Intelligent vehicle sensor fusion perception and decision-making method and system based on low earth orbit satellite communication and high-precision positioning

Abstract

Disclosed are an intelligent vehicle sensor fusion perception and decision-making method and system based on Low Earth Orbit satellite communication and high-precision positioning, wherein sensing data of an intelligent vehicle is collected by using Low Earth Orbit satellite communication and high-precision positioning technology, and a cloud infrastructure is used to perform spatiotemporal consistency processing comprising latency compensation, semantic information matching, and multi-observation joint state estimation, so as to improve perception accuracy in complex scenarios, and reduce hazardous decision-making behaviors and high takeover rates of intelligent vehicles caused by perception deficits. Next, a behavior guidance strategy is constructed on the basis of trajectory data from a large number of vehicles, so as to provide a macro-level decision-making basis for an intelligent vehicle fleet. A multi-agent reinforcement learning algorithm allows for the intelligent vehicle fleet to perform ego-vehicle optimal decision-making, thereby achieving efficient traffic flow management and fleet-wide collaborative optimization. The present invention provides a brand-new technical solution for dynamic traffic flow management in complex traffic scenarios, which not only improves traffic safety and efficiency, but also provides important support for intelligent traffic systems and autonomous driving applications.

Description

A method and system for intelligent vehicle fusion perception and decision-making based on low-orbit satellite communication and high-precision positioning Technical Field

[0001] This invention relates to the field of intelligent transportation systems, and in particular to a method and system for realizing the fusion of perception information and collaborative decision-making of intelligent vehicles in complex traffic scenarios by utilizing low-orbit satellite communication and high-precision positioning technology. Background Technology

[0002] With the increasing complexity and intelligence of urban traffic, traditional traffic management systems face enormous challenges, especially in dynamic traffic environments. The rapid development of Intelligent Transportation Systems (ITS) and autonomous driving technologies has driven the intelligent upgrading of traffic management. Improving traffic flow efficiency, enhancing traffic safety, and optimizing traffic decisions have become current research hotspots in the transportation field. Intelligent vehicles, as a core component of autonomous driving systems, possess enormous potential in real-time perception, decision-making, and collaboration. However, due to the complexity and dynamic changes in traffic environments, the perception and decision-making capabilities of intelligent vehicles still face many technical challenges.

[0003] The rapid development of low-Earth orbit (LEO) satellite communication technology in recent years has provided intelligent vehicles with an efficient, stable, and high-precision positioning and communication solution. Compared with traditional positioning and navigation systems, LEO satellite systems have advantages such as low latency, high bandwidth, and high-precision positioning, enabling intelligent vehicles to provide more accurate positioning data and real-time communication support. This offers new possibilities for collaborative decision-making and multi-source information fusion among intelligent vehicle groups. However, how to effectively utilize LEO satellite communication and positioning technology to achieve efficient collaboration among intelligent vehicles in complex traffic scenarios remains a pressing problem to be solved.

[0004] Therefore, how to leverage low-Earth orbit satellite communication and high-precision positioning technologies, combine various heterogeneous sensing information, and achieve collaborative decision-making among intelligent vehicle groups through intelligent algorithms has become a key technology for improving traffic flow efficiency, optimizing traffic safety, and enhancing the performance of intelligent transportation systems. This invention proposes a fusion perception and decision-making method for intelligent vehicles based on low-Earth orbit satellite communication and high-precision positioning, aiming to provide a new systematic perception and decision-making solution for intelligent vehicles in complex traffic scenarios, thereby effectively improving the overall efficiency and safety of intelligent transportation systems. Summary of the Invention

[0005] This invention aims to provide a method and system for intelligent vehicle fusion perception and decision-making based on low-orbit satellite communication and high-precision positioning. This system can achieve the fusion and optimized decision-making of multi-source perception information in complex traffic scenarios, thereby improving overall traffic efficiency. The method mainly includes the following two key steps:

[0006] Fusion of heterogeneous sensing information: By integrating sensing data from multiple heterogeneous intelligent vehicles, a comprehensive understanding of traffic scenarios can be achieved.

[0007] Collaborative Decision-Making and Game Theory: Based on scenario understanding, optimize the game decision-making between intelligent vehicles and non-intelligent vehicles, while realizing collaborative decision-making within the intelligent vehicle system, and ultimately build an intelligent traffic management system.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] S1 The first aspect of this invention provides a method for fusion and optimization of sensing results based on low-Earth orbit satellite communication and high-precision positioning capabilities, comprising two parts: information acquisition and cloud information processing.

[0010] The information acquisition section described in S11 is used to collect and transmit environmental perception data from heterogeneous intelligent vehicles. In complex traffic scenarios, a communication link between the cloud and each intelligent vehicle is established using low-orbit satellites. Addressing the heterogeneous differences in sensor configurations and algorithms among different intelligent vehicles, this invention employs a post-fusion approach to integrate the perceived information.

[0011] The information collection includes the following two parts: sensor data collection and data encoding and transmission:

[0012] The perception data acquisition includes, but is not limited to, perception data of heterogeneous intelligent vehicles (such as adjacent vehicle status, 3D detection boxes, semantic information, etc.); and high-precision vehicle status information (such as position, heading angle, speed, acceleration, etc.) calculated by multiple sets of low-orbit satellite propagation signals and ground monitoring station signals.

