A multi-target adaptive dynamic obstacle inflation method and system for complex mixed driving scenarios in a park

Abstract

The application discloses a kind of multi-objective adaptive dynamic obstacle inflation methods and systems for complex mixed running scenarios of park, it is related to unmanned vehicle environment perception and safety control technical field, the method includes: based on multi-source state information constructs obstacle state model;Based on obstacle state model constructs multi-objective risk assessment model, assesses the comprehensive risk of obstacle to unmanned vehicle operating space;Under the constraint of comprehensive risk assessment result, adaptive correction is carried out to basic inflation parameter;The direction related asymmetric adjustment is carried out to the basic inflation parameter after correction, and constructs elliptical dynamic inflation area;Elliptical dynamic inflation area is updated in real time, and the elliptical dynamic inflation area after updating is as the risk occupied space of obstacle output.The application can significantly improve the traffic efficiency under the premise of not reducing safety in the narrow channel of park, mixed running area and high interaction scene.

Description

Technical Field

[0001] This invention relates to the field of environmental perception and safety control technology for unmanned vehicles, and more specifically, to a multi-target adaptive dynamic obstacle expansion method and system for complex mixed traffic scenarios in industrial parks. Background Technology

[0002] With the gradual application of autonomous driving technology in semi-enclosed environments such as industrial parks, factories, and campuses, autonomous vehicles in these environments often need to operate alongside various traffic participants, including pedestrians, bicycles, and low-speed motor vehicles, under low-speed conditions. To ensure vehicle safety, obstacle inflation is widely used in the environmental perception and safety control processes of autonomous vehicles. This method involves constructing safe restricted zones around obstacles to limit the drivable space of the autonomous vehicle and provide safety constraints for path planning, behavior decision-making, and motion control modules. Therefore, the rationality of the inflation model directly affects the operational safety and traffic efficiency of the autonomous vehicle system.

[0003] In existing autonomous vehicle systems for industrial parks, fixed radius expansion or simple adaptive expansion methods are commonly used. Fixed radius expansion typically involves superimposing a fixed safety distance outside the geometric contour of obstacles to form a circular or near-circular restricted area, which is simple to implement and has low computational overhead. The simple adaptive expansion method, on the other hand, introduces factors such as obstacle speed or distance from the vehicle to linearly adjust the expansion radius, thereby improving safety in dynamic scenarios.

[0004] However, in complex mixed-traffic environments within parks, the aforementioned methods still have certain limitations. The fixed-radius expansion model is a static modeling approach, which struggles to reflect the differences in motion states of various obstacles. When the motion state of obstacles changes, it cannot adjust the safety zone in a timely manner, easily leading to overly conservative safety constraints in densely populated pedestrian areas. Furthermore, it may suffer from insufficient risk coverage in dynamic mixed-traffic scenarios. Simple adaptive methods have a relatively singular adjustment dimension, failing to comprehensively consider factors such as obstacle type, relative motion relationships, and environmental semantic information, making it difficult to achieve a refined expression of dynamic risks.

[0005] In addition, the road environment in the park is usually characterized by low speed, dense intersections, narrow passages and frequent interaction of traffic participants. Most existing expansion methods are based on isotropic circular models to construct safe areas, which do not fully consider the impact of obstacle movement direction and potential movement trend on risk distribution. This results in the safe area maintaining the same scale in all directions, thus generating more ineffective expansion space in non-primary risk directions and reducing the efficiency of road space utilization.

[0006] Therefore, in complex park scenarios, how to construct a dynamic expansion model that can comprehensively reflect the movement state, relative relationship and scene semantic information of obstacles while ensuring the safe operation of unmanned vehicles, so that the safe zone can be adaptively adjusted according to the movement trend of obstacles and form differentiated risk coverage in different directions, thereby achieving a balance between safety and traffic efficiency, has become an urgent technical problem to be solved in the field of environmental perception and safety control of unmanned vehicles in parks.

[0007] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a multi-objective adaptive dynamic obstacle expansion method and system for complex mixed-traffic scenarios in industrial parks. This solves the problems mentioned in the background that existing obstacle expansion methods for unmanned vehicles in industrial parks generally suffer from static expansion models, single risk characterization dimensions, difficulty in reflecting the real movement risks of dynamic obstacles, and insufficient adaptability to complex operational scenarios in industrial parks.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] According to one aspect of the present invention, a multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks is provided, comprising:

[0011] S1. Obtain multi-source state information of obstacles in the park environment, and construct an obstacle state model based on the multi-source state information. The multi-source state information includes at least the physical size, position, speed, acceleration, relative motion state with the unmanned vehicle, and semantic category of the park scene where the obstacle is located.

[0012] S2. Construct a multi-objective risk assessment model based on the obstacle state model, which includes safety objectives and traffic efficiency objectives, and use the multi-objective risk assessment model to assess the comprehensive risk of obstacles to the unmanned vehicle's operating space;

[0013] S3. Under the constraints of the comprehensive risk assessment results, the basic expansion parameters are calculated based on the physical dimensions of the obstacle, and the basic expansion parameters are adaptively corrected by combining the obstacle's movement speed, acceleration, and relative motion state with the unmanned vehicle.

[0014] S4. Based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, the corrected basic expansion parameters are adjusted asymmetrically in a direction-dependent manner, and an elliptical dynamic expansion region is constructed.

[0015] S5. Update the elliptical dynamic expansion region in real time according to a preset cycle or a trigger mechanism based on the change of obstacle state, and output the updated elliptical dynamic expansion region as the risk-occupied space of the obstacle.

[0016] According to another aspect of the present invention, a multi-objective adaptive dynamic obstacle expansion system for complex mixed-traffic scenarios in industrial parks is also provided, the system comprising:

[0017] The state modeling module is used to acquire multi-source state information of obstacles in the park environment and build an obstacle state model based on the multi-source state information. The multi-source state information includes at least the physical size, position, speed, acceleration, relative motion state with the unmanned vehicle, and semantic category of the park scene where the obstacle is located.

[0018] The scenario risk assessment module is used to construct a multi-objective risk assessment model based on the obstacle state model, which includes safety objectives and traffic efficiency objectives, and to use the multi-objective risk assessment model to assess the comprehensive risk of obstacles to the operating space of unmanned vehicles.

[0019] The adaptive correction module is used to calculate the basic expansion parameters based on the physical dimensions of the obstacle under the constraints of the comprehensive risk assessment results, and to adaptively correct the basic expansion parameters by combining the obstacle's movement speed, acceleration, and relative motion state with the unmanned vehicle.

[0020] The elliptical risk modeling module is used to perform direction-related asymmetric adjustments on the corrected basic expansion parameters based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, and to construct an elliptical dynamic expansion region.

[0021] The risk area output module is used to update the elliptical dynamic expansion area in real time according to a preset period or based on the triggering mechanism of obstacle state changes, and output the updated elliptical dynamic expansion area as the risk-occupied space of the obstacle.

[0022] The beneficial effects of this invention are as follows:

[0023] 1. This invention uses a direction-dependent elliptical risk region to describe obstacle risks, making safety constraints more forward-looking in the main direction of obstacle movement, while reducing unnecessary expansion space in the lateral direction. In narrow passages, mixed pedestrian and vehicle areas, and high-interaction scenarios, it can significantly improve traffic efficiency without compromising safety, thus having higher engineering application value. By introducing traffic efficiency and risk event frequency to construct operational status indicators and driving dynamic changes in scenario correction coefficients, the risk region can be adjusted in real time between safety and traffic efficiency according to environmental changes, thereby avoiding the insufficient adaptability caused by static or single adjustment of risk models in traditional methods.

[0024] 2. This invention extracts scene feature parameters such as the degree of road space constraint, obstacle density, and motion interaction complexity to construct a unified scene risk index model. It also adaptively determines the safety weight and traffic efficiency weight through a normalized mapping function, so that the risk modeling results can dynamically balance between safety requirements and traffic efficiency according to the risk level of different scenes, thereby avoiding the problem of over-conservatism caused by a single safety constraint in traditional methods.

[0025] 3. In the process of obstacle expansion modeling, this invention introduces an elliptical risk region related to the direction of motion, so that the major axis of the ellipse is arranged along the direction of obstacle motion and the minor axis is perpendicular to the direction of motion. The ellipse scale is adaptively corrected in combination with the obstacle's motion speed and dynamic state, so that the constructed risk region can more realistically reflect the future motion trend of the obstacle and the distribution of potential collision risks.

[0026] 4. This invention introduces a risk forward shift mechanism to improve risk prediction capabilities. By calculating the movement trend of obstacles within the prediction time scale, the center position of the elliptical risk area is shifted forward in a direction-related manner, so that the risk area can cover the space position that the obstacle may occupy in the future in advance, thereby improving the autonomous vehicle's ability to perceive potential collision risks in advance.

