Intersection trajectory planning method for autonomous vehicle based on interaction intensity

Abstract

The application discloses an automatic driving vehicle intersection trajectory planning method based on interaction intensity, comprising the following steps: acquiring the state information of a host vehicle, surrounding vehicles and a map in an intersection scene; generating a multi-modal prediction trajectory based on the historical trajectory of the surrounding vehicles; calculating the aggressiveness weight of the surrounding vehicles according to the recent driving behavior; constructing a local influence field based on the multi-modal trajectory of the host vehicle and the surrounding vehicles; calculating the interaction intensity according to the influence field overlap degree, the relative motion direction and the aggressiveness weight; embedding the interaction intensity into the trajectory planning to generate a candidate trajectory set; and selecting the optimal trajectory as the planning result after comprehensive scoring. The application quantifies the dynamic coupling relationship between the host vehicle and the surrounding vehicles through the interaction intensity, solves the problem that the existing method is difficult to depict the intention, driving style and multi-vehicle interaction of the surrounding vehicles, and improves the safety, traffic efficiency and interaction adaptability of automatic driving in the intersection scene.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and specifically to a method for planning the intersection trajectory of autonomous vehicles based on interaction intensity. Background Technology

[0002] At intersections, the most conflict-prone areas in urban road traffic, vehicles need to continuously and dynamically interact with surrounding vehicles when performing actions such as going straight, turning left, turning right, yielding, and merging. Especially in human-machine hybrid driving environments, autonomous vehicles and human-driven vehicles coexist for a long time. The driving intentions, driving styles, and risk preferences of human-driven vehicles differ significantly, leading to a high degree of uncertainty in the trajectory planning of autonomous vehicles in intersection scenarios.

[0003] Currently, most autonomous driving trajectory planning methods rely on criteria such as minimum safe distance, collision time, lane constraints, or rule bases. While these methods can meet basic collision avoidance requirements, they primarily focus on the geometric relationships and kinematic states between vehicles, making it difficult to effectively characterize the potential intentions and driving style differences of surrounding vehicles. Figure 1 As shown, Figure 1 (a) is the traditional TTC method. Figure 1 (b) To consider intent and style modeling, when a left-turning vehicle is currently moving at a low speed and does not exhibit obvious turning behavior, traditional time-of-collision (TTC) based methods would assume that the collision time with oncoming straight-ahead vehicles tends to infinity, thus misjudging the conflict risk as low. However, the vehicle actually has a clear intention to turn left, and its expected trajectory poses a potential conflict with oncoming straight-ahead vehicles. If the vehicle's driving style is aggressive, it is more likely to accelerate and overtake, significantly increasing the actual conflict risk. Therefore, relying solely on underlying kinematic indicators is insufficient to fully reflect the interaction process and can easily lead to an underestimation of interaction risk. Summary of the Invention

[0004] The purpose of this invention is to provide an intersection trajectory planning method for autonomous vehicles based on interaction intensity, which solves the problem that existing autonomous driving trajectory planning methods are unable to fully characterize the intentions, driving styles and dynamic interaction relationships of surrounding vehicles.

[0005] This invention constructs an interaction intensity index, which integrates the multimodal predicted trajectories, driving aggression, and local influence fields of surrounding vehicles to quantify the dynamic coupling relationship between the autonomous driving master vehicle and surrounding vehicles. This interaction intensity is then embedded into the trajectory planning process, thereby generating safer, more efficient, smoother, and more interactively adaptable autonomous vehicle trajectories.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for planning the intersection trajectory of autonomous vehicles based on interaction intensity, comprising the following steps:

[0007] S1. Obtain the status information of the autonomous driving master vehicle, surrounding vehicles, and map environment in the intersection scenario;

[0008] S2. Based on the historical trajectory information of surrounding vehicles and map environment information, generate multimodal predicted trajectories of surrounding vehicles in the future planning time domain;

[0009] S3. Calculate the driving aggression weights of surrounding vehicles based on their recent driving behavior characteristics.