[0013] The data encoding and transmission process involves compressing and encoding the collected data, then uploading it to a cloud server for storage and processing via the communication capabilities of low-Earth orbit satellites. The transmission process employs efficient encoding methods (such as entropy-based compression algorithms) to reduce bandwidth while ensuring real-time performance and data integrity.

[0014] The cloud information processing part described in S12 is used to receive perception status information collected and released by heterogeneous intelligent vehicles in the regulatory scenario through low-orbit satellite communication in the cloud server, and to achieve target-level fusion based on high-precision positioning, so as to achieve more accurate scene understanding in highly uncertain traffic scenarios.

[0015] The cloud processing component primarily employs strategies including time delay compensation, semantic association matching, and multi-observation joint state trajectory estimation.

[0016] The time delay compensation, based on the multi-source perception results of heterogeneous intelligent vehicles, performs delay compensation for all observed targets to be estimated. The specific steps are as follows:

[0017] Motion prediction model construction: Based on the motion state (position, velocity, acceleration, etc.) of the target object, a motion prediction model for each vehicle is constructed, such as using a uniform speed model or a uniform acceleration model.

[0018] Delay time calculation: Based on the data transmission delay and the sensing link delay, calculate the possible motion state delay Δt of the target object;

[0019] State prediction: Using a motion prediction model, the current accurate position of the target object is predicted to obtain the compensated state variables.

[0020] The semantic association matching, in the context of an observation scenario, involves matching targets within the overlapping areas of the observation ranges of multiple intelligent vehicles with sensing capabilities. It mainly comprises two parts: basic matching based on intersection-union ratio (IU) and semantic feature-enhanced matching, as detailed below:

[0021] The basic matching based on cross-union ratio (CUNR) utilizes the high-precision positioning capability of low-orbit satellites. Based on the spatial positioning of the sensing carrier, it performs three-dimensional spatial projection on the sensing target and uses the cross-union ratio matching strategy to achieve basic spatial matching, taking the driving scenario as a unit, and constructs the preliminary matching results of heterogeneous intelligent vehicle perception.

[0022] The semantic feature-enhanced matching utilizes high-level semantic information to achieve high-order joint matching. This includes target association feature extraction, matching algorithm design, and dynamic weight adjustment, as detailed below:

[0023] The target association feature extraction combines the vehicle's location information and semantic information, including color, vehicle type, etc., to construct a joint feature vector F = [P, S], where P represents spatial location features and S represents semantic labels;

[0024] The matching algorithm is designed to use either a location-based nearest neighbor matching algorithm or a semantically information-based weighted matching method to match the perception data of each vehicle with pre-labeled information in the map. The matching degree calculation formula can be expressed as:

[0025] S 匹配 =αIoU+βS 相似度

[0026] Where α and β are weight parameters, IoU is the aforementioned basic matching score based on intersection-union ratio, and S 相似度 The matching similarity score is obtained by encoding and decoding the joint feature vector basis F using a neural network.

[0027]

[0028] Where f ω For neural network models, The semantic joint feature vectors of N vehicles to be observed are collected.

[0029] The dynamic weight adjustment adjusts the weight of each information source according to environmental conditions (such as lighting and weather). For example, color information is more reliable and is given a larger weight when the lighting is good; location information is more reliable and its weight is adjusted accordingly when the weather is bad.

[0030] The multi-observation joint state estimation, based on delay compensation and semantic matching strategies, introduces a multi-observation joint state estimation strategy that integrates perception data from multiple intelligent vehicles and other observation sources to form a unified state estimate of targets in traffic scenarios. Delay compensation unifies the spatiotemporal reference, and a target association algorithm based on semantic information matching optimizes the observation consistency of the state estimate. Furthermore, using Kalman filtering or unscented Kalman filtering methods, all compensated and matched observation data are fused into the unified state estimate of the target.

[0031] This method improves the accuracy and stability of global perception by integrating multi-dimensional information such as spatiotemporal consistency, semantic information, and motion state of the observation data. This joint approach is particularly effective in environments with high observation noise, as it can dynamically adjust observation weights and prioritize more reliable sources, thereby improving the confidence of the estimation results. Ultimately, it achieves the output of a joint temporally and spatially consistent state sequence of dynamic and static obstacles on a scene-by-scene basis.

[0032] S2. A second aspect of the present invention provides a group behavior guidance and intelligent vehicle decision-making strategy based on a unified scene perspective of heterogeneous multi-intelligent vehicle fusion perception. This mainly includes group behavior guidance from a unified scene perspective and intelligent vehicle system decision-making based on group behavior guidance, as detailed below:

[0033] S21 describes a unified scenario perspective for guiding group behavior. This invention, based on the joint perception and processing results of traffic management areas in the cloud, collects rich traffic movement status data, realizes cloud-based movement strategy aggregation, and improves trajectory prediction at the joint scenario level for traffic participants, thereby achieving joint cognition of their behavior and constructing a unified scenario perspective framework for guiding group behavior. Through comprehensive perception and reasonable inference, the specific steps are as follows:

[0034] Data collection: Utilizing cloud-based sensing results, we extensively collect driving data involving various types of vehicles, including but not limited to vehicle driving trajectories, speed changes, and steering behavior.

[0035] Data preprocessing: The collected data is cleaned, denoised, and standardized to the same time base. A dynamic delay compensation strategy is used to correct for delays caused by network transmission, ensuring the spatiotemporal consistency of all data.

