A multi-vehicle interaction test method based on reachable space compression

Abstract

The application discloses a kind of based on reachable space compression multi-vehicle interaction test method, comprising: obtaining the vehicle state information of main vehicle and multiple interactive vehicles in multi-vehicle interaction test scene, the reachable area of main vehicle in prediction period is predicted based on vehicle kinematics model, and angular and radial discrete are carried out, and fan-shaped reachable space grid is formed;The motion trajectory of each interactive vehicle is predicted, the space occupation area is determined according to the vehicle contour of each predicted time, and the space occupation area is mapped to fan-shaped reachable space grid, the nearest occupation radial distance in each angular sector from main vehicle is determined, and the remaining reachable space of main vehicle after being occupied is obtained;Determine reachable space compression index, and determine the test result of multi-vehicle interaction test scene according to the comparison result of reachable space compression index and preset boundary criterion.The application quantifies the influence of multi-vehicle interaction on the reachable space of main vehicle from the angle of space occupation.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle testing technology, specifically relating to a multi-vehicle interaction testing method based on reachability space compression. Background Technology

[0002] Simulation testing, as a crucial technical means for the research and verification of autonomous driving systems, can efficiently construct complex traffic scenarios and repeatedly execute tests without the need for real roads and vehicles, thereby evaluating the performance of autonomous driving systems in perception, decision-making, and control. Currently, simulation testing methods focusing on vehicle interaction capabilities mainly include rule-driven simulation, data-driven simulation, and spatial evaluation methods based on lane or trajectory constraints.

[0003] Rule-driven simulation methods typically generate test scenarios by manually setting behavioral rules or scripts for traffic participants, such as pre-setting the acceleration / deceleration behavior, entry timing, or driving trajectory of target vehicles. While these methods are simple to implement and highly controllable, the behavior patterns of the target objects are heavily reliant on human experience, and the interaction process lacks dynamic game-theoretic characteristics. In complex interaction scenarios, rule-driven behavior often exhibits significant predictability; the tested vehicle can respond using fixed strategies, making it difficult to expose the potential flaws of the decision-making algorithm in extreme or atypical situations.

[0004] Data-driven simulation methods collect real road data and perform trajectory playback or statistical modeling to reproduce the interactive behavioral characteristics in real traffic flow. While these methods have certain advantages in terms of distribution consistency and behavioral realism, their testing capabilities are significantly limited by the data coverage. High-risk, low-probability extreme interaction scenarios occur infrequently in real data, making it difficult to systematically cover them through pure data playback. Furthermore, the behavior of traffic participants in trajectory playback simulations is usually not adjustable online, thus limiting their effectiveness in capability boundary exploration and stress testing.

[0005] Furthermore, some simulation testing methods have introduced space occupancy-based evaluation metrics to quantify the impact of the target vehicle on the driving space of the tested vehicle. However, existing space compression or occupancy assessment methods are mostly derived from same-direction driving scenarios such as ramp merging and lane changing, and usually use lane lines, lane center lines, or lane topology as spatial constraints. In scenarios such as intersections, unprotected left turns, and multi-directional flow, the differences in vehicle movement directions are significant, lane boundaries are unclear or even missing, and the applicability of the above methods is limited.

[0006] For example, Chinese invention patent application CN122675637A discloses a method for generating autonomous driving interaction test scenarios based on a bird's-eye view. This method generates candidate future control sequences and candidate future trajectories for multiple traffic participants based on bird's-eye view images, vehicle states, historical trajectories, and test targets. It then filters the target future trajectory by combining constraints such as physical feasibility, maps, collisions, traffic rules, and test indicators. The scheme also constructs multi-directional interaction pressure indicators based on the time difference of arrival at conflict points, the minimum distance between vehicles, and the coverage of sectors in the direction of the traffic participants.

[0007] However, its directional coverage mainly reflects the distribution of traffic participants in different directions, making it difficult to further characterize the distance differences between different traffic participants in each direction and their actual degree of obstruction to the movable space of the tested vehicle; in addition, the vehicle size in this scheme is mainly used for safety radius and collision constraint evaluation, and there is still a lack of direct quantification of the superimposed impact of the geometric occupancy of multiple traffic participants on the overall movable space of the tested vehicle.

[0008] On the other hand, existing simulation evaluation systems still largely rely on information such as vehicle mass points, trajectory lines, distances between vehicles, or conflict times to evaluate the degree of interaction at the spatial modeling level, while insufficiently considering the impact of the actual space occupied by vehicle geometry. In multi-vehicle interaction scenarios, even if different test scenarios have similar minimum distances between vehicles, conflict times, or directional coverage, differences in the location, size, and spatial distribution of each vehicle can lead to significant variations in the actual remaining motion space of the tested vehicles, which existing evaluation indicators cannot uniformly represent.

[0009] In summary, existing simulation testing schemes still have the following shortcomings in complex interaction scenarios: the generation of interactive behaviors depends to some extent on manual rules or historical data; some spatial evaluation methods are highly dependent on lane structure or preset paths, and their applicability is limited in laneless and multi-directional interaction scenarios; existing evaluation indicators based on inter-vehicle distance, conflict time, or directional coverage are difficult to fully reflect the impact of vehicle geometric occupancy on the movable space of the tested vehicle; and in multi-vehicle interaction scenarios, the spatial obstruction and their superimposed effects generated by different traffic participants are also difficult to quantify uniformly. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a multi-vehicle interaction testing method based on reachable space compression.

[0011] The objective of this invention can be achieved through the following technical solutions: This invention provides a multi-vehicle interaction testing method based on reachability space compression, comprising the following steps: Obtain vehicle status information of the main vehicle and multiple interacting vehicles in a multi-vehicle interaction test scenario; Based on the vehicle state information and vehicle kinematics model of the main vehicle, the reachable area of ​​the main vehicle during the prediction period is predicted, and the reachable area is discretized along the angular and radial directions to obtain a fan-shaped reachable space grid composed of multiple angular sectors and multiple radial intervals. Based on the vehicle state information of each interactive vehicle and the vehicle kinematic model, the predicted trajectory of each interactive vehicle within the predicted time period is predicted, and the spatial occupancy area at each predicted time is determined based on the vehicle outline of each interactive vehicle at each predicted time of the predicted trajectory. Each of the space-occupied areas is mapped to the sector-shaped reachable space grid, and for each angular sector, the nearest occupancy radial distance to the main vehicle is determined in each of the space-occupied areas mapped to that angular sector, and the remaining reachable space of the main vehicle after being occupied by the interactive vehicle is determined based on each of the occupancy radial distances. The reachability compression index of the main vehicle is determined based on the reachable space area of ​​the sector-shaped reachable space grid when it is not occupied by the interactive vehicle and the area of ​​the remaining reachable space. The test results of the multi-vehicle interaction test scenario are determined based on the comparison results between the reachable space compression index and the preset boundary criteria.