[0027] 5. This invention uses a periodic update or event-triggered update mechanism to dynamically adjust the risk-occupied area of ​​obstacles in real time and outputs it in the form of a standardized risk area, providing continuous and stable safety constraints for the unmanned vehicle path planning, behavior decision-making and motion control modules, thereby improving the overall operational stability of the system. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in a park, according to an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram of a multi-objective adaptive dynamic obstacle expansion system for complex mixed-traffic scenarios in a park, according to an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of the obstacle risk area generation process according to an embodiment of the present invention;

[0032] Figure 4 This is a final expansion region diagram according to an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram of the overall process according to an embodiment of the present invention.

[0034] In the picture:

[0035] 1. State modeling module; 2. Scenario risk assessment module; 3. Adaptive correction module; 4. Elliptic risk modeling module; 5. Risk area output module. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0037] In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0038] According to an embodiment of the present invention, a multi-objective adaptive dynamic obstacle expansion method and system for complex mixed-traffic scenarios in industrial parks are provided.

[0039] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in a park, according to an embodiment of the present invention, includes:

[0040] S1. Obtain multi-source state information of obstacles in the park environment, and construct an obstacle state model based on the multi-source state information. The multi-source state information includes at least the physical size, position, speed, acceleration, relative motion state with the unmanned vehicle, and semantic category of the park scene where the obstacle is located.

[0041] S2. Construct a multi-objective risk assessment model based on the obstacle state model, which includes safety objectives and traffic efficiency objectives, and use the multi-objective risk assessment model to assess the comprehensive risk of obstacles to the unmanned vehicle's operating space;

[0042] S3. Under the constraints of the comprehensive risk assessment results, the basic expansion parameters are calculated based on the physical dimensions of the obstacle, and the basic expansion parameters are adaptively corrected by combining the obstacle's movement speed, acceleration, and relative motion state with the unmanned vehicle.

[0043] S4. Based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, the corrected basic expansion parameters are adjusted asymmetrically in a direction-dependent manner, and an elliptical dynamic expansion region is constructed.

[0044] S5. Update the elliptical dynamic expansion region in real time according to a preset cycle or a trigger mechanism based on the change of obstacle state, and output the updated elliptical dynamic expansion region as the risk-occupied space of the obstacle.

[0045] In this optional embodiment, multi-source state information of obstacles in the park environment is obtained. The multi-source state information includes at least the physical size, position, speed, acceleration, relative motion state with the autonomous vehicle, and semantic category of the park scene in which the obstacle is located. Constructing an obstacle state model based on the multi-source state information includes:

[0046] S11. Import the park digital map into the unmanned vehicle controller or the device that communicates with the unmanned vehicle controller in advance, use a unified global coordinate system to describe the park map, and establish a unified spatial coordinate reference for the unmanned vehicle's own position, obstacle position and scene semantic information.

[0047] S12. Acquire data from satellite positioning, inertial measurement unit, and odometer outputs. After time synchronization processing, use extended Kalman filter algorithm for fusion processing. Combined with semantic constraints of the park map and key position correction mechanism, obtain the global position coordinates, heading angle, and linear velocity of the unmanned vehicle in the global coordinate system.

[0048] S13. Use vehicle-mounted lidar, cameras and millimeter-wave radar to detect and identify obstacles in the park environment, and obtain the physical size and location information of the obstacles;

[0049] S14. Calculate the velocity and acceleration of the obstacle using the time difference method based on the positional changes of the obstacle in continuous time frames.

[0050] S15. Calculate the relative distance between the obstacle and the unmanned vehicle based on the obstacle's position and the unmanned vehicle's position; calculate the relative speed between the obstacle and the unmanned vehicle based on the obstacle's speed and the unmanned vehicle's speed; and determine the obstacle's motion direction angle based on the continuous change in the obstacle's position.

[0051] S16. Use a point-in-polygon detection algorithm to perform spatial semantic determination on the current position of the unmanned vehicle, match the current position of the unmanned vehicle with the scene semantic region in the park map, and determine the semantic category of the park scene where the obstacle is located.

[0052] S17. Integrate the physical dimensions, position, speed, acceleration, direction angle of movement of obstacles, relative distance between obstacles and autonomous vehicles, relative speed between obstacles and autonomous vehicles, and semantic categories of the park scene to form complete multi-source state information and establish a unified obstacle state model.

[0053] In this optional embodiment, a multi-objective risk assessment model is constructed based on the obstacle state model, incorporating both safety and traffic efficiency objectives. This multi-objective risk assessment model is then used to evaluate the comprehensive risk posed by obstacles to the autonomous vehicle's operating space, including:

[0054] S21. Perform feature analysis and normalization on the obstacle state vector to construct a risk feature representation with unified dimensions;

[0055] S22. Based on the relative distance and relative speed between the obstacle and the unmanned vehicle, calculate the safety risk index, and based on the area of ​​the obstacle risk area and the area of ​​the unmanned vehicle's local planning area, calculate the space occupancy index. Normalize the safety risk index to ensure that the safety risk index and the space occupancy index are consistent in dimension.

[0056] S23. Based on the semantic annotation of the park map, divide the scene units, extract the road space constraint index, obstacle density index, and motion interaction complexity index and normalize them. Based on the weighted fusion of the normalized road space constraint index, obstacle density index, and motion interaction complexity index, obtain the scene risk index.

[0057] S24. Determine the value range of safety weight and traffic efficiency weight based on the semantic category of the park scene where the obstacle is located, and calculate the initial value of the weight based on the scene risk index and the normalized mapping function.

[0058] S25. Calculate the traffic efficiency index based on the real-time average speed of the unmanned vehicle and the planned expected speed, and dynamically update the safety weight and traffic efficiency weight based on the traffic efficiency index so that the sum of the safety weight and traffic efficiency weight remains constant.

[0059] S26. Using a weighted summation method, the safety risk indicators and space occupancy indicators are integrated with the updated safety weights and traffic efficiency weights to obtain a comprehensive risk assessment result of obstacles on the operating space of unmanned vehicles.

[0060] In this optional embodiment, under the constraint of the comprehensive risk assessment results, the basic expansion parameters are calculated based on the physical dimensions of the obstacle, and the basic expansion parameters are adaptively corrected by combining the obstacle's movement speed, acceleration, and relative motion state with the unmanned vehicle.

[0061] S31. Calculate the basic expansion scale of the obstacle based on the equivalent geometric radius of the obstacle, the equivalent radius of the unmanned vehicle, and the preset minimum safety distance.

[0062] S32. Based on the braking performance of the unmanned vehicle, the system perception and control delay, and the upper limit of the park's operating speed, a preset minimum safe distance is set.

[0063] S33. Adaptively adjust the basic expansion scale by combining the relative velocity and acceleration of the obstacle with the corresponding velocity correction coefficient and acceleration correction coefficient.

[0064] S34. Based on the correction relationship, obtain the dynamic expansion scale that adapts to the motion state of the obstacle, and increase the expansion scale when the obstacle accelerates and approaches, and suppress the expansion scale amplification when the motion risk is low.

[0065] In this optional embodiment, based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, the corrected basic expansion parameters are adjusted asymmetrically in a direction-dependent manner, and an elliptical dynamic expansion region is constructed, including:

[0066] S41. Based on historical operation data, calibrate the initial scene semantic correction coefficients corresponding to the semantic categories of each scene, and dynamically adjust the scene semantic correction coefficients according to the real-time operation status of the scene.

[0067] S42. The dynamic expansion scale after motion state correction is corrected by using the dynamically adjusted scene semantic correction coefficient to obtain the equivalent expansion scale.

[0068] S43. Determine the direction of the major axis of the ellipse based on the direction of obstacle movement, and calculate the parameters of the major and minor axes of the ellipse by combining the equivalent expansion scale, obstacle movement speed, forward risk amplification coefficient and risk prediction time window.

[0069] S44. Based on the obstacle's movement speed, the risk prediction time scale, and the weights of safety and traffic efficiency, calculate the risk forward shift along the movement direction and determine the geometric center of the elliptical risk region.

[0070] S45. Construct a rotation matrix based on the major axis, minor axis parameters, geometric center, and motion direction angle of the ellipse, and generate the non-central elliptical risk-occupying area corresponding to the obstacle;

[0071] S46. The non-central elliptical risk-occupied area is output as the risk-occupied space of the obstacle, and is used as the spatial safety constraint for path planning and obstacle avoidance control.

[0072] In this optional embodiment, the initial scene semantic correction coefficients corresponding to each scene semantic category are calibrated based on historical operating data, and the scene semantic correction coefficients are dynamically adjusted according to the real-time operating status of the scene, including:

[0073] S411. Construct a local risk field for each scene unit and extract a set of high-risk areas. Perform parametric modeling on the high-risk areas and extract spatial scale feature parameters.

[0074] S412. Construct an optimization objective function constrained by safety and traffic efficiency, comprehensively evaluate the degree of risk coverage and space occupancy, and determine the optimal risk scale for each scenario unit.

[0075] S413. Based on the optimal risk scale and the basic risk scale under the standard reference scenario, determine the initial scenario semantic correction coefficients corresponding to the semantic categories of each scenario.