[0036] Prediction Model Construction and Training: A scene-level trajectory prediction model based on deep learning is adopted. Inputs include historical state variables (e.g., location, speed), semantic information (e.g., vehicle type, color), and scene context (e.g., road network structure, traffic signals). Techniques such as Graph Neural Networks (GNN), Long Short-Term Memory Networks (LSTM), or attention mechanisms are used to model the dynamic and static interactions between multiple vehicles and map elements in the scene. Supervised learning methods are employed to minimize the error between the predicted trajectory and the actual trajectory, and the probability distribution of the prediction results is output.

[0037] Predictive model enhancement: This invention implements three types of predictive model enhancement strategies, mainly targeting different driving scenarios, different types of traffic participants, and different driving conditions, as detailed below:

[0038] The method involves using monitoring data collected in the cloud to train specific prediction patterns for different driving scenarios, thereby enhancing the model's adaptability to specific intersections or road sections.

[0039] The approach targets different types of traffic participants: the motion characteristics and semantic information of different types of vehicles are integrated into the model, and the trajectory distribution of different types of vehicles (such as buses, private cars, construction vehicles, etc.) is inferred through joint modeling.

[0040] The method addresses different driving conditions by collecting and analyzing driving data under various weather conditions (such as sunny, rainy, and snowy days), optimizing model parameters, and ensuring that the model can accurately predict the trajectory of non-intelligent vehicles under various environmental conditions.

[0041] S22 describes intelligent vehicle collaborative decision-making based on group behavior guidance. Building upon scenario-level predictive behavior guidance, this invention achieves collaborative decision-making for a group of intelligent vehicles through low-Earth orbit satellite communication. This group collaborative decision-making not only focuses on individual optimality but also improves overall traffic efficiency through collaborative cooperation, based on a full understanding of the behavior of non-intelligent vehicle groups. It mainly includes collaborative decision-making strategy formulation, optimization and real-time adjustment of the impact of non-intelligent groups on decisions, as detailed below:

[0042] The collaborative decision-making strategy is formulated as follows: The cloud receives real-time location information and target positions of intelligent vehicles, combined with predicted trajectories of non-intelligent vehicles; a multi-agent reinforcement learning (MARL) algorithm is used to formulate the collaborative decision-making strategy. Within the MARL framework, each intelligent vehicle acts as an agent, interacting and collaborating with other intelligent vehicles based on its own objectives (such as safe and efficient traffic flow) and environmental perception; by sharing information and negotiating priorities, they jointly optimize traffic flow. For example, in an intersection scenario, intelligent vehicles negotiate and determine the passage order to avoid conflicts. The strategy optimization process can be represented as:

[0043]

[0044] Where π is the strategy, γ t Let be the discount factor, R be the reward function, and s be the discount factor. t Let a be a state variable. t Here, T represents the motion quantity, and T represents the trajectory time step. This represents the expected average return that the strategy can achieve over the long term.

[0045] The impact of non-intelligent vehicles in the game-theoretic decision-making process: During collaborative decision-making, intelligent vehicles fully consider the predicted trajectories of non-intelligent vehicles and adjust their own decision-making strategies to achieve safe interaction. For example, if a non-intelligent vehicle is predicted to suddenly change lanes, the intelligent vehicle will slow down or adjust its route in advance to ensure safe passage. Specifically, path adjustment can be achieved through dynamic programming methods.

[0046] The decision optimization and real-time adjustment utilize real-time communication and high-precision positioning via low-Earth orbit satellites. This allows intelligent vehicle swarms to acquire real-time traffic information and actual driving performance, dynamically optimizing decision-making strategies. In emergencies such as traffic accidents and road closures, intelligent vehicles can quickly adjust their strategies to adapt to the new environment. For example, they can replan their routes or change traffic flow. Simultaneously, during training, reinforcement learning algorithms (such as Q-learning and policy gradient algorithms) are used to dynamically optimize the policy network parameters, making future decisions more adaptable to actual traffic conditions. The optimization goal is to maximize overall traffic efficiency and safety.

[0047] Based on the above method, the present invention also proposes an intelligent vehicle fusion perception and decision-making system based on low-orbit satellite communication and high-precision positioning, including a perception part and a decision-making part. The perception part can realize the fusion of heterogeneous perception information in S1 above; the decision-making part can realize collaborative decision-making and game playing in S2 above.

[0048] The beneficial effects of this invention are:

[0049] (1) This invention combines high-precision positioning of low-orbit satellites with a fusion strategy of perception information from heterogeneous intelligent vehicles, which can effectively improve the accuracy and stability of global perception in complex traffic scenarios. Through real-time and accurate data fusion, reliable scenario-based data support is provided for collaborative decision-making of intelligent vehicles, reducing decision-making biases caused by perception errors, thereby significantly improving the safety and efficiency of the traffic system.

[0050] (2) By constructing a scenario-based trajectory prediction model based on deep learning and performing personalized optimizations for different driving scenarios, traffic participant types, and diverse weather conditions, the model becomes more aligned with actual needs. This optimization strategy effectively improves the reliability of vehicle behavior prediction, enhances the vehicle's interaction capabilities in dynamic traffic environments, and further improves the decision-making accuracy and responsiveness of intelligent vehicles in complex traffic scenarios.