[0012] Furthermore, the vehicle status information includes the vehicle's position, longitudinal speed, and heading angle; The control inputs to the vehicle kinematics model include longitudinal acceleration and steering input; The vehicle kinematics model adopts a discrete-time single-track model. Based on the vehicle's position, longitudinal velocity, and heading angle at the current sampling time, as well as the longitudinal acceleration, steering input, vehicle wheelbase, and sampling period, the vehicle's position, longitudinal velocity, and heading angle at the next sampling time are determined, and the vehicle state sequence is recursively obtained within the prediction period according to the sampling period.

[0013] Furthermore, the step of predicting the reachable area of ​​the main vehicle within the prediction period based on the vehicle state information and vehicle kinematic model, and discretizing the reachable area along the angular and radial directions to obtain a fan-shaped reachable space grid composed of multiple angular sectors and multiple radial intervals, specifically includes: A local polar coordinate system for the main vehicle is established with the current centroid position of the main vehicle as the pole and the current heading of the main vehicle as the zero angle direction. The maximum reachable radial distance of the reachable area is determined based on the current longitudinal speed of the main vehicle, the preset upper limit of longitudinal acceleration, and the predicted time period. The maximum lateral angle of the reachable area is determined based on the minimum turning radius that characterizes the steering capability of the main vehicle; The angular sector is divided into multiple equal-angle sectors within the angular range defined by the maximum lateral angle, and the radial interval is divided into multiple radial intervals within the radial range defined by the maximum reachable radial distance, to obtain the sector-shaped reachable space grid.

[0014] Furthermore, the step of predicting the predicted trajectory of each interactive vehicle within the predicted time period based on the vehicle state information and the vehicle kinematic model, and determining the spatial occupancy area at each predicted time based on the vehicle outline corresponding to each predicted time of the predicted trajectory, specifically includes: For each of the interactive vehicles, based on the vehicle state information and control input of the interactive vehicle at the current sampling time, the vehicle kinematics model is used to recursively obtain the predicted states corresponding to multiple prediction times within the prediction period, and the predicted trajectory is formed by the multiple predicted states. The vehicle's external attitude at the corresponding prediction time is determined based on its position and heading angle in each of the predicted states. Determine the positions of the center point and four corner points of the vehicle outline in the vehicle's local coordinate system, and transform the center point and four corner points to the global coordinate system according to the position and heading angle at the corresponding prediction time; The spatial occupancy area at the corresponding prediction time is determined based on the center point and four corner points transformed to the global coordinate system.

[0015] Furthermore, the step of mapping each of the space-occupied areas to the sector-shaped reachable space grid, and for each angular sector, determining the nearest occupancy radial distance to the main vehicle in each of the space-occupied areas mapped to that angular sector, and determining the remaining reachable space of the main vehicle after it is occupied by the interacting vehicle based on each of the occupancy radial distances, specifically includes: Establish a local polar coordinate system for the main vehicle with its current position as the pole and its current heading as the zero-angle direction; Transform the vehicle outline points of each interactive vehicle at each prediction time to the local polar coordinate system of the main vehicle to obtain the angular position and radial distance of each vehicle outline point relative to the main vehicle. Based on the angular position and radial distance of each vehicle outline point, the vehicle outline points located within the reachable area are mapped to the corresponding angular sector and radial interval; For each angular sector, the radial distances of all interactive vehicles mapped to that angular sector at all prediction times are compared, and the minimum radial distance is determined as the occupied radial distance of that angular sector; when there is no vehicle outline point in an angular sector, the maximum reachable radial distance is determined as the radial reachable distance of that angular sector. The remaining reachable space is determined by using the occupied radial distance or the radial reachable distance corresponding to each angular sector as the radial boundary of that angular sector.

[0016] Furthermore, determining the reachability compression index of the main vehicle based on the reachability area of ​​the sector-shaped reachability grid when it is not occupied by the interactive vehicle and the area of ​​the remaining reachability space specifically includes: Each of the angular sectors is set to have the same angular width. Based on the angular width of the angular sector, the maximum reachable radial distance of the reachable area, and the occupied radial distance corresponding to each angular sector, the reachable space area and the remaining reachable space area are determined respectively. The areas of the reachable space and the remaining reachable space are defined by the following formulas: The reachable space compression index is determined based on the reachable space area and the remaining reachable space area. The reachable space compression index is calculated using the following formula: in, This represents the reachable space area of ​​the sector-shaped reachable space grid when it is not occupied by the interactive vehicle; This represents the area of ​​the remaining reachable space; Indicates the number of angular sectors; Indicates the angular width of each of the aforementioned angular sectors; This represents the maximum reachable radial distance of the reachable region; Indicates the first The occupied radial distance corresponding to the angular sector, when the angular sector... When the corner sector is not occupied Take the maximum achievable radial distance; This indicates a pre-set positive stability constant; This represents the reachable space compression index.

[0017] Furthermore, before determining the test result of the multi-vehicle interaction test scenario based on the comparison result of the reachable space compression index and the preset boundary criterion, the method further includes: Obtain the target path information of the main vehicle and each of the interactive vehicles, and determine potential conflict points based on the intersection relationship between the target path of the main vehicle and the target paths of each of the interactive vehicles. The path distances from the main vehicle and each of the interactive vehicles to the potential conflict point along the corresponding target path are determined respectively. When the path distance corresponding to the main vehicle is not greater than a first preset distance and the path distance corresponding to at least one of the interactive vehicles is not greater than a second preset distance, the current reachable space compression index is marked as valid. Based on the potential conflict point, the blocking area is determined according to the preset blocking distance threshold. When the main vehicle and any of the interactive vehicles are both located in the same blocking area at the same test time, the interactive vehicle is identified as the blocking interactive vehicle, and an interactive blocking evaluation index corresponding to the blocking interactive vehicle is generated. Based on the vehicle status information of the main vehicle and the blocking interaction vehicle, as well as their respective target path information, the estimated arrival time of the two vehicles to the same potential conflict point is determined. The absolute difference between the two estimated arrival times is determined as the potential interaction time difference. The temporal state of the interaction blocking evaluation index is determined based on the comparison result between the potential interaction time difference and the preset interaction time threshold.

[0018] Furthermore, it also includes building collaborative interaction enhancement metrics, specifically including: For each interactive vehicle, its spatial occupancy area at each prediction time is mapped to the sector reachable space grid to determine the radial distance of the vehicle's occupancy in each angular sector. An interactive vehicle that meets any of the following conditions will be included in the valid participation set: the radial distance occupied by the interactive vehicle in at least one angular sector is less than the difference between the maximum reachable radial distance and the preset radial blocking margin; the interactive vehicle will be identified as the blocking interactive vehicle. The number of interactive vehicles in the effective participation set is counted, and the collaborative interaction enhancement index is determined based on the number of interactive vehicles. The collaborative interaction enhancement index satisfies the following: in, This represents the enhanced collaborative interaction index; Indicates the preset collaborative scaling factor; This indicates the number of interactive vehicles in the valid participant set.

