An unmanned path planning method and system based on interactive trajectory prediction

Abstract

The application discloses an unmanned path planning method and system based on interactive trajectory prediction, and belongs to the technical field of intelligent driving path planning and vehicle motion decision-making. The method first generates a candidate trajectory of an unmanned vehicle. A region to be driven in front of the vehicle is divided into a near-field region and a far-field region, and the division boundary is a dynamic self-adaptive boundary. For the far-field region, interactive objects are screened and interactive trajectory prediction is performed according to the arrival time sequence and spatial overlap relationship. A reasoning post-processing layer independent of the interactive trajectory prediction removes sequentially inconsistent prediction trajectories and occupationally conflicting prediction trajectories according to the yielding order consistency and the occupancy conflict exclusivity, obtains a scene-consistent prediction result, and extracts a passable space-time interval. For the near-field region, scene clipping is performed on the basis of the scene-consistent prediction result, a safe response driving mode is generated in advance, and matching execution is performed. The application is suitable for unmanned path planning in a strong interactive traffic scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving path planning and vehicle motion decision-making technology, and more specifically, to an unmanned driving path planning method and system based on interactive trajectory prediction. Background Technology

[0002] In highly interactive traffic scenarios, autonomous vehicles need to predict the future movements of surrounding traffic participants and plan their own trajectories accordingly. Existing interactive trajectory prediction methods typically use candidate trajectories of the autonomous vehicle as conditions to predict the trajectories of surrounding traffic participants, obtaining a set of predicted trajectories, and then perform path planning based on the prediction results. To improve the consistency of prediction results at the scene level, existing technologies mainly employ two approaches: The first approach encodes scene consistency constraints into the model parameters through a loss function during the training phase of the prediction model, enabling the model to intrinsically output scene-consistent prediction results; the second approach constructs a scene interaction graph or uses a joint prediction model to jointly model the predicted trajectories of multiple agents during the prediction process, thereby improving the scene-level consistency of multi-agent prediction modalities. Furthermore, existing technologies also include planning schemes that separate the near-field and far-field regions. These schemes typically divide the area in front of the vehicle into near-field and far-field regions based on distance, using higher-precision prediction or stronger tracking weights for the near-field region and lower-precision prediction or weaker tracking weights for the far-field region, thus allocating computational resources differently at different distances.

[0003] However, the existing technology still has the following shortcomings: (1) The scenario consistency constraint of the existing technology depends on the model training stage, which requires modification or retraining of the prediction model, resulting in high engineering deployment costs and difficulty in adapting to existing interactive trajectory prediction models; although the joint prediction model can improve consistency in the prediction process, it also requires modification of the model itself; (2) The existing technology lacks an explicit filtering mechanism for decision-making contradictions between agents in the prediction results; when two predicted trajectories give opposite right-of-way claims, the passable spatiotemporal interval extracted based on such contradictions may contain false passable information, leading to safety risks in planning decisions; (3) The existing technology has near-field and far-field constraints. In the field layering scheme, the near field region still relies on probabilistic trajectory prediction, which outputs multiple possible predicted trajectories. There are inconsistencies in the right-of-way claims among the multiple predicted trajectories, which leads to the need for repeated weighing of multiple possibilities in the near field decision-making. The decision-making is delayed and it is difficult to meet the requirements of real-time performance and certainty in the near field region. The prediction of the far field region is also not linked with the near field recognition results. The near field perception information is not used to constrain the far field prediction. (4) The near field planning and far field planning of the existing technology are independent of each other. The near field recognition results do not participate in the far field prediction and consistency verification. The far field prediction results do not provide a reference for the near field decision-making. There is a lack of cross-layer information coupling and the overall robustness of the system is insufficient. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned problems existing in the prior art and to greatly improve its technical effect on the basis of the original technology; this invention proposes an autonomous driving path planning method based on interactive trajectory prediction, the method comprising: S1. Obtain the autonomous vehicle's state, road topology, and historical trajectories of surrounding traffic participants to generate candidate trajectories for the autonomous vehicle. S2. Obtain a complete set of preset behavior patterns. This set is a pre-constructed collection of behavior patterns covering all behavior types of traffic participants. Each behavior pattern is defined by a behavior type identifier and its corresponding kinematic parameter range. S3. Based on the autonomous vehicle's travel distance, divide the area ahead of the vehicle into a near-field region and a far-field region. The far-field region extends along the autonomous vehicle's planned path, extending from the vehicle's current position to a preset distance ahead. The near-field region is the area along the vehicle's travel direction and at a distance from the vehicle's current position. The near-field region is a region within a preset near-field distance range, and the far-field region is the remaining area of ​​the driving area outside the near-field region; the boundary between the near-field region and the far-field region is a dynamically adaptive boundary; S4, for the far-field region; for each candidate trajectory of the autonomous vehicle, interactive objects are selected based on arrival time sequence and spatial overlap relationship to construct a trajectory interaction scenario; using the candidate trajectory of the autonomous vehicle in the trajectory interaction scenario as a condition, interactive trajectory prediction is performed on the interactive objects to obtain a predicted trajectory set; subsequently, the inference post-processing layer, independent of the interactive trajectory prediction, removes the predicted trajectory set from the predicted trajectory set based on the quantitative judgment rules of yielding order consistency and occupancy conflict mutual exclusion. Excluding contradictory predicted trajectories, a consistent scene prediction result is obtained; the post-processing layer of inference filters the predicted trajectory set according to manually preset rules and preset thresholds; then, the consistent scene prediction result is written back to the candidate trajectory of the autonomous vehicle, and the passable spatiotemporal interval of each candidate trajectory of the autonomous vehicle is extracted using the consistent scene prediction result as input; S5, for the near-field region; the near-field region is the region that has previously completed interactive trajectory prediction and scene consistency filtering as the far-field region. Based on the obtained consistent scene prediction result, the entire set of preset behavior patterns is cropped to obtain the behavior pattern subset corresponding to the near-field region; the behavior pattern subset is enumerated. For each behavior pattern, a corresponding safe response driving mode for the autonomous vehicle is pre-generated, and each safe response driving mode includes a corresponding spatiotemporal constraint boundary. When the autonomous vehicle actually drives to the near-field area, the corresponding safe response driving mode is matched and executed according to the actual behavior pattern of the traffic participant and the predicted trajectory of the traffic participant in the consistent prediction result of the scene. S6, as the autonomous vehicle moves forward, the far-field area is continuously updated. The spatiotemporal constraint boundary included in the safe response driving mode of the near-field area is used as a constraint condition for the yielding order consistency determination, occupancy conflict mutual exclusion determination, and passable spatiotemporal interval extraction of the far-field area.The identified traffic participant behavior patterns in the near-field region are used as feedforward constraints to input the interactive trajectory prediction in the far-field region, causing the prediction modality space of the interactive object to be clipped to a subset of modalities corresponding to the behavior patterns. The modality subset refers to the set of prediction modalities corresponding to the behavior patterns in the multimodal trajectory output by the interactive trajectory prediction. S7, the safe response driving mode in the near-field region and the trajectory determined by the far-field region based on the traversable spatiotemporal interval are spatiotemporally connected and fused to output the final planned trajectory. The spatiotemporal connection and fusion adopts an asymmetric arbitration rule, with the deterministic planning result of the near-field region having higher priority than the interactive prediction planning result of the far-field region. The deterministic planning result includes the safe response driving mode and its spatiotemporal constraint boundaries, while the interactive prediction planning result includes the updated trajectory determined by the far-field region based on the traversable spatiotemporal interval.

[0005] Furthermore, when dividing the area in front of the vehicle into a near-field area and a far-field area, the preset distance and the preset near-field distance range are determined based on the vehicle's current speed. The dynamic adaptive adjustment method for the boundary between the near-field area and the far-field area is as follows: the position of the dynamic adaptive boundary is dynamically adjusted according to the current speed of the autonomous vehicle, the current road speed limit, and the density of surrounding traffic participants. The higher the vehicle speed, the higher the road speed limit, and / or the greater the density of traffic participants, the larger the range of the near-field area; the lower the vehicle speed and / or the smaller the density of traffic participants, the smaller the range of the near-field area is. The method for scene clipping of the entire set of preset behavior patterns is as follows: the entire set of preset behavior patterns is clipped based on the current road topology, traffic rules, and interaction object types to obtain a subset of behavior patterns corresponding to the near-field area. As the autonomous vehicle moves forward, the dynamic adaptive boundary continues to move forward, and the far-field area transforms into the near-field area after the autonomous vehicle moves forward, forming an updated far-field area in front of the autonomous vehicle. This cycle forms a closed-loop progression in the rolling time domain.

[0006] Furthermore, the quantitative judgment rule for yielding order consistency includes: extracting conflict areas in the trajectory interaction scenario, where the conflict area refers to the area where the predicted trajectories of two or more interactive objects overlap spatially and the arrival time difference corresponding to the overlapping area is less than a preset time threshold; for each predicted trajectory in the predicted trajectory set, time sampling is performed along the predicted trajectory to obtain multiple trajectory sampling points and their corresponding sampling times; each trajectory sampling point is spatially matched with the conflict area, and the sampling time corresponding to the trajectory sampling point that first falls into the conflict area is determined as the arrival time of the interactive object corresponding to the predicted trajectory to the conflict area; the interactive objects are sorted in ascending order according to the arrival time to generate a predicted passage order sequence; the predicted passage order sequence is subjected to time reversal detection, where time reversal means that for the same conflict area, the relative passage order of the same pair of interactive objects in two different predicted trajectories is opposite; if time reversal exists, the pair of predicted trajectories is determined to violate yielding order consistency, and the pair of predicted trajectories is removed from the predicted trajectory set.