[0076] S414. Within the sliding time window, statistically analyze the traffic efficiency index and the frequency of risk events of the scene unit, and construct a comprehensive operation status function;

[0077] S415. Based on the comprehensive operating state function, risk sensitivity coefficient and efficiency adjustment coefficient, the initial scene semantic correction coefficient is dynamically adjusted, and the adjusted scene semantic correction coefficient is subject to interval constraints.

[0078] In this optional embodiment, the risk forward shift offset along the direction of movement is calculated based on the obstacle's movement speed, the risk prediction timescale, and the weights of safety and traffic efficiency, and the geometric center of the elliptical risk region is determined, including:

[0079] S441. Calculate the obstacle's motion direction angle based on the position difference of the obstacle at consecutive moments, and verify the consistency of the motion direction angle in conjunction with the geometric orientation of the park's roads.

[0080] S442. Calculate the risk forward shift along the direction of movement based on the obstacle's movement speed, risk prediction time scale, basic offset ratio coefficient, and safety and traffic efficiency weights.

[0081] S443. Superimpose the current position of the obstacle with the forward offset of the risk to obtain the geometric center position of the non-central elliptical risk area.

[0082] In this optional embodiment, the elliptical dynamic expansion region is updated in real time according to a preset period or a triggering mechanism based on changes in obstacle state, and the updated elliptical dynamic expansion region is output as the risk-occupied space of the obstacle, including:

[0083] S51. By periodically acquiring the latest state information of obstacles, detect changes in the movement state of obstacles and the semantic category of the scene in which they are located;

[0084] S52. If a change in obstacle status or scene semantics is detected, the parameters of the multi-objective risk assessment model and the scene semantic correction coefficients are recalculated to update the risk area.

[0085] S53. Based on the updated multi-objective risk assessment model parameters and scenario semantic correction coefficients, regenerate the corresponding dynamic risk area to achieve dynamic evolution of the risk space.

[0086] S54. Smooth the dynamic risk area to ensure the continuity and rationality of the risk space, and use the principle of maximum risk fusion to spatially fuse the smoothed dynamic risk area in order to complete the calibration of the overall risk space.

[0087] S55. The completed risk space output will serve as the environmental constraint basis for path planning and decision-making.

[0088] According to another embodiment of the invention, such as Figure 2 As shown, a multi-objective adaptive dynamic obstacle expansion system for complex mixed-traffic scenarios in industrial parks is also provided. This system includes:

[0089] State modeling module 1 is used to acquire multi-source state information of obstacles in the park environment and construct an obstacle state model based on the multi-source state information. The multi-source state information includes at least the physical size, position, speed, acceleration, relative motion state with the unmanned vehicle, and semantic category of the park scene where the obstacle is located.

[0090] Scene risk assessment module 2 is used to construct a multi-objective risk assessment model that includes safety and traffic efficiency objectives based on the obstacle state model, and to use the multi-objective risk assessment model to assess the comprehensive risk of obstacles to the unmanned vehicle's operating space;

[0091] The adaptive correction module 3 is used to calculate the basic expansion parameters based on the physical size of the obstacle under the constraint of the comprehensive risk assessment results, and to adaptively correct the basic expansion parameters by combining the obstacle's movement speed, acceleration and relative motion state with the unmanned vehicle.

[0092] Elliptical risk modeling module 4 is used to perform direction-related asymmetric adjustments on the corrected basic expansion parameters based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, and to construct an elliptical dynamic expansion region.

[0093] The risk area output module 5 is used to update the elliptical dynamic expansion area in real time according to a preset period or a trigger mechanism based on the change of obstacle status, and output the updated elliptical dynamic expansion area as the risk-occupied space of the obstacle.

[0094] To facilitate understanding of the above technical solutions of the present invention, the following provides a detailed description of the multi-objective adaptive dynamic obstacle expansion of the present invention in practical applications for complex mixed-traffic scenarios in industrial parks.

[0095] Step 1: Obstacle State Information Acquisition and Modeling. In this invention, to achieve spatial correlation between obstacle state information and the park's operating environment, a digital map of the park is pre-imported into the unmanned vehicle controller or its communication-connected computing unit. The park map is described using a unified global coordinate system and includes scene semantic annotation information such as road structure, traffic area boundaries and intersections, narrow passages, and densely populated areas. By importing the park map, a unified spatial coordinate reference is established in the unmanned vehicle control system for the unified expression of the unmanned vehicle's own position, obstacle positions, and scene semantic information.

[0096] Specifically, in this invention, in order to realize the position correction and continuous update of the unmanned vehicle in the global coordinate system of the park, a multi-source fusion positioning algorithm based on extended Kalman filtering is constructed, and combined with the semantic constraints of the park map and the key position correction mechanism, the stable output of the global position and attitude of the unmanned vehicle is realized.

[0097] During the operation of the autonomous vehicle, the positioning module periodically acquires the following data: the absolute position coordinates of the autonomous vehicle output by the satellite positioning receiver module or the regional positioning beacon; the linear acceleration and angular velocity output by the inertial measurement unit; the displacement increment information output by the odometer; and the digital map data of the park and scene semantic annotation information. The above data is then input into the positioning fusion module after time synchronization processing.

[0098] Using the centroid of the autonomous vehicle as a reference point, a state vector is established in the unified global coordinate system of the park:

[0099] ;

[0100] in: This represents the global position coordinates of the autonomous vehicle; Indicates the heading angle; This indicates the speed of the driverless vehicle.

[0101] Specifically, in this invention, the Extended Kalman Filter (EKF) algorithm is used to fuse absolute positioning information and inertial prediction information. The EKF uses the obstacle state vector as the estimation object, which includes position, velocity, and direction of motion information. The system uses the raw measurement data output by the sensing module as the observation input, including obstacle position and velocity information detected by lidar or visual sensors.

[0102] During the filtering process, the obstacle-based kinematic model predicts and updates the vehicle's state; combined with real-time observation data, the prediction results are corrected to obtain a smoother and more continuous state estimation result. The state estimation result serves as input for subsequent risk modeling and elliptical region construction. In terms of data processing, the perceived data undergoes time synchronization and outlier removal before being input into the filtering algorithm to ensure data consistency and stability. Regarding parameter settings, the noise covariance parameter during the filtering process is set based on sensor accuracy and the actual operating environment, and can be adjusted during system operation based on data statistical characteristics to improve state estimation accuracy. To determine the semantic region of the park scene where the autonomous vehicle is currently located, after completing the extended Kalman filter localization fusion, the current state estimation vector of the autonomous vehicle is obtained:

[0103] ;

[0104] in, This indicates the position coordinates of the autonomous vehicle within the unified global coordinate system of the park. Indicates the heading angle of the autonomous vehicle. This represents the linear velocity of the autonomous vehicle. Based on the above state estimation results, the current position of the autonomous vehicle is extracted:

[0105] ;

[0106] in, Indicates the current location of the driverless vehicle; This represents the position coordinates of the autonomous vehicle within the unified global coordinate system of the park. This position point is then used as the detection point and spatially matched with the scene semantic region in the park map. To achieve region determination, this invention employs a point-within-a-polygon detection algorithm to perform spatial semantic determination of the autonomous vehicle's current position. The scene semantic region is represented as a polygon in the park map, and its boundary is composed of a set of sequentially connected vertex coordinates. For any point to be determined... The region assignment is determined using the ray method: when the number of intersections between the ray and the polygon boundary is odd, the assignment point is located inside the polygon; when the number is even, the assignment point is located outside the polygon. For special cases such as the ray coinciding with a vertex or parallel to the boundary, a unified determination rule is used to ensure computational stability.

[0107] Specifically, the polygon data originates from the scene semantic annotation results of the park map. Input point data undergoes coordinate unification processing before judgment to ensure spatial matching consistency and achieve real-time association between the autonomous vehicle's position and scene units. The aforementioned position, attitude, and scene semantic information collectively serve as the input basis for the subsequent multi-objective risk assessment model, used to calculate the relative positions of obstacles. Relative velocity and relative distance This provides a spatial and motion reference for subsequent calculation of basic expansion parameters and dynamic correction mechanisms. By calculating the changes in obstacle position across consecutive time frames, the velocity and acceleration information of the obstacles are obtained. Velocity and acceleration can be calculated using a time-difference-based method, with the following calculation form:

[0108] ;

[0109] ;

[0110] in, The sampling time interval is represented by t; the current sampling time is represented by t. Indicates the previous sampling time; Indicates the first The position of each obstacle in the global coordinate system; Indicates the first The speed of movement of each obstacle; Indicates the first The acceleration of the obstacle. The relative velocity between the obstacle and the autonomous vehicle. The velocity vector is determined based on the relative relationship between the obstacle's velocity and the autonomous vehicle's velocity. Let the autonomous vehicle's velocity vector in the global coordinate system be... , No. The velocity vectors of the obstacles are Then the relative velocity vector is expressed as:

[0111] ;

[0112] To characterize the tendency of an obstacle to approach or move away from the autonomous vehicle, the relative velocity vector is projected onto the direction of the line connecting the obstacle and the autonomous vehicle, resulting in a scalar form of relative velocity:

[0113]

[0114] in, Indicates the first The location of the obstacle Indicates the location of the driverless car. Indicates the first The relative speed between the obstacle and the driverless vehicle; This represents the relative velocity vector of the i-th obstacle relative to the autonomous vehicle. It effectively reflects the approach speed of the obstacle relative to the autonomous vehicle and is used for adaptive adjustment of the subsequent dynamic expansion scale.