[0019] Furthermore, it also includes: Each of the reachable space compression index, the interaction blocking evaluation index corresponding to each of the blocking interaction vehicles, and the collaborative interaction enhancement index is assigned a corresponding evaluation weight. The weighted reachable space compression index, the interaction blocking evaluation index, and the collaborative interaction enhancement index are then fused to obtain a comprehensive interaction reward signal. The integrated interactive reward signal is distributed to the interactive vehicles participating in the strategy update; For each of the interactive vehicles, a basic driving reward is determined based on the vehicle's driving progress, longitudinal speed, lateral deviation, heading deviation, control input changes, and terminal status. The basic driving reward is then fused with the comprehensive interactive reward signal allocated to the interactive vehicle to obtain the strategy update reward for that interactive vehicle. The reward is updated according to the strategy to update the behavior strategy of the corresponding interactive vehicle, and the control input corresponding to the next test moment is determined according to the updated behavior strategy.

[0020] Furthermore, determining the test result of the multi-vehicle interaction test scenario based on the comparison result between the reachable space compression index and the preset boundary criterion specifically includes: At each test moment, the reachable space compression index is compared with the preset compression threshold in the preset boundary criterion; When the reachable space compression index marked as valid reaches or exceeds the preset compression threshold, the vehicle status information of the master vehicle and each of the interactive vehicles, the predicted trajectory, the occupied radial distance, the reachable space compression index, the interaction blocking evaluation index, and the comprehensive interaction reward signal are recorded during the current test. The current multi-vehicle interaction test scenario is determined as the test scenario that reaches the preset boundary criterion, and the test scenario library is updated using the recorded test process data. When the reachable space compression index does not meet the preset compression threshold and does not reach the preset test termination condition, the vehicle status information of each interactive vehicle is updated according to the control input, and the steps of predicting the predicted trajectory of each interactive vehicle, determining the space occupancy area, determining the occupancy radial distance, and determining the reachable space compression index are re-executed. When the preset test termination condition is met and the reachable space compression index does not meet the preset compression threshold, the multi-vehicle interaction test scenario is determined as a test scenario that has not met the preset boundary criterion.

[0021] Compared with the prior art, the present invention has the following advantages: (1) Existing spatial evaluation methods rely heavily on lane lines, lane centerlines, road topology, or preset reference paths. In scenarios such as intersections, unprotected left turns, and multi-directional merging, it is difficult to use a unified spatial scale to evaluate the impact of traffic participants from different directions on the main vehicle. This invention determines the reachable area within the prediction period based on the vehicle state information and vehicle kinematics model of the main vehicle, and forms a fan-shaped reachable spatial grid along the angular and radial directions. This unifies the vehicle interaction relationships under different road structures into the relationship between the movement direction and reachable distance relative to the main vehicle, reduces the dependence of spatial evaluation on specific lane structures, and enables scenarios such as intersections, unprotected left turns, and multi-directional vehicle merging to use a unified spatial expression and evaluation scale, thereby improving the comparability of test results between different traffic scenarios.

[0022] (2) Existing evaluation methods based on vehicle center point, single predicted trajectory, or minimum distance between vehicles are insufficient to fully reflect the impact of vehicle size and attitude on the actual movable space of the tested vehicle. Even if two traffic participants have the same center distance from the host vehicle, different vehicle sizes and headings may result in significantly different actual obstructions. This invention determines the space occupancy area based on the vehicle outline of each interactive vehicle at multiple prediction times and maps each space occupancy area to the fan-shaped reachable space grid of the host vehicle. For each angular sector, the nearest occupancy radial distance to the host vehicle is determined from the space occupancy formed by all interactive vehicles at all prediction times, and the remaining reachable space is determined accordingly. Therefore, it can not only determine whether there are interactive vehicles in a certain direction, but also further determine the extent to which the actual movable distance in that direction is compressed. For test scenarios with the same directional coverage but different vehicle radial positions, vehicle sizes, or attitudes, different space occupancy results can still be obtained, thereby improving the ability to distinguish complex multi-vehicle interaction states.

[0023] (3) Existing methods that use minimum distance between vehicles, conflict time, or directional coverage as the main evaluation indicators usually focus on a single local relationship and are difficult to uniformly represent the degree of restriction of the overall motion space of the main vehicle after multiple interactive vehicles interact. For example, when a vehicle forms a close-range obstruction in a single direction and multiple vehicles form a moderate obstruction from multiple directions, it is difficult to accurately distinguish the overall interaction intensity of the two based solely on the minimum distance. This invention determines the reachable space compression index based on the reachable space area when it is not occupied by interactive vehicles and the remaining reachable space area after the predicted space is occupied. This index includes both the obstruction range and obstruction depth in each direction, so that the spatial impact generated by different numbers, directions, and distances of interactive vehicles can be continuously quantified on the same scale. This allows for a more direct characterization of the changes in the main vehicle's motion degrees of freedom during the multi-vehicle interaction process, providing a unified evaluation basis for intensity comparison and capability boundary determination in complex interaction scenarios.

[0024] (4) Existing spatial distance evaluation methods tend to include vehicles that are geographically close but do not actually have traffic conflicts in the effective interactions, making it difficult to distinguish between high traffic density and real high-intensity interactions. This invention further determines potential conflict points based on the target paths of the main vehicle and interacting vehicles, and determines the interaction blocking evaluation by combining the path distance from the vehicle to the potential conflict point, the blocking area, and the expected arrival time difference. This reduces the interference of neighboring vehicles unrelated to actual traffic competition on the evaluation results, allowing the spatial compression evaluation to focus more on the interaction process with actual path conflicts.

[0025] (5) Existing multi-vehicle evaluation methods directly use the total number of surrounding vehicles or the number of directional coverages to measure the group effect, which may include vehicles that do not actually affect the main vehicle's movement space. This invention determines the effective participation set based on the actual radial obstruction formed by each interacting vehicle on the fan-shaped reachable space and the interaction blocking state, and constructs a collaborative interaction enhancement index based on the number of effective participating vehicles. Therefore, it can distinguish between the two situations where there are multiple vehicles in the scene and multiple vehicles actually jointly restricting the main vehicle's movement space, and more accurately characterize the group effect generated by multi-vehicle collaborative blocking.

[0026] (6) Existing high-intensity interaction scenarios, if mainly achieved by shortening distances or enhancing aggression by a single vehicle, are prone to unreasonable rapid acceleration, sharp turns, and other behaviors. This invention combines reachable space compression indicators, interaction blocking evaluation indicators, and collaborative interaction enhancement indicators into a comprehensive interaction reward signal, and combines the driving progress, speed, deviation, control input changes, and terminal status of the interactive vehicle to form a strategy update reward. Therefore, while maintaining the basic driving rationality of the interactive vehicle, the test pressure comes more from multi-vehicle spatial collaboration and path blocking relationships, reducing the situation where test pressure is generated by unreasonable aggressive behavior of a single vehicle.