[0007] Specifically, for interactive objects whose behavior patterns have been identified in the near-field area, a deterministic consistency criterion is used to perform time-sequence reversal detection based on the deterministic sorting results of the arrival times. For interactive objects whose behavior patterns have not been identified in the near-field area, a probabilistic consistency criterion is used to perform time-sequence reversal detection based on the confidence interval of the probability distribution of the arrival times. Simultaneously, when the actual behavior pattern of a traffic participant in the near-field area has been identified and the identification confidence of the actual behavior pattern is higher than a preset identification confidence threshold, the passage order corresponding to the actual behavior pattern is used as a locking condition. The predicted trajectory in the updated far-field area is forced to match the passage order of the traffic participant with the locking condition; inconsistent predicted trajectories are directly discarded. When the identification confidence is not higher than the preset identification confidence threshold, the locking is not performed, and the passage order is determined automatically by the interactive trajectory prediction in the updated far-field area. The preset identification confidence threshold is dynamically adjusted according to the scene congestion level.

[0008] Furthermore, the quantification rule for determining the mutual exclusion of occupancy conflicts includes: discretizing the drivable space in the trajectory interaction scenario in three dimensions along the road longitudinal direction, road transverse direction, and time dimension to obtain a set of spatiotemporal units, each spatiotemporal unit being uniquely identified by a longitudinal position interval, a transverse position interval, and a time interval; for each predicted trajectory in the predicted trajectory set, performing time sampling along the predicted trajectory to determine the occupancy time window of the interactive object corresponding to the predicted trajectory in each spatiotemporal unit; for the spatiotemporal unit corresponding to the near-field region, superimposing a safety margin expansion amount determined by the safe response driving mode onto the occupancy time window; the safety margin expansion amount is based on... The current speed of the autonomous vehicle, the type of the interaction object, and the current road friction coefficient are coupled to determine the safety margin expansion. The higher the vehicle speed, the higher the risk level of the interaction object type, and / or the lower the road friction coefficient, the greater the safety margin expansion. A spatiotemporal occupancy matrix is ​​constructed, where rows correspond to spatiotemporal units and columns correspond to interaction objects. The matrix elements are the occupancy time windows of the interaction object in the spatiotemporal unit after safety margin expansion. It is detected whether there are at least two overlapping occupancy time windows of interaction objects in the same spatiotemporal unit. If there is an overlap, it is determined that the predicted trajectory violates the mutual exclusion of occupancy conflicts, and the predicted trajectory is removed from the predicted trajectory set.

[0009] Further, the step of writing back the scene consistency prediction result to the autonomous vehicle candidate trajectory, and extracting the passable spatiotemporal interval of each autonomous vehicle candidate trajectory using the scene consistency prediction result as input, includes: for each autonomous vehicle candidate trajectory, establishing a time alignment mapping relationship between the trajectory space of the candidate trajectory and the trajectory space of each predicted trajectory in the scene consistency prediction result; according to the time alignment mapping relationship, mapping the predicted trajectory corresponding to the candidate trajectory condition in the scene consistency prediction result back to the trajectory space where the candidate trajectory is located, forming a passable state sequence of the candidate trajectory under the constraints of the scene consistency prediction result; sampling along the spatial dimension of the candidate trajectory at equal or adaptive intervals to obtain multiple sampling positions; for each sampling position, determining the time window occupied by any predicted trajectory in the scene consistency prediction result in the time dimension of the sampling position according to the passable state sequence, and determining the complement of the time window of the sampling position that is not occupied in the time dimension as the sampling position. The passable time intervals of all sampled locations are combined in spatial order to obtain the passable spatiotemporal interval of the candidate trajectory. Furthermore, the deterministic passable time interval determined by the near-field region based on the safe response driving mode and the probabilistic passable time interval extracted by the far-field region based on the scene consistency prediction result are spatially spliced ​​along the vehicle's travel direction to form a passable spatiotemporal interval covering the entire route. The spatiotemporal constraint boundary contained in the safe response driving mode of the near-field region is injected as a hard constraint into the extraction process of the updated probabilistic passable time interval of the far-field region, so that the updated probabilistic passable time interval of the far-field region automatically meets the requirements of the spatiotemporal constraint boundary at the near-field boundary. At the splicing interface, the deterministic passable time interval of the near-field region is given priority, and the probabilistic passable time interval of the far-field region is only used as a backup constraint. When the two conflict, the deterministic passable time interval of the near-field region is used as the arbitration benchmark.

[0010] Furthermore, the present invention also includes closed-loop feedback iteration: when performing yielding order consistency and occupancy conflict mutual exclusion determination, a consistency metric is calculated; the consistency metric includes at least one of the number of conflict areas, the number of time sequence reversals, and the spatiotemporal unit occupancy conflict rate; results where the consistency metric is lower than a preset threshold are fed back to the interactive object screening and interactive trajectory prediction steps, expanding the screening range of interactive objects or increasing the number of predicted trajectories of the interactive object in the interactive trajectory prediction, and re-executing the interactive trajectory prediction and consistency determination until the consistency metric meets the preset convergence condition or reaches the maximum number of iterations; when the near-field area detects a change in the behavior pattern of a traffic participant, the switching amplitude of the behavior pattern switch is calculated, the switching amplitude being jointly determined by the traffic participant's speed change rate and heading angle change rate. The heading angle change rate is defined as the rate of change of the heading angle of traffic participants over time. If the switching amplitude exceeds a preset behavior change threshold, the updated far-field interactive trajectory prediction is re-executed, with behavior mode switching serving as the highest priority iteration trigger condition. If the switching amplitude does not exceed the preset behavior change threshold, the updated far-field interactive trajectory prediction is not re-executed. The arbitration state of the asymmetric arbitration rule is configured with a hysteresis condition. When the arbitration state switches from far-field to near-field, the conflict level between the updated far-field and near-field must be lower than a preset switching threshold for multiple consecutive planning cycles. When the arbitration state switches from near-field to far-field, the conflict level between the near-field and the updated far-field must be higher than a preset trigger threshold, and the trigger threshold must be greater than the switching threshold.

[0011] Furthermore, it also includes prediction uncertainty quantification: obtaining the uncertainty measure of each prediction trajectory in the prediction trajectory set, the uncertainty measure including the probability distribution of the corresponding prediction trajectory, confidence score, and cognitive uncertainty and accidental uncertainty quantified through deep evidence learning; in the yielding order consistency determination, calculating the probability distribution of the arrival time of each interactive object to the conflict area, and determining that there is an order contradiction if and only if the confidence intervals of the arrival time probability distributions of two interactive objects do not overlap at a preset confidence level, and the prediction trajectories of the two interactive objects give opposite relative passage orders; in the occupancy conflict mutual exclusion determination, calculating the occupancy probability of each interactive object in each spatiotemporal unit, and determining that the occupancy conflict mutual exclusion is violated if and only if the occupancy probability of at least two interactive objects simultaneously exceeds a preset probability threshold; the probability occupancy information is transmitted as a soft constraint to the extraction of the passable spatiotemporal interval; wherein, in the near field area, based on the identified and determined behavior pattern, the cognitive uncertainty at the behavior level of the traffic participant is eliminated, and only the accidental uncertainty brought about by motion state measurement is retained; the updated far field area retains complete cognitive uncertainty and accidental uncertainty quantification; the near field area and the updated far field area adopt a mutually distinguishable uncertainty quantification strategy.

[0012] Furthermore, it also includes the calculation and adjustment of scene consistency confidence: based on the proportion of predicted trajectory pairs with detected sequential contradictions to the total number of predicted trajectory pairs, the proportion of spatiotemporal units with occupancy conflicts to the total number of spatiotemporal units, and the confidence of near-field region behavior pattern recognition, the scene consistency confidence score Q is calculated, and the formula for calculating Q is: Where, N 检 The logarithm of the predicted trajectory that detects inconsistencies in order is dimensionless; N 总 V is the total logarithm of predicted trajectory combinations, dimensionless; 冲 V represents the number of spatiotemporal units where occupancy conflicts occur; it is dimensionless. 总 C represents the total number of spatiotemporal units, dimensionless. 近 λ1, λ2, and λ3 are the confidence scores for near-field region behavior pattern recognition, which are dimensionless and range from [0,1]; λ1, λ2, and λ3 are preset weight coefficients, which are dimensionless and all greater than zero; Q is the scene consistency confidence score, which is dimensionless; exp is the natural exponential function; and the confidence score C for near-field region behavior pattern recognition is... 近 Directly attenuate the global scenario consistency confidence Q, C 近 The lower the value of Q, the lower the safety margin, and the greater the safety margin across all scenarios. The safety margin is dynamically adjusted based on Q when extracting passable time and space intervals. The higher the value of Q, the smaller the safety margin; the lower the value of Q, the larger the safety margin.

[0013] Furthermore, it also includes prediction error boundary propagation: obtaining the prediction error boundary of each prediction trajectory in the scene consistency prediction results; when extracting the passable spatiotemporal interval, shrinking the error boundary of the passable time interval for each sampling location, deducting the time deviation corresponding to the prediction error boundary from the passable time interval to obtain the shrunken passable time interval; wherein, the shrinkage amplitude of the updated far-field region error boundary is dynamically adjusted according to the scene consistency confidence score Q, the higher Q, the smaller the shrinkage amplitude, and the lower Q, the larger the shrinkage amplitude; the shrinkage amplitude of the near-field region error boundary is independently determined according to the near-field safety level, the near-field safety level is graded according to the type of traffic participant and the conflict risk level, different levels correspond to different error boundary shrinkage amplitudes, which are not dynamically adjusted with the scene consistency confidence score Q; the conflict risk level includes close-range crossing risk, sudden stop risk when following another vehicle, and lane change intrusion risk.