[0115] Specifically, the unmanned vehicles in the park use onboard LiDAR, cameras, and millimeter-wave radar to detect and identify obstacles in the operating environment. For each detected obstacle, its basic state information is acquired, and a unified obstacle state model is established based on the obtained parameters. The obstacle state model can be represented as:

[0116] ;

[0117] in: Indicates the first The position of each obstacle in the global coordinate system; Indicates the first The speed of movement of each obstacle; Indicates the first The acceleration of the obstacle; Indicates the first The direction angle of movement of each obstacle; Indicates the first The relative distance between an obstacle and the autonomous vehicle's point mass; Indicates the first The relative speed between the obstacle and the driverless vehicle; Indicates the first The semantic category of the park scene where the obstacle is located.

[0118] It should be clarified that in this invention, the vehicle-mounted LiDAR, camera, and millimeter-wave radar are only used to acquire information on the spatial location, geometric contour, and motion state of obstacles. This invention sets prominent warning signs at the data acquisition end to fulfill its obligation to inform users, and obtains the user's separate consent through the vehicle terminal. For data acquired by the camera, this invention adopts a technical strategy of "acquisition-processing-destruction": the system only performs real-time contour extraction, position estimation, and motion parameter calculation on targets in the image, and technically prohibits the storage of original image data. All processing is completed locally at the edge, without facial recognition, identity recognition, or behavior recognition, nor generating information that can be used to identify specific individuals. The data processing flow ensures that the identity or trajectory of a specific individual cannot be recovered. Therefore, the obstacle perception and state modeling involved in this invention pertains to the anonymization and de-identification of environmental targets, and is only used for autonomous vehicle operation safety assessment and spatial risk modeling, and does not involve the collection, storage, or use of personal identity information.

[0119] Step 2: Construction of a Multi-Objective Risk Assessment Model. To comprehensively characterize the impact of obstacles on the operational safety and traffic efficiency of autonomous vehicles in complex mixed-traffic scenarios within the park, a multi-objective risk assessment model is constructed based on the obstacle state model established in Step 1. This model is used to quantitatively assess the risk level of single or multiple obstacles and provides constraints for subsequent calculations of dynamic expansion parameters. The input to the multi-objective risk assessment model is the obstacle state vector:

[0120] ;

[0121] in,: Indicates the first The position of each obstacle in the global coordinate system; Indicates the first The speed of movement of each obstacle; Indicates the first The acceleration of the obstacle; Indicates the first The direction angle of movement of each obstacle; Indicates the first The relative distance between an obstacle and the autonomous vehicle's point mass; Indicates the first The relative speed between the obstacle and the driverless vehicle; Indicates the first The semantic category of the park scene where each obstacle is located. All state variables originate from the vehicle-mounted perception system, positioning module, and park map semantic information, and are all within the same global coordinate system. The output of the multi-objective risk assessment model is the obstacle risk assessment result:

[0122] ;

[0123] Indicates the first The comprehensive risk level of each obstacle to the current operating state of the autonomous vehicle is indicated by a higher numerical value. The multi-objective risk assessment model employs a hierarchical processing structure. Its processing flow includes: feature analysis and normalization of the obstacle state vectors to construct a risk feature representation with unified dimensions; and calculation of safety risk indicators. With space occupancy index The above indicators are weighted and fused based on time-varying weighting coefficients to obtain a comprehensive risk assessment result. The above processing steps logically correspond to three stages: feature analysis, risk assessment, and multi-objective fusion.

[0124] The multi-objective evaluation function is constructed. The multi-objective risk assessment model adopts a weighted summation form, and its comprehensive evaluation function is expressed as follows:

[0125] ;

[0126] in: Indicates the first Safety risk indicators for each obstacle; Indicates the first The space occupancy index of each obstacle is used to characterize the degree to which the risk expansion area compresses the space accessible to autonomous vehicles. , All are time-varying weighting coefficients, and satisfy the following conditions: .

[0127] To ensure the feasibility of the multi-objective risk assessment model, the calculation methods for the safety risk index and traffic efficiency index are explained below:

[0128] Safety risk indicators Taking into account the spatial distance and relative motion between the obstacle and the autonomous vehicle, it is defined as:

[0129] ;

[0130] in: The distance between the autonomous vehicle and the obstacle. For relative velocity, For distance attenuation scale parameters, This refers to the speed adjustment coefficient. Distance attenuation scale parameter. Determined based on the braking capability and system response time of the autonomous vehicle. Obtained through calibration using historical operational data. Space occupancy index. Defined as:

[0131] ;

[0132] in: The area of ​​the risk zone corresponding to the obstacle. This refers to the area designated for the autonomous vehicle's local planning. This represents the spatial range of the autonomous vehicle's local path planning, which is dynamically determined based on the vehicle's operating status. To ensure consistency in the dimensions of different indicators, safety risk indicators are normalized before being weighted, ensuring they fall within the same numerical range as traffic efficiency indicators. In different scenarios, the emphasis on safety and traffic efficiency by autonomous vehicles in the park varies significantly. Therefore, by establishing a mapping relationship between scenario semantic categories and weight coefficients, the risk assessment results are matched with actual operational needs.

[0133] Specifically, this invention quantifies the operational risks of different park scenarios by constructing a scenario risk index model. First, based on the semantic annotation information of the park map, the operating environment is divided into scenario units, and a set of scenario semantic categories is constructed:

[0134] ;

[0135] Each scene unit This invention addresses areas with specific traffic characteristics, including intersections, narrow passages, densely populated areas, and general traffic areas. Based on this, to achieve a unified quantitative description of risk levels across different scenarios, this invention constructs a scenario risk index. The scenario risk index is not based on experience, but rather on modeling the risk formation mechanism of autonomous vehicle operation.

[0136] Specifically, the main factors affecting the operational risks of unmanned vehicles include: (1) spatial traffic constraints: the narrower the road, the lower the obstacle avoidance capability; (2) obstacle density: the more targets there are, the higher the probability of potential conflicts; (3) motion interaction complexity: multi-directional motion increases uncertainty.

[0137] Based on the above analysis, road space constraint indicators were selected respectively. Obstacle density index and motion interaction complexity index The scenario risks are characterized. Three indicators (i.e., three different dimensions) describe risk characteristics from three different dimensions: spatial constraints, conflict probability, and behavioral complexity, exhibiting good complementarity and relative independence. Scene semantic features are extracted based on park map information and real-time perception data. Road spatial constraint indicators... It is used to describe the degree to which road space restricts the operation of autonomous vehicles, and is characterized by the ratio of the passable road width to the width of the autonomous vehicle:

[0138] ;

[0139] in, Indicates the current passable width of the road. This indicates the width of the autonomous vehicle's body. When... The smaller the value, the narrower the road space and the higher the risk level of the scenario. Obstacle density index Used to describe the density of obstacle distribution in the current scene:

[0140] ;

[0141] in, This indicates the number of obstacles detected within the current scene area. This represents the area of ​​the statistical region. Motion interaction complexity index. This is used to describe the complexity of interactions in different motion directions within a scene. In this invention, the motion directions are divided into 8 directional intervals (each 45° is a category), resulting in:

[0142] ;

[0143] Since stronger road space constraints lead to higher risks, their reciprocal form is used to positively represent the degree of space constraint. Based on the above characteristic parameters, a scenario risk index is constructed. Its calculation expression is:

[0144] ;

[0145] in, Normalized road space constraint index ; Normalized obstacle density index ; Normalized motion interaction complexity index The normalization of each indicator is as follows:

[0146] ;

[0147] ;

[0148] ;

[0149] The weighting coefficients satisfy the following constraints:

[0150] ;

[0151] Regarding parameter settings, statistical analysis was conducted based on historical operational data to examine the correlation between each indicator and actual risk events (including sudden deceleration, emergency braking, and near-collision events) to determine the contribution of each indicator in risk assessment. Road space constraints have the most direct impact on traffic capacity, and their weight range is [range missing]. Obstacle density has a secondary impact on path planning and obstacle avoidance complexity, and its weight range is [value missing]. Motion interaction complexity is considered an auxiliary risk factor, and its weight range is [range missing]. .

[0152] Specifically, in practical applications, the aforementioned weighting coefficients can be adaptively adjusted based on the characteristics of the park's operating environment and historical data statistics. Consistency processing is then performed while meeting normalization constraints to ensure the comparability and stability of the combined effects of various risk factors on the scenario risk index. Through this parameter setting method, a unified quantitative expression of key risk factors such as road space constraints, obstacle density, and motion interaction complexity is achieved, thereby constructing a scenario risk index with consistent dimensions and engineering interpretability. .