[0051] (3) The intelligent vehicle collaborative decision-making mechanism based on the multi-agent reinforcement learning (MARL) algorithm can fully consider the behavioral characteristics of various traffic participants and the overall traffic efficiency. Through the collaborative cooperation among multiple agents, traffic flow management is optimized, overall traffic efficiency is improved, traffic congestion and accident risks are reduced, and the traffic system is promoted to develop in a more efficient and intelligent direction.

[0052] (4) This invention constructs a game and cooperation system between intelligent and non-intelligent vehicles through real-time collaborative interaction between the cloud and intelligent vehicles. This system provides an innovative technical means for dynamic traffic flow management in complex traffic scenarios, helping to more accurately predict and regulate traffic flow and ensure rapid response in emergencies or complex traffic situations. It provides important technical support for the comprehensive promotion of intelligent transportation systems and the implementation of autonomous driving technology, and has broad application prospects and commercial value. Attached Figure Description

[0053] Figure 1 is a schematic diagram of the key steps in an embodiment of the present invention;

[0054] Figure 2 is a flowchart of the perception result fusion and process optimization based on low-orbit satellite communication and high-precision positioning capabilities in an embodiment of the present invention.

[0055] Figure 3 is a flowchart of the group behavior guidance and intelligent vehicle collaborative decision-making strategy based on a unified scene perspective of multi-intelligent vehicle fusion perception according to an embodiment of the present invention. Detailed Implementation

[0056] The invention will now be further described with reference to the accompanying drawings.

[0057] As shown in Figure 1, which illustrates the steps of an embodiment of the present invention, the present invention includes two aspects:

[0058] S1 In a first aspect, the present invention provides a method for post-fusion and process optimization of sensing results based on low-Earth orbit satellite communication and high-precision positioning capabilities, as shown in Figure 2, and the specific content is as follows:

[0059] S11 Information Acquisition: In complex traffic scenarios, this invention establishes a communication link between the cloud and intelligent vehicles via low-Earth orbit (LEO) satellites. Due to differences in sensor configurations and autonomous driving algorithms among different intelligent vehicles, a post-fusion technique is employed to integrate the perception information of each vehicle to achieve comprehensive and accurate scene perception. Specifically, each intelligent vehicle collects perception data of the current scene, including state parameters of neighboring vehicles (such as speed, acceleration, and heading angle), 3D bounding boxes (identifying the spatial position and outline of objects), and semantic information (such as vehicle type, color, and road signs). Simultaneously, the vehicle receives multiple sets of LEO satellite signals and, combined with a high-precision positioning algorithm, calculates its own high-precision positioning information (such as position coordinates, heading angle, speed, and acceleration). This state data, after compression and encoding, is rapidly uploaded to the cloud via LEO satellite communication technology, supporting subsequent data fusion processing.

[0060] Taking an urban intersection as an example, multiple heterogeneous intelligent vehicles drive at the intersection, each equipped with various sensors such as LiDAR, cameras, and millimeter-wave radar, continuously detecting the status and semantic information of surrounding non-intelligent vehicles. Simultaneously, the vehicles use low-Earth orbit (LEO) satellite signals and ground station signals to calculate their own positioning, speed, and other status information. After encoding and processing, this sensing data and positioning information are transmitted in real time to the cloud via LEO satellite links.

[0061] S12 Cloud Information Processing: After receiving perception information uploaded by multiple intelligent vehicles, the cloud fuses target-level data based on high-precision positioning, thereby improving the accuracy of understanding complex traffic scenarios. This process employs a series of key strategies, including time delay compensation, semantic association matching, and multi-observation joint state trajectory, as detailed below:

[0062] Time Delay Compensation: Given the dynamic nature of traffic scenarios and the latency characteristics of sensing information transmission, delay compensation is required for the sensing results of each intelligent vehicle. A motion prediction model is constructed based on the target object's motion state, such as position, velocity, and acceleration. This model, combined with the transmission time and the inter-system delay, accurately predicts the target object's motion state at the current moment to offset the delay error. Let the target vehicle's position at time t0 be (x0, y0), and its velocity be... acceleration is After a delay of Δt, at the cloud reception time t1=t0+Δt, taking the uniform acceleration kinematics formula as an example, its predicted position (x1,y1) can be obtained from the following formula:

[0063]

[0064] Speed ​​prediction value for:

[0065]

[0066] This predictive compensation effectively improves the timeliness and accuracy of perceived information.

[0067] Semantic matching: In multi-vehicle perception scenarios, accurate target matching is required for overlapping areas of the observation range. Utilizing the high-precision positioning capabilities of low-Earth orbit satellites, combined with traditional multi-sensor fusion matching algorithms (such as the Intersection over Union (IOU) matching strategy), initial target association can be achieved. Building upon this, an enhanced matching strategy based on semantic information is incorporated, combining extracted semantic information such as color and vehicle type with vehicle positioning data to construct a joint feature containing spatial location and semantic labels. Through location-based nearest neighbor matching or semantic information-based weighted matching methods, the perception data of each vehicle is matched with pre-labeled map information. For the same observed target, multi-source information such as color, type, and location is comprehensively considered, and weights are dynamically assigned based on the credibility of each information source for fusion. The calculation method for the joint matching score is as follows:

[0068] By combining the vehicle's location information and semantic information, including color and vehicle type, a joint feature vector F = [P, S] is constructed, where P represents spatial location features and S represents semantic labels.

