rate of 1e-3 and 20,000 iterations. The Mean Squared Error (MSE) is used as the error function for training. The iteration ends when the error is less than 5e-6, and the network model parameters are locked after training. The model is then tested using a test set. When the defined error requirement is met, the corresponding anthropomorphic car-following network model is generated.

[0106] A data-closed-loop driven anthropomorphic system and method can be used to generate anthropomorphic model data from raw collected data. The generated anthropomorphic model can be used for the anthropomorphic optimization and improvement of ACC (Adaptive Cruise Control) functions.

[0107] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0108] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the division of units described above is merely a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. The aforementioned units may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven intelligent vehicle anthropomorphic decision-making and planning system, characterized in that: It includes a data acquisition vehicle, vehicle-side multi-source sensors and sensor controllers, vehicle-side data acquisition system, vehicle-side large-capacity storage system, vehicle-side power supply system, multi-source heterogeneous data preprocessing module, behavior scene mining module, correlation feature filtering and extraction module, behavior style classification and evaluation module, anthropomorphic model parameter identification module, anthropomorphic network training module, and various supporting software and hardware equipment. This includes natural driving data collection, multi-source heterogeneous data preprocessing, driving behavior scenario mining, behavioral feature selection and extraction, behavioral style classification and evaluation of advantages and disadvantages, anthropomorphic model parameter identification, and anthropomorphic network model training. It is represented in the form of data streams, corresponding to raw data, preprocessed data, behavioral scene data, behavioral feature data, standardized labeled data, anthropomorphic rule data, and anthropomorphic model data, respectively. The corresponding functional modules are: multi-source heterogeneous data preprocessing module, behavior scene mining module, association feature filtering and extraction module, behavior style classification and evaluation module, anthropomorphic model parameter identification module, and anthropomorphic network training module. The behavior scene mining module is used to mine scenes from the generated preprocessed data, including common behavior scenes such as following other cars, changing lanes, turning, making U-turns, passing through intersections, entering and exiting ramps, and passing through tunnels, as well as dangerous behavior scenes such as going downhill, high-speed obstacle avoidance, high-speed cornering, and emergency avoidance. In the behavior scenario mining module, the execution methods include: First, define the start and end points of the behavior. The start point is defined as the moment when the vehicle information, target object information, and traffic environment information meet the triggering conditions of the behavior scenario. Among them, the vehicle information refers to the vehicle's motion state, including lateral and longitudinal vehicle speeds, yaw rate, and acceleration / deceleration. The constraints are as follows: In the formula, These represent the minimum longitudinal speed of the vehicle, the actual longitudinal speed of the vehicle, and the maximum longitudinal speed of the vehicle that meet the triggering conditions of the behavior scenario, respectively. These represent the minimum lateral speed of the vehicle, the actual lateral speed of the vehicle, and the maximum lateral speed of the vehicle that meet the triggering conditions of the behavior scenario, respectively. These are the minimum yaw rate of the vehicle that meets the triggering conditions of the behavior scenario, the actual yaw rate of the vehicle, and the maximum yaw rate of the vehicle that meets the triggering conditions of the behavior scenario, respectively. These are the minimum longitudinal acceleration of the vehicle that meets the triggering conditions of the behavior scenario, the actual longitudinal acceleration of the vehicle, and the maximum longitudinal acceleration of the vehicle that meets the triggering conditions of the behavior scenario, respectively. These represent the minimum lateral acceleration of the vehicle, the actual lateral acceleration of the vehicle, and the maximum lateral acceleration of the vehicle that meet the triggering conditions of the behavior scenario, respectively. The target information focuses on the six main targets after target tracking is completed, including those directly in front, directly behind, to the left front, to the left rear, to the right front, and to the right rear. This information is determined by the target type, the relative lateral and longitudinal velocities between the vehicle and each target, and the relative lateral and longitudinal distances. In the formula, The type of the target object; These represent the minimum relative longitudinal speed between the vehicle and the target object, the actual measured value of the relative longitudinal speed between the vehicle and the target object, and the maximum relative longitudinal speed between the vehicle and the target object, respectively, when the behavior scene trigger conditions are met. These represent the minimum relative lateral velocity between the vehicle and the target object, the actual measured value of the relative lateral velocity between the vehicle and the target object, and the maximum relative lateral velocity between the vehicle and the target object, respectively, when the behavior scene trigger conditions are met. These represent the minimum relative longitudinal distance between the vehicle and the target object, the actual measured relative longitudinal distance between the vehicle and the target object, and the maximum relative longitudinal distance between the vehicle and the target object, respectively, when the behavior scene triggering conditions are met. These represent the minimum relative lateral distance between the vehicle and the target object, the actual measured relative lateral distance between the vehicle and the target object, and the maximum relative lateral distance between the vehicle and the target object, respectively, when the behavior scene trigger conditions are met. Environmental information pertains to road traffic signs and lane markings, and constraints on road traffic signs include... The system identifies the vehicle's category and the relative lateral and longitudinal distances between the vehicle and traffic signs. For lane line constraints, it refers to the lane line type and the relative distance between the vehicle and the lane line. When the above conditions are met, it is defined as the starting point of the behavior scene. The ending point of the behavior scene, combined with the scene characteristics, is defined as the end of the scene when one or more of the above conditions are not met. Furthermore, after completing large-scale data collection from the vehicle, the collected data is uniformly preprocessed through a multi-source heterogeneous data preprocessing module. This includes ensuring synchronization of text and video data, filtering out outliers, and supplementing missing data. After data preprocessing, the preprocessed data is extracted using a behavior scene mining module to extract car-following behavior. The constraints for the car-following behavior scene description are defined as follows: In the formula: D ego W represents the width of the vehicle; L represents the width of the lane; l L r v represents the distance from the origin of the coordinate system to the left and right lane lines. ego X represents the vehicle's speed. n X represents the relative longitudinal distance between the vehicle and an object within the same lane. obj ,Y obj These represent the relative longitudinal distance and relative lateral distance to the target object being followed; D obj v is the width of the target object. obj X represents the speed of the target vehicle. max ,X min These represent the relative longitudinal distance thresholds between the vehicle and the target object; THW represents the headway; THW min The following car start is defined as the time distance threshold for the car front. When the above constraints are met, the following car start is defined. The segment with a duration of more than 20 seconds is set as the required following car segment. After the constraints and start and end points are defined, a batch processing program for following car behavior scene mining is written in PyCharm based on Python to extract the preprocessed data. The extracted data is manually verified and the constraints and defined thresholds are optimized.