[0153] Specifically, in this invention, the weighting coefficients in the multi-objective risk assessment function and A hierarchical update mechanism is adopted, combining scene constraints, computational generation, and dynamic correction. This is based on the semantic category of the park scene where the obstacle is located. The range of values ​​for the security weight α is determined from the preset scenario weight table (as shown in Table 1):

[0154] ;

[0155] And obtain the corresponding traffic efficiency weight range:

[0156] ;

[0157] in:

[0158] ;

[0159] The range of values ​​for the traffic efficiency weight is set as follows: This avoids system oscillations or excessive reduction in security.

[0160] Specifically, based on the scenario risk index Determine the initial values ​​of the weighting coefficients. The scenario risk index reflects the overall risk level of the current scenario; the higher the value, the higher the scenario risk. Based on this, determine the initial values ​​of the safety weights using a normalized mapping function:

[0161] ;

[0162] And obtain the traffic efficiency weights based on the constraints: .

[0163] in, and These represent the minimum and maximum values ​​of the scenario risk index, respectively. and These represent the minimum and maximum values ​​of the safety weight, respectively. The weight ratio between safety and traffic efficiency can be adaptively determined based on the risk characteristics of different scenarios, thus ensuring that the risk assessment results match the safety and traffic requirements of the actual operating environment. During the system initialization phase, the semantic category of the park scenario where the obstacle is located is used... The initial values ​​of the weight coefficients are determined using the above mapping relationship:

[0164] ;

[0165] ;

[0166] in, It indicates the semantic category of the scene in the park, such as road passage area, intersection area, pedestrian mixed area or loading and unloading operation area, etc. This represents the weight of the security objectives during system initialization; This represents the weight of the traffic efficiency target during system initialization. The security and traffic efficiency requirements differ across semantic categories in different scenarios, therefore their initial weight ranges also differ. The weight settings for different scenarios are shown in Table 1.

[0167] A traffic efficiency feedback index is constructed. To achieve adaptive adjustment, this invention introduces a traffic efficiency index:

[0168] ;

[0169] in: Real-time average speed of driverless vehicles; : Expected speed in planning; .when At that time, the efficiency was normal; At times, traffic efficiency decreases; when When the value is much less than 1, congestion or excessive conservatism occurs. When traffic efficiency decreases, the risk assessment model (i.e., the multi-objective risk assessment model) is dynamically updated by adjusting the weight parameters to appropriately improve traffic efficiency while ensuring safety. This enables adaptive risk control for unmanned vehicles in different operating scenarios within the park.

[0170] ;

[0171] ;

[0172] in: To adjust the gain coefficient.

[0173] Specifically, as shown in Table 1, the weight of the goal of improving traffic efficiency is as follows in low-risk, high-traffic-demand scenarios such as main arterial routes. In high-risk scenarios such as intersections and densely populated areas, the weight of safety objectives should be increased. In space-constrained scenarios such as narrow passages, the system appropriately increases the weight of safety while suppressing the weight of traffic efficiency to prevent the expansion area from excessively occupying the effective passage space. Through this mechanism, when the system is in a state of low efficiency for a long period of time, the weight of the traffic efficiency target will be automatically increased to suppress the excessive expansion of the expansion area, thus achieving a dynamic balance between safety and efficiency.

[0174] Table 1 Preset Scene Weight Table

[0175] Scene semantic category Scene type Scene feature description Safety weight α Traffic efficiency weight β Explanation of Weight Setting Principles <![CDATA[S1]]> Main road Wide roads, good visibility, and low obstacle density 0.3～0.4 0.6～0.7 Prioritizing traffic efficiency while meeting basic safety constraints <![CDATA[S2]]> Branch road / subway The road is narrow, and pedestrians and vehicles share the road. 0.5～0.6 0.4～0.5 A balance between safety and efficiency <![CDATA[S3]]> Intersection area Multiple inflows, frequent interactions, and high uncertainty 0.7～0.8 0.2～0.3 Safety first <![CDATA[S4]]> Narrow passage Limited passage width and insufficient lateral space 0.6～0.7 0.3～0.4 Suppress lateral risks and avoid excessive occupation of passage space. <![CDATA[S5]]> Densely populated areas Pedestrians are highly random and frequently start and stop. 0.75～0.85 0.15～0.25 High safety margin preferred

[0176] Step 3: Calculation of basic expansion parameters and adaptive correction of motion state. Under the constraints of the multi-objective risk assessment results, the basic expansion parameters are first calculated based on the physical dimensions of the obstacle. The basic expansion scale can be expressed as:

[0177] ;

[0178] in: The equivalent geometric radius of the obstacle; Let be the equivalent radius of the autonomous vehicle; This is a preset minimum safe distance, used to ensure a basic safety margin for unmanned vehicles operating at low speeds within the park. Minimum safe distance The settings can be based on the braking performance of the autonomous vehicle, the system's perception and control latency, and the upper limit of the park's operating speed. The value ranges from 0.3m to 0.8m. Based on this, the motion state of the obstacle is introduced to adaptively correct the expansion scale. The corrected expansion scale is expressed as:

[0179] ;

[0180] in: This is a speed correction factor; This is the acceleration correction factor. When the obstacle is accelerating towards the target, the expansion scale is increased accordingly; when the motion risk is low, unnecessary amplification is suppressed.

[0181] Step 4: Scene semantics and direction-related non-central elliptical expansion modeling. After completing the adaptive expansion correction based on the obstacle motion state, this invention further introduces park scene semantic information, obstacle motion direction information, and risk forward shift mechanism to perform asymmetric modeling of the risk-occupied area of ​​the obstacle, thereby constructing a non-central elliptical expansion region that can reflect the dynamic risk distribution characteristics.

[0182] The risk forward shift mechanism refers to appropriately shifting the geometric center of the risk-occupied area of ​​an obstacle forward along its direction of movement, based on the obstacle's speed and direction of movement, so that the risk area more accurately reflects the high-risk space that the obstacle may occupy in the future time window.

[0183] Specifically, in this invention, by combining obstacle movement direction information and park scene semantic information, the elliptical scale parameters and center position of the risk area are jointly modeled: the direction of the major axis of the ellipse is determined by the direction of movement, so that the risk area has a larger coverage area in the direction of movement; the risk forward shift offset is calculated according to the obstacle speed and prediction time coefficient, and the center of the ellipse is translated from the current position of the obstacle to the predicted risk position; at the same time, the expansion scale is dynamically adjusted by the scene semantic correction coefficient, so that the risk area under different park scenarios can match the actual operating environment, thereby forming a non-central elliptical risk modeling method with directional correlation, spatial forward shift and scene adaptive characteristics.

[0184] It should be noted that in the complex mixed-traffic environment of a park, the risk distribution of dynamic obstacles exhibits significant directionality, with the risk in the area in front of the obstacle being significantly higher than in the lateral and rear areas. Traditional isotropic circular expansion models, due to uniform expansion in all directions, easily lead to excessive lateral space occupation, reducing passage efficiency. In contrast, this invention constructs a non-central elliptical risk region extending along the direction of movement (i.e., a non-central elliptical risk region model), effectively reducing ineffective expansion space while ensuring safety, and improving passage capacity in narrow passages and highly interactive scenarios.

[0185] By using the non-central elliptical expansion region as the risk-occupying space output for obstacles, a directionally sensitive and forward-looking safety constraint is provided for subsequent path planning and obstacle avoidance control. Based on this, for scene units... The scene semantic correction coefficient is not set manually, but is data-driven calibration based on historical operating data.

[0186] Specifically, in each scene unit Within this framework, a local risk field is constructed and a set of high-risk areas is extracted:

[0187] ;

[0188] Based on this, high-risk areas Parametric modeling is performed to represent it as an equivalent risk area model with directionality, and its spatial scale characteristic parameters are extracted.

[0189] Define candidate risk scale Furthermore, an optimization objective function constrained by safety and traffic efficiency is constructed. By comprehensively evaluating the risk coverage and space occupancy in historical samples, the optimal risk scale is determined.

[0190] ;

[0191] in, This represents the evaluation function, used to comprehensively measure the trade-off between insufficient risk coverage and excessive space occupancy under a given risk scale 𝑅. This represents the optimal risk scale determined by optimizing the objective function for the i-th scene unit based on historical data.

[0192] Evaluation function It is used to measure the overall effect under different risk scales, and its construction is based on the spatial distribution characteristics of risk areas in historical operational data.

[0193] Specifically, through the candidate risk scale The coverage relationship between the generated risk areas and historical high-risk areas is analyzed and evaluated from the following two aspects:

[0194] (1) Risk coverage level, used to characterize the degree to which high-risk areas are effectively covered;

[0195] (2) Space occupancy level, used to characterize the occupancy of the actual passable space by the risk area.

