Intelligent automobile sky-ground integrated driving data acquisition, processing and collaborative decision-making system

Abstract

The invention discloses a sky-ground integrated driving data acquisition, processing and collaborative decision-making system for an intelligent automobile. The sky-ground integrated driving data acquisition, processing and collaborative decision-making system comprises a motion database acquisition subsystem and a scene adaptive collaborative decision-making enhancement subsystem, in a database acquisition part, vehicle-mounted equipment combines a low-orbit satellite and a ground monitoring station to observe data, high-precision real-time acquisition of vehicle trajectory data is realized, the trajectory data is cleaned and clustered, and the data quality and availability are improved; in the collaborative decision enhancement part, on the basis of preprocessed data, a rule filtering framework, a data-driven scoring module and a multi-supervision regression loss training strategy are combined, and high-precision and diversified guidance of vehicle movement behaviors in different driving scenes is achieved. According to the invention, through collaborative optimization, high-quality trajectory data is collected and fully utilized as an enhanced sample, and the safety requirement of intelligent traffic is met.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation systems, and particularly to an integrated space-ground driving data acquisition, processing, and collaborative decision-making system for intelligent vehicles, covering technologies such as vehicle motion state monitoring, data preprocessing, trajectory data enhancement, and driving behavior guidance. Background Art

[0002] With the rapid development of intelligent transportation systems and autonomous driving technologies, accurately collecting a large amount of high-precision driving state data has become one of the important links in improving vehicle intelligence and traffic safety. Currently, traditional vehicle data collection mainly relies on global positioning and navigation systems and ground base station observations. However, these systems have certain limitations in high-precision positioning and real-time communication. Especially in complex traffic environments and high-dynamic scenarios, it is difficult to provide sufficiently accurate and reliable data support. In addition, ground monitoring strategies also have deficiencies in terms of coverage, anti-interference ability, and data update frequency, restricting their application in high-precision driving behavior analysis and guidance.

[0003] On the other hand, in terms of intelligent vehicle decision-making and driving behavior guidance, existing in-vehicle models usually rely on vehicle's own sensors to build a fast decision-making system. Since the accuracy and breadth of training data collection directly affect the training quality and performance of the model, traditional data collection methods are difficult to meet the requirements of intelligent transportation systems for high-precision, real-time, and diverse driving behaviors. The fast system decision-making of vehicles often requires more external information support in complex environments, and existing methods fail to fully combine the guidance of the slow system, resulting in insufficient safety and rationality in the decision-making process.

[0004] Therefore, there is an urgent need for a driving data collection method that can combine low-earth orbit satellite communication and high-precision positioning technology to improve the accuracy, breadth, and real-time performance of trajectory state data collection. At the same time, intelligent transportation systems also need to build a high-precision cloud slow decision-making system to realize the behavior decision-making guidance of the fast decision-making system of intelligent vehicles based on the cloud, so as to provide a safer and more reasonable driving behavior guidance. Summary of the Invention

[0005] In view of this, the present invention provides an integrated space-ground driving data acquisition, processing, and collaborative decision-making system for intelligent vehicles. By introducing low-earth orbit satellite communication and high-precision positioning technology, it solves the accuracy and real-time problems of traditional data collection methods. At the same time, through the collaboration of the motion behavior database and the cloud slow decision-making system, it guides the fast system decision-making process and improves the safety and rationality of decision-making. The present invention not only improves the quality and usability of trajectory data, but also significantly enhances the diversity and accuracy of driving behavior guidance, providing strong technical support for the development of intelligent transportation systems.

[0006] The present invention achieves the above technical objectives through the following technical means.