[0014] Furthermore, this invention also provides an autonomous driving path planning system based on interactive trajectory prediction. This system implements the aforementioned autonomous driving path planning method based on interactive trajectory prediction, comprising: a candidate trajectory generation module; used to acquire the autonomous vehicle's state, road topology, and historical trajectories of surrounding traffic participants to generate candidate trajectories for the autonomous vehicle; and a spatial layering and partitioning module; used to divide the area to be driven in front of the vehicle into a near-field region and a far-field region according to the autonomous vehicle's travel distance, wherein the near-field region is a segment of area within a preset near-field distance range along the vehicle's travel direction and the distance between the vehicle and its current position, and the far-field region is the remaining area of ​​the area to be driven outside the near-field region; the near-field region and the... The boundaries between the far-field regions are dynamically adaptive boundaries, adjusted dynamically based on the current speed of the autonomous vehicle, the current road speed limit, and the density of surrounding traffic participants. An interactive scenario construction module is used to select interactive objects for each candidate trajectory of the autonomous vehicle within the far-field regions based on arrival time and spatial overlap, constructing trajectory interaction scenarios. An interactive prediction module is used to perform interactive trajectory prediction on the interactive objects based on the candidate trajectories of the autonomous vehicles in the trajectory interaction scenarios, obtaining a set of predicted trajectories. An inference post-processing module, independent of the interactive prediction module, is used to process the predicted trajectory set according to manually preset rules and thresholds, following quantitative judgment rules for yielding order consistency and occupancy conflict mutual exclusion. The system performs filtering to remove contradictory predicted trajectories, resulting in a consistent prediction result. A near-field deterministic planning module is used to perform scene trimming on the entire set of preset behavior patterns, combining the current road topology, traffic rules, and interaction object types, to obtain a subset of behavior patterns. It enumerates each behavior pattern in the subset and pre-generates a corresponding safe response driving mode for each behavior pattern, with each safe response driving mode containing corresponding spatiotemporal constraint boundaries. When the autonomous vehicle actually travels to the near-field area, it matches and executes the corresponding safe response driving mode based on the actual behavior pattern of the traffic participant and the predicted trajectory of that participant in the consistent prediction result. A spatiotemporal interval extraction module is used to extract the scene... The consistent scene prediction results are written back to the candidate trajectories of autonomous vehicles. The passable spatiotemporal intervals of each candidate trajectory of the autonomous vehicle are extracted using the consistent scene prediction results as input. The spatiotemporal constraint boundary output by the near-field deterministic planning module is used as a constraint condition input to the spatiotemporal interval extraction module. The hierarchical fusion decision module is used to spatiotemporally connect and fuse the safe response driving mode in the near-field region with the trajectory determined based on the passable spatiotemporal interval in the far-field region, and output the final planned trajectory. The hierarchical fusion decision module is configured with arbitration priority rules. The deterministic planning result of the near-field region has higher priority than the updated interactive prediction planning result of the far-field region. In the event of a spatiotemporal conflict, the deterministic planning result of the near-field region shall prevail.The spatiotemporal constraint boundaries output by the near-field deterministic planning module are fed back to the interactive scene construction module and the inference post-processing module, achieving cross-module constraint linkage. The identified traffic participant behavior patterns output by the near-field deterministic planning module are used as feedforward constraints input to the interactive prediction module, causing the prediction modality space of the interactive object to be pruned into a subset of modalities corresponding to the behavior patterns.

[0015] The beneficial effects of this invention are as follows: This invention proposes an autonomous driving path planning method and system based on interactive trajectory prediction. The advantages are: 1. This invention uses a post-processing layer for inference, independent of interactive trajectory prediction, to filter the predicted trajectory set according to manually preset rules and thresholds. Scene consistency verification is implemented in the inference stage using rule-based filtering, eliminating the need to modify or retrain the interactive trajectory prediction model. This allows for adaptation to existing interactive trajectory prediction models, reducing engineering deployment costs. 2. Based on quantitative judgment rules for yielding order consistency and occupancy conflict mutual exclusion, this invention removes contradictory predicted trajectories from the predicted trajectory set, obtaining scene-consistent prediction results. These results are then written back to the candidate trajectories of the autonomous vehicle to extract passable spatiotemporal intervals. This eliminates false passable information caused by decision-making contradictions between predicted trajectories, improving the safety of planning decisions. 3. This invention separates the near-field region from the far-field region. The near-field region does not rely on probabilistic trajectory prediction, but instead uses behavior pattern enumeration and pre-generated safe response driving modes. When a vehicle actually drives into the near-field region, the corresponding safe response driving mode is directly matched and executed based on the actual behavior pattern of the traffic participants. This eliminates the need to repeatedly weigh multiple possible predicted trajectories, solving the decision lag problem caused by the reliance on probabilistic prediction in the near-field region. 4. This invention uses the spatiotemporal constraint boundary of the near-field region as a constraint condition to input the yielding order consistency determination, occupancy conflict mutual exclusion determination, and passable spatiotemporal interval extraction in the far-field region. The identified behavior patterns in the near-field region are used as feedforward constraints to input the interactive trajectory prediction in the far-field region. This allows the near-field identification results to participate in far-field prediction and verification, and the consistent prediction results of the far-field scene provide a basis for near-field processing, forming cross-layer coupling constraints and improving the overall robustness of the system. Attached Figure Description

[0016] Figure 1 This is a flowchart of an autonomous driving path planning method based on interactive trajectory prediction according to the present invention.

[0017] Figure 2 This is a schematic diagram of an autonomous driving path planning system based on interactive trajectory prediction according to the present invention. Detailed Implementation

[0018] The following combination Figure 1 and Figure 2The specific embodiments of the present invention will be described in detail below. The following embodiments are used to illustrate how the technical solutions of the present invention are implemented. Those skilled in the art can replace or adjust the specific parameters, module settings and implementation methods in the following embodiments without departing from the concept of the present invention.

[0019] The autonomous driving path planning method of this embodiment runs on the onboard computing platform of the autonomous vehicle. The onboard computing platform includes a perception sensor group, a domain controller, and a vehicle control execution unit. The perception sensor group includes a lidar, millimeter-wave radar, a forward-looking camera, a surround-view camera, and a GNSS / IMU combined positioning unit, etc., for acquiring the autonomous vehicle's status, road topology, and historical trajectories of surrounding traffic participants. The domain controller includes a GPU computing unit and a CPU computing unit for running the path planning method of this invention. The vehicle control execution unit includes a longitudinal control actuator and a lateral control actuator for controlling the vehicle's movement according to the final planned trajectory. The output of the perception sensor group is transmitted to the domain controller via an onboard Ethernet, and the domain controller runs the path planning method of this invention and outputs the final planned trajectory to the vehicle control execution unit.

[0020] like Figure 1 The diagram shows a flowchart of an autonomous driving path planning method based on interactive trajectory prediction according to the present invention. The flowchart includes steps S1 to S7. Step S1 involves generating vehicle state and candidate trajectories; Step S2 involves acquiring the entire set of preset behavior patterns; Step S3 involves near-field and far-field region division; Step S4 involves far-field interactive trajectory prediction, scene consistency filtering, and extraction of passable spatiotemporal intervals; Step S5 involves near-field behavior pattern matching and safety response generation; Step S6 involves near-field information feedforward and far-field rolling update; and Step S7 involves asymmetric arbitration fusion and planned trajectory output. Figure 2 The diagram shows a schematic of an autonomous driving path planning system based on interactive trajectory prediction according to the present invention. The schematic shows that the system of the present invention includes: a candidate trajectory generation module A10, a spatial hierarchical partitioning module A20, an interactive scene construction module A30, an interactive prediction module A40, an inference post-processing module A50, a near-field deterministic planning module A60, a spatiotemporal interval extraction module A70, and a hierarchical fusion decision module A80. The modules interact with each other through the data bus inside the domain controller.

[0021] In step S1, the candidate trajectory generation module A10 acquires the autonomous vehicle's state, road topology, and historical trajectories of surrounding traffic participants to generate candidate trajectories for the autonomous vehicle. Specifically, the candidate trajectory generation module A10 reads the vehicle's speed, acceleration, yaw rate, steering wheel angle, and brake pedal opening via the vehicle controller area network (CAN) bus, obtains the vehicle's absolute position and heading angle via the GNSS / IMU combined positioning unit, and takes the state sequence within the most recent 0.5 to 1 second time window as the current state input. The system reads lane geometry information (including lane centerline equation, lane width and curvature), lane connection relationships (including forks, merging and intersection connections), traffic signs and traffic markings from a pre-set high-precision map database within a range of 50 to 200 meters in front of the vehicle. The candidate trajectory generation module A10 detects surrounding traffic participants (including vehicles, pedestrians and cyclists) through a group of perception sensors and obtains the tracking trajectory of each traffic participant through a multi-target tracking algorithm. Each tracking trajectory includes the position coordinates, speed and heading angle at each sampling time, and the tracking trajectory duration window is 2 to 3 seconds.