[0010] S4. Construct a local influence field for vehicles based on the trajectory of the autonomous driving master vehicle and the multimodal predicted trajectories of surrounding vehicles;

[0011] S5. Calculate the interaction intensity based on the degree of overlap of the local influence fields between the autonomous driving master vehicle and surrounding vehicles, the relative motion direction, and the driving aggression weight.

[0012] S6. Embed the interaction intensity into the autonomous vehicle trajectory planning process to generate a candidate trajectory set;

[0013] S7. The candidate trajectory set is comprehensively scored, and the candidate trajectory with the highest score is selected as the planned trajectory of the autonomous vehicle in the intersection scenario.

[0014] Furthermore, step S1 obtains the state information of the autonomous driving master vehicle, surrounding vehicles, and map environment in the intersection scenario, including constructing the current environmental state. :

[0015] ;

[0016] in, Indicates the status of the autonomous driving master vehicle. Indicates the first The status of surrounding vehicles. Represents high-precision map information;

[0017] The vehicle status for:

[0018] ;

[0019] in, Indicates the vehicle's location. Indicates speed, Indicates acceleration. Indicates the heading angle;

[0020] The high-precision map information Structured modeling is performed using lane diagrams.

[0021] Furthermore, in step S2, the multimodal trajectory is predicted. for:

[0022] ;

[0023] in, Indicates the first The first of the surrounding vehicles Predicted trajectory, This indicates the probability corresponding to the predicted trajectory;

[0024] ;

[0025] .

[0026] The aforementioned multimodal predicted trajectory is used to characterize the potential intentions of surrounding vehicles, such as going straight, turning left, turning right, yielding, or cutting in.

[0027] Furthermore, step S3 calculates the driving aggression weights of surrounding vehicles by obtaining the driving behavior sequences of surrounding vehicles over a recent period and extracting speed, acceleration, jerk, and their fluctuation features; these features are then input into a pre-trained driving style recognition model to obtain the first... Driving style embedding vector of surrounding vehicles Based on the aforementioned first Driving style embedding vector of surrounding vehicles Calculate the first Weight of driving aggression of surrounding vehicles :

[0028] ;

[0029] in, The cluster centers representing aggressive driving styles; Indicates the temperature coefficient; This represents the cluster center of driving style obtained from offline training;

[0030] when A larger value indicates that surrounding vehicles are more likely to exhibit aggressive cutting-edge, rapid acceleration, or strong interactive behaviors; when A smaller value indicates that the behavior of surrounding vehicles is relatively stable or conservative.

[0031] Furthermore, step S4, constructing the vehicle's local influence field, includes building a local coordinate system with reference to the vehicle's current position and its predicted trajectory direction. This local coordinate system is the Frenet coordinate system, and for any spatial point... In the local coordinate system, it is represented as:

[0032] ;

[0033] in, Indicates the longitudinal distance along the centerline of the lane. Indicates lateral offset relative to the lane centerline;

[0034] Define the vehicle in the Frenet coordinate system. exist The influence field strength on spatial point P at time t is:

[0035] ;

[0036] in, The current vehicle speed, For the width of the driving lane, For the time window with the greatest impact;

[0037] This is a speed amplification term, used to control the amplification of the influence of speed. This is the speed amplification factor;

[0038] The distance attenuation term represents the natural decay of the vehicle's influence on a spatial point with distance. It is defined as a two-dimensional Gaussian function, and the specific formula is as follows:

[0039] ;

[0040] in, and These are the Gaussian decay scales in the longitudinal and transverse directions, respectively;

[0041] This is a directional deviation penalty term, used to suppress misjudgments caused by the inconsistency between the vehicle's current heading and the orientation of a spatial point. The specific formula is as follows:

[0042] ;

[0043]

[0044] in, It is the direction angle from the vehicle's current position to spatial point P. For the vehicle's yaw angle, This is the direction deviation adjustment coefficient, which controls the sensitivity to changes in direction. It is set here. rad represents the distance a point in space deviates from its current heading by more than approximately [a certain value]. At a certain rad, its influence will significantly decrease. This penalty term constrains the dominant direction of the vehicle's influence, causing it to be mainly distributed along the vehicle's intended direction, suppressing unnecessary interference in the rear region.