[0007] An integrated space-ground driving data acquisition, processing and collaborative decision-making system for intelligent vehicles, comprising:

[0008] A motion database acquisition subsystem based on high-precision positioning and communication of low-earth orbit satellites, comprising an in-vehicle device data acquisition module and a data preprocessing module. The in-vehicle device data acquisition module is used to obtain the real-time trajectory data of the intelligent vehicle, and the data preprocessing module cleans and clusters the trajectory data; the clustered trajectory data is stored in partitions to form a motion database and is constructed into a trajectory bag-of-words library;

[0009] A scenario-adaptive collaborative decision-making enhancement subsystem for behavior decision-making guidance, comprising a rule filtering framework, a data-driven scoring module, a multi-supervised regression loss training strategy and a decision-making guidance module; the rule filtering framework uses basic driving rules to perform scenario-adaptive screening on the motion trajectories of target vehicles in different scenarios in the motion database; the data-driven scoring module, on the basis of rule screening, adopts a data-driven soft scoring strategy to quantify the suitability of each candidate trajectory in the current scenario; the multi-supervised regression loss training strategy is used to incorporate enhanced trajectories into the training of the decision-making model; the decision-making guidance module is used to guide the fast system of the autonomous vehicle by the decision-making model.

[0010] In the above technical solution, the rule filtering framework includes a road map constraint, an adjacent agent constraint and a self-motion constraint, and the intersection of the trajectory sets that meet the above three constraints is used as the candidate trajectory set.

[0011] In the above technical solution, the road map constraint is: transform the trajectories in the trajectory bag-of-words library into the local coordinate system of the target agent, and filter out the trajectories that do not exceed the drivable area during driving. Combine the lane topology map to identify the set of candidate lanes that the target agent may reach in the future through the historical trajectory of the target agent, and eliminate the trajectories not within the set of candidate lanes.

[0012] In the above technical solution, the adjacent agent constraint is: considering the spatial interaction between the trajectory and other agents, designing a collision detection condition. At any time step, if the enhanced trajectory collides with the trajectories of other adjacent agents in the scenario, it is regarded as an unreasonable trajectory.

[0013] In the above technical solution, the self-motion constraint is to set constraint conditions for the differences in the initial velocity and heading angle observed by the target agent:

[0014] Δv = ||v v (0)-v init ||, Δθ = ||θ v (0)-θ init ||

[0015]

[0016] Among them, Δv represents the difference between the observed speed of the target agent and the initial speed of the enhanced trajectory, and Δθ represents the difference between the observed heading angle of the target agent and the initial heading angle of the enhanced trajectory. represents the set of trajectories that conform to the self-motion constraints, ν v (0) and θ v (0) are the initial speed and initial heading angle of the enhanced trajectory v, respectively, v init and θ init are the observed moment speed and heading angle of the target agent, respectively, ∈ v and ∈ θ are the tolerances of speed and heading angle, respectively.

[0017] In the above technical solution, the data-driven scoring module includes the construction and training of the data-driven scoring module. The construction of the data-driven scoring module includes a scene encoder and a score decoder. When training the data-driven scoring module, negative samples are considered, the candidate trajectory set is introduced into the sample set, and a scene-based behavior partitioning strategy is designed.

[0018] In the above technical solution, the scene encoder includes: using a multi-layer perceptron and an attention interaction mechanism to realize the encoding and interaction of scene elements, and using a multi-layer perceptron and a max-pooling operation to compress the time dimension, retaining the key spatio-temporal features of each trajectory, and realizing the encoding of the candidate trajectory set.

[0019] In the above technical solution, the score decoder adopts a Transformer-based structure. The target agent embedding compresses the time dimension through max-pooling and fuses the scene context information through the self-attention mechanism; the updated agent embedding is concatenated with the embedding features of each candidate trajectory, and the score of each candidate trajectory is output through a multi-layer perceptron.

[0020] In the above technical solution, the scene-based behavior partitioning strategy is specifically as follows:

[0021] Obtain the set of lane nodes passed by each candidate trajectory;

[0022] Judge the relationship of all candidate trajectory node sets; through set partitioning operation, the candidate trajectory set is divided into multiple non-overlapping and non-empty subsets, that is, the node sets with inclusion relationships are classified into the same behavior type, and each subset is assigned a category label;

[0023] After obtaining the category label, the candidate trajectories that belong to the same behavior label as the true trajectory of the target agent are used as negative samples for the training of the behavior data-driven scoring module.