[0069] The matching algorithm employs either a location-based nearest neighbor matching algorithm or a semantically information-based weighted matching method to match the perceived data of each vehicle with pre-labeled information in the map. The matching degree calculation formula can be expressed as:

[0070] S 匹配 =αIoU+βS 相似度

[0071] Where α and β are weight parameters, IoU is the aforementioned basic matching score based on intersection-union ratio, and S 相似度 The matching similarity score is obtained by encoding and decoding the joint feature vector basis F using a neural network.

[0072]

[0073] Where f ω For neural network models, Let N be the semantic joint feature vectors collected from the N vehicles to be observed.

[0074] Multi-observation Joint State Estimation: Based on delay compensation and semantic matching strategies, a multi-observation joint state estimation strategy is introduced, fusing perception data from multiple intelligent vehicles and other observation sources to form a unified state estimate of the target in the traffic scene. Delay compensation provides a unified spatiotemporal reference, and semantic matching optimizes observation consistency. On this basis, Kalman filtering or unscented Kalman filtering methods are used to integrate the compensated and matched observation data into the unified state estimate of the target. Taking intelligent transportation scenarios as an example, Kalman filtering plays a core role in multi-observation joint state estimation, effectively fusing multi-source observation data and providing accurate system state estimates. For example, in an intelligent transportation scenario, assuming n intelligent vehicles observe a target vehicle, the system state vector x... k It contains state information such as the target vehicle's position, velocity, and acceleration, with a dimension of m. Observation vector z k,i Let p be the observation of the target vehicle by the i-th intelligent vehicle. State transition matrix F k It describes the transition relationship of the system state between adjacent time steps. For example, in a uniform motion model, if the state vector x k =[x,y,v x ,v y ] T (x, y are position coordinates, v) x ,v y (where the velocity component is), then:

[0075]

[0076] Where Δt is the time interval, representing the change in position based on velocity, while the velocity remains constant within the time interval Δt. Observation matrix H k,i Mapping the system state to the observation space; for example, if the observations are only the position information of the target vehicle, then:

[0077]

[0078] Process noise covariance Q k The uncertainty of the system model itself is characterized by factors such as road surface bumps and minor fluctuations in the power system during vehicle operation, leading to uncertainties in state changes; the observation noise covariance R0 k,i It reflects errors in the observation process, such as observation deviations caused by sensor accuracy limitations and environmental interference (such as the effects of light, rain, and snow on cameras and radar).

[0079] The prediction steps of Kalman filtering:

[0080]

[0081] here It is a prediction of the current state based on the state estimate of the previous time step. It is the covariance matrix of the predicted state, reflecting the uncertainty of the prediction, P k-1 The vector represents the state vector at time k-1, with the subscripts k and k-1 indicating time k and time k-1, respectively.

[0082] Update steps:

[0083]

[0084] Among them, R k,i To observe the noise covariance, K k,i The Kalman gain is a weighted average of predicted and observed values, dynamically adjusted based on the predicted and observed covariances. For observation sources of varying reliability, the covariance differs, and the Kalman gain changes accordingly, affecting the final state estimate. The system favors more reliable information. For example, when a smart vehicle's sensors have high accuracy (small observation noise covariance), its observations have a greater weight in the fusion process.

[0085] Compared to single-source estimation, multi-observation joint state estimation has significant advantages. Single-source estimation relies solely on data from a single vehicle or sensor, limited by its local perspective and inherent errors. For example, a single camera may miss the target or misjudge its position and velocity under direct sunlight. In contrast, multi-observation joint state estimation integrates information from multiple vehicles and sensors, utilizing the dynamic weighting mechanism of Kalman filtering to combine the advantages of each, suppress noise, and improve the accuracy and stability of state estimation. This provides reliable global state information for intelligent vehicle decision-making, significantly enhancing the operational efficiency of intelligent transportation systems in complex scenarios.

[0086] The second aspect of S2 provides a unified scenario perspective for guiding group behavior and collaborative decision-making strategies for intelligent vehicles based on the fusion perception of multiple intelligent vehicles, as shown in Figure 3. The specific details are as follows:

[0087] S21 Unified Scene Perspective for Group Behavior Guidance. After achieving perception fusion, the cloud further utilizes these perception results to collect data and guide the group behavior of various traffic participants, aiming to build an intelligent traffic management system from a unified scene perspective. First, the cloud extensively collects driving data involving various vehicles from real traffic scenarios, including information such as speed, acceleration, driving trajectory, and turn signal signals. After preprocessing, this raw data is unified to the same time base, and a dynamic latency compensation strategy is used to correct the latency caused by network transmission, ensuring that all data has spatiotemporal consistency and providing a high-quality data foundation for subsequent model training.