[0045] Furthermore, step S5 calculates the interaction strength, including setting the autonomous driving master vehicle at time... The local influence field is , No. The surrounding vehicles were in the first The local influence field under the predicted trajectory is The spatial overlap between the two is:

[0046] ;

[0047] in, This represents the effective threshold of the influence field.

[0048] Furthermore, step S5, calculating the interaction strength, also includes calculating the direction adjustment factor between the host vehicle and surrounding vehicles:

[0049] ;

[0050] in, Indicates the direction of the vehicle's speed. Indicates the surrounding vehicles in the first Velocity direction under the predicted trajectory;

[0051] Next, the autonomous driving master vehicle and the first The surrounding vehicles were in the first The instantaneous interaction intensity under the predicted trajectory is:

[0052] ;

[0053] in, This indicates the weight of driving aggression.

[0054] Furthermore, considering the uncertainty of the future intentions of surrounding vehicles, step S5, calculating the interaction intensity, also includes probabilistically weighted fusion of the instantaneous interaction intensity under different predicted trajectory modes:

[0055] ;

[0056] Next, the summation over all surrounding vehicles is used to obtain the overall instantaneous interaction intensity:

[0057] ;

[0058] Then, to facilitate numerical comparisons across different scenarios, the overall instantaneous interaction intensity is normalized:

[0059] ;

[0060] in, Indicates the compression factor. Indicates the center parameter of the interaction strength;

[0061] Finally, within the future planning time domain, candidate trajectories The cumulative interaction strength is:

[0062] ;

[0063] in, Representing candidate trajectories At any moment The corresponding normalized instantaneous interaction intensity.

[0064] Furthermore, step S6, generating candidate trajectories, includes constructing the discrete action space of the autonomous vehicle, assuming the master vehicle's control actions are:

[0065] ;

[0066] in, Indicates longitudinal acceleration. Indicates the steering angle;

[0067] Discretize the longitudinal acceleration and steering angle separately to obtain the motion space:

[0068] ;

[0069] Next, a Monte Carlo tree search guided by interaction intensity is used to generate candidate trajectories for the states in the search tree. The following action The selection criteria are as follows:

[0070] ;

[0071] in, Indicates the value of a state action. This represents the prior probability of an action. Indicates the number of times the state has been accessed. Indicates the number of times the state action is accessed. Indicates the exploration coefficient. This represents the interaction intensity penalty coefficient. Indicates the execution of an action The intensity of interaction in the subsequent corresponding state;

[0072] The prior probability of the action can be obtained by fusing manual priors and learned priors:

[0073] ;

[0074] in, This represents the learning prior obtained from the multimodal prediction results. This represents the artificial prior knowledge obtained based on smooth driving rules. This represents the dynamic fusion coefficient.

[0075] Furthermore, step S7 performs a comprehensive score on the candidate trajectory set and selects the candidate trajectory with the highest score as the planned trajectory of the autonomous vehicle in the intersection scenario, including constructing an instant reward function for the candidate trajectory:

[0076] ;

[0077] in, Indicates a security reward. This indicates a reward for improving traffic flow. Indicates a comfort reward. These represent the corresponding weights;

[0078] The security reward for:

[0079] ;

[0080] in, Indicates the collision penalty item. This indicates a penalty for driving out of the permitted driving area. Indicates the normalized instantaneous interaction strength;

[0081] Traffic efficiency reward for:

[0082] ;

[0083] in, This indicates the distance the main vehicle travels along the target path. Indicates the road speed limit;

[0084] The comfort reward for:

[0085] ;

[0086] in, This represents the change in acceleration. This indicates the change in steering angle. Indicates lateral offset. This indicates a deviation in heading.

[0087] Next, a value assessment network is used to estimate the value of the latter part of the candidate trajectory. The overall score is:

[0088] ;

[0089] in, This represents the value of the latter part of the trajectory output by the value assessment network. This represents the cumulative interaction intensity penalty coefficient;

[0090] Finally, the candidate trajectory with the highest score is selected as the planned trajectory for the autonomous vehicle:

[0091] ;

[0092] in, Represents the set of candidate trajectories. This represents the optimal trajectory.