[0027] (7) Fixed-trajectory testing is prone to the failure of the original interaction relationship after the master vehicle changes its decision. In this invention, the predicted trajectory, space occupancy and reachable space compression indicators are continuously updated according to the vehicle status during the test, and the corresponding test process data and test scenario library are recorded when the preset boundary criteria are reached. This enables the interaction test to iterate with the real-time response of the master vehicle and continuously explore the interaction boundary corresponding to different decision behaviors; at the same time, the reachable space compression and related evaluation data can be used for both the updating of the interaction vehicle strategy and the evaluation of test results, so that the strategy training and simulation test adopt a more consistent evaluation basis. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the initial reference trajectory and the optimized trajectory after interactive capture when the test scenario is an intersection in an embodiment of the present invention; Figure 3 is a schematic diagram of sector grid space compression of the host vehicle and projection of predicted trajectories of interactive vehicles in the embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following combines Figures 1 to 3 to further describe a multi-vehicle interaction testing method based on reachable space compression. The following parameter values are used to illustrate the calculation process and data flow relationship, and do not constitute a limitation on the scope of protection. The method includes two parts: offline modeling and online testing. The offline modeling mainly completes the establishment of multi-vehicle interaction test scenarios, vehicle kinematics models, the local reachable space of the host vehicle, and grid parameters; the online testing mainly completes short-term trajectory prediction of interactive vehicles, mapping of vehicle contour space occupation, reachable space compression calculation, potential conflict and blockage judgment, multi-vehicle collaborative evaluation, strategy update, boundary judgment, and test process data recording. Figure 1 presents the overall processing flow, where S1 to S9 in the flow correspond to the data processing sequence described below.

[0030] S1. Establish a multi-vehicle interaction test scenario according to test requirements and road structure, and initialize the state of each vehicle. The test scenario includes at least one host vehicle AP and a plurality of interactive vehicles BV. The host vehicle is used to carry the tested autonomous driving system, and the interactive vehicles are used to form dynamic traffic interaction with the host vehicle. The set of interaction participants is expressed as: wherein, represents the set of interaction participants, represents the host vehicle, represents the th interactive vehicle, represents the number of interactive vehicles, .

[0031] Each vehicle is initialized with position, longitudinal speed, heading angle and target path information. The target path is expressed using a global path point sequence, a lane center reference line, a polyline trajectory output by a planner, or a continuous parameter curve. In intersection tests, the target path at least covers the preparatory section of the vehicle before entering the conflict area, the inner section of the conflict area, and the recovery section after leaving the conflict area.

[0032] as shown in Figure 2As shown, in a specific intersection scenario, 11 represents the master vehicle (AP), 12 represents the interacting vehicle (BV1), and 13 represents the interacting vehicle (BV2). 21 represents the initial and intelligent driving trajectories of the master vehicle, 22 represents the initial reference trajectory of interacting vehicle BV1, 23 represents the initial reference trajectory of interacting vehicle BV2, 32 represents the optimized encirclement trajectory of interacting vehicle BV1 after the interaction strategy update, and 33 represents the optimized encirclement trajectory of interacting vehicle BV2 after the interaction strategy update. CP1 represents the interaction point between AP and BV1, and CP2 represents the interaction point between AP and BV2. This scenario allows two interacting vehicles to enter the master vehicle's decision space from different directions, facilitating simultaneous observation of radial obstruction, multi-directional cooperation, and path conflicts.

[0033] S2. Establish vehicle state, control input, and state recursion relationships. Each vehicle at discrete sampling times... The state vector is defined as: in, and This represents the horizontal and vertical coordinates of the vehicle's center of gravity in the global coordinate system. Indicates the vehicle's heading angle. Indicates the longitudinal speed of the vehicle. This represents the transpose of a vector.

[0034] The vehicle control input vector is defined as: in, Indicates longitudinal acceleration. This indicates the equivalent steering input for the front wheels.

[0035] The vehicle motion is modeled using a simplified single-track kinematics model. In continuous time, the vehicle state satisfies: in, and Let represent the first-order time derivatives of the global coordinate positions, respectively. Indicates the rate of change of heading angle. Indicates the rate of change of longitudinal velocity. This indicates the vehicle's wheelbase.

[0036] Discrete prediction uses a fixed sampling period Performing a first-order forward recursion, the state transition relation is represented as: in, Let represent the discrete state transition function that satisfies the vehicle nonholonomic constraints. The corresponding discrete update relation is: in, Indicates the sampling period. This represents the vehicle's wheelbase. In the actual simulation, vehicle physical boundaries are also applied to the speed, acceleration, and steering inputs, with the speed limited to... Acceleration limit is Steering input is limited to Control variables that exceed the boundary are truncated before state recursion to prevent the learned policy from outputting unexecutable vehicle actions.

[0037] The primary vehicle and the interacting vehicle use the same monorail model, with vehicle wheelbase and motion boundaries configured separately for each vehicle type. This implementation provides unified kinematic constraints for both primary and interacting vehicle predictions, facilitating comparison of their future spatial relationships within the same prediction timeframe. For low-speed, small-angle scenarios, constant acceleration and constant heading angular velocity models can also be used instead of the monorail model, provided that the predicted state still outputs the three core quantities: position, longitudinal velocity, and heading angle, and uses the same time reference as subsequent spatial mappings.

[0038] S3. Establish a local polar coordinate system for the main vehicle and construct a sector-shaped reachable space grid. The current centroid position of the main vehicle at the test moment is taken as the pole, and the current heading of the main vehicle is taken as the zero-angle direction. The prediction time period is denoted as... Based on the current longitudinal velocity and preset longitudinal acceleration limit Estimate the maximum reachable radial distance within the forecast period: in, Indicates the maximum achievable radial distance. Indicates the current longitudinal speed of the main vehicle. Indicates the length of the forecast period. This indicates the maximum preset longitudinal acceleration of the main vehicle.

[0039] To ensure that the maximum lateral angle corresponds to the vehicle's actual steering capability, one implementation first obtains the minimum turning radius based on the vehicle's wheelbase and maximum steering input: Then, determine the total angle of the reachable sector based on the maximum forward travel arc length corresponding to the predicted time period: in, Indicates the minimum turning radius. Indicates the maximum steering input. This represents the total angle of the reachable region of the sector. This represents the maximum permissible angle set by the scene or vehicle motion boundaries. Here, a circular arc approximation is used to convert the vehicle's extreme steering capability into the spatial angles on the left and right sides in front of the main vehicle. The sector angle range is written as... When the minimum turning radius is directly given by the vehicle controller, that minimum turning radius is used directly and is not recalculated. .

[0040] Divide the angle range equally. Each angular sector will have a radial range. Average score The radial intervals have angular and radial resolutions as follows: in, Indicates the number of angular sectors. Indicates the number of radial intervals. Indicates the angle width from a single angle to the sector. This represents the radial width of a single radial interval.

[0041] The fan-shaped grid set is represented as: in, Indicates the first The first angular sector and the first A grid cell is formed by radial intervals. Each grid cell corresponds to a directional interval. and distance interval This vehicle-centric discrete approach does not depend on the existence of continuous lane lines within the intersection; the free movement space of the vehicle in different road configurations is projected onto the same angle-radial coordinate frame.

[0042] S4. Predict the short-term trajectory of each interactive vehicle within the prediction period and establish the vehicle's external space occupancy. For the first... The interactive vehicles proceed forward recursively from their current state according to the discrete state equations. The number of prediction steps is denoted as... Then the predicted trajectory satisfies: in, Indicates the first Interactive vehicles in the first The state at the predicted time. This indicates the corresponding control input. This indicates the number of prediction steps. The total prediction time usually satisfies... This ensures that the reachable space of the main vehicle and the space occupied by the interacting vehicles have the same prediction time range.