[0196] Based on this, the evaluation function This characterizes the comprehensive trade-off between the two types of indicators, used to depict the balance between safety and traffic efficiency achieved by the risk scale. In practical implementation, discrete scale traversal or numerical optimization methods can be used to solve the objective function. Based on the optimal risk scale, the scene semantic correction coefficient is determined:

[0197] ;

[0198] in, Indicates the first The basic risk scale is determined under the standard reference scenario, which is a typical traffic environment with low risk, low density and stable operation (such as the main road scenario), and is obtained by statistical analysis of historical operation data.

[0199] Specifically, in practical applications, different scenario units correspond to The value can be determined based on historical data statistics, and its range is as follows: .

[0200] During the operation of autonomous vehicles, considering the dynamic changes in scene states, a parameter update mechanism based on the operating state is introduced. To accurately represent the scene operating state, a dynamic parameter update mechanism based on scene topology is constructed.

[0201] Specifically, within the sliding time window, for each scene unit Statistical operational status indicators, including traffic efficiency indicators and the frequency of risk events, are used to construct a comprehensive operational status function. Its expression is:

[0202] ;

[0203] in, This represents the normalized traffic efficiency index. This represents the normalized frequency of risk events. The fusion weighting coefficient has a range of values. .

[0204] Risk event frequency sub-indicator Defined as:

[0205] ;

[0206] in, Indicates the scene unit within the sliding time window. The number of risk events detected in the data. This represents the total number of observations or interaction events within the corresponding time window. The ratio, after normalization, ranges from [value missing]. Risk events include, but are not limited to, behavioral events that reflect potential safety risks, such as close approach, sudden deceleration to avoid collisions, and path conflicts.

[0207] During the operation of the autonomous vehicle, the park's scene status changes dynamically over time. To ensure the real-time performance and stability of the scene correction coefficient, the aforementioned dynamic update mechanism based on the operating status uses a sliding time window to adjust the traffic efficiency index. Risk event frequency index It will be continuously updated.

[0208] During the operation of the autonomous vehicle, the dynamic changes in the scene state are considered, and the scene correction coefficient is dynamically adjusted based on the operating state to achieve adaptive adjustment of the risk area scale.

[0209] Based on this, the scene correction coefficient is dynamically adjusted:

[0210] ;

[0211] in, and These are the risk sensitivity coefficient and the efficiency adjustment coefficient, respectively. Indicates the first Scene unit at time The dynamic scene semantic correction coefficient. As the frequency of risk events increases... Increase the size of the area to expand the risk zone and improve safety; when traffic efficiency decreases... This reduces the risk area, thereby improving traffic efficiency; when the system is operating well... Keep it near the baseline value. To ensure parameter stability, [the following applies]. Apply interval constraints:

[0212] ;

[0213] Based on the semantic correction coefficients for the above scenarios, the dynamic inflation scale is corrected to obtain the equivalent inflation scale:

[0214] ;

[0215] in, This represents the equivalent inflation scale after scene semantic correction.

[0216] Construction of direction-related elliptical scale parameters: Considering that obstacles have a higher potential risk in front of their direction of movement, this invention uses a non-central elliptical risk region model to describe / represent the risk-occupied area (i.e., the non-central elliptical risk region), in which the major axis of the ellipse is consistent with the direction of obstacle movement, and the minor axis is perpendicular to the direction of movement.

[0217] Specifically, the major and minor axis parameters of the ellipse are defined as follows:

[0218] ;

[0219] ;

[0220] in: The length of the major axis of the ellipse representing the direction of movement along the obstacle; The length of the minor axis of the ellipse perpendicular to the direction of the obstacle's movement; Indicates the speed of the obstacle; This represents the forward risk amplification factor; This is the risk prediction time window, used to characterize the system's ability to make forward-looking assessments of the future movement trends of obstacles.

[0221] Specifically, in this invention, the forward risk amplification factor... The settings can be configured based on the movement type of the obstacles and the park's operational experience. For highly mobile obstacles such as pedestrians and bicycles, The value range is 0.1 to 0.3 for low-speed motor vehicles. The value ranges from 0.05 to 0.15. This allows the elliptical region to have a larger coverage area in front of the obstacle's direction of movement, reflecting the space the obstacle may occupy in the future.

[0222] Center offset modeling of non-central elliptical risk regions: It should be noted that the elliptical region constructed in this invention describes the high-risk occupied space of an obstacle over a future period, rather than the current geometric shape of the obstacle. Therefore, the current position of the obstacle is not necessarily located at the geometric center of the ellipse. To address this, an elliptical center offset based on the obstacle's motion state is introduced to shift the elliptical risk region forward in a direction-related manner. The elliptical center position is defined as:

[0223] ;

[0224] in: Indicates the current position of the obstacle; Indicates the geometric center of the elliptical risk region after offset; This represents the risk offset along the direction of obstacle movement. To achieve unified modeling, detected obstacles are abstracted as point mass objects.

[0225] Specifically, the spatial contour information of obstacles is obtained through the vehicle-mounted environmental perception system, and their geometric center or bounding box center is extracted as the current position. Then, combined with the autonomous vehicle's positioning information, it is converted to the park's global coordinate system as a spatial reference for risk modeling. Obstacle movement direction angle. It is obtained by calculating the position difference at consecutive time intervals:

[0226] ;

[0227] in, and These represent the positions of the obstacle in the global coordinate system at the current and previous moments, respectively. Furthermore, the motion direction angle can be combined with the geometric orientation of roads or passageways in the park map for consistency verification, thereby improving the stability and reliability of the direction estimation. The offset is defined as:

[0228] ;

[0229] in: The speed of the obstacle's movement; For risk prediction time scale; For the direction angle of motion A defined unit direction vector; For security weights; As the traffic efficiency weight, and satisfying ; , This is the weighting adjustment coefficient; It is the base offset ratio coefficient.

[0230] Regarding parameter selection, The value ranges from 0.5 to 1.2, and is used to control the proportional relationship between the forward movement distance of the risk area and the predicted movement distance of the obstacle; The settings are based on the system response time and the autonomous vehicle's operating speed, typically ranging from 0.5 to 2 seconds. When the scenario risk is high, the safety weight is increased. Increase the offset; when the traffic efficiency requirement is high, increase the efficiency weight. By appropriately reducing the offset, the risk area can be adaptively shifted forward.

[0231] Prediction time scale This represents the system's forward prediction time for the future movement trend of obstacles, and its value can be set according to the operating speed of the autonomous vehicle and the system response time.

[0232] The offset mechanism shifts the elliptical risk area forward slightly in the direction of obstacle movement, thereby more accurately reflecting the spatial location that the obstacle may occupy in the future.

[0233] Obtaining ellipse scale parameters , and the center of the ellipse After that, the The risk area occupied by the obstacle Represented as:

[0234] ;

[0235] in: Represents any point in space; Represents the transpose of a matrix; This represents a symmetric positive definite matrix that describes the shape and orientation of an ellipse.

[0236] To ensure that the major axis of the ellipse is aligned with the direction of obstacle movement, a method based on the direction angle is introduced. Rotation matrix:

[0237] ;

[0238] Based on this, elliptical matrix Defined as:

[0239] ;

[0240] in: The major axis of the ellipse representing the direction of movement along the obstacle; The minor axis of the ellipse is perpendicular to the direction of motion; Represents a rotation matrix formed by the directions of movement of obstacles; This represents the transpose of the rotation matrix. During the operation of the unmanned vehicle in the park, the environmental perception system detects dynamic obstacles in real time and outputs the spatial position, velocity, and direction of motion of the obstacles. This information is input into the non-central elliptical risk region model to calculate the risk-occupied area corresponding to each obstacle. The final expansion region is shown as Figure 4 As shown.

[0241] Specifically, to enable the non-central elliptical risk region model to more accurately reflect the impact range of dynamic obstacles in the park on the operational safety of unmanned vehicles, this invention further constructs a model parameter calibration method based on historical operational data. This method uses data-driven determination of key parameters in the model (i.e., the non-central elliptical risk region model) through risk field construction, ellipse fitting, and parameter optimization. An obstacle risk field dataset is constructed based on historical environmental perception data collected during the operation of unmanned vehicles in the park. The environmental perception data is acquired by the onboard perception system, including obstacle information output by LiDAR, visual sensors, and the positioning module, specifically including the spatial location of obstacles. Speed ​​of movement , direction angle of motion This includes obstacle category information. It also incorporates the autonomous vehicle's own operational status data, such as vehicle position, planned trajectory, and historical near-collision or obstacle avoidance behavior records. Based on this, a discrete risk grid is constructed in the surrounding spatial area, using the obstacle's current position as a reference. A risk weight is calculated for each grid point, thus forming a two-dimensional risk field.

[0242] ;

[0243] in, This represents the local two-dimensional risk field corresponding to the i-th obstacle; Indicates the location of spatial grid points. This indicates the risk weight corresponding to that location. Risk values ​​exceeding a preset threshold are extracted from the risk field. High-risk point set:

[0244] ;

[0245] High-risk point set These samples serve as supervised samples for training the model (i.e., the non-centered elliptical risk region model), characterizing the spatial distribution of high-risk areas during actual operation. An ellipse fit is applied to this set to obtain ellipse parameters that cover the main high-risk regions. The ellipse parameters include the ellipse center... Major axis length minor axis length and direction angle Ellipse fitting can be achieved using the least squares method, principal component analysis, or optimization-based isoparameter estimation methods. To ensure that the elliptical region covers the main risk areas while avoiding excessive expansion that could affect passage space, this invention constructs the following training objective function:

[0246] ;

[0247] in:

[0248] ;

[0249] Used to measure the extent to which high-risk points are not covered by the ellipse.