[0022] Specifically, the candidate trajectory generation module A10 generates candidate trajectories based on the vehicle's current state and road topology using a decoupled sampling method for lateral and longitudinal directions based on the Frenet coordinate system. The candidate trajectory generation module A10 projects the vehicle onto the Frenet coordinate system corresponding to the road reference line to obtain the longitudinal position s0 and the lateral position d0. In the longitudinal dimension, using the vehicle's current speed v0 as a reference, the target speed is sampled in the speed range [v0-5m / s, v0+5m / s] with a step size of 1m / s, and in the acceleration range [-3m / s², +2m / s²]... Within [d0 / s²], target acceleration is sampled with a step size of 0.5 m / s², and the prediction time domain is 5 seconds; in the lateral dimension, the target lateral offset is sampled with a step size of 0.5 m within the lateral offset interval [d0-3.5m, d0+3.5m]; the sampling results of the longitudinal and lateral dimensions are combined, and each candidate trajectory is generated by fitting a fifth-order polynomial, satisfying the continuity constraints of position, velocity, and acceleration; candidate trajectories that do not satisfy the kinematic constraints and road boundary constraints are eliminated, where the kinematic constraints include: maximum curvature constraint and maximum acceleration constraint; the candidate trajectory set {T} is retained. 候 (1), T 候 (2), ..., T 候 (N)}; where, T 候 (i) represents the i-th candidate trajectory in the candidate trajectory set, where i is the index of the candidate trajectory, ranging from 1 to N; T 候(i) It consists of a series of trajectory points arranged in chronological order. Each trajectory point contains the position coordinates, speed and heading angle of the vehicle at that moment; N is the number of candidate trajectories, which represents the total number of candidate trajectories retained after elimination, usually 50 to 200.

[0023] In step S2, the preset behavior pattern set is a pre-constructed set of behavior patterns that covers all behavior types of traffic participants. Each behavior pattern is defined by a behavior type identifier and the corresponding kinematic parameter range. Specifically, the preset behavior pattern set is constructed offline during the system initialization phase and stored in the storage unit of the domain controller. When constructing the preset behavior pattern set, the behavior types of traffic participants are extracted from traffic rules, including following other vehicles, driving freely, changing lanes, overtaking, slowing down to yield, stopping to yield, turning left, turning right, going straight, and crossing the road. The kinematic parameter ranges for each behavior type were statistically analyzed from historical traffic data, which included publicly available driving datasets and real-vehicle data. The kinematic parameters included speed range, acceleration range, heading angle rate of change range, following distance range, and lateral speed range. For example, for following behavior, the statistically obtained kinematic parameter ranges were: speed 0 m / s to 30 m / s, acceleration -4 m / s² to 2 m / s², heading angle rate of change -0.05 rad / s to 0.05 rad / s, and following distance 0.8 seconds to 3.0 seconds. For lane-changing behavior, the statistically obtained kinematic parameter ranges were: speed 5 m / s to 35 m / s, lateral speed 0.2 m / s to 2.0 m / s, and heading angle rate of change 0.05 rad to 0.5 rad. The behavior type identifiers for each behavior type were combined with their corresponding kinematic parameter ranges to form a complete set of preset behavior patterns.

[0024] In step S3, the spatial layering and partitioning module A20 divides the area to be driven in front of the autonomous vehicle into a near-field area and a far-field area based on the distance traveled by the autonomous vehicle. The area to be driven is the region extending along the planned path of the autonomous vehicle, within a preset distance from the vehicle's current position. The preset distance of the area to be driven is determined based on the vehicle's current speed; the higher the speed, the larger the preset distance. The near-field area is a section of the area where the distance between the vehicle's current position and the vehicle's direction of travel is within a preset near-field distance range. The near-field distance is determined based on the vehicle's current speed; the higher the speed, the larger the near-field distance. The far-field area is the remaining area of ​​the area to be driven outside the near-field area. The boundary between the near-field area and the far-field area is a dynamically adaptive boundary.

[0025] Specifically, during vehicle operation, the spatial layering and partitioning module A20 recalculates the near-field distance according to the planning period (100 milliseconds) to achieve dynamic adaptive adjustment; the dynamic adaptive adjustment method is as follows: near-field distance D 近Take the preset minimum near-field distance D min The larger of the calculated values; where D 近 D represents the spatial length of the near-field region along the vehicle's direction of travel, in meters. min This represents the preset minimum near-field distance, used to prevent the near-field area from being too short; it is set to 10 meters. The calculated value is the vehicle's current speed v. 当 Multiplied by the preset near-field time constant T 近 Then multiply by the density correction factor; where v 当 This indicates the vehicle's current speed, expressed in meters per second; T 近 This represents the preset near-field time constant, taken as 2 seconds; the density correction factor is 1 plus the density influence coefficient α and the density ρ of surrounding traffic participants. 参 The product of ρ and α, where α ranges from 0.1 to 0.3 and is used to adjust the degree of influence of traffic participant density on the near-field region length; 参 The unit is vehicles / 100 meters. When the density of traffic participants is high, the density correction factor is greater than 1, and the near-field area expands; when the density of traffic participants is low, the density correction factor is close to 1, and the near-field area shrinks; for example, when v 当 = 10m / s, T 近 = 2s, α=0.2, ρ 参 When = 3 vehicles / 100m, D 近 =max(10,10×2×(1+0.2×3))=32m.

[0026] In step S4, the interactive scene construction module A30, the interactive prediction module A40, the inference post-processing module A50, and the spatiotemporal interval extraction module A70 work together to perform interactive trajectory prediction, scene consistency filtering, and passable spatiotemporal interval extraction in the far field region.

[0027] Specifically, the interaction scenario construction module A30 filters interaction objects based on arrival time sequence and spatial overlap for each candidate trajectory of the autonomous vehicle, and constructs trajectory interaction scenarios; more specifically, the interaction scenario construction module A30 constructs interaction scenarios for each candidate trajectory T. 候 (i) Calculate the predicted extension of the candidate trajectory and the historical trajectories of each traffic participant in the time dimension, and perform short-term extrapolation based on the constant speed model or constant acceleration model to obtain the coarse predicted trajectory T of traffic participant j. 粗 (j); where T 粗 (j) represents the rough predicted trajectory of traffic participant j in the near future, where j is the traffic participant's index, and T... 粗 (j) consists of a series of trajectory points arranged in chronological order; the interactive scene construction module A30 calculates the candidate trajectory T. 候 (i) and coarse predicted trajectory T 粗(j) Spatial overlap relationship, if the spatial distance between the two trajectories at any time is less than the preset overlap threshold d 阈 If d , then it is determined that there is spatial overlap; where d 阈 To determine whether two trajectories are spatially close, a distance of 3 meters is used. For traffic participants with spatial overlap, the interaction scene construction module A30 calculates the arrival time of the traffic participant to the overlapping area. If the difference between this arrival time and the arrival time of the vehicle to the overlapping area is less than a preset time threshold t... 阈 Then the traffic participant will be included in the set of interaction objects O. 交 (i); where t 阈 Used to determine whether there is a possibility of temporal interaction between a traffic participant and the vehicle, a time interval of 5 seconds is taken; O 交 (i) indicates that the candidate trajectory T 候 (i) A set of traffic participants with interactive relationships; the interactive scenario construction module A30 consists of the set of interactive objects O included in the module. 交 (i) Construct the trajectory interaction scenario SC(i); where SC(i) represents the interaction between the candidate trajectory T and the trajectory T. 候 (i) The corresponding trajectory interaction scenario, including the candidate trajectory and the set of interaction objects O 交 (i) shows the historical trajectory and coarse predicted trajectory of each interactive object.

[0028] The interactive prediction module A40 uses the candidate trajectories of autonomous vehicles in the trajectory interaction scenario as conditions to perform interactive trajectory prediction on the interactive objects, obtaining a set of predicted trajectories. In this embodiment, the interactive prediction module A40 uses a conditional variational autoencoder (CVAE) as the interactive trajectory prediction model. The encoder network of the CVAE includes two layers of long short-term memory networks with 64 and 32 hidden units, respectively; the decoder network includes two layers of long short-term memory networks with 32 and 64 hidden units, respectively; the model input includes the historical trajectory of each interactive object, the road topology encoding vector, and the candidate trajectory T. 候 (i) The future trajectory encoding vector; the model output is the set of predicted trajectories for each interactive object, with each predicted trajectory accompanied by a probability value; for interactive object j, the set of predicted trajectories is T. 预 (j)= {(τ j (1), p j (1)), (τ j (2), p j (2)), ..., (τ) j (K), p j (K))};wherein, T 预 (j) represents the set of predicted trajectories for interactive object j; τ j (k) represents the k-th predicted trajectory of the interactive object j, which consists of a series of trajectory points arranged in chronological order; pj (k) represents the probability value corresponding to the k-th predicted trajectory; K represents the number of prediction modes, i.e., the number of predicted trajectories output by the interactive prediction module, which is 6; k is the index of the prediction mode, which ranges from 1 to K; each probability value satisfies Σp j (k) = 1; The model is trained using a public driving dataset. After training, the model parameters are fixed and are not updated during the inference phase.

[0029] Furthermore, the inference post-processing module A50 is independent of the interactive prediction module A40 and does not participate in the training process of the interactive trajectory prediction model. It only uses manually preset rules and thresholds to filter the predicted trajectory set. The inference post-processing module A50 performs yield order consistency determination and occupancy conflict mutual exclusion determination. The inference post-processing module A50 is an independent software module running on the domain controller, implemented in the C++ coding language, and filters the predicted trajectory set output by the interactive prediction module A40 through preset rules and thresholds. The inference post-processing module A50 communicates with the interactive prediction module A40 through a data interface. The inference post-processing module A50 does not participate in the model training process of the interactive prediction module A40, and there is no need to modify or retrain the interactive trajectory prediction model. It can be adapted to existing interactive trajectory prediction models, thereby reducing engineering deployment costs.