[0043] The vehicle's geometric outline is represented using a rectangular bounding box (BBox). In the vehicle's local coordinate system, at least the vehicle's center point and four corner points are taken as basic sampling points. Let any local sampling point be... The vehicle's global position and heading at the corresponding predicted time are respectively , and Then the coordinates of the sampling point after transformation to the global coordinate system are: in, and This represents the global coordinates of the vehicle's outer contour sampling points. This represents the longitudinal coordinate of the sampling point relative to the vehicle's center of mass. This represents the lateral coordinates of the sampling point relative to the vehicle's centroid. The four corner points of the rectangle are determined by the vehicle's length. and vehicle width Determined as Then, relative to the center point of the vehicle Together they form a basic set of sampling points.

[0044] To reduce boundary mapping errors caused by using only five discrete points, this embodiment also follows the four sides of the rectangle according to a value not greater than... The sampling interval is used to supplement boundary points, and the center point, corner points, and boundary points are used together as input for spatial occupancy mapping. The basic five-point mode has a smaller computational load and is suitable for scenarios with a large number of interactive vehicles or limited online computing resources; the boundary encryption mode improves the geometric coverage integrity of slender vehicles, angled vehicles, and vehicles with high grid resolution.

[0045] Transform the vehicle's outline sampling points in the global coordinate system to the main vehicle's local polar coordinate system. Let the main vehicle's current centroid be... , heading The translational distance of any sampling point relative to the main vehicle is , Then the radial distance and relative azimuth are: in, This represents the radial distance of the sampling point relative to the main vehicle. This indicates the azimuth angle relative to the forward reference of the main vehicle. Represents the arctangent function with quadrant discrimination. This means normalizing the angle to .

[0046] when and At that time, the sampling point enters the area accessible by the main vehicle. The corresponding angular sector index and radial bucket index are as follows: in, Indicates the angular sector index. Represents radial bucket index, rounding function This indicates rounding down. The calculated indices are then limited to a specific range. as well as Within the range.

[0047] For each interactive vehicle, establish a radial array for vehicle occupancy. Initialized to In raster quantization mode, when the... The sampling points of the interactive vehicles are mapped to the first The first angular sector and the first When there is a radial bucket, update the first blocking position of the sector according to the following formula: in, Indicates the first Interactive vehicles in the first The radial distance occupied by each vehicle within the sector is determined by its angle. Initialization. This indicates that the interactive vehicle did not create an obstruction in the corresponding direction.

[0048] After aggregating the occupancy results of all interactive vehicles and all predicted times, the 1st The final radial distance occupied by each corner sector is: in, Indicates the first The radial distance occupied by each corner sector to the nearest main vehicle. When all All equal to hour, The corresponding direction remains unobstructed during the predicted period.

[0049] Figure 3 This diagram illustrates one result for the reachable sector of the main vehicle, the predicted trajectories of interacting vehicles, and occupancy mapping. The sector region in front of the main vehicle consists of multiple angular sectors. After an interacting vehicle enters this sector region, the radial boundary in the corresponding direction is shortened. Compared to simply counting the presence of vehicles in a single direction, this approach... At the same time, both directional information and blocking depth information are preserved.

[0050] S5. Determine the remaining reachable space based on the occupied radial distance of each angular sector, and calculate the reachable space compression index. In the absence of interacting vehicles occupying the space, the area of ​​the sector-shaped reachable space is defined as: in, This indicates the accessible space area when not occupied by interacting vehicles. Indicates the number of angular sectors. Indicates the angle width of the sector. Indicates the maximum achievable radial distance.

[0051] After considering the predicted occupancy of all interactive vehicles, the remaining accessible space area is defined as: in, Indicates the remaining accessible space area. Indicates the first The occupied radial distance corresponding to each angular sector; the unoccupied angular sector is taken as... .

[0052] The reachability compression metric at the current test moment is defined as: in, This represents the reachability space compression index. This represents a preset positive stability constant, used to avoid zero denominators in numerical implementations. lie in Within the range, a larger value indicates a more significant overall compression of the main vehicle's free movement space by multiple interacting vehicles during the prediction period. Since area calculations include both angular range and radial depth, single-direction near-distance obstruction and multi-direction medium-distance obstruction will result in different... This avoids the problem that the overall degrees of freedom of motion cannot be described when only the closest distance is used.

[0053] S6. Identify potential conflict points based on the target path and calculate the time difference between the blocking state and the potential interaction. The target path of the main vehicle is denoted as... , No. The target path of the interactive vehicles is denoted as In one implementation, the path is represented by polyline points sorted by driving direction. Two-dimensional intersection detection is performed on each segment of the two paths, and the intersection point that is within the effective forward path range of both vehicles and has the smallest distance along the path is identified as the current potential conflict point. If the two paths do not have a strict geometric intersection, but the distance between the nearest segments is less than a preset conflict distance... The midpoint of the nearest point pair is then used as an approximate potential conflict point. This approach takes into account situations where centerlines do not completely intersect but vehicle outlines may still compete.

[0054] To prevent vehicles far from the conflict zone from falsely triggering the reachability compression index, in one specific setting, only when the path distance from the primary vehicle to the potential conflict point is falsely triggered will the reachability compression index be applied. Not greater than the first preset distance And the path distance from at least one interacting vehicle to the same potential conflict point Not greater than the second preset distance At that time, the current Marked as valid. The condition is expressed as: in, Indicates the distance along the route from the main vehicle to the potential point of conflict. Indicates the first The distance between the two vehicles and the potential point of conflict. Indicates the first preset distance. This indicates the second preset distance. In a set of intersection parameters, and .

[0055] Establish a blocking zone around each potential conflict point, with the blocking radius denoted as . When the main vehicle and the first All interactive vehicles at the same test time meet the requirement that the distance to the same potential conflict point is less than At that time, mark the interactive vehicle as a blocking interactive vehicle and generate a blocking reward. In a set of intersections, The values ​​are set within the range of 5 to 10 m; the specific values ​​are determined based on the conflict zone scale, vehicle length, and expected lead time.

[0056] Potential Interaction Time Difference (PET) is used to further describe the degree of synchronization between the arrival times of the two vehicles at the same potential conflict point. If the primary vehicle and the second... The estimated arrival times of the two interactive vehicles are as follows: and ,but: in, Indicates the main vehicle and the first The potential interaction time difference between two interacting vehicles regarding the same potential point of conflict. Indicates the estimated arrival time of the main vehicle. Indicates the first The estimated arrival time of the interactive vehicles is calculated in two ways. The first approach uses a constant speed approximation, where the estimated arrival time is obtained by dividing the remaining distance traveled by the current speed, and a preset minimum speed is used to avoid division by zero at low speeds. The second approach calculates the estimated arrival time by solving a one-dimensional uniformly accelerated displacement equation based on the current longitudinal acceleration, which is more suitable for scenarios where the interactive vehicles are significantly accelerating or braking.