[0024] In the above technical solution, during the training process of the decision model, the trajectory regression loss is calculated based on the loss function of the Laplace distribution:

[0025]

[0026] Wherein, represents the trajectory regression loss, y gt is the true trajectory, and respectively represent the trajectory and uncertainty output by the decision model for calculating the loss function, α e is the weight corresponding to each enhanced trajectory, N e represents the number of enhanced trajectories used for training, is the enhanced trajectory used for training, and are respectively the position and uncertainty vector output by the decision model, y is the true value of the trajectory used for calculating the loss function, T f is the future time step, y t is the true position at time t, and are respectively the trajectory and uncertainty with the minimum final distance error output by the decision model and corresponding to the enhanced trajectory.

[0027] The beneficial effects of the present invention are as follows:

[0028] (1) The present invention utilizes the low-orbit satellite communication and high-precision positioning system, significantly improving the positioning accuracy and data update frequency of the vehicle motion trajectory. The data fusion strategy with the in-vehicle state monitoring sensor ensures the precise positioning and high-frequency state monitoring of the vehicle in a complex traffic environment, overcoming the deficiencies of the traditional ground monitoring and positioning system in terms of coverage and anti-interference ability, and greatly enhancing the accuracy and real-time performance of trajectory data collection.

[0029] (2) By adopting the scenario-adaptive collaborative decision-making enhancement strategy, the present invention can, based on high-quality motion trajectory data, combine the rule filtering framework, data-driven scoring module, and multi-supervised regression loss training strategy to provide precise and reasonable decision guidance for driving behaviors. The rule filtering framework ensures that the trajectory data conforms to road constraints and vehicle dynamics conditions, and the data-driven scoring module optimizes decision support to ensure the safety and rationality of driving behaviors in various complex traffic situations. This strategy effectively improves the safety and reliability of the fast system decision-making.

[0030] (3) Through the collaborative optimization of the cloud slow decision-making system and the vehicle-mounted fast decision-making system in the present invention, it is ensured that the slow decision-making system can provide accurate behavior guidance for the fast decision-making system in different driving scenarios. The slow decision-making system not only enhances the diversity of decision-making results, but also improves the stability and adaptability of the decision-making process, enabling intelligent vehicles to make safer and more reasonable decisions in complex traffic environments, thereby optimizing the overall driving behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the intelligent vehicle heaven and earth integrated driving data acquisition, processing and collaborative decision-making system according to the disclosed embodiment of the present invention;

[0032] Figure 2 Schematic diagram of part of the process of driving data acquisition according to the disclosed embodiment of the present invention;

[0033] Figure 3 Schematic diagram of the process of the motion decision enhancement system according to the disclosed embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] In the following, exemplary embodiments of the present application will be described with reference to the accompanying drawings. For clarity and conciseness, not all features of the actual embodiments are described in the specification. However, it should be understood that many specific embodiment-specific decisions may be made during the development of any such actual embodiment to achieve the specific goals of the developer, and these decisions may vary with different embodiments. The protection scope of the present invention is not limited to the following embodiments.

[0035] It should be understood that the present application is not limited to the described implementation forms only due to the description with reference to the accompanying drawings. Where feasible, different embodiments can be combined, replaced, borrowed, and features in different embodiments can be omitted without changing the idea proposed by the present invention.

[0036] The fast decision-making system in the present invention is good at handling simple tasks and can efficiently handle the vast majority of autonomous driving decision-making tasks in conventional scenarios, such as following a vehicle, changing lanes, traffic light recognition, etc., to ensure the smoothness and real-time nature of driving. The slow decision-making system, through a deeper understanding and learning of the environment, combines historical data, logical reasoning, and complex analysis to form a more comprehensive decision-making ability, suitable for dealing with complex traffic scenarios such as emergencies, unstructured roads, unknown obstacles, etc. Although the computational efficiency is low, it can provide more accurate and safe decision-making guarantees. The fast and slow decision-making systems cooperate with each other to ensure that autonomous driving takes into account safety, real-time nature, and stability in different environments.