[0030] In the above embodiments, specifically, the yielding order consistency determination includes the following operations: the inference post-processing module A50 extracts the conflict region of the trajectory interaction scenario SC(i). The conflict region refers to the region where the predicted trajectories of at least two interactive objects overlap spatially and the arrival time difference corresponding to the overlapping region is less than a preset time threshold, which is 3 seconds here; for each predicted trajectory in the predicted trajectory set, the inference post-processing module A50 performs time sampling along the predicted trajectory at a sampling interval of 0.1 seconds to obtain multiple trajectory sampling points and their corresponding sampling times. The trajectory sampling points are spatially matched with the conflict region, and the sampling time corresponding to the trajectory sampling point that first falls into the conflict region is determined as the arrival time t of the interactive object to the conflict region. 到 (j, k); where t 到 (j, k) represents the arrival time of the k-th predicted trajectory of interactive object j to the conflict area, in seconds; the inference post-processing module A50 sorts each interactive object in ascending order according to the arrival time to generate a predicted passage order sequence; the inference post-processing module A50 performs time reversal detection on the predicted passage order sequence. Time reversal means that for the same conflict area, the relative passage order of the same pair of interactive objects in two different predicted trajectories is reversed; if time reversal exists, the inference post-processing module A50 determines that the pair of predicted trajectories violates the yielding order consistency and removes the pair of predicted trajectories from the predicted trajectory set.

[0031] In addition, the inference post-processing module A50 uses a deterministic consistency criterion to perform time-series reversal detection based on the deterministic ranking of arrival times for interactive objects whose behavior patterns have been identified in the near-field region; for interactive objects whose behavior patterns have not been identified in the near-field region, it uses a probabilistic consistency criterion to perform time-series reversal detection based on the confidence interval of the probability distribution of arrival times; when the near-field deterministic planning module A60 identifies the actual behavior pattern of a traffic participant and the identification confidence level is higher than the preset identification confidence level threshold of 0.8, the inference post-processing module A50 allows the passage corresponding to the actual behavior pattern. The order is used as a locking condition. The predicted trajectory of the traffic participant in the updated far-field area must be consistent with the locking condition. Inconsistent predicted trajectories are directly eliminated. When the identification confidence is not higher than the preset identification confidence threshold, the inference post-processing module A50 does not perform locking and lets the interactive trajectory prediction of the updated far-field area determine the passage order itself. The preset identification confidence threshold is dynamically adjusted according to the scene congestion level. When the scene congestion level is high, the preset identification confidence threshold is increased to 0.9. When the scene congestion level is low, the preset identification confidence threshold is decreased to 0.7.

[0032] In the above embodiments, specifically, the occupancy conflict mutual exclusion determination includes the following operations: the inference post-processing module A50 discretizes the drivable space in the trajectory interaction scene SC(i) in three dimensions along the road longitudinal direction, road transverse direction, and time dimension. The longitudinal discretization interval Δs is 1 meter, the transverse discretization interval Δd is 0.5 meters, and the time discretization interval Δt is 0.2 seconds. Specifically, Δs represents the discretization interval along the road longitudinal direction, Δd represents the discretization interval along the road transverse direction, and Δt represents the discretization interval along the time dimension. After discretization, a set of spatiotemporal units is obtained, and each spatiotemporal unit is uniquely identified by the longitudinal position interval, the transverse position interval, and the time interval. For each predicted trajectory in the predicted trajectory set, the inference post-processing module A50 performs time sampling along the predicted trajectory to determine the occupancy time window of the interactive object corresponding to the predicted trajectory in each spatiotemporal unit. Block A50 superimposes a safety margin inflation determined by the safe response driving mode onto the occupancy time window of the spatiotemporal unit corresponding to the near-field region. The safety margin inflation is determined by the coupling of the current vehicle speed, the type of interaction object, and the current road friction coefficient. The higher the vehicle speed, the higher the risk level of the interaction object type, and / or the lower the road friction coefficient, the larger the safety margin inflation. The inference post-processing module A50 constructs a spatiotemporal occupancy matrix. The rows of the spatiotemporal occupancy matrix correspond to spatiotemporal units, and the columns correspond to interaction objects. The matrix elements are the occupancy time windows of the interaction object in the spatiotemporal unit after safety margin inflation. The inference post-processing module A50 detects whether there are at least two overlapping occupancy time windows of interaction objects in the same spatiotemporal unit. If there is an overlap, it determines that the predicted trajectory violates the mutual exclusion of occupancy conflicts and removes the predicted trajectory from the predicted trajectory set.

[0033] Furthermore, the inference post-processing module A50 retains the predicted trajectory that simultaneously satisfies yield order consistency and occupancy conflict mutual exclusion as the scene consistency prediction result; thus, scene consistency verification is implemented in the inference stage using a rule-based filtering method, without the need to modify or retrain the interactive trajectory prediction model, and can be adapted to existing interactive trajectory prediction models, reducing engineering deployment costs.

[0034] Furthermore, the spatiotemporal interval extraction module A70 writes back the scene consistency prediction results to the candidate trajectories of the autonomous vehicle, and extracts the passable spatiotemporal intervals of each candidate trajectory of the autonomous vehicle using the scene consistency prediction results as input; specifically, for each candidate trajectory T, the spatiotemporal interval extraction module A70 extracts the passable spatiotemporal intervals. 候(i) Establish a time alignment mapping relationship between the trajectory space of the candidate trajectory and the trajectory space of each predicted trajectory in the scene consistent prediction result; according to the time alignment mapping relationship, map the predicted trajectory corresponding to the candidate trajectory condition in the scene consistent prediction result back to the trajectory space where the candidate trajectory is located, forming a sequence of passable states of the candidate trajectory under the constraints of the scene consistent prediction result; sample along the spatial dimension of the candidate trajectory at a preset interval, here 1 meter, to obtain multiple sampling positions; for each sampling position, determine the time window occupied by any predicted trajectory in the scene consistent prediction result in the time dimension of the sampling position according to the passable state sequence, and determine the complement of the time window not occupied in the time dimension of the sampling position as the passable time interval of the sampling position; combine the passable time intervals of all sampling positions in spatial order to obtain the passable spatiotemporal interval of the candidate trajectory; since the passable spatiotemporal interval is extracted with the scene consistent prediction result as input, rather than with the original predicted trajectory set output by interactive trajectory prediction as input, the passable spatiotemporal interval eliminates false passable information caused by decision contradictions between predicted trajectories, thus improving the security of planning decisions.

[0035] In step S5, the near-field deterministic planning module A60 performs behavior pattern matching and safety response generation for the near-field region. The near-field region is the area that has previously undergone interactive trajectory prediction and scene consistency filtering as a far-field region. Based on the obtained scene consistency prediction results, the near-field deterministic planning module A60 performs scene clipping on the entire set of preset behavior patterns to obtain a subset of behavior patterns corresponding to the near-field region. The clipping method combines the current road topology, traffic rules, and interaction object types to select behavior patterns that match the scene of the near-field region from the entire set of preset behavior patterns. For example, when the near-field region is a straight-line following scenario, the clipped behavior pattern subset is following, free driving, and slowing down to yield; when the near-field region is an intersection scenario, the clipped behavior pattern subset is stopping to yield, turning left, turning right, going straight, and slowing down to yield. Specifically, the near-field deterministic planning module A60 performs scene matching and safety response generation for the near-field region. The planning module A60 sets each behavior mode as a subset of the pattern. For each behavior mode, a corresponding safe response driving mode for the autonomous vehicle is pre-generated. Each safe response driving mode includes corresponding spatiotemporal constraint boundaries. Taking the following behavior mode as an example, the corresponding safe response driving mode is to maintain a time distance from the vehicle in front of it that is not less than the preset minimum safe time distance (1.2 seconds), with a speed adjustment range within [-2m / s², +1m / s²]. The spatiotemporal constraint boundary is that the longitudinal position of the vehicle in the near field area must not exceed the position of the rear bumper of the vehicle in front minus the preset minimum safe distance (5 meters). Taking the deceleration and yielding behavior mode as an example, the corresponding safe response driving mode is to reduce the speed to below the preset yielding speed (2m / s) before reaching the conflict area, and stop and wait if necessary. The spatiotemporal constraint boundary is that the time when the vehicle arrives at the conflict area must not be earlier than the time when the interactive object passes through the conflict area.

[0036] Furthermore, when the autonomous vehicle actually enters the near-field area, the near-field deterministic planning module A60 uses a group of perception sensors to detect the actual behavior patterns of each traffic participant in real time. Based on the actual behavior patterns of the traffic participants and the predicted trajectory of the traffic participants in the scene consistency prediction results, it matches and executes the corresponding safety response driving mode from a list of pre-generated safety response driving modes. The matching rule is as follows: if the actual behavior pattern is consistent with the predicted trajectory in the scene consistency prediction results, the safety response driving mode corresponding to that behavior pattern is directly matched; if they are inconsistent, the actual behavior pattern is taken as the standard, and the safety response driving mode corresponding to the actual behavior pattern is matched from the pre-generated safety response driving modes. Thus, the near-field area does not rely on probabilistic trajectory prediction, but instead uses behavior pattern enumeration and pre-generation of safety response driving modes. When the vehicle actually enters the near-field area, it directly matches and executes the corresponding safety response driving mode based on the actual behavior patterns of the traffic participants, without having to repeatedly weigh multiple possible predicted trajectories, thus solving the decision lag problem caused by relying on probabilistic prediction in the near-field area.

[0037] In step S6, as the autonomous vehicle moves forward, the far-field region is continuously updated; the near-field region is also continuously updated, except that the original far-field region becomes the near-field region. The update cycle is consistent with the planning cycle (100 milliseconds). In each planning cycle, the spatial layering and partitioning module A20 recalculates the boundary between the near-field region and the far-field region. The part of the far-field region that is closer to the vehicle becomes the near-field region after the vehicle moves forward, and the updated far-field region is formed in front of the autonomous vehicle. The spatiotemporal constraint boundary of the safe response driving mode output by the near-field deterministic planning module A60 is used as a constraint condition input for the yielding order consistency determination, occupancy conflict mutual exclusion determination, and passable spatiotemporal interval extraction of the updated far-field region. For example, if the currently generated safe response driving mode in the near-field region requires maintaining a safe distance of 5 meters from the vehicle in front, then this safe distance constraint is used as a hard constraint condition when extracting the passable spatiotemporal interval of the far-field region, so that the probabilistic passable time interval of the far-field region automatically meets the safe distance requirement at the near-field boundary.