[0057] when At that time, the blocking evaluation enters the time synchronization state. In a set of training parameters, To ensure PET's continuous participation in reward adjustment, the reward blocking mechanism can also be constructed using the following formula: in, This indicates the blocking status; it is set to 1 when blocking occurs and 0 otherwise. Indicates the basic blocking reward. This represents the PET adjustment factor. This indicates a preset interaction time threshold. This continuous format ensures that the closer two vehicles are to synchronously entering the conflict zone, the higher the interaction evaluation, while preventing distant vehicles that have not yet entered the blocking zone from being mistakenly counted as effective blocking.

[0058] S7. Establish an effective participant set based on the actual impact of each interacting vehicle on the master vehicle's space and path, forming a multi-vehicle collaborative evaluation, and use the evaluation results for updating the interacting vehicle strategy. For the first... A vehicle that participates in an interaction will be added to the valid participant set if it meets any of the following conditions: at least one angular sector satisfies Or the interactive vehicle is already blocked. Here This indicates a preset radial blocking margin used to filter out spurious engagements caused by boundary sampling errors or extremely slight grid intrusions. In one setting, Take a depth of 0.5 to 1.5 m.

[0059] The number of vehicles that effectively participated is recorded as To avoid numerical runaway in collaborative evaluation as the number of vehicles increases linearly, the collaborative interaction enhancement index adopts a logarithmic growth form: in, Indicators representing enhanced collaborative interaction Indicates the preset collaborative scaling factor. Indicates the number of vehicles that are effectively participating. Let represent the natural logarithm function. This processing allows the second and third vehicles that actually participate in blocking to still increase the group's collaborative evaluation, but the marginal bonus of the newly added vehicles gradually decreases, avoiding excessive rewards simply by increasing the number of background vehicles.

[0060] The overall collaborative reward is derived from the fusion of spatial compression contribution, blocking reward, and collaborative interaction enhancement indicators. The original fusion form is as follows: in, Indicates overall collaborative reward, Indicates the weight of spatial compression evaluation. Indicates the weight of the blocking evaluation. Indicates the first Accessible space compression index corresponding to each vehicle owner. Indicates the first The main vehicle and the first Interruption rewards between interacting vehicles This represents an enhanced collaborative interaction metric. In the current implementation testing with a single master vehicle, the master vehicle summation term only contains the AP object, naturally degenerating the formula into a single master vehicle form. During parallel stress testing with multiple master vehicles, each master vehicle establishes its own locally reachable space, and then the summations are performed using the same formula.

[0061] The overall collaborative reward is distributed among the interactive vehicles participating in the learning process. A simple distribution method is equal allocation: in, This represents the collaborative reward received by each learning and interactive vehicle. This represents the number of interactive vehicles participating in the strategy update, excluding the master vehicle. Evenly distributing homogeneous interactive vehicles suitable for parameter sharing encourages each vehicle to focus on the overall containment outcome. Another implementation allocates rewards based on each vehicle's marginal spatial contribution and blocking state, first removing the [unclear - likely a vehicle name]. Each interacting vehicle recalculates its remaining reachable space to obtain its marginal contribution, which is then normalized and used as the allocation coefficient. This method makes it easier to distinguish the individual contributions of different interacting vehicles and is suitable for scenarios with heterogeneous vehicles or participants with significantly different roles.

[0062] To prevent interactive vehicles from engaging in illogical driving maneuvers by solely focusing on compressing the space of the main vehicle, each learning interactive vehicle also calculates a basic driving reward: in, Indicates basic driving bonus; Indicates progress rewards along the target path; Indicates speed reward; Indicates the lateral deviation term; Indicates the heading deviation term; Indicates the smoothness control term; Indicates a terminal item; , , , and Each represents a weight.

[0063] In one numerical implementation, the progress reward is normalized using the increment of the cumulative mileage along the target path at adjacent test times; the speed reward is constructed based on the deviation between the current speed and the reference speed; the lateral deviation and heading deviation terms are negativeed based on the absolute values ​​of the path lateral and heading errors, respectively; the control smoothness term is negativeed based on the changes in acceleration and steering input at adjacent times; the terminal term provides a positive reward for successfully completing the interaction task and a negative reward for collisions, reversals, or continuous stoppages. With this structure, the interactive vehicle needs to achieve spatial compression or path blocking while maintaining normal driving, rather than relying on unrestrained rapid acceleration and sharp turns to obtain a high interaction rating.

[0064] The final reward returned to the learning interaction vehicle at each step is represented as follows: in, Indicates a reward for strategy updates. This represents a numerical truncation function. and These represent the lower and upper bounds of the reward, respectively. Reward truncation limits the magnitude of the single-step reward in extreme scenarios, improving the stability of training values.

[0065] Early training improvement , or At least one of these factors encourages interactive vehicles to more actively explore multi-vehicle cooperative spatial relationships; gradually reducing the relative weight of the cooperative component in the later stages of training allows the strategy to focus more on fine-grained individual motion control. A linear decay method is used according to the training progress. Calculate collaborative weights ,in Indicates the weights at the start of training. This indicates the weights at the end of training.

[0066] In the first specific implementation, the interactive vehicle strategy employs a near-end strategy to optimize the PPO algorithm. The input observations consist of four types of information: the first type includes the interactive vehicle's own speed, heading, path progress, and control input from the previous time step; the second type includes the master vehicle's relative position, relative speed, and relative heading; and the third type is the normalized sector radial vector. , , and blocking state; the fourth category is the normalized distance along the route from the current interacting vehicle to the potential conflict point and The location distance is divided by . The speed is divided by the preset speed limit, and the heading uses sine and cosine dual-channel encoding to reduce angle variations. and Discontinuity at the point.

[0067] The PPO policy network employs a two-layer fully connected network with 256 neurons per layer, and ReLU activation is used in the hidden layers. The policy outputs the Gaussian distribution mean values ​​corresponding to the acceleration and steering inputs, respectively, with the logarithmic standard deviation of the two action dimensions used as trainable parameters. Sampled actions are compressed using a hyperbolic tangent function and linearly mapped to the vehicle's permissible acceleration and steering ranges. The value network uses an independent two-layer 256-unit fully connected network and outputs a scalar state value. During training, the discount factor is set to 0.99, the generalized advantage estimation parameter to 0.95, the PPO pruning parameter to 0.2, and the learning rate to [missing value]. Each round of trajectory data is trained in 10 mini-batch training cycles, with each mini-batch containing 256 samples. Multiple interactive vehicles of the same type share the same set of policy parameters but retain their own independent observations and cyclic states. Parameter sharing reduces the number of model parameters as the number of vehicles increases and encourages different interactive vehicles to form consistent collaborative rules.

[0068] In another specific implementation, the interactive vehicle strategy employs the Soft Actor-Critic algorithm. The policy network still uses a two-layer 256-unit fully connected structure, outputting continuous acceleration and steering actions compressed with hyperbolic tangent; simultaneously, two independent Q-networks and corresponding target networks are set up, with the experience replay capacity set to [value missing]. The batch size is set to 256, the discount factor to 0.99, the target network soft update coefficient to 0.005, and the entropy temperature parameter is automatically adjusted. This method utilizes off-policy experience replay to improve sample reuse rate, making it more suitable for training environments with high simulation costs or those requiring frequent reuse of critical interaction data. Both PPO and SoftActor-Critic are used as policy optimization implementation methods, while the data processing chain of space compression, blocking, and collaborative evaluation remains unchanged.