[0250] ;

[0251] This is used to constrain the area of ​​the ellipse to prevent excessive expansion of the risk region. Among them, This is a weighting coefficient used to balance risk coverage and passage space utilization. Represents the overall training objective function; This indicates a penalty for high-risk points not being effectively covered by the ellipse. This represents the elliptical area constraint term, used to suppress excessive expansion of the region. By minimizing the objective function, key parameters in the non-central elliptical risk region model are optimized. The optimized parameters constitute the parameter set:

[0252] ;

[0253] in, This is the forward risk amplification factor. For risk prediction time window, This is the risk shift time coefficient. This represents the set of parameters to be optimized. Parameter optimization can be achieved using existing numerical optimization methods, but iterative updates using the Adam adaptive gradient optimization algorithm are preferred.

[0254] ;

[0255] Among them, w This represents the set of parameters to be optimized. Indicates the learning rate. and These represent the first-order moment estimate and the second-order moment estimate of the gradient, respectively. To prevent the use of tiny constants with a denominator of zero, the update methods for the first-order moment estimate and the second-order moment estimate are as follows:

[0256] ;

[0257] ;

[0258] in, and These are the decay coefficients. Regarding parameter settings, the learning rate... The value range is 0.001 to 0.01. Take 0.9, Take 0.999, Pick Weighting coefficient The value ranges from 0.1 to 1.0, and the risk threshold is... Determined based on historical risk distribution statistics. This is a first-moment estimate of the gradient; This is the second moment estimate of the gradient; loss function Regarding parameters The gradient of the loss function. By setting the parameters as described above, optimization efficiency can be improved while ensuring convergence stability. Once the loss function converges or reaches the preset number of iterations, the set of parameters that minimizes the loss function is obtained. This data is then used as the fundamental parameters for the non-central elliptical risk region model and applied during the online operation of autonomous vehicles. Building upon the aforementioned fundamental parameter calibration based on the risk field distribution, this invention introduces a parameter calibration mechanism based on historical operational behavior to further enhance the model parameters' adaptability to actual operational behavior.

[0259] Specifically, based on historical trajectory data of autonomous vehicles, the trajectories are labeled with safety criteria, distinguishing between safe passage trajectories and high-risk trajectories. A non-central elliptical risk region model is used to assess the risk response of different trajectories during operation, and risk consistency constraints are constructed to ensure the model has a consistent risk discrimination capability for trajectories of different safety levels. Based on this, the parameter set... Calibration and adjustments are performed to ensure that the parameters not only meet the basic spatial risk modeling requirements but also better reflect actual operational behavior characteristics. To guarantee parameter stability, constraints are introduced between the calibration and basic parameters to prevent deviations from reasonable ranges. The optimization process can also be implemented using existing numerical optimization methods. During the operation of the autonomous vehicle, the park scene state changes dynamically over time. To ensure the real-time performance and stability of the scene correction coefficients, the aforementioned dynamic update mechanism based on operational state can use a sliding time window to adjust the traffic efficiency index. It will be continuously updated.

[0260] Specifically, when a decrease in traffic efficiency or an increase in the frequency of risk events is detected, the running state function... The corresponding reduction will thus decrease the scene correction coefficient. The model is automatically updated based on existing data to expand the scope of risk coverage; when the scenario is running smoothly, When the value approaches 1, the correction coefficient approaches the initial value, thus avoiding unnecessary expansion. This enables adaptive adjustment of the scene correction coefficient under different operating conditions, improving the model's adaptability to dynamic environments.

[0261] Specifically, to ensure the continuity and stability of parameter changes, the traffic efficiency index is smoothed using a sliding time window statistical method or an exponential weighted average. During real-time operation, the system adjusts the data based on the real-time positions of obstacles. ,speed and direction of movement Calculate ellipse parameters Construct obstacle risk zones This serves as a spatial safety constraint for path planning and obstacle avoidance decisions. During the path planning phase, the autonomous vehicle planning algorithm considers the non-central elliptical risk area... As a space safety constraint, it is introduced into the existing trajectory optimization framework for modeling and solving. A comprehensive risk index is obtained based on a multi-objective risk assessment function. It is used to rank the risks of different obstacles and use them as risk weights in path planning to guide autonomous vehicles to prioritize avoiding high-risk areas.

[0262] It should be noted that path planning methods may include covariance optimization (CHOMP) or model predictive control (MPC), etc. The key to this invention lies in the construction of the risk region and its constraint expression in the planning.

[0263] Step 5: Dynamic Expansion Region Update and Output. During the operation of the autonomous vehicle, the controller dynamically updates the risk area according to a preset time period or a trigger mechanism based on changes in obstacle state. Specifically, it periodically acquires the latest state information of obstacles and detects their motion state and scene semantic category. The changes; when the update conditions are met, steps 1 to 4 are re-executed to adjust the risk model parameters and scenario correction coefficients. The calculation is updated. Based on this, the dynamic risk areas corresponding to each obstacle at the current moment are generated. Furthermore, dynamic evolution modeling of the risk space is achieved through continuous time updates or smooth replacements.

[0264] Specifically, in multi-obstacle scenarios, the risk areas of each obstacle are spatially fused to obtain the overall risk-occupied space:

[0265] ;

[0266] in, The space occupied by the overall risk at time t is represented. When risk areas of different obstacles overlap, the maximum risk fusion principle based on risk value is used for processing. For any point in the space... Its final risk characterization is defined as:

[0267] ;

[0268] in, Let i be the risk distribution corresponding to the i-th obstacle. For the i-th obstacle at position ( ) and the risk intensity function at time t.

[0269] Ultimately, the integrated overall risk distribution or its corresponding risk-occupied area is output as a unified environmental constraint to the path planning and decision-making module of the autonomous vehicle system, providing risk constraints and cost inputs for path generation.

[0270] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks, characterized in that, include: S1. Obtain multi-source state information of obstacles in the park environment, and construct an obstacle state model based on the multi-source state information. The multi-source state information includes at least the physical size, position, speed, acceleration, relative motion state with the unmanned vehicle, and semantic category of the park scene where the obstacle is located. S2. Construct a multi-objective risk assessment model based on the obstacle state model, which includes safety objectives and traffic efficiency objectives, and use the multi-objective risk assessment model to assess the comprehensive risk of obstacles to the unmanned vehicle's operating space; S3. Under the constraints of the comprehensive risk assessment results, the basic expansion parameters are calculated based on the physical dimensions of the obstacle, and the basic expansion parameters are adaptively corrected by combining the obstacle's movement speed, acceleration, and relative motion state with the unmanned vehicle. S4. Based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, the corrected basic expansion parameters are adjusted asymmetrically in a direction-dependent manner, and an elliptical dynamic expansion region is constructed. S5. Update the elliptical dynamic expansion region in real time according to a preset cycle or a trigger mechanism based on the change of obstacle state, and output the updated elliptical dynamic expansion region as the risk-occupied space of the obstacle.

2. The multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks according to claim 1, characterized in that, The acquisition of multi-source state information of obstacles in the park environment includes at least the physical size, position, speed, acceleration, relative motion state with the autonomous vehicle, and semantic category of the park scene in which the obstacle is located. Constructing an obstacle state model based on this multi-source state information includes: S11. Import the park digital map into the unmanned vehicle controller or the device that communicates with the unmanned vehicle controller in advance, use a unified global coordinate system to describe the park map, and establish a unified spatial coordinate reference for the unmanned vehicle's own position, obstacle position and scene semantic information. S12. Acquire data from satellite positioning, inertial measurement unit, and odometer outputs. After time synchronization processing, use extended Kalman filter algorithm for fusion processing. Combined with semantic constraints of the park map and key position correction mechanism, obtain the global position coordinates, heading angle, and linear velocity of the unmanned vehicle in the global coordinate system. S13. Use vehicle-mounted lidar, cameras and millimeter-wave radar to detect and identify obstacles in the park environment, and obtain the physical size and location information of the obstacles; S14. Calculate the velocity and acceleration of the obstacle using the time difference method based on the positional changes of the obstacle in continuous time frames. S15. Calculate the relative distance between the obstacle and the unmanned vehicle based on the obstacle's position and the unmanned vehicle's position; calculate the relative speed between the obstacle and the unmanned vehicle based on the obstacle's speed and the unmanned vehicle's speed; and determine the obstacle's motion direction angle based on the continuous change in the obstacle's position. S16. Use a point-in-polygon detection algorithm to perform spatial semantic determination on the current position of the unmanned vehicle, match the current position of the unmanned vehicle with the scene semantic region in the park map, and determine the semantic category of the park scene where the obstacle is located. S17. Integrate the physical dimensions, position, speed, acceleration, direction angle of movement of obstacles, relative distance between obstacles and autonomous vehicles, relative speed between obstacles and autonomous vehicles, and semantic categories of the park scene to form complete multi-source state information and establish a unified obstacle state model.