[0038] Furthermore, the identified traffic participant behavior patterns output by the near-field deterministic planning module A60 are used as feedforward constraints input to the interactive prediction module A40, causing the prediction modality space of the interactive object to be pruned into a subset of modalities corresponding to the behavior patterns. A modality subset refers to the set of prediction modes corresponding to the behavior patterns in the multimodal trajectories output by the interactive trajectory prediction. For example, if the actual behavior pattern of the vehicle ahead is identified as deceleration and yielding in the near-field region, then the modality subset corresponding to this behavior pattern is the deceleration trajectory mode. The interactive prediction module A40 prunes the prediction modality space of this vehicle to retain only the deceleration trajectory mode and removes other irrelevant prediction modes. Thus, the near-field identification results participate in far-field prediction and verification, and the consistent prediction results of the far-field scene provide a basis for near-field processing, forming cross-layer coupling constraints and improving the overall robustness of the system.

[0039] In addition, in step S6, closed-loop feedback iteration is also performed. Specifically, when performing yielding order consistency and occupancy conflict mutual exclusion determination, the inference post-processing module A50 calculates a consistency metric. The consistency metric includes at least one of the number of conflict regions, the number of time sequence reversals, and the spatiotemporal unit occupancy conflict rate. The inference post-processing module A50 feeds back the result that the consistency metric is lower than a preset threshold to the interactive scene construction module A30 and the interactive prediction module A40. The interactive scene construction module A30 expands the selection range of interactive objects, or the interactive prediction module A40 increases the number of predicted trajectories of the interactive object in the interactive trajectory prediction, and re-executes the interactive trajectory prediction and consistency determination until the consistency metric meets the preset convergence condition or reaches the maximum number of iterations.

[0040] When the near-field deterministic planning module A60 detects a change in the behavior pattern of a traffic participant, it calculates the magnitude of this change. The magnitude is jointly determined by the rate of change of the traffic participant's speed and the rate of change of its heading angle, where the rate of change of the heading angle is the rate of change of the traffic participant's heading angle over time. If the magnitude of the change exceeds a preset behavior change threshold, the updated far-field interactive trajectory prediction is re-executed, with the behavior pattern change serving as the highest priority iteration trigger condition. If the magnitude of the change does not exceed the preset behavior change threshold, the updated far-field interactive trajectory prediction is not re-executed.

[0041] In step S7, the hierarchical fusion decision module A80 performs spatiotemporal connection and fusion between the safe response driving mode in the near-field region and the trajectory determined based on the passable spatiotemporal interval in the far-field region, outputting the final planned trajectory. The hierarchical fusion decision module A80 also spatially stitches the deterministic passable time interval determined by the safe response driving mode in the near-field region with the probabilistic passable time interval extracted based on the scene consistency prediction result in the far-field region along the vehicle's travel direction. At the stitching interface, the deterministic passable time interval of the near-field region is given priority, while the probabilistic passable time interval of the far-field region is only used as a backup constraint. When a conflict arises, the deterministic passable time interval of the near-field region is used as the arbitration benchmark. Specifically, the spatiotemporal connection and fusion adopts an asymmetric arbitration rule, with the deterministic planning result of the near-field region having higher priority than the interactive result of the far-field region. The predictive planning results include: deterministic planning results including safe driving modes and their spatiotemporal constraints; interactive predictive planning results including the updated far-field region and the trajectory determined based on the traversable spatiotemporal interval; the arbitration state is configured with hysteresis conditions: when the arbitration state switches from far-field to near-field, the conflict level between the updated far-field and near-field must be lower than the preset switching threshold of 0.3 for three consecutive planning cycles; when the arbitration state switches from near-field to far-field, the conflict level between the near-field and the updated far-field must be higher than the preset trigger threshold of 0.7, and the trigger threshold must be greater than the switching threshold; the hierarchical fusion decision module A80 outputs the fused trajectory as the final planned trajectory to the vehicle control execution unit, which generates longitudinal control commands and lateral control commands based on the final planned trajectory to control the vehicle to travel along the final planned trajectory.

[0042] Furthermore, when extracting passable spatiotemporal intervals, prediction uncertainty quantification and prediction error boundary propagation are also performed. The spatiotemporal interval extraction module A70 obtains the uncertainty measure of each predicted trajectory in the predicted trajectory set. The uncertainty measure includes the probability distribution, confidence score, and cognitive uncertainty and accidental uncertainty quantified through deep evidence learning for the corresponding predicted trajectory. In the yielding order consistency determination, the inference post-processing module A50 calculates the probability distribution of the arrival time of each interactive object to the conflict area. A conflict is determined to exist if and only if the confidence intervals of the arrival time probability distributions of two interactive objects do not overlap under a preset confidence level of 95%, and the predicted trajectories of the two interactive objects give opposite relative passing orders. Sequence contradiction; In the determination of occupancy conflict mutual exclusion, the inference post-processing module A50 calculates the occupancy probability of each interactive object in each spatiotemporal unit. If and only if the occupancy probability of at least two interactive objects simultaneously exceeds the preset probability threshold, it is determined that the occupancy conflict mutual exclusion is violated; The inference post-processing module A50 transmits the probability occupancy information as a soft constraint to the extraction of the passable spatiotemporal interval; In the near field area, based on the identified and determined behavior pattern, the cognitive uncertainty at the behavioral level of the traffic participant is eliminated, and only the random uncertainty brought about by the motion state measurement is retained; The updated far field area retains the complete quantification of cognitive uncertainty and random uncertainty; The near field area and the updated far field area adopt a mutually distinguishable uncertainty quantification strategy.

[0043] Furthermore, the spatiotemporal interval extraction module A70 obtains the prediction error boundary of each predicted trajectory in the scene consistency prediction result and performs error boundary shrinkage on the passable time interval of each sampling location. The time deviation corresponding to the prediction error boundary is subtracted from the passable time interval to obtain the shrunken passable time interval. The shrinkage magnitude of the updated far-field region error boundary is dynamically adjusted according to the scene consistency confidence score Q. The higher the Q, the smaller the shrinkage magnitude; the lower the Q, the larger the shrinkage magnitude. The shrinkage magnitude of the near-field region error boundary is determined independently according to the near-field safety level. The near-field safety level is classified according to the type of traffic participant and the conflict risk level. Different levels correspond to different error boundary shrinkage magnitudes, which are not dynamically adjusted with the scene consistency confidence score Q. The conflict risk level includes close-range crossing risk, sudden stop risk when following another vehicle, and lane change intrusion risk.

[0044] The scenario consistency confidence score Q is calculated as follows: Among them, N 检 The logarithm of the predicted trajectory that detects inconsistencies in order is dimensionless; N 总 V is the total logarithm of predicted trajectory combinations, dimensionless; 冲 V represents the number of spatiotemporal units where occupancy conflicts occur; it is dimensionless. 总C represents the total number of spatiotemporal units, dimensionless. 近 λ1, λ2, and λ3 are the confidence scores for near-field behavior pattern recognition, dimensionless and ranging from [0, 1]; λ1, λ2, and λ3 are preset weight coefficients, dimensionless and all greater than zero (in this embodiment, they are 2, 1.5, and 1, respectively); Q is the scene consistency confidence score, dimensionless; exp is the natural exponential function; the physical meaning of this formula is: when the proportion of sequence inconsistencies and occupancy conflicts increases, Q decays exponentially; when the confidence score for near-field behavior pattern recognition C... 近 As Q decreases, it also decays exponentially; the higher Q is, the more reliable the consistent prediction results for the scenario, and the smaller the safety margin when extracting the passable spatiotemporal interval; the lower Q is, the less reliable the prediction results, and the larger the safety margin; C 近 The near-field deterministic planning module A60 outputs the classification confidence score based on behavior pattern recognition. Behavior pattern recognition employs a Support Vector Machine (SVM) classifier. The input consists of the kinematic features of traffic participants, including a four-dimensional feature vector of velocity, acceleration, rate of change of heading angle, and lateral offset. The output is the classification probability of each behavior pattern, with the highest probability being taken as C. 近 .

[0045] In an alternative implementation, when extracting the passable spatiotemporal interval, the spatial dimension sampling of the candidate trajectory can adopt an adaptive spacing; for example, where the curvature of the candidate trajectory is large, the sampling spacing is reduced to 0.5 meters when the curvature is greater than 0.05 rad / m, and where the curvature is small, the sampling spacing is increased to 2 meters when the curvature is not greater than 0.05 rad / m, so as to reduce the amount of computation while ensuring the accuracy of the spatiotemporal interval where the curvature is large.

[0046] In another alternative implementation, when dynamically adapting the boundary adjustment, the density influence coefficient α can be set as a dynamic adjustment parameter. When the density of traffic participants is greater than the preset high density threshold (5 vehicles / 100m), it is set to 0.3, and when the density of traffic participants is not greater than the preset high density threshold, it is set to 0.15, so as to expand the near-field area in high-density scenarios.

[0047] In another alternative implementation, the interactive trajectory prediction model can employ a prediction model based on a Transformer model with a self-attention mechanism, a prediction model based on a graph neural network, or a prediction model based on a long short-term memory network, as long as it can output a set of multimodal predicted trajectories for the interactive object, conditioned on the candidate trajectories of the autonomous vehicle.