[0069] S8. Perform boundary checks based on the current interaction evaluation results. At each test moment, the effective... With preset compression threshold Compare and record the blocking status, PET, And collision state. A basic boundary criterion is... In more stringent interaction boundary tests, the boundary criterion also requires the presence of at least one vehicle blocking the interaction, or at least one... This allows us to distinguish between scenarios with high traffic density but no actual path conflicts and scenarios where the primary vehicle's decision space is realistically constrained.

[0070] If the boundary criteria are not met and the test does not reach the termination condition, the control inputs of each interactive vehicle are updated according to the policy network output, the simulation is advanced for one sampling cycle, and then S4 to S8 are re-executed. Termination conditions include reaching the maximum test duration, a collision between the main vehicle or an interactive vehicle, a vehicle leaving the valid test area, and the simulation platform voluntarily terminating. If the boundary criteria are met before termination, the current scene is marked as having reached the interaction boundary; if the boundary criteria are not met when the termination condition is reached, it is marked as not having reached the interaction boundary.

[0071] S9. Record the interaction process and update the test scenario library. When the interaction boundary is reached, record the initial state, time-by-time state, control input, target path, predicted trajectory, vehicle outline occupancy points, and each sector of the main vehicle and all interacting vehicles. , , , Potential conflict points, PET, blocking status, effective participation set, The system includes overall collaborative rewards, strategy version, random seed, and boundary judgment results. The scenario library assigns a unique scenario number to each test and simultaneously saves the road geometry version and vehicle parameter version. Subsequent reproductions are reloaded using the same version and random seed.

[0072] In the test mode without online training, the predicted trajectory of interactive vehicles is derived from pre-trained policies, traffic simulators, measured trajectory deformation results, or external trajectory planners. The system still performs geometric occupancy, space compression, and conflict blocking calculations in S4 to S6, and completes boundary judgment and data recording in S8 and S9, but does not perform policy parameter updates. In this way, the same spatial evaluation chain serves both interactive policy training and offline retesting of fixed policies, avoiding inconsistencies in evaluation scales caused by using different indicators in the training and testing phases.

[0073] Through the above implementation methods, the reachable area of ​​the main vehicle, the future geometrical occupancy of interactive vehicles, the first obstruction position in each direction, the remaining reachable space, and the reachable space compression index form a continuous data chain. The spatial model is based on the main vehicle's own mobility capabilities, and does not require interactive vehicles to travel in the same direction as the main vehicle, nor does it require continuous lane lines in the conflict area; vehicle length, width, and heading directly affect the radial boundaries of each sector through outline mapping; path conflict, PET, and blocking status further filter vehicles that are only spatially adjacent but do not actually compete for passage; effective participation set and logarithmic collaborative enhancement distinguish vehicles that actually participate in compression or blocking from other vehicles in the scene. The resulting evaluation quantity can be used simultaneously for testing boundary judgment and interactive vehicle strategy optimization, and can completely reproduce the dynamic process of forming a high-intensity interactive state through the scene library.

[0074] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-vehicle interaction testing method based on reachability space compression, characterized in that, Includes the following steps: Obtain vehicle status information of the main vehicle and multiple interacting vehicles in a multi-vehicle interaction test scenario; Based on the vehicle state information and vehicle kinematics model of the main vehicle, the reachable area of ​​the main vehicle during the prediction period is predicted, and the reachable area is discretized along the angular and radial directions to obtain a fan-shaped reachable space grid composed of multiple angular sectors and multiple radial intervals. Based on the vehicle state information of each interactive vehicle and the vehicle kinematic model, the predicted trajectory of each interactive vehicle within the predicted time period is predicted, and the spatial occupancy area at each predicted time is determined based on the vehicle outline of each interactive vehicle at each predicted time of the predicted trajectory. Each of the space-occupied areas is mapped to the sector-shaped reachable space grid, and for each angular sector, the nearest occupancy radial distance to the main vehicle is determined in each of the space-occupied areas mapped to that angular sector, and the remaining reachable space of the main vehicle after being occupied by the interactive vehicle is determined based on each of the occupancy radial distances. The reachability compression index of the main vehicle is determined based on the reachable space area of ​​the sector-shaped reachable space grid when it is not occupied by the interactive vehicle and the area of ​​the remaining reachable space. The test results of the multi-vehicle interaction test scenario are determined based on the comparison results between the reachable space compression index and the preset boundary criteria.

2. The multi-vehicle interaction testing method based on reachability space compression according to claim 1, characterized in that, The vehicle status information includes the vehicle's position, longitudinal speed, and heading angle; The control inputs to the vehicle kinematics model include longitudinal acceleration and steering input; The vehicle kinematics model adopts a discrete-time single-track model. Based on the vehicle's position, longitudinal velocity, and heading angle at the current sampling time, as well as the longitudinal acceleration, steering input, vehicle wheelbase, and sampling period, the vehicle's position, longitudinal velocity, and heading angle at the next sampling time are determined, and the vehicle state sequence is recursively obtained within the prediction period according to the sampling period.

3. The multi-vehicle interaction testing method based on reachability space compression according to claim 1, characterized in that, The process involves predicting the reachable area of ​​the main vehicle within a prediction time period based on the vehicle state information and vehicle kinematic model, and discretizing the reachable area along the angular and radial directions to obtain a fan-shaped reachable space grid composed of multiple angular sectors and multiple radial intervals. Specifically, this includes: A local polar coordinate system for the main vehicle is established with the current centroid position of the main vehicle as the pole and the current heading of the main vehicle as the zero angle direction. The maximum reachable radial distance of the reachable area is determined based on the current longitudinal speed of the main vehicle, the preset upper limit of longitudinal acceleration, and the predicted time period. The maximum lateral angle of the reachable area is determined based on the minimum turning radius that characterizes the steering capability of the main vehicle; The angular sector is divided into multiple equal-angle sectors within the angular range defined by the maximum lateral angle, and the radial interval is divided into multiple radial intervals within the radial range defined by the maximum reachable radial distance, to obtain the sector-shaped reachable space grid.

4. The multi-vehicle interaction testing method based on reachability space compression according to claim 1, characterized in that, The step of predicting the trajectory of each interactive vehicle within the prediction period based on the vehicle state information and the vehicle kinematic model, and determining the spatial occupancy area at each prediction time based on the vehicle outline corresponding to each prediction time of the prediction trajectory, specifically includes: For each of the interactive vehicles, based on the vehicle state information and control input of the interactive vehicle at the current sampling time, the vehicle kinematics model is used to recursively obtain the predicted states corresponding to multiple prediction times within the prediction period, and the predicted trajectory is formed by the multiple predicted states. The vehicle's external attitude at the corresponding prediction time is determined based on its position and heading angle in each of the predicted states. Determine the positions of the center point and four corner points of the vehicle outline in the vehicle's local coordinate system, and transform the center point and four corner points to the global coordinate system according to the position and heading angle at the corresponding prediction time; The spatial occupancy area at the corresponding prediction time is determined based on the center point and four corner points transformed to the global coordinate system.