[0093] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0094] 1. Significantly improves driving safety in complex intersection scenarios: This application constructs a local influence field of the vehicle and calculates the spatial overlap between the influence fields of the main vehicle and surrounding vehicles, which can accurately quantify potential collision risks. When the interaction intensity increases, the planning system will proactively adopt a more conservative avoidance strategy, effectively avoiding collisions in high-risk scenarios such as unprotected left turns and mixed traffic.

[0095] Furthermore, by incorporating the multimodal predicted trajectories of surrounding vehicles, it considers various future motion possibilities rather than relying solely on a single trajectory assumption. This enhances the vehicle's ability to respond to sudden changes in intent (such as other vehicles suddenly cutting in or overtaking), reducing safety hazards caused by uncertainty.

[0096] 2. Effectively Improve Intersection Traffic Efficiency: The interaction intensity index in this application not only reflects risk but also incorporates the driving style (aggression weight) and movement trend (relative direction of movement) of surrounding vehicles. The driver can identify behavioral characteristics such as aggressive vehicles potentially passing quickly and conservative vehicles potentially yielding, thereby rationally determining the timing of merging and weaving while ensuring safety, reducing unnecessary waiting and deceleration. Simultaneously, the combination of staged trajectory planning and Monte Carlo tree search can efficiently explore multiple candidate trajectories and guide the search direction through interaction intensity, avoiding a decrease in traffic efficiency due to excessive caution, thus achieving a balance between safety and efficiency.

[0097] 3. Enhanced driving smoothness and ride comfort: In traditional methods, the vehicle often takes abrupt and sudden deceleration or steering actions after sensing a risk, resulting in an uneven driving experience. This application embeds both cumulative and instantaneous interaction intensity into the planning process, enabling the vehicle to predict the gradual changes in interaction pressure, thereby generating smooth and gradual acceleration and curvature trajectories, reducing phenomena such as sudden braking and sharp turns.

[0098] The trajectory value assessment network comprehensively considers smoothness indicators and interaction intensity constraints when scoring, ensuring that the final planned trajectory meets both interaction safety requirements and human comfort requirements.

[0099] 4. Highly Adaptable Interaction and Human-like Driving Behavior: This application dynamically identifies the driving styles of surrounding vehicles through a weighted approach to driving aggression, and adaptively adjusts the planning strategy of the primary vehicle accordingly. For example, when encountering aggressive vehicles, the primary vehicle appropriately increases its yielding tendency; when encountering conservative vehicles, it can more confidently utilize gaps. This adaptability makes the behavior of autonomous vehicles more similar to that of experienced human drivers, making them easier for surrounding road users to understand and accept.

[0100] Meanwhile, by constructing an influence field using the Frenet coordinate system, the relative position and motion relationships under complex road geometry conditions such as curves and irregular intersections can be naturally described, ensuring the universal adaptability of the method under different intersection topologies.

[0101] 5. Solving the challenge of modeling dynamic coupling relationships among multiple vehicles: Existing methods often simplify the independent motion between vehicles or only consider paired interactions, making it difficult to characterize the mutual influence and coupling between multiple vehicles. This application constructs a holistic quantitative index of interaction intensity by integrating spatial overlap, relative motion direction, intention probability, and aggressiveness weight. This index reflects the comprehensive force state of the master vehicle in multi-party games, providing a unified mathematical framework for decision-making and planning under multi-vehicle interactions.

[0102] Meanwhile, this interaction strength can be embedded into various existing planning frameworks (such as sampling methods, optimization methods, and learning-based methods), and has good scalability and integration capabilities. Attached Figure Description

[0103] Figure 1 This is a comparison chart of the traditional TTC method and the modeling of intent and style in interaction risk assessment in the background technology.

[0104] Figure 2 The flowchart of the intersection trajectory planning method for autonomous vehicles based on interaction intensity of the present invention is as follows. Figure 1 ;

[0105] Figure 3 The flowchart of the intersection trajectory planning method for autonomous vehicles based on interaction intensity of the present invention is as follows. Figure 2 ;

[0106] Figure 4 This is a schematic diagram illustrating the interaction intensity calculation considering intent and driving style in this invention;

[0107] Figure 5 This is a comparison chart of performance indicators under different module combinations in this invention;

[0108] Figure 6 This is a dynamic programming simulation (non-interactive closed-loop simulation). Orange represents the expert trajectory, and blue represents the planned trajectory.