[0048] The specific values ​​described in the above embodiments are all exemplary values, and those skilled in the art can adjust them according to the actual application scenario and the specific parameters of the vehicle platform; the method of the present invention can be implemented by software program, stored in the readable storage medium of the vehicle computing platform, and executed by the processor.

Claims

1. An autonomous driving path planning method based on interactive trajectory prediction, characterized in that, The method includes: S1: Obtain the autonomous vehicle status, road topology, and historical trajectories of surrounding traffic participants to generate candidate trajectories for the autonomous vehicle. S2, Obtain the complete set of preset behavior patterns; the complete set of preset behavior patterns is a pre-constructed set of behavior patterns covering all behavior types of traffic participants; each behavior pattern is defined by a behavior type identifier and the corresponding range of kinematic parameters. S3. Based on the travel distance of the autonomous vehicle, the area to be driven in front of the vehicle is divided into a near-field area and a far-field area. The area to be driven is the area extending along the planned path of the autonomous vehicle, within a preset distance from the current position of the vehicle. The near-field area is a segment of the area where the distance between the vehicle's current position and the vehicle's travel direction is within a preset near-field distance range. The far-field area is the remaining area of ​​the area to be driven outside the near-field area. The boundary between the near-field area and the far-field area is a dynamic adaptive boundary. S4, for the far-field region; for each candidate trajectory of the autonomous vehicle, interactive objects are selected based on arrival time and spatial overlap to construct a trajectory interaction scenario; using the candidate trajectories of the autonomous vehicle in the trajectory interaction scenario as conditions, interactive trajectory prediction is performed on the interactive objects to obtain a set of predicted trajectories; subsequently, the inference post-processing layer, independent of the interactive trajectory prediction, removes the predicted trajectories with contradictory order from the set of predicted trajectories based on the quantitative judgment rules of yielding order consistency and occupancy conflict mutual exclusion to obtain a scene-consistent prediction result; the inference post-processing layer filters the set of predicted trajectories based on manually preset rules and preset thresholds; then, the scene-consistent prediction result is written back to the candidate trajectory of the autonomous vehicle, and the passable spatiotemporal interval of each candidate trajectory of the autonomous vehicle is extracted using the scene-consistent prediction result as input; S5, for the near-field region; the near-field region is the region that has previously completed interactive trajectory prediction and scene consistency filtering as the far-field region. Based on the obtained scene consistency prediction results, the entire set of preset behavior patterns is cropped to obtain a subset of behavior patterns corresponding to the near-field region; each behavior pattern in the subset of behavior patterns is enumerated, and a safe response driving mode corresponding to the autonomous vehicle is pre-generated for each behavior pattern. Each safe response driving mode includes a corresponding spatiotemporal constraint boundary; when the autonomous vehicle actually drives to the near-field region, the corresponding safe response driving mode is matched and executed according to the actual behavior pattern of the traffic participant and combined with the predicted trajectory of the traffic participant in the scene consistency prediction results. S6, as the autonomous vehicle moves forward, the far-field region is continuously updated; the spatiotemporal constraint boundaries of the safety response driving mode in the near-field region are used as constraint conditions to input the yielding order consistency determination, occupancy conflict mutual exclusion determination, and passable spatiotemporal interval extraction in the far-field region; the traffic participant behavior patterns identified in the near-field region are used as feedforward constraints to input the interactive trajectory prediction in the far-field region, so that the prediction modality space of the interactive object is clipped to the modality subset corresponding to the behavior pattern; the modality subset refers to the set of prediction modes corresponding to the behavior pattern in the multimodal trajectory output by the interactive trajectory prediction. S7, the safe response driving mode in the near field area and the trajectory determined by the far field area based on the traversable spatiotemporal interval are spatiotemporally connected and fused to output the final planned trajectory; the spatiotemporal connection and fusion adopts an asymmetric arbitration rule, and the deterministic planning result of the near field area has a higher priority than the interactive predictive planning result of the far field area; the deterministic planning result includes the safe response driving mode and its spatiotemporal constraint boundary, and the interactive predictive planning result includes the updated trajectory determined by the far field area based on the traversable spatiotemporal interval.

2. The method according to claim 1, characterized in that, When dividing the area in front of the vehicle into a near-field area and a far-field area, the preset distance and the preset near-field distance range are determined based on the vehicle's current speed. The dynamic adaptive adjustment method for the boundary between the near-field area and the far-field area is as follows: the position of the dynamic adaptive boundary is dynamically adjusted according to the current speed of the autonomous vehicle, the current road speed limit, and the density of surrounding traffic participants. The higher the vehicle speed, the higher the road speed limit, and / or the greater the density of traffic participants, the larger the range of the near-field area; the lower the vehicle speed and / or the smaller the density of traffic participants, the smaller the range of the near-field area is. The method for scene clipping of the entire set of preset behavior patterns is as follows: the entire set of preset behavior patterns is clipped based on the current road topology, traffic rules, and interaction object types to obtain a subset of behavior patterns corresponding to the near-field area. As the autonomous vehicle moves forward, the dynamic adaptive boundary continues to move forward, and the far-field region transforms into the near-field region after the autonomous vehicle moves forward. An updated far-field region is formed in front of the autonomous vehicle, and this cycle forms a closed-loop progression in the rolling time domain.

3. The method according to claim 2, characterized in that, The quantitative judgment rule for yielding order consistency includes: extracting conflict areas in the trajectory interaction scenario, where a conflict area refers to an area where the predicted trajectories of two or more interactive objects overlap spatially and the arrival time difference corresponding to the overlapping area is less than a preset time threshold; for each predicted trajectory in the predicted trajectory set, time sampling is performed along the predicted trajectory to obtain multiple trajectory sampling points and their corresponding sampling times; each trajectory sampling point is spatially matched with the conflict area, and the sampling time corresponding to the trajectory sampling point that first falls into the conflict area is determined as the arrival time of the interactive object corresponding to the predicted trajectory to the conflict area; the interactive objects are sorted in ascending order according to the arrival time to generate a predicted passage order sequence; the predicted passage order sequence is subjected to time reversal detection, where time reversal means that for the same conflict area, the relative passage order of the same pair of interactive objects in two different predicted trajectories is opposite; if time reversal exists, the pair of predicted trajectories is determined to violate yielding order consistency and is removed from the predicted trajectory set; Specifically, for interactive objects whose behavior patterns have been identified in the near-field area, a deterministic consistency criterion is used to perform time-sequence reversal detection based on the deterministic sorting results of the arrival times. For interactive objects whose behavior patterns have not been identified in the near-field area, a probabilistic consistency criterion is used to perform time-sequence reversal detection based on the confidence interval of the probability distribution of the arrival times. Simultaneously, when the actual behavior pattern of a traffic participant in the near-field area has been identified and the identification confidence of the actual behavior pattern is higher than a preset identification confidence threshold, the passage order corresponding to the actual behavior pattern is used as a locking condition. The predicted trajectory in the updated far-field area is forced to match the passage order of the traffic participant with the locking condition; inconsistent predicted trajectories are directly discarded. When the identification confidence is not higher than the preset identification confidence threshold, the locking is not performed, and the passage order is determined automatically by the interactive trajectory prediction in the updated far-field area. The preset identification confidence threshold is dynamically adjusted according to the scene congestion level.

4. The method according to claim 1, characterized in that, The quantitative determination rule for occupancy conflict mutual exclusion includes: discretizing the drivable space in the trajectory interaction scenario in three dimensions along the road longitudinal direction, road transverse direction, and time dimension to obtain a set of spatiotemporal units, each spatiotemporal unit being uniquely identified by a longitudinal position interval, a transverse position interval, and a time interval; for each predicted trajectory in the predicted trajectory set, time sampling is performed along the predicted trajectory to determine the occupancy time window of the interactive object corresponding to the predicted trajectory in each spatiotemporal unit; for the spatiotemporal unit corresponding to the near-field region, a safety margin expansion amount determined by the safe response driving mode is superimposed on the occupancy time window; the safety margin expansion amount is based on the unmanned The current vehicle speed, the type of the interaction object, and the current road friction coefficient are coupled to determine the safety margin inflation. The higher the vehicle speed, the higher the risk level of the interaction object type, and / or the lower the road friction coefficient, the greater the safety margin inflation. A spatiotemporal occupancy matrix is ​​constructed, where rows correspond to spatiotemporal units and columns correspond to interaction objects. The matrix elements are the occupancy time windows of the interaction object in the spatiotemporal unit after safety margin inflation. It is detected whether there are at least two overlapping occupancy time windows of interaction objects in the same spatiotemporal unit. If there is an overlap, it is determined that the predicted trajectory violates the mutual exclusion of occupancy conflicts, and the predicted trajectory is removed from the predicted trajectory set.

5. The method according to claim 1, characterized in that, The step of writing back the scene consistency prediction results to the candidate trajectories of autonomous vehicles, and extracting the passable spatiotemporal intervals of each candidate trajectory using the scene consistency prediction results as input, includes: for each candidate trajectory, establishing a time alignment mapping relationship between the trajectory space of the candidate trajectory and the trajectory space of each predicted trajectory in the scene consistency prediction results; according to the time alignment mapping relationship, mapping the predicted trajectory corresponding to the conditions of the candidate trajectory in the scene consistency prediction results back to the trajectory space where the candidate trajectory is located, forming a passable state sequence of the candidate trajectory under the constraints of the scene consistency prediction results; sampling along the spatial dimension of the candidate trajectory at equal or adaptive intervals to obtain multiple sampling positions; for each sampling position, determining the time window occupied by any predicted trajectory in the scene consistency prediction results in the time dimension of the sampling position according to the passable state sequence, and determining the complement of the time window of the sampling position that is not occupied in the time dimension as the sampling position. The passable time intervals are determined by combining the passable time intervals of all sampling locations in spatial order to obtain the passable spatiotemporal interval of the candidate trajectory. Furthermore, the deterministic passable time interval determined by the near-field region based on the safe response driving mode and the probabilistic passable time interval extracted by the far-field region based on the scene consistency prediction result are spatially spliced ​​along the vehicle's travel direction to form a passable spatiotemporal interval covering the entire journey. The spatiotemporal constraint boundary contained in the safe response driving mode of the near-field region is injected as a hard constraint into the extraction process of the updated probabilistic passable time interval of the far-field region, so that the updated probabilistic passable time interval of the far-field region automatically meets the requirements of the spatiotemporal constraint boundary at the near-field boundary. At the splicing interface, the deterministic passable time interval of the near-field region is preferentially accepted, and the probabilistic passable time interval of the far-field region is only used as a backup constraint. When the two conflict, the deterministic passable time interval of the near-field region is used as the arbitration benchmark.