2. The data-driven intelligent vehicle anthropomorphic decision-making and planning system according to claim 1, characterized in that: The data acquisition vehicle serves as the foundation, and is equipped with multi-source sensors and controllers, a vehicle-side data acquisition system, a vehicle-side large-capacity storage system, and a vehicle-side power supply system. A sensor controller includes the controller corresponding to the sensor.

3. The data-driven intelligent vehicle anthropomorphic decision-making and planning system according to claim 1, characterized in that: Vehicle-mounted multi-source sensors include functional cameras, LiDAR, millimeter-wave radar, GNSS inertial navigation equipment, and high-definition cameras; The functional cameras include front-view and rear-view cameras, which can collect information on objects, traffic signs, and lane lines in front of and behind the test vehicle. Object information includes the type of object, the relative lateral and longitudinal distances between the object and the vehicle, and the relative lateral and longitudinal speeds. Traffic sign information includes the relative distances of speed limit signs, traffic lights, zebra crossings, and stop lines to the vehicle. Lane line information includes the type of lane line, the color of the lane line, and the distance between the vehicle and the lane line. The lidar includes front left, front right, rear left, and rear right lidars, used to collect information on targets to the front and rear of the vehicle, including the relative lateral and longitudinal distances, relative lateral and longitudinal velocities, and relative lateral and longitudinal accelerations between the target and the vehicle. Millimeter-wave radar includes left front, left rear, right front, and right rear millimeter-wave radars, used to collect information on targets in the front and rear sides, including the relative lateral and longitudinal distances, relative lateral and longitudinal velocities, and relative lateral and longitudinal accelerations between the target and the vehicle. The target information is fused and characterized by millimeter-wave radar, lidar, and functional cameras. GNSS inertial navigation equipment is used to collect information on the test vehicle's positioning, heading angle, and road curvature. The high-definition cameras include front-facing, rear-facing, side-facing, and in-vehicle driver cameras, which respectively capture synchronous high-definition video of the driver from the front, rear, side, and inside the vehicle. The vehicle-side data acquisition system includes hardware and software. The hardware consists of a multi-interface high-performance industrial control computer and a high-definition display screen inside the vehicle for real-time monitoring. The software includes data acquisition software that can access signals from various sensors. The acquisition software is used to calibrate various sensors, configure cameras, monitor and display signals, fuse multi-source data, and automatically save data. The vehicle-side high-capacity data storage system includes NAS devices and switches deployed at the vehicle end. The NAS devices are connected to a multi-interface high-performance industrial control computer via the switches to store the collected data. The vehicle-side power supply system consists of a power conversion module and an inverter, which is used to power various sensors and controllers in the vehicle, vehicle-side data acquisition system hardware, and vehicle-side large-capacity storage system. The inverter is connected to the vehicle battery and is used to power multi-interface high-performance industrial control computers, NAS devices, and switches. The power conversion module is connected to the vehicle's battery and is used to power various sensors and controllers.

4. The data-driven intelligent vehicle anthropomorphic decision-making and planning system according to claim 1, characterized in that: The multi-source heterogeneous data preprocessing module is used to preprocess the raw data collected from the vehicle, which includes whole vehicle data, target object data, lane line data, traffic sign data, vehicle posture data, text data, and high-definition video data from different perspectives.

5. The data-driven intelligent vehicle anthropomorphic decision-making and planning system according to claim 1, characterized in that: The associated feature filtering and extraction module is used to filter and extract the associated feature data of the behavioral scenarios obtained through mining.

6. The data-driven intelligent vehicle anthropomorphic decision-making and planning system according to claim 1, characterized in that: The behavioral stylistic classification and evaluation module is used to perform stylistic classification and evaluation of the extracted behavioral feature data. Stylistic classification refers to the classification of behavioral styles based on various unsupervised learning methods. Clustering and labeling are performed, and the merits and demerits of various driving behaviors are evaluated by formulating evaluation criteria and labeling the merits and demerits of the behaviors.

7. The data-driven intelligent vehicle anthropomorphic decision-making and planning system according to claim 1, characterized in that: The anthropomorphic model parameter identification module is used to identify the parameters of the behavioral rule model through various optimization algorithms. The rule model refers to a rule-based model with a large number of parameters. The anthropomorphic model parameter identification module can convert the standardized labeled data into anthropomorphic rule data.

8. The data-driven intelligent vehicle anthropomorphic decision-making and planning system according to claim 1, characterized in that: The anthropomorphic network training module is used to input various standardized labeled data into the constructed neural network model. Through model training, an anthropomorphic network is built, which can be used to directly predict specific behaviors.

Images