5. The multi-vehicle interaction testing method based on reachability space compression according to claim 1, characterized in that, The process of mapping each of the space-occupied areas to the sector-shaped reachable space grid, and for each angular sector, determining the nearest occupancy radial distance to the main vehicle in each of the space-occupied areas mapped to that angular sector, and determining the remaining reachable space of the main vehicle after it is occupied by the interacting vehicle based on each of the occupancy radial distances, specifically includes: Establish a local polar coordinate system for the main vehicle with its current position as the pole and its current heading as the zero-angle direction; Transform the vehicle outline points of each interactive vehicle at each prediction time to the local polar coordinate system of the main vehicle to obtain the angular position and radial distance of each vehicle outline point relative to the main vehicle. Based on the angular position and radial distance of each vehicle outline point, the vehicle outline points located within the reachable area are mapped to the corresponding angular sector and radial interval; For each angular sector, the radial distances of all interactive vehicles mapped to that angular sector at all prediction times are compared, and the minimum radial distance is determined as the occupied radial distance of that angular sector; when there is no vehicle outline point in an angular sector, the maximum reachable radial distance is determined as the radial reachable distance of that angular sector. The remaining reachable space is determined by using the occupied radial distance or the radial reachable distance corresponding to each angular sector as the radial boundary of that angular sector.

6. The multi-vehicle interaction testing method based on reachability space compression according to claim 1, characterized in that, The step of determining the reachability compression index of the main vehicle based on the reachability area of ​​the sector-shaped reachability grid when it is not occupied by the interactive vehicle and the area of ​​the remaining reachability space specifically includes: Each of the angular sectors is set to have the same angular width. Based on the angular width of the angular sector, the maximum reachable radial distance of the reachable area, and the occupied radial distance corresponding to each angular sector, the reachable space area and the remaining reachable space area are determined respectively. The areas of the reachable space and the remaining reachable space are defined by the following formulas: The reachable space compression index is determined based on the reachable space area and the remaining reachable space area. The reachable space compression index is calculated using the following formula: in, This represents the reachable space area of ​​the sector-shaped reachable space grid when it is not occupied by the interactive vehicle; This represents the area of ​​the remaining reachable space; Indicates the number of angular sectors; Indicates the angular width of each of the aforementioned angular sectors; This represents the maximum reachable radial distance of the reachable region; Indicates the first The occupied radial distance corresponding to the angular sector, when the angular sector... When the corner sector is not occupied Take the maximum achievable radial distance; This indicates a pre-set positive stability constant; This represents the reachable space compression index.

7. The multi-vehicle interaction testing method based on reachability space compression according to claim 1, characterized in that, Before determining the test result of the multi-vehicle interaction test scenario based on the comparison result of the reachable space compression index and the preset boundary criterion, the method further includes: Obtain the target path information of the main vehicle and each of the interactive vehicles, and determine potential conflict points based on the intersection relationship between the target path of the main vehicle and the target paths of each of the interactive vehicles. The path distances from the main vehicle and each of the interactive vehicles to the potential conflict point along the corresponding target path are determined respectively. When the path distance corresponding to the main vehicle is not greater than a first preset distance and the path distance corresponding to at least one of the interactive vehicles is not greater than a second preset distance, the current reachable space compression index is marked as valid. Based on the potential conflict point, the blocking area is determined according to the preset blocking distance threshold. When the main vehicle and any of the interactive vehicles are both located in the same blocking area at the same test time, the interactive vehicle is identified as the blocking interactive vehicle, and an interactive blocking evaluation index corresponding to the blocking interactive vehicle is generated. Based on the vehicle status information of the main vehicle and the blocking interaction vehicle, as well as their respective target path information, the estimated arrival time of the two vehicles to the same potential conflict point is determined. The absolute difference between the two estimated arrival times is determined as the potential interaction time difference. The temporal state of the interaction blocking evaluation index is determined based on the comparison result between the potential interaction time difference and the preset interaction time threshold.

8. A multi-vehicle interaction testing method based on reachability space compression according to claim 7, characterized in that, It also includes building collaborative interaction enhancement metrics, specifically including: For each interactive vehicle, its spatial occupancy area at each prediction time is mapped to the sector reachable space grid to determine the radial distance of the vehicle's occupancy in each angular sector. An interactive vehicle that meets any of the following conditions will be included in the valid participation set: the radial distance occupied by the interactive vehicle in at least one angular sector is less than the difference between the maximum reachable radial distance and the preset radial blocking margin; the interactive vehicle will be identified as the blocking interactive vehicle. The number of interactive vehicles in the effective participation set is counted, and the collaborative interaction enhancement index is determined based on the number of interactive vehicles. The collaborative interaction enhancement index satisfies the following: in, This represents the enhanced collaborative interaction index; Indicates the preset collaborative scaling factor; This indicates the number of interactive vehicles in the valid participant set.

9. A multi-vehicle interaction testing method based on reachability space compression according to claim 8, characterized in that, Also includes: Each of the reachable space compression index, the interaction blocking evaluation index corresponding to each of the blocking interaction vehicles, and the collaborative interaction enhancement index is assigned a corresponding evaluation weight. The weighted reachable space compression index, the interaction blocking evaluation index, and the collaborative interaction enhancement index are then fused to obtain a comprehensive interaction reward signal. The integrated interactive reward signal is distributed to the interactive vehicles participating in the strategy update; For each of the interactive vehicles, a basic driving reward is determined based on the vehicle's driving progress, longitudinal speed, lateral deviation, heading deviation, control input changes, and terminal status. The basic driving reward is then fused with the comprehensive interactive reward signal allocated to the interactive vehicle to obtain the strategy update reward for that interactive vehicle. The reward is updated according to the strategy to update the behavior strategy of the corresponding interactive vehicle, and the control input corresponding to the next test moment is determined according to the updated behavior strategy.

10. A multi-vehicle interaction testing method based on reachability space compression according to claim 9, characterized in that, The step of determining the test result of the multi-vehicle interaction test scenario based on the comparison result of the reachable space compression index and the preset boundary criterion specifically includes: At each test moment, the reachable space compression index is compared with the preset compression threshold in the preset boundary criterion; When the reachable space compression index marked as valid reaches or exceeds the preset compression threshold, the vehicle status information of the master vehicle and each of the interactive vehicles, the predicted trajectory, the occupied radial distance, the reachable space compression index, the interaction blocking evaluation index, and the comprehensive interaction reward signal are recorded during the current test. The current multi-vehicle interaction test scenario is determined as the test scenario that reaches the preset boundary criterion, and the test scenario library is updated using the recorded test process data. When the reachable space compression index does not meet the preset compression threshold and does not reach the preset test termination condition, the vehicle status information of each interactive vehicle is updated according to the control input, and the steps of predicting the predicted trajectory of each interactive vehicle, determining the space occupancy area, determining the occupancy radial distance, and determining the reachable space compression index are re-executed. When the preset test termination condition is met and the reachable space compression index does not meet the preset compression threshold, the multi-vehicle interaction test scenario is determined as a test scenario that has not met the preset boundary criterion.

Images