6. The method according to claim 1, characterized in that, It also includes closed-loop feedback iteration: when performing yielding order consistency and occupancy conflict mutual exclusion determination, a consistency metric is calculated; the consistency metric includes at least one of the number of conflict regions, the number of time sequence reversals, and the spatiotemporal unit occupancy conflict rate; the result of the consistency metric being lower than a preset threshold is fed back to the interactive object screening and interactive trajectory prediction steps, expanding the screening range of interactive objects or increasing the number of predicted trajectories of the interactive object in the interactive trajectory prediction, and re-executing the interactive trajectory prediction and consistency determination until the consistency metric meets the preset convergence condition or reaches the maximum number of iterations; When the near-field region detects a change in the behavior pattern of a traffic participant, the magnitude of the change in behavior pattern is calculated. The magnitude of the change is determined by the rate of change of the traffic participant's speed and the rate of change of the heading angle. The rate of change of the heading angle is the rate of change of the traffic participant's heading angle over time. If the magnitude of the change exceeds a preset behavior change threshold, the updated interactive trajectory prediction in the far-field region is re-executed, with the behavior pattern change serving as the highest priority iteration trigger condition. If the switching magnitude does not exceed the preset behavior change threshold, the re-execution of the updated far-field interactive trajectory prediction will not be triggered. The arbitration status configuration of the asymmetric arbitration rule is equipped with a hysteresis condition; When the arbitration status switches from far field to near field, the conflict level between the updated far field and near field must be lower than the preset switching threshold for multiple consecutive planning cycles. When the arbitration status switches from near-field to far-field, the conflict between the near-field and the updated far-field must be greater than a preset trigger threshold, and the trigger threshold must be greater than the switching threshold.

7. The method according to claim 1, characterized in that, It also includes prediction uncertainty quantification: obtaining the uncertainty measure of each prediction trajectory in the prediction trajectory set, the uncertainty measure including the probability distribution of the corresponding prediction trajectory, confidence score, and cognitive uncertainty and accidental uncertainty quantified by deep evidence learning; in the yielding order consistency determination, calculating the probability distribution of the arrival time of each interactive object to the conflict area, and determining that there is an order contradiction if and only if the confidence intervals of the arrival time probability distributions of two interactive objects do not overlap at a preset confidence level, and the prediction trajectories of the two interactive objects give opposite relative passage orders; in the occupancy conflict mutual exclusion determination, calculating the occupancy probability of each interactive object in each spatiotemporal unit, and determining that the occupancy conflict mutual exclusion is violated if and only if the occupancy probability of at least two interactive objects simultaneously exceeds a preset probability threshold. The probability occupancy information is passed as a soft constraint to the extraction of the passable spatiotemporal interval; wherein, within the near field area, based on the identified and determined behavior pattern, the cognitive uncertainty at the behavioral level of the traffic participant is eliminated, and only the accidental uncertainty brought about by the motion state measurement is retained; The updated far-field region retains complete quantification of cognitive uncertainty and accidental uncertainty; the near-field region and the updated far-field region employ a mutually distinguishable uncertainty quantification strategy.

8. The method according to claim 1, characterized in that, It also includes the calculation and adjustment of scene consistency confidence: based on the proportion of predicted trajectory pairs with detected sequential contradictions to the total number of predicted trajectory pairs, the proportion of spatiotemporal units with occupancy conflicts to the total number of spatiotemporal units, and the confidence of the near-field region behavior pattern recognition, the scene consistency confidence score Q is calculated, and the formula for calculating Q is: Where, N 检 The logarithm of the predicted trajectory that detects inconsistencies in order is dimensionless; N 总 V is the total logarithm of predicted trajectory combinations, dimensionless; 冲 V represents the number of spatiotemporal units where occupancy conflicts occur; it is dimensionless. 总 C represents the total number of spatiotemporal units, dimensionless. 近 λ1, λ2, and λ3 are the confidence scores for near-field region behavior pattern recognition, which are dimensionless and range from [0,1]; λ1, λ2, and λ3 are preset weight coefficients, which are dimensionless and all greater than zero; Q is the scene consistency confidence score, which is dimensionless; exp is the natural exponential function; and the confidence score C for near-field region behavior pattern recognition is... 近 Directly attenuate the global scenario consistency confidence Q, C 近 The lower the value of Q, the lower the safety margin, and the greater the safety margin across all scenarios. The safety margin is dynamically adjusted based on Q when extracting passable time and space intervals. The higher the value of Q, the smaller the safety margin; the lower the value of Q, the larger the safety margin.

9. The method according to claim 8, characterized in that, It also includes prediction error boundary propagation: obtaining the prediction error boundary of each prediction trajectory in the scene consistency prediction results; when extracting the passable spatiotemporal interval, the error boundary of the passable time interval at each sampling location is shrunk, and the time deviation corresponding to the prediction error boundary is deducted from the passable time interval to obtain the shrunk passable time interval; wherein, the error boundary shrinkage amplitude of the updated far-field region is dynamically adjusted according to the scene consistency confidence score Q, the higher Q, the smaller the shrinkage amplitude, and the lower Q, the larger the shrinkage amplitude; the error boundary shrinkage amplitude of the near-field region is independently determined according to the near-field safety level, the near-field safety level is graded according to the type of traffic participants and the conflict risk level, and different levels correspond to different error boundary shrinkage amplitudes, which are not dynamically adjusted with the scene consistency confidence score Q; the conflict risk level includes close-range crossing risk, sudden stop risk when following another vehicle, and lane change intrusion risk.

10. An autonomous driving path planning system based on interactive trajectory prediction, characterized in that, The system is used to implement the autonomous driving path planning method based on interactive trajectory prediction as described in claim 1, and the system includes: Candidate trajectory generation module; used to acquire the autonomous vehicle's status, road topology, and historical trajectories of surrounding traffic participants, and generate candidate trajectories for the autonomous vehicle. The spatial layering and partitioning module is used to divide the area in front of the autonomous vehicle into a near-field area and a far-field area based on the distance traveled by the autonomous vehicle. The near-field area is a region within a preset near-field distance range along the vehicle's direction of travel and its current position. The far-field area is the remaining area of ​​the area to be driven outside the near-field area. The boundary between the near-field area and the far-field area is a dynamic adaptive boundary, which is dynamically adjusted according to the current speed of the autonomous vehicle, the current road speed limit, and the density of surrounding traffic participants. An interactive scene construction module is used to filter interactive objects for each candidate trajectory of an autonomous vehicle in the far-field region based on arrival time sequence and spatial overlap relationship, and construct trajectory interactive scenes. Interactive prediction module; used to perform interactive trajectory prediction on interactive objects based on candidate trajectories of autonomous vehicles in trajectory interaction scenarios, and obtain a set of predicted trajectories; The inference post-processing module, independent of the interactive prediction module, is used to filter the predicted trajectory set according to the manually preset rules and preset thresholds, and according to the quantitative judgment rules of yielding order consistency and occupancy conflict mutual exclusion, to remove the predicted trajectories with contradictory order, and obtain the scene consistent prediction result. The near-field deterministic planning module is used to combine the current road topology, traffic rules, and interaction object types to perform scene clipping on the entire set of preset behavior patterns to obtain a subset of behavior patterns; enumerate each behavior pattern in the subset of behavior patterns, and pre-generate a safe response driving mode corresponding to the autonomous vehicle for each behavior pattern, with each safe response driving mode containing a corresponding spatiotemporal constraint boundary; when the autonomous vehicle actually drives to the near-field area, it matches and executes the corresponding safe response driving mode according to the actual behavior pattern of the traffic participant and the predicted trajectory of the traffic participant in the scene consistent prediction result. The spatiotemporal interval extraction module is used to write back the scene consistent prediction results to the candidate trajectories of the autonomous vehicle, and extract the passable spatiotemporal intervals of each candidate trajectory of the autonomous vehicle using the scene consistent prediction results as input; the spatiotemporal constraint boundary output by the near-field deterministic planning module is used as the constraint condition input to the spatiotemporal interval extraction module. The hierarchical fusion decision module is used to spatiotemporally connect and fuse the safe driving mode in the near-field area with the trajectory determined based on the passable spatiotemporal interval in the far-field area, and output the final planned trajectory. The hierarchical fusion decision module is configured with arbitration priority rules, in which the deterministic planning result of the near-field area has higher priority than the updated interactive prediction planning result of the far-field area. In the event of spatiotemporal conflict, the deterministic planning result of the near-field area shall prevail. The spatiotemporal constraint boundary output by the near-field deterministic planning module is fed back to the interactive scene construction module and the inference post-processing module to realize cross-module constraint linkage. The identified traffic participant behavior patterns output by the near-field deterministic planning module are used as feedforward constraints input to the interactive prediction module, so that the prediction modal space of the interactive object is clipped to the modal subset corresponding to the behavior pattern.

Images