[0109] Figure 7 The planning results of different models in the same scene (the top is the starting position, the middle is the position at a certain moment in the process, and the bottom is the ending position; orange is the actual trajectory). Detailed Implementation

[0110] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0111] Reference Figure 2 As shown, a preferred embodiment of the present invention, an intersection trajectory planning method for autonomous vehicles based on interaction intensity, includes the following steps:

[0112] S1. Obtain the status information of the autonomous driving master vehicle, surrounding vehicles, and map environment in the intersection scenario;

[0113] It includes constructing the current environmental state. :

[0114] ;

[0115] in, Indicates the status of the autonomous driving master vehicle. Indicates the first The status of surrounding vehicles. Represents high-precision map information;

[0116] The vehicle status for:

[0117] ;

[0118] in, Indicates the vehicle's location. Indicates speed, Indicates acceleration. Indicates the heading angle;

[0119] The high-precision map information Structured modeling is performed using lane diagrams.

[0120] S2. Based on the historical trajectory information of surrounding vehicles and map environment information, generate multimodal predicted trajectories of surrounding vehicles in the future planning time domain;

[0121] The multimodal predicted trajectory for:

[0122] ;

[0123] in, Indicates the first The first of the surrounding vehicles Predicted trajectory, This indicates the probability corresponding to the predicted trajectory;

[0124] ;

[0125] .

[0126] The aforementioned multimodal predicted trajectory is used to characterize the potential intentions of surrounding vehicles, such as going straight, turning left, turning right, yielding, or cutting in.

[0127] S3. Calculate the driving aggression weights of surrounding vehicles based on their recent driving behavior characteristics.

[0128] This involves acquiring the driving behavior sequence of surrounding vehicles over a recent period and extracting speed, acceleration, jerk, and their fluctuation features; inputting these features into a pre-trained driving style recognition model to obtain the first... Driving style embedding vector of surrounding vehicles Based on the aforementioned first Driving style embedding vector of surrounding vehicles Calculate the first Weight of driving aggression of surrounding vehicles :

[0129] ;

[0130] in, The cluster centers representing aggressive driving styles; Indicates the temperature coefficient; This represents the cluster center of driving style obtained from offline training;

[0131] when A larger value indicates that surrounding vehicles are more likely to exhibit aggressive cutting-edge, rapid acceleration, or strong interactive behaviors; when A smaller value indicates that the behavior of surrounding vehicles is relatively stable or conservative.

[0132] S4. Construct a local influence field for vehicles based on the trajectory of the autonomous driving master vehicle and the multimodal predicted trajectories of surrounding vehicles;

[0133] It includes constructing a local coordinate system with reference to the vehicle's current position and its predicted trajectory direction. This local coordinate system is the Frenet coordinate system, and for any spatial point... In the local coordinate system, it is represented as:

[0134] ;

[0135] in, Indicates the longitudinal distance along the centerline of the lane. Indicates lateral offset relative to the lane centerline;

[0136] Define the vehicle in the Frenet coordinate system. exist The influence field strength on spatial point P at time t is:

[0137] ;

[0138] in, The current vehicle speed, For the width of the driving lane, For the time window with the greatest impact;

[0139] This is a speed amplification term, used to control the amplification of the influence of speed. This is the speed amplification factor;

[0140] The distance attenuation term represents the natural decay of the vehicle's influence on a spatial point with distance. It is defined as a two-dimensional Gaussian function, and the specific formula is as follows:

[0141] ;

[0142] in, and These are the Gaussian decay scales in the longitudinal and transverse directions, respectively;

[0143] This is a directional deviation penalty term, used to suppress misjudgments caused by the inconsistency between the vehicle's current heading and the orientation of a spatial point. The specific formula is as follows:

[0144] ;

[0145] ;

[0146] in, It is the direction angle from the vehicle's current position to spatial point P. For the vehicle's yaw angle, This is the direction deviation adjustment coefficient, which controls the sensitivity to changes in direction. It is set here. rad represents the distance a point in space deviates from its current heading by more than approximately [a certain value]. At a certain rad, its influence will significantly decrease. This penalty term constrains the dominant direction of the vehicle's influence, causing it to be mainly distributed along the vehicle's intended direction, suppressing unnecessary interference in the rear region.