[0037] The present invention aims to provide an integrated space-ground driving data acquisition, processing, and collaborative decision-making system for intelligent vehicles. Through in-vehicle motion state monitoring devices and high-precision positioning and communication capabilities based on low-Earth orbit satellites, accurate, real, and scene-constrained driving trajectory data can be acquired. At the same time, a hybrid-scenario adaptive motion data enhancement is built based on a real motion database to guide the behavior decision-making of the intelligent vehicle fast system and improve decision-making security.

[0038] The first part of the present invention provides a motion database acquisition subsystem based on high-precision positioning and communication of low-Earth orbit satellites, including an in-vehicle device data acquisition module and a data preprocessing module.

[0039] The in-vehicle device data acquisition module includes, but is not limited to, an inertial measurement unit (IMU), a wheel speed sensor, and a high-precision positioning and navigation system connected to the data interfaces of low-Earth orbit satellites (LEO satellites) and ground monitoring stations.

[0040] The low-Earth orbit satellite and the ground monitoring station respectively provide satellite observables and navigation enhancement signals. Through the least squares or Kalman filtering algorithm, combined with error correction, the real-time high-precision positioning and navigation information of the vehicle is solved.

[0041] The in-vehicle device data acquisition module realizes the precise and high-frequency monitoring of the vehicle's real-time motion state by collecting information such as the vehicle's acceleration, angular velocity, and wheel speed in real time, and coordinating the error correction data of the low-Earth orbit satellite communication link and the ground monitoring station. Using the collected sensor (such as accelerometers, gyroscopes, wheel speed sensors, etc.) data, through optimization or filtering fusion algorithms, the real-time state estimation of key motion parameters such as the vehicle's attitude, position, speed, and acceleration is realized; by combining with high-precision positioning and navigation information, high-frequency (for example, 10Hz or higher frequency) state updates can be achieved, and real-time trajectory data of the intelligent vehicle can be obtained.

[0042] The data preprocessing module is used for cleaning and clustering of trajectory data, and adopts an adaptive algorithm for dynamic adjustment according to the special requirements under different regions and scene conditions.

[0043] Data cleaning includes screening and complementing operations; original trajectory data screening: first, the obtained original trajectory data is screened to remove outliers and noise. Through smoothing processing based on Kalman filtering and outlier detection algorithms, the error data caused by sensor errors or environmental factors is effectively removed; for the case of missing or interrupted data, interpolation algorithms or complementing methods based on adjacent data points are used to fill the trajectory data to ensure the integrity of the trajectory data.

[0044] Data clustering specifically involves: dividing and clustering the cleaned trajectory data according to multi-dimensional features such as the vehicle's motion behavior, driving environment, and speed pattern. Through the DBSCAN clustering algorithm, trajectories with similar features are classified and stored in the database for subsequent analysis and modeling. This process helps to discover the vehicle's motion patterns in different regions or scenarios and provides high-quality training data for the subsequent model.

[0045] Meanwhile, the system adopts a multi-level data structure storage technology to partition and store the clustered trajectory data according to different spatial resolutions and time scales, forming a motion database and constructing it into a trajectory bag-of-words library to support subsequent efficient querying and analysis. Through these data preprocessing steps, the quality and usability of the original trajectory data can be maximally improved.

[0046] The second part of the present invention provides a scenario-adaptive collaborative decision-making enhancer subsystem for behavior decision guidance, including a rule filtering framework, a data-driven scoring module, a multi-supervised regression loss training strategy, and a decision guidance module.

[0047] The rule filtering framework uses basic driving rules to adaptively screen the motion trajectories of target vehicles in different scenarios for the real motion database obtained by the first part of the present invention through low-earth orbit satellite communication and high-precision positioning capabilities, so as to strengthen the diversity of motion behaviors. It includes road map constraints, adjacent agent constraints, and self-motion constraints.

[0048] The road map constraint means that the enhanced trajectory of the target agent must follow the restrictions of the road topology and drivable area. Specifically: First, a lane topology graph representing road constraints is constructed In the lane topology graph, the lane centerline is used as a node, and the connection between lanes is represented by a directed edge; then, the trajectories in the trajectory bag-of-words library are mapped to the local coordinate system of the target agent, and the trajectories that do not exceed the drivable area during driving are screened out

[0049]

[0050] Among them, v represents the enhanced trajectory, represents the trajectory bag-of-words library, DriveableArea represents the trajectory bag-of-words library, t represents the time step, and v(t) represents the spatial coordinates of the enhanced trajectory at time t;

[0051] Finally, combined with the lane topology graph The set of candidate lanes that the target agent may reach in the future is identified through the historical trajectory of the target agent Trajectories not in the set of candidate lanes are excluded; through this screening process, the remaining trajectories are ensured to meet the road map constraints while retaining the diversity of the lane topology:

[0052]

[0053] Among them, represents the set of trajectories that conform to the road map constraints.

[0054] The adjacent agent constraints are specifically as follows: The interaction between agents significantly affects the rationality of the trajectory. It is necessary to consider the spatial interaction between the trajectory and other agents to avoid collisions. The collision detection conditions are as follows:

[0055]

[0056] Among them, Dist(·,·) is the Euclidean distance, and d min is the minimum safety distance threshold. represents the set of trajectories of all adjacent agents, and v i (t) represents the spatial coordinates of the enhanced trajectory of the i-th adjacent agent at time t, and v i represents the trajectory of the i-th adjacent agent;

[0057] Furthermore, by calculating the minimum Euclidean distance and time occupancy conflict between trajectories, trajectories with collision risks are filtered out. At any time step, if the enhanced trajectory collides with the trajectories of other adjacent agents in the scene, it is regarded as an unreasonable trajectory:

[0058]

[0059] Among them, represents the set of trajectories that conform to the adjacent agent constraints.

[0060] The self-motion constraints are specifically as follows: The trajectory of the target agent needs to satisfy the constraints of dynamics and the smoothness of motion to ensure physical feasibility. The enhanced trajectory must satisfy the limitations of speed, acceleration, and steering angle, and avoid sudden changes. Since the trajectories in the trajectory bag-of-words library are from reality, this limitation is implicitly satisfied.

[0061] Furthermore, it is necessary to ensure that the initial state of the enhanced trajectory is consistent with the observed state of the target agent to ensure the natural connection of the trajectory. Certain constraint conditions are set for the differences in the initial speed and heading angle observed for the target agent:

[0062] Δv = ||v v (0) - v init ||, Δθ = ||θ v (0) - θ init ||

[0063]

[0064] Among them, v v(0) and θ v (0) is the initial velocity and heading angle of the enhanced trajectory v, respectively, and v init and θ init are the velocity and heading angle of the target agent at the observation moment, respectively, and ∈ v and ∈ θ are the tolerances of velocity and heading angle, respectively. Δv represents the difference between the observed velocity of the target agent and the initial velocity of the enhanced trajectory, and Δθ represents the difference between the observed heading angle of the target agent and the initial heading angle of the enhanced trajectory. represents the set of trajectories that conform to the self-motion constraints.

[0065] The finally retained set of candidate trajectories is the intersection of the sets of trajectories that satisfy three constraint conditions:

[0066]

[0067] Based on the rule filtering framework, hierarchical filtering of the trajectory bag-of-words library can be realized, and the diversity and rationality of the enhanced trajectories of the target agent in the current scenario can be improved.

[0068] The data-driven scoring module is used to further adopt a data-driven soft scoring strategy on the basis of rule screening to quantify the suitability of each candidate trajectory in the current scenario. It includes the construction and training of the data-driven scoring module.

[0069] The construction of the data-driven scoring module includes a scene encoder and a score decoder.

[0070] The scene encoder is used to extract the historical trajectory embedding features and scene context information of the target agent, and uses a multi-layer perceptron and an attention interaction mechanism to realize the encoding and interaction of scene elements:

[0071]

[0072] where X is the observation state of the target agent at the historical time step T h within, M is the surrounding road map information, is the encoded feature of the road map polygon, is the encoded feature of the agent at each moment, N represents the number of agents, M represents the number of road map elements, and D represents the feature dimension.

[0073] At the same time, for the encoding of the candidate trajectory set, a multi-layer perceptron and a max pooling operation are used to compress the time dimension and retain the key spatio-temporal features of each trajectory:

[0074]

[0075] where Mp represents the max pooling operation, and φa is a multi - layer perceptron, is the spatio - temporal embedding feature of the candidate trajectory set, is the number of candidate trajectories.

[0076] The evaluation decoder adopts a Transformer - based structure; taking the nth agent as an example, the target agent embedding will compress the time dimension through max - pooling and fuse the scene context information through the self - attention mechanism:

[0077]

[0078] a n ←MSA(q,k,v=[a1,…,a N ,s1,…,s M )

[0079] where MSA(q,k,v) is the multi - head self - attention module, and q, k, v represent the query, key, and value parameters respectively, and T h represents the historical time.

[0080] The updated agent embedding a n will be concatenated with the embedding features of each candidate trajectory, and the scores of each candidate trajectory will be output through a multi - layer perceptron. These scores reflect the relevance and feasibility of the trajectory in the current context:

[0081]

[0082]

[0083] where φ score is the multi - layer perceptron for scoring, and r n,k represents the normalized scoring result of the k - th candidate trajectory in and represents the spatio - temporal embedding feature of the k - th candidate trajectory of

[0084] The training of the data - driven scoring module is used to solve the problem of missing negative samples during the training of the data - driven scoring module. The future trajectory recorded by the target agent in the real trajectory data is used as the positive sample S + , and there is a lack of diverse negative samples S - .

[0085] To strengthen the acquisition of training negative samples, the candidate trajectory set is introduced into the sample set, and a scene - based behavior partitioning strategy is designed to avoid classification bias caused by unbalanced sample behaviors.

[0086] Furthermore, the scenario-based behavior partitioning strategy means that in the current scenario, the candidate trajectory set will be grouped according to the road topology structure to identify different driving behaviors under a specific road topology, specifically including:

[0087] First, obtain the set of lane nodes passed by each candidate trajectory, which reflects the road topology route it follows;

[0088] Secondly, judge the relationship of all candidate trajectory node sets; through the set partitioning operation (SetPartitioning), the candidate trajectory set is divided into multiple non-overlapping and non-empty subsets, that is, the node sets with inclusion relationships are grouped into the same behavior type, and each subset will be assigned a class label c e :

[0089]

[0090] Among them, represents that the two sets belong to the inclusion relationship, is an empty set, v l represents any candidate trajectory, and f represents the label.

[0091] After obtaining the class label, the candidate trajectories that belong to the same behavior label as the true trajectory of the target agent are used as negative samples S - , rather than all candidate trajectories, for the training of the behavior data-driven scoring module. This can improve the model's ability to judge the optimal trajectory under similar behaviors and reduce the impact of high-frequency behaviors on low-frequency behaviors.

[0092] Furthermore, the cross-entropy function is used as the training loss function to complete the training of the data-driven scoring module:

[0093]

[0094] Among them, N S represents the total number of trajectories in the training sample, and c k represents the k-th trajectory in the training sample.

[0095] The multi-supervised regression loss training strategy is used to effectively integrate the enhanced trajectories into the training of the decision-making model and improve the model's ability of diverse and high-precision decision guidance. The present invention designs a query-based decoder to introduce the multi-supervised regression loss training strategy into the enhanced training of the decision-making model, specifically as follows:

[0096] Each query vector independently extracts information from the scene context through the attention mechanism in each iteration and uses the self-attention mechanism for intra-modal interaction to ensure that the model can effectively capture the interactions between different driving behaviors.

[0097] The updated queries will be used to decode the positions and uncertainties of the multimodal trajectories; according to the minimum mean displacement error criterion, the queries with the least error are simultaneously found for the ground truth trajectory and the enhanced trajectory from the generated trajectories; the selected queries will account for the regression loss and participate in the backpropagation optimization to promote the update of the model parameters.

[0098] To avoid the occurrence of the "mode collapse" phenomenon, that is, all query vectors tend to generate the same trajectory mode, during the loss calculation process, the original ground truth trajectory in the scene will be given priority to ensure that the real trajectory dominates the training process, and each query vector corresponds to only one unique trajectory result, thus avoiding the training instability caused by inconsistent generated trajectory patterns.

[0099] Furthermore, a loss function based on the Laplace distribution is used to calculate the trajectory regression loss to realize the training of the decision-making model:

[0100]

[0101] where Laplace is the probability density function of the Laplace distribution, α e is the weight corresponding to each enhanced trajectory, is the enhanced trajectory for training, y gt is the ground truth trajectory, and are the position and uncertainty vectors output by the decision-making model respectively, N e represents the number of enhanced trajectories for training, T f represents the future time step, y t represents the real position at time t, and represent the position and uncertainty vectors at time t.

[0102] The decision guidance module is specifically used for the decision-making model to guide the fast system of the autonomous vehicle, including:

[0103] Scene adaptive collaborative decision generation: Based on high-quality motion trajectory data, the slow decision-making system analyzes different driving scenarios and generates multiple candidate decision paths; through the data-driven scoring module, the suitability and safety of each candidate decision path are evaluated, and the most suitable decision result is selected.

[0104] Decision guidance transmission to the fast system: The slow decision-making system transmits the selected decision-making scheme to the in-vehicle fast system. During this process, the fast system makes a quick response based on the real-time sensor data and traffic environment changes; however, the guidance of the slow decision-making system will continue to be used as a reference to ensure that the fast system can maintain a high decision-making safety and rationality during the execution process.

[0105] Decision feedback and model update: During actual driving, the real-time decisions of the fast system generate feedback data, which are analyzed by the slow decision-making system. Then, the decision-making model is iteratively optimized based on the latest data. Through this collaborative mechanism, the system continuously improves the decision-making process to cope with the changing traffic environment.

[0106] Adaptive optimization and improvement of decision stability: In multiple iterations and decision feedbacks, the slow decision-making system adaptively adjusts its decision rules and strategies. Combining with the execution of the in-vehicle fast system, it gradually improves the decision stability and intelligence level of the overall system. Through this process, the slow system can effectively guide the fast system to ensure the efficiency and safety of intelligent driving.

[0107] The above-mentioned functions of the decision guidance module are all prior arts, and the decision-making model can be an existing neural network-based architecture.

[0108] The described embodiments are the preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Without departing from the essence of the present invention, any obvious improvements, substitutions or variations that those skilled in the art can make fall within the protection scope of the present invention.

Claims

1. An integrated space-earth intelligent vehicle driving data acquisition, processing and collaborative decision-making system, characterized in that, Including: A motion database acquisition subsystem based on high-precision positioning and communication of low-earth orbit satellites, including a vehicle-mounted device data acquisition module and a data preprocessing module. The vehicle-mounted device data acquisition module is used to obtain the real-time trajectory data of intelligent vehicles, and the data preprocessing module cleans and clusters the trajectory data; the clustered trajectory data is stored in partitions to form a motion database and is constructed into a trajectory bag-of-words library; A scene adaptive collaborative decision-making enhancement subsystem for behavior decision-making guidance, including a rule filtering framework, a data-driven scoring module, a multi-supervised regression loss training strategy, and a decision-making guidance module; the rule filtering framework uses basic driving rules to perform scene adaptive screening on the motion trajectories of target vehicles in different scenes by the motion database; The data-driven scoring module, on the basis of rule screening, adopts a data-driven soft scoring strategy to quantify the suitability of each candidate trajectory in the current scene; the multi-supervised regression loss training strategy is used to incorporate enhanced trajectories into the training of the decision-making model; The decision-making guidance module is used to guide the fast system of autonomous vehicles by the decision-making model.

2. The integrated space-ground driving data acquisition, processing and collaborative decision-making system for intelligent vehicles according to claim 1, wherein The rule filtering framework includes road map constraints, adjacent agent constraints, and self-motion constraints, and the intersection of the trajectory sets that meet the above three constraints is used as the candidate trajectory set.

3. The integrated space-earth intelligent vehicle driving data acquisition, processing and collaborative decision-making system according to claim 2, wherein The road map constraint is: transform the trajectories in the trajectory bag-of-words library into the local coordinate system of the target agent, and filter out the trajectories that do not exceed the drivable area during driving. Combining with the lane topology map, identify the candidate lane sets that the target agent may reach in the future through the historical trajectory of the target agent, and eliminate the trajectories that are not within the candidate lane sets.

4. The integrated space-earth intelligent vehicle driving data acquisition, processing and collaborative decision-making system according to claim 2, characterized in that The adjacent agent constraint is: considering the spatial interaction between the trajectory and other agents, design a collision detection condition. At any time step, if the enhanced trajectory collides with the trajectories of other adjacent agents in the scene, it is regarded as an unreasonable trajectory.

5. The integrated space-earth intelligent vehicle driving data acquisition, processing and collaborative decision-making system according to claim 2, wherein The self-motion constraint is: set constraint conditions for the differences in the initial speed and heading angle observed by the target agent: Δv = ||v v (0) - v init ||, Δθ = ||θ v (0) - θ init || where, Δν represents the difference between the observed velocity of the target agent and the initial velocity of the enhanced trajectory, and Δθ represents the difference between the observed heading angle of the target agent and the initial heading angle of the enhanced trajectory. represents the set of trajectories that comply with its own motion constraints, ν v (0) and θ v (0) are the initial velocity and the initial heading angle of the enhanced trajectory v, respectively, v init and θ init are the velocity and the heading angle of the target agent at the observation moment, respectively, ∈ v and ∈ θ are the tolerances of the velocity and the heading angle, respectively.

6. The integrated space-earth intelligent vehicle driving data acquisition, processing and collaborative decision-making system according to claim 2, wherein The data-driven scoring module includes the construction of the data-driven scoring module and the training of the data-driven scoring module. The construction of the data-driven scoring module includes a scene encoder and a score decoder. When training the data-driven scoring module, negative samples are considered, the candidate trajectory set is introduced into the sample set, and a scene-based behavior division strategy is designed.

7. The integrated space-ground driving data acquisition, processing and collaborative decision-making system for intelligent vehicles according to claim 6, wherein The scene encoder includes: using a multi-layer perceptron and an attention interaction mechanism to realize the encoding and interaction of scene elements, and using a multi-layer perceptron and a max-pooling operation to compress the time dimension, retaining the key spatio-temporal features of each trajectory, and realizing the encoding of the candidate trajectory set.

8. The integrated space-earth intelligent vehicle driving data acquisition, processing and collaborative decision-making system according to claim 6, wherein The score decoder adopts a Transformer-based structure. The target agent embedding compresses the time dimension through max-pooling and fuses the scene context information through the self-attention mechanism; the updated agent embedding is concatenated with the embedding features of each candidate trajectory, and the score of each candidate trajectory is output through a multi-layer perceptron.

9. The integrated space-earth intelligent vehicle driving data acquisition, processing and collaborative decision-making system according to claim 6, characterized in that, The scene-based behavior division strategy is specifically: Obtain the set of lane nodes passed by each candidate trajectory; Judge the relationship of all candidate trajectory node sets; through the set partitioning operation, partition the candidate trajectory set into multiple non-overlapping and non-empty subsets, that is, the node sets with inclusion relationships are grouped into the same behavior type, and each subset is assigned a category label; After obtaining the category label, use the candidate trajectories that belong to the same behavior label as the true trajectory of the target agent as negative samples for the training of the behavior data-driven scoring module.

10. The integrated space-ground driving data acquisition, processing and collaborative decision-making system for intelligent vehicles according to claim 1, characterized in that, During the training process of the decision-making model, calculate the trajectory regression loss based on the loss function of the Laplace distribution: Among them, represents the trajectory regression loss, and y gt is the true trajectory, and respectively represent the trajectory and uncertainty output by the decision model for calculating the loss function. α e is the weight corresponding to each augmented trajectory, and N e represents the number of augmented trajectories used for training. is the augmented trajectory used for training, and respectively represent the position and uncertainty vector output by the decision model. y is the true value of the trajectory used for calculating the loss function, and T f is the future time step, and y t is the true position at time t. and respectively represent the trajectory and uncertainty with the smallest final distance error output by the decision model and corresponding to the augmented trajectory.

Images