[0088] Based on preprocessed data, a deep learning-based scene-level trajectory prediction model is constructed. The model's input includes multi-dimensional information, covering vehicle historical states (such as speed and position changes over a past period), semantic information (such as vehicle type, color, and purpose), and scene context (such as road network structure, traffic signal status, and road slope / curvature). In terms of model architecture, this embodiment employs Graph Neural Networks (GNNs), sequence networks (such as Long Short-Term Memory networks (LSTMs), and attention mechanisms to model the interaction between vehicle historical states and multimodal elements in the traffic scene. For example, GNNs can abstract the traffic scene into a graph structure, where nodes represent key vehicle and road nodes, and edges represent the connections between them. Through feature transfer and updates of nodes and edges, the mutual influence between vehicles is accurately captured. LSTMs, leveraging their advantages in time-series data processing, memorize historical trajectory features and predict future trajectory trends. The attention mechanism focuses on vehicles or scene areas that have a key impact on trajectory prediction, thereby improving prediction accuracy. This model outputs the probability distribution of joint predicted trajectories and multimodal predicted trajectories for all non-intelligent vehicles, providing a basis for intelligent vehicle group game-theoretic decision-making and collaborative planning.

[0089] To enable the predictive model to adapt to complex and ever-changing traffic environments, the powerful communication and interaction capabilities of low-Earth orbit satellites are leveraged to optimize and enhance it for different driving scenarios, types of traffic participants, and driving conditions.

[0090] In different driving scenarios, the cloud collects non-intelligent vehicle driving data from different monitored intersections (such as urban main road intersections, suburban curves, highway ramps, etc.), and conducts targeted model training for the characteristics of each scenario to ensure that the model output can fit the unique traffic flow characteristics and driving trajectory patterns of each intersection.

[0091] For different types of traffic participants, the model integrates motion characteristics and semantic information. For example, considering the characteristics of buses that frequently stop at stations and have slow starts, and the characteristics of engineering vehicles that have relatively fixed routes and limited speeds, the model accurately predicts the trajectory distribution of specific types of vehicles by jointly modeling location and vehicle type.

[0092] The cloud collects driving behavior data under various weather conditions, such as sunny, rainy, snowy, and foggy days, to verify and test the model, ensuring that it can still accurately predict the trajectory of non-intelligent vehicles under complex weather conditions.

[0093] S22 Intelligent Vehicle Collaborative Decision-Making Based on Group Behavior Guidance. Building upon scenario-level predictive behavior guidance for vehicles in specific driving scenarios, this invention aims to achieve collaborative decision-making for all intelligent vehicle groups linked via low-Earth orbit satellite communication. The decision-making of intelligent vehicles should not be limited to individual optimality but should achieve overall traffic efficiency through collaborative cooperation, based on a full understanding of the behavior of other vehicle groups. Specific strategies are as follows:

[0094] Collaborative Decision-Making Strategy Formulation: Based on scenario-level trajectory prediction, this invention enables collaborative decision-making among all intelligent vehicles in the current monitoring scenario via the cloud, thereby overcoming the limitations of individual optimality and maximizing overall traffic efficiency. The specific process is as follows:

[0095] The cloud receives real-time location information and destination locations uploaded by all intelligent vehicles and uses a multi-agent reinforcement learning (MARL) algorithm to formulate collaborative decision-making strategies for the intelligent vehicles. Within the MARL framework, each intelligent vehicle is considered an agent, interacting and cooperating with other agents based on its own objectives (such as safety, efficient passage, and optimal energy consumption) and its perception of the environment. For example, in an intersection scenario, intelligent vehicles negotiate their respective passage order through low-orbit satellite communication, comprehensively considering factors such as vehicle speed, distance to the intersection, and turning direction to rationally allocate right-of-way, avoid collisions, and ensure smooth traffic flow at the intersection.

[0096] In collaborative decision-making, intelligent vehicles fully consider the behavior and predicted trajectories of non-intelligent vehicles, treating them as key decision variables. By deeply analyzing the possible behaviors of non-intelligent vehicles (such as sudden lane changes, emergency braking, and illegal turns) and their potential impact on traffic flow, intelligent vehicles adjust their decision-making strategies in a timely manner to achieve safe and harmonious interaction with non-intelligent vehicles. For example, if a non-intelligent vehicle is predicted to suddenly cut into the lane, the intelligent vehicle will slow down in advance or slightly adjust its driving route to leave a safe buffer space and ensure driving safety.

[0097] Leveraging the real-time communication and high-precision positioning capabilities of low-Earth orbit satellites, intelligent vehicle swarms can continuously optimize their decision-making strategies based on real-time traffic information (such as traffic accidents, temporary road closures, sudden changes in traffic flow, and other emergencies) and actual driving performance. When encountering emergencies, intelligent vehicles quickly adjust their decisions and adapt rapidly to new traffic environments. Simultaneously, through continuous interaction with other intelligent vehicles, they continuously improve the collaborative decision-making mechanism, gradually enhancing overall traffic efficiency and safety. During training, classic reinforcement learning algorithms (such as Q-learning and policy gradient algorithms) are employed to update and optimize decision-making strategies based on dynamically changing driving conditions, adjusting parameters in the policy network to ensure future decisions better align with actual traffic conditions, thus achieving self-learning and adaptive evolution of the intelligent transportation system. The specific process of the MARL algorithm is as follows:

[0098] First, during the initialization phase, each intelligent vehicle, acting as an independent intelligent agent, sets its own initial state. This includes vehicle position, speed, orientation, and surrounding environment perception information, while also setting an initial strategy. This strategy determines the next action, such as acceleration, deceleration, or turning, based on the vehicle's current state. The cloud acts as the coordination center, constructing a global environmental state containing state information of all intelligent vehicles.