[0147] S5. Calculate the interaction intensity based on the degree of overlap of the local influence fields between the autonomous driving master vehicle and surrounding vehicles, the relative motion direction, and the driving aggression weight.

[0148] This includes setting up an autonomous driving master vehicle at any time. The local influence field is , No. The surrounding vehicles were in the first The local influence field under the predicted trajectory is The spatial overlap between the two is:

[0149] ;

[0150] in, This represents the effective threshold of the influence field.

[0151] Calculate the direction adjustment factor between the main vehicle and surrounding vehicles:

[0152] ;

[0153] in, Indicates the direction of the vehicle's speed. Indicates the surrounding vehicles in the first Velocity direction under the predicted trajectory;

[0154] Calculate the autonomous driving master vehicle and the first The surrounding vehicles were in the first The instantaneous interaction intensity under the predicted trajectory is:

[0155] ;

[0156] in, This indicates the weight of driving aggression.

[0157] Considering the uncertainty of the future intentions of surrounding vehicles, the instantaneous interaction intensity under different predicted trajectory modes is probabilistically weighted and fused:

[0158] ;

[0159] Summing over all surrounding vehicles yields the overall instantaneous interaction intensity:

[0160] ;

[0161] To facilitate numerical comparisons across different scenarios, the overall instantaneous interaction intensity was normalized:

[0162] ;

[0163] in, Indicates the compression factor. Indicates the center parameter of the interaction strength;

[0164] Candidate trajectories within the future planning time domain The cumulative interaction strength is:

[0165] ;

[0166] in, Representing candidate trajectories At any moment The corresponding normalized instantaneous interaction intensity.

[0167] S6. Embed the interaction intensity into the autonomous vehicle trajectory planning process to generate a candidate trajectory set;

[0168] It includes constructing the discrete action space of autonomous vehicles, assuming the master vehicle's control actions are:

[0169] ;

[0170] in, Indicates longitudinal acceleration. Indicates the steering angle;

[0171] Discretize the longitudinal acceleration and steering angle separately to obtain the motion space:

[0172] ;

[0173] Next, a Monte Carlo tree search guided by interaction intensity is used to generate candidate trajectories for the states in the search tree. The following action The selection criteria are as follows:

[0174] ;

[0175] in, Indicates the value of a state action. This represents the prior probability of an action. Indicates the number of times the state has been accessed. Indicates the number of times the state action is accessed. Indicates the exploration coefficient. This represents the interaction intensity penalty coefficient. Indicates the execution of an action The intensity of interaction in the subsequent corresponding state;

[0176] The prior probability of the action can be obtained by fusing manual priors and learned priors:

[0177] ;

[0178] in, This represents the learning prior obtained from the multimodal prediction results. This represents the artificial prior knowledge obtained based on smooth driving rules. This represents the dynamic fusion coefficient.

[0179] S7. The candidate trajectory set is comprehensively scored, and the candidate trajectory with the highest score is selected as the planned trajectory of the autonomous vehicle in the intersection scenario.

[0180] This includes constructing an instant reward function for candidate trajectories:

[0181] ;

[0182] in, Indicates a security reward. This indicates a reward for improving traffic flow. Indicates a comfort reward. These represent the corresponding weights;

[0183] The security reward for:

[0184] ;

[0185] in, Indicates the collision penalty item. This indicates a penalty for driving out of the permitted driving area. Indicates the normalized instantaneous interaction strength;

[0186] Traffic efficiency reward for:

[0187] ;

[0188] in, This indicates the distance the main vehicle travels along the target path. Indicates the road speed limit;

[0189] The comfort reward for:

[0190] ;

[0191] in, This represents the change in acceleration. This indicates the change in steering angle. Indicates lateral offset. This indicates a deviation in heading.

[0192] Next, a value assessment network is used to estimate the value of the latter part of the candidate trajectory. The overall score is:

[0193] ;

[0194] in, This represents the value of the latter part of the trajectory output by the value assessment network. This represents the cumulative interaction intensity penalty coefficient;

[0195] Finally, the candidate trajectory with the highest score is selected as the planned trajectory for the autonomous vehicle:

[0196]

[0197] in, Represents the set of candidate trajectories. This represents the optimal trajectory.

[0198] The present invention will be further described below with reference to the embodiments.

[0199] In one specific embodiment, such as Figure 3-4 As shown, when an autonomous vehicle approaches an intersection, it first obtains the status of the main vehicle, the status of surrounding vehicles, and intersection map information through onboard sensors, vehicle-to-infrastructure (V2I) devices, or a map module. The system uses the position, speed, acceleration, and heading angle of surrounding vehicles over the past few seconds as historical input, and combines this with lane center lines, traffic signals, and drivable area information to generate multimodal predicted trajectories for surrounding vehicles over the next few seconds.

[0200] For each surrounding vehicle, the system further extracts its recent speed, acceleration, and jerk fluctuation features, and inputs them into the driving style recognition model to obtain a driving style embedding vector. Jerk is the rate of change of acceleration with respect to time, also known as jerkiness, and is generally described as being related to the comfort of vehicle driving. The driving aggression weight is calculated based on the distance between this embedding vector and the cluster center of aggressive driving styles. If a vehicle has a higher aggression level, it is assigned a higher weight in the interaction intensity calculation.

[0201] Subsequently, the system constructs local influence fields based on the predicted trajectories of the main vehicle and surrounding vehicles. The field strength of the local influence field decreases with increasing spatial distance and is distributed along the direction of the vehicle's intended driving direction. When there is a significant overlap between the influence fields of the main vehicle and those of the surrounding vehicles, and the two vehicles are moving in opposite directions, while the surrounding vehicles exhibit high driving aggression, the system calculates a higher interaction intensity.

[0202] During the trajectory planning phase, the system discretizes the longitudinal acceleration and steering angle, forming a finite action space. The Monte Carlo tree search module, when expanding candidate actions, comprehensively selects actions based on state-action value, prior action probability, and interaction intensity penalty. Actions that would cause the main vehicle to enter a high-interaction-intensity region have a lower search priority; actions that are safe, stable, and conducive to task progress have a higher search priority.

[0203] After generating a set of candidate trajectories, the system further invokes a value evaluation network to estimate the value of the later stages of the candidate trajectories. The final score of the candidate trajectory simultaneously considers the immediate reward, the value of the later stages of the trajectory, and the cumulative interaction intensity penalty. The system selects the trajectory with the highest score as the execution trajectory of the autonomous vehicle within the current planning cycle.

[0204] During vehicle operation, the system performs rolling replanning according to a preset cycle. When it detects that surrounding vehicles suddenly accelerate, change lanes, have significantly changed intent probabilities, or experience a rapid increase in interaction intensity, the system can trigger a new planning process in advance, thereby improving the dynamic response capability of autonomous vehicles in complex intersection scenarios.

[0205] The effectiveness of the above-mentioned intersection trajectory planning method for autonomous vehicles based on interaction intensity will be verified below.

[0206] (1) Experimental Environment and Verification Scenarios: In one embodiment, an autonomous driving closed-loop simulation platform was used to verify the method of the present invention. The simulation data selected typical urban intersection scenarios, including three types of scenarios: unprotected left turns, straight-through intersections, and right turns. 800 segments were selected for each type of scenario, totaling 2400 simulation sequences. The selected scenarios covered cross intersections, T-junctions, Y-junctions, and complex intersections, and included various traffic organization forms such as signalized and unsignalized traffic control.

[0207] During the simulation, the autonomous vehicle performs closed-loop replanning at fixed intervals. Within each planning cycle, the system first acquires the status of the master vehicle, the status of surrounding vehicles, and high-precision map information; then it generates the predicted future trajectories of surrounding vehicles and calculates the driving aggression of surrounding vehicles; next, it calculates the interaction intensity based on the overlapping relationship of the influence field; finally, the master vehicle trajectory is output by the staged planning module guided by the interaction intensity.

[0208] To verify the applicability of the method of the present invention under different interaction conditions, tests were conducted using both non-interactive closed-loop simulation and interactive closed-loop simulation.

[0209] (2) Evaluation indicators: The effectiveness of the present invention will be verified using the following evaluation indicators:

[0210] a. Collision-Free Rate: Measures the system's ability to avoid at-fault collisions through proactive decision-making. This metric only counts "at-fault collisions" that should have been caused by the lead vehicle's planning errors and are theoretically avoidable (including rear-end collisions, side collisions, and collisions caused by accidental lane changes or intersection entry). It is calculated as the percentage of scenarios where no at-fault collision occurred out of the total number of scenarios, then converted into a percentage score. A higher score indicates stronger safety robustness of the model in the simulation environment.

[0211] b. Safety Margin: Reflects the vehicle's ability to maintain reasonable timing intervals during dynamic interactions. It is calculated as the proportion of frames where the minimum TTC (Time To Call) is greater than the safety threshold (0.95 seconds). If the planning results contain too many short-term potential conflicts (i.e., too low a TTC), the score will decrease. A higher score indicates that the system is more proactive and has greater safety redundancy in complex interactions.

[0212] c. Route Compliance Rate: This assesses whether vehicles consistently travel within the passable area. If a vehicle's boundary deviates from the passable area on the map by more than 0.3 meters, the frame is considered a violation. This metric calculates the proportion of compliant trajectories across all frames, measuring the model's adherence to road geometry and traffic constraints.

[0213] d. Comfort: Evaluate the smoothness of the trajectory in terms of dynamic characteristics such as acceleration, yaw rate, and jerk. Compare the corresponding variables for each frame with default thresholds empirically derived from expert trajectory datasets (minimum longitudinal acceleration = -4.05 m / s²). 2 Maximum longitudinal acceleration = 2.40 m / s² 2 The absolute value of the maximum yaw acceleration is 1.93 rad / s². 2 The maximum absolute value of the yaw rate is 0.95 rad / s, and the maximum absolute value of the longitudinal jerk is 4.13 m / s². 3 Maximum jerk = 8.37 m / s² 3 This indicator determines whether a frame exceeds the limits. If any variable in a frame exceeds the threshold, it is considered an uncomfortable frame. This indicator calculates the percentage of frames that meet all comfort constraints; a higher score indicates better trajectory smoothness and a better riding experience.

[0214] e. Task Progress: This metric measures the degree to which the vehicle progresses along the reference path within the planning period. Specifically, it is calculated by dividing the projected distance of the trajectory endpoint along the target path by the total length of the path to obtain the progress percentage. A higher value indicates better task execution efficiency and effectively reflects the model's control over its ability to achieve the target.

[0215] All the above metrics are calculated independently at the scene level, statistically analyzing key performance indicators across the corresponding trajectories and averaging them across multiple scenes to obtain the final score. Each metric is uniformly mapped to a percentage score of 0-100, with higher values ​​indicating better model performance.

[0216] (3) Non-interactive closed-loop simulation verification: Under non-interactive closed-loop simulation conditions, the method of the present invention was compared with several existing trajectory planning methods. The results are shown in Table 1. The results show that the method of the present invention achieves better results in terms of collision-free rate, comfort, and task advancement. Among them, the collision-free rate of the method of the present invention reaches 96.31%, and the task advancement reaches 92.72%, which is significantly better than the Urban Driver, IDM, Gameformer, and MBAPPE methods.

[0217] Compared to Urban Driver, the collision-free rate of the method in this invention is improved from 63.02% to 96.31%, and the task progress is improved from 80.63% to 92.72%, indicating that the present invention, by introducing interaction intensity modeling, can effectively reduce potential collision risks and improve intersection traffic efficiency. Compared to PLUTO, the method in this invention shows better performance in terms of comfort and safety margin, indicating that the method can generate smoother trajectories with higher safety redundancy while maintaining high traffic efficiency.

[0218] Table 1 Comparison of non-interactive closed-loop simulation results

[0219]

[0220] Note: 1) The simulation covers three scenarios: left turn