3. The multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks according to claim 1, characterized in that, The multi-objective risk assessment model, which incorporates safety and traffic efficiency objectives based on the obstacle state model, is used to assess the comprehensive risk posed by obstacles to the autonomous vehicle's operating space. S21. Perform feature analysis and normalization on the obstacle state vector to construct a risk feature representation with unified dimensions; S22. Based on the relative distance and relative speed between the obstacle and the unmanned vehicle, calculate the safety risk index, and based on the area of ​​the obstacle risk area and the area of ​​the unmanned vehicle's local planning area, calculate the space occupancy index. Normalize the safety risk index to ensure that the safety risk index and the space occupancy index are consistent in dimension. S23. Based on the semantic annotation of the park map, divide the scene units, extract the road space constraint index, obstacle density index, and motion interaction complexity index and normalize them. Based on the weighted fusion of the normalized road space constraint index, obstacle density index, and motion interaction complexity index, obtain the scene risk index. S24. Determine the value range of safety weight and traffic efficiency weight based on the semantic category of the park scene where the obstacle is located, and calculate the initial value of the weight based on the scene risk index and the normalized mapping function. S25. Calculate the traffic efficiency index based on the real-time average speed of the unmanned vehicle and the planned expected speed, and dynamically update the safety weight and traffic efficiency weight based on the traffic efficiency index so that the sum of the safety weight and traffic efficiency weight remains constant. S26. Using a weighted summation method, the safety risk indicators and space occupancy indicators are integrated with the updated safety weights and traffic efficiency weights to obtain a comprehensive risk assessment result of obstacles on the operating space of unmanned vehicles.

4. The multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks according to claim 3, characterized in that, The formula for calculating the security risk index is as follows: ； In the formula, This represents the safety risk index of the i-th obstacle; Indicates the distance between the driverless vehicle and the obstacle; Represents relative velocity; This represents the distance attenuation scale parameter; This indicates the speed adjustment coefficient.

5. A multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks, as described in claim 1, is characterized in that... The process of calculating basic expansion parameters based on the physical dimensions of obstacles, under the constraints of the comprehensive risk assessment results, and adaptively correcting these parameters by considering the obstacle's speed, acceleration, and relative motion with the autonomous vehicle, includes: S31. Calculate the basic expansion scale of the obstacle based on the equivalent geometric radius of the obstacle, the equivalent radius of the unmanned vehicle, and the preset minimum safety distance; S32. Based on the braking performance of the unmanned vehicle, the system perception and control delay, and the upper limit of the park's operating speed, a preset minimum safe distance is set. S33. Adaptively adjust the basic expansion scale by combining the relative velocity and acceleration of the obstacle with the corresponding velocity correction coefficient and acceleration correction coefficient. S34. Based on the correction relationship, obtain the dynamic expansion scale that adapts to the motion state of the obstacle, and increase the expansion scale when the obstacle accelerates and approaches, and suppress the expansion scale amplification when the motion risk is low.

6. The multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks according to claim 1, characterized in that, The process of adjusting the modified base expansion parameters asymmetrically based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, and constructing an elliptical dynamic expansion region, includes: S41. Based on historical operation data, calibrate the initial scene semantic correction coefficients corresponding to the semantic categories of each scene, and dynamically adjust the scene semantic correction coefficients according to the real-time operation status of the scene. S42. The dynamic expansion scale after motion state correction is corrected by using the dynamically adjusted scene semantic correction coefficient to obtain the equivalent expansion scale. S43. Determine the direction of the major axis of the ellipse based on the direction of obstacle movement, and calculate the parameters of the major and minor axes of the ellipse by combining the equivalent expansion scale, obstacle movement speed, forward risk amplification coefficient and risk prediction time window. S44. Based on the obstacle's movement speed, the risk prediction time scale, and the weights of safety and traffic efficiency, calculate the risk forward shift along the movement direction and determine the geometric center of the elliptical risk region. S45. Construct a rotation matrix based on the major axis, minor axis parameters, geometric center, and motion direction angle of the ellipse, and generate the non-central elliptical risk-occupying area corresponding to the obstacle; S46. The non-central elliptical risk-occupied area is output as the risk-occupied space of the obstacle, and is used as the spatial safety constraint for path planning and obstacle avoidance control.

7. A multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks, as described in claim 6, is characterized in that... The process of calibrating the initial scene semantic correction coefficients corresponding to each scene semantic category based on historical operational data, and dynamically adjusting the scene semantic correction coefficients according to the real-time operational status of the scene, includes: S411. Construct a local risk field for each scene unit and extract a set of high-risk areas. Perform parametric modeling on the high-risk areas and extract spatial scale feature parameters. S412. Construct an optimization objective function constrained by safety and traffic efficiency, comprehensively evaluate the degree of risk coverage and space occupancy, and determine the optimal risk scale for each scenario unit. S413. Based on the optimal risk scale and the basic risk scale under the standard reference scenario, determine the initial scenario semantic correction coefficients corresponding to the semantic categories of each scenario. S414. Within the sliding time window, statistically analyze the traffic efficiency index and the frequency of risk events of the scene unit, and construct a comprehensive operation status function; S415. Based on the comprehensive operating state function, risk sensitivity coefficient and efficiency adjustment coefficient, the initial scene semantic correction coefficient is dynamically adjusted, and the adjusted scene semantic correction coefficient is subject to interval constraints.

8. A multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks, as described in claim 6, is characterized in that... The process of calculating the risk forward shift along the direction of movement and determining the geometric center of the elliptical risk region based on the obstacle's movement speed, the risk prediction time scale, and the weights of safety and traffic efficiency includes: S441. Calculate the obstacle's motion direction angle based on the position difference of the obstacle at consecutive moments, and verify the consistency of the motion direction angle in conjunction with the geometric orientation of the park's roads. S442. Calculate the risk forward shift along the direction of movement based on the obstacle's movement speed, risk prediction time scale, basic offset ratio coefficient, and safety and traffic efficiency weights. S443. Superimpose the current position of the obstacle with the forward offset of the risk to obtain the geometric center position of the non-central elliptical risk area.

9. A multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks, as described in claim 1, is characterized in that... The step of updating the elliptical dynamic expansion region in real time according to a preset period or a triggering mechanism based on changes in obstacle state, and outputting the updated elliptical dynamic expansion region as the risk-occupied space of the obstacle, includes: S51. By periodically acquiring the latest state information of obstacles, detect changes in the movement state of obstacles and the semantic category of the scene in which they are located; S52. If a change in obstacle status or scene semantics is detected, the parameters of the multi-objective risk assessment model and the scene semantic correction coefficients are recalculated to update the risk area. S53. Based on the updated multi-objective risk assessment model parameters and scenario semantic correction coefficients, regenerate the corresponding dynamic risk area to achieve dynamic evolution of the risk space. S54. Smooth the dynamic risk area to ensure the continuity and rationality of the risk space, and use the principle of maximum risk fusion to spatially fuse the smoothed dynamic risk area in order to complete the calibration of the overall risk space. S55. The completed risk space output will serve as the environmental constraint basis for path planning and decision-making.

10. A multi-objective adaptive dynamic obstacle expansion system for complex mixed-traffic scenarios in industrial parks, used to implement the multi-objective adaptive dynamic obstacle expansion method for complex mixed-traffic scenarios in industrial parks as described in any one of claims 1-9, characterized in that, The system includes: The state modeling module is used to acquire multi-source state information of obstacles in the park environment and construct an obstacle state model based on the multi-source state information. The multi-source state information includes at least the physical size, position, speed, acceleration, relative motion state with the unmanned vehicle, and semantic category of the park scene where the obstacle is located. The scenario risk assessment module is used to construct a multi-objective risk assessment model based on the obstacle state model, which includes safety objectives and traffic efficiency objectives, and to use the multi-objective risk assessment model to assess the comprehensive risk of obstacles to the operating space of unmanned vehicles. The adaptive correction module is used to calculate the basic expansion parameters based on the physical dimensions of the obstacle under the constraints of the comprehensive risk assessment results, and to adaptively correct the basic expansion parameters by combining the obstacle's movement speed, acceleration, and relative motion state with the unmanned vehicle. The elliptical risk modeling module is used to perform direction-related asymmetric adjustments on the corrected basic expansion parameters based on the semantic category of the park scene where the obstacle is located and the relative relationship between the obstacle's movement direction and the autonomous vehicle's driving direction, and to construct an elliptical dynamic expansion region. The risk area output module is used to update the elliptical dynamic expansion area in real time according to a preset period or based on the triggering mechanism of obstacle state changes, and output the updated elliptical dynamic expansion area as the risk-occupied space of the obstacle.

Images