[0099] At each decision moment t, each intelligent vehicle, based on its current state... and strategy Choose an action The information about this action is then uploaded to the cloud in real time via low-Earth orbit satellite communication. For example, if intelligent vehicle A is at an intersection and a non-intelligent vehicle in front of it slows down, based on its own strategy, it decides to slow down and make a slight rightward adjustment, and uploads this action information.

[0100] After receiving the actions of all intelligent vehicles, the cloud updates the global environmental state S. t+1 Based on traffic rules, road conditions, and the actions of other vehicles, the instant reward for each intelligent vehicle is calculated. The reward function takes into account multiple factors, such as safety factors (high reward for avoiding collisions, penalty for approaching dangerous distances), efficiency factors (reward for fast passage, penalty for frequent start-stop), and cooperation factors (reward for good cooperation with other intelligent vehicles, penalty for conflict and interference).

[0101] Subsequently, each intelligent vehicle is based on the new global environmental state S t+1 and the instant rewards received Update its own policy using the policy gradient algorithm During the training process, the above decision-making, interaction, and update steps are repeated multiple times. The intelligent vehicle continuously optimizes its own strategy and gradually learns to cooperate with other intelligent vehicles in complex traffic scenarios to improve overall traffic efficiency.

[0102] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent vehicle fusion perception and decision-making based on low-orbit satellite communication and high-precision positioning, characterized in that, include: S1 Heterogeneous Perception Information Fusion: By integrating perception data from multiple heterogeneous intelligent vehicles, a comprehensive understanding of traffic scenarios can be achieved. S2 Collaborative Decision-Making and Game Theory: Based on a comprehensive understanding of traffic scenarios, optimize game decision-making between intelligent and non-intelligent vehicles, while simultaneously achieving collaborative decision-making within the intelligent vehicle system.

2. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 1, characterized in that, The specific implementation of S1 includes: The collection and transmission of environmental perception data from the S11 heterogeneous intelligent vehicle; specifically: In complex traffic scenarios, a communication link between the cloud and various intelligent vehicles is built using low-orbit satellites. To address the heterogeneous differences in sensor configurations and algorithms among different intelligent vehicles, a post-fusion approach is adopted to integrate the perceived information. The information collection includes sensor data collection, data encoding, and transmission: The perception data acquisition includes perception data of heterogeneous intelligent vehicles: adjacent vehicle status, three-dimensional detection boxes, semantic information; and high-precision vehicle status information calculated through multiple sets of low-orbit satellite propagation signals and ground monitoring station signals: position, heading angle, speed, and acceleration. The data encoding and transmission process involves compressing and encoding the collected data, then uploading it to a cloud server for storage and processing via the communication capabilities of low-Earth orbit satellites. The transmission process employs efficient encoding methods to reduce bandwidth while ensuring real-time performance and integrity.

3. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 2, characterized in that, The specific implementation of S1 also includes: The S12 cloud server processes information; specifically: The system receives perception status information collected and released by heterogeneous intelligent vehicles in the regulatory scenario via low-orbit satellite communication in the cloud server, and achieves target-level fusion based on high-precision positioning to realize accurate scene understanding in highly uncertain traffic scenarios. The cloud-based processing strategies include time delay compensation, semantic association matching, and multi-observation joint state trajectory estimation. The time delay compensation is based on the multi-source perception results of heterogeneous intelligent vehicles, and performs delay compensation for all observed targets to be estimated. The semantic association matching refers to the target matching work performed on the overlapping parts of the observation range of multiple intelligent vehicles with perception capabilities in the observation scenario. The multi-observation joint state estimation, based on delay compensation and semantic association matching, introduces a multi-observation joint state estimation strategy, which integrates perception data from multiple intelligent vehicles and other observation sources to form a unified state estimate of targets in traffic scenarios. Based on delay compensation and a unified spatiotemporal benchmark, a target association algorithm based on semantic information matching optimizes the observation consistency of the state estimate. On this basis, Kalman filtering or unscented Kalman filtering is used to fuse all compensated and matched observation data into the unified state estimate of the target.

4. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 3, characterized in that, The specific design of the time delay compensation is as follows: Constructing motion prediction models: Construct motion prediction models for each vehicle based on the motion state of the target object, using either a uniform speed model or a uniform acceleration model; Calculate the delay time: Based on the data transmission delay and the sensing link delay, calculate the possible motion state delay time Δt of the target object; State prediction: Using a motion prediction model, the current accurate position of the target object is predicted to obtain the compensated state variables.

5. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 4, characterized in that, The semantic association matching includes two parts: basic matching based on intersection-union ratio and semantic feature-enhanced matching, as detailed below: The basic matching based on cross-union ratio (CUNR) utilizes the high-precision positioning capability of low-orbit satellites. Based on the spatial positioning of the sensing carrier, it performs three-dimensional spatial projection on the sensing target and uses the cross-union ratio matching strategy to achieve basic spatial matching, taking the driving scenario as a unit, and constructs the preliminary matching results of heterogeneous intelligent vehicle perception. The semantic feature-enhanced matching, which utilizes high-level semantic information to achieve high-order joint matching, includes target association feature extraction, matching algorithm design, and dynamic weight adjustment, as detailed below: The associated feature extraction combines the vehicle's location information and semantic information, including color and vehicle type, to construct a joint feature vector F = [P, S], where P represents spatial location features and S represents semantic labels; The matching algorithm design employs a location-based nearest neighbor matching algorithm or a semantic information-based weighted matching method to match the perception data of each vehicle with pre-labeled information in the map. The matching degree calculation formula can be expressed as: S 匹配 =αIoU+βS 相似度 Where α and β are weight parameters, IoU is the aforementioned basic matching score based on intersection-union ratio, and S 相似度 The matching similarity score is obtained by encoding and decoding the joint feature vector basis F using a neural network. Where f ω For neural network models, The semantic joint feature vectors of N vehicles to be observed are collected. The dynamic weight adjustment adjusts the weight of each information source according to environmental conditions. When the lighting is good, color information is more reliable and is given a larger weight; when the weather is bad, location information is more reliable and its weight is adjusted accordingly.

6. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 1, characterized in that, The implementation of S2 includes: S21 Group Behavior Guidance from a Unified Scenario Perspective: Based on the joint perception and processing results of traffic management areas in the cloud, rich traffic movement status data is collected to achieve cloud-based aggregation of movement strategies. This, in turn, enhances trajectory prediction at the joint scenario level for traffic participants, enabling joint cognition of their behavior. A unified scenario perspective framework for group behavior guidance is constructed to achieve comprehensive perception and reasonable inference. Specific steps are as follows: Data collection: Utilizing cloud-based sensing results, extensively collect driving data involving various types of vehicles, including vehicle driving trajectory, speed changes, and steering behavior; Data preprocessing: The collected data is cleaned and denoised, and unified to the same time base. A dynamic delay compensation strategy is used to correct the delay caused by network transmission and ensure the spatiotemporal consistency of all data. Construction and training of the prediction model: A scene-level trajectory prediction model based on deep learning is adopted. The input includes historical state quantities, semantic information and scene context. The dynamic and static interaction relationships of multiple vehicles and map elements in the scene are modeled using graph neural networks (GNN), long short-term memory networks (LSTM) or attention mechanisms. Supervised learning methods are used to minimize the error between the predicted trajectory and the real trajectory and output the probability distribution of the prediction results. Predictive model enhancement: This invention implements three types of predictive model enhancement strategies, including those for different driving scenarios, different types of traffic participants, and different driving conditions, as detailed below: The method involves using monitoring data collected in the cloud to train specific prediction patterns for different driving scenarios, thereby enhancing the model's adaptability to specific intersections or road sections. The approach for different types of traffic participants involves fusing the motion characteristics and semantic information of different types of vehicles into the model, and inferring the trajectory distribution of different types of vehicles through joint modeling. The method addresses different driving conditions by collecting and analyzing driving data under various weather conditions, optimizing model parameters, and ensuring that the model can accurately predict the trajectory of non-intelligent vehicles under various environmental conditions.

7. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 6, characterized in that, The implementation of S2 also includes: S22 is an intelligent vehicle collaborative decision-making system based on group behavior guidance. Building upon scenario-level predictive behavior guidance, it achieves collaborative decision-making among intelligent vehicle groups via low-Earth orbit satellite communication. This group collaborative decision-making not only focuses on individual optimality but also improves overall traffic efficiency through collaborative cooperation, based on a full understanding of the behavior of non-intelligent vehicle groups. This includes formulating collaborative decision-making strategies, assessing the impact of non-intelligent groups, optimizing decisions, and real-time adjustments, as detailed below: The collaborative decision-making strategy is formulated as follows: the cloud receives real-time positioning information and target location of intelligent vehicles, and combines them with the predicted trajectory of non-intelligent vehicles; a multi-agent reinforcement learning (MARL) algorithm is used to formulate a collaborative decision-making strategy. The judgment of the impact of non-intelligent groups: In the process of collaborative decision-making, intelligent vehicles fully consider the predicted behavior trajectory of non-intelligent vehicles and adjust their own decision-making strategies to achieve safe interaction; The aforementioned decision optimization and real-time adjustment: With the help of real-time communication and high-precision positioning of low-orbit satellites, intelligent vehicle swarms can obtain traffic information and actual driving effects in real time and dynamically optimize decision strategies; in the event of emergencies such as traffic accidents and road control, intelligent vehicles can quickly adjust decision strategies to adapt to the new environment, including replanning driving routes or changing traffic flow.

8. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 7, characterized in that, The specific details of formulating a collaborative decision-making strategy using the MARL algorithm are as follows: Each intelligent vehicle, acting as an intelligent agent, interacts and collaborates with other intelligent vehicles based on its own goals and environmental perception. By sharing information and negotiating priorities, they jointly optimize traffic flow. For example, in an intersection scenario, intelligent vehicles determine the passage order through communication negotiation to avoid conflicts. This strategy optimization process can be represented as: Where π is the strategy, γ is the discount factor, R is the reward function, and s t Let a be a state variable. t This refers to the amount of motion.

9. The intelligent vehicle fusion perception and decision-making method based on low-orbit satellite communication and high-precision positioning according to claim 7, characterized in that, During the training of the decision network, reinforcement learning algorithms are used to dynamically optimize the network parameters, making future decisions more adaptable to actual traffic conditions. The optimization goal is to maximize overall traffic efficiency and safety.

10. An intelligent vehicle fusion perception and decision-making system based on low-orbit satellite communication and high-precision positioning, characterized in that, It includes a perception part and a decision-making part. The perception part is capable of fusing heterogeneous perception information as described in claims 2-5. The decision-making part is capable of collaborative decision-making and game theory as described in claims 6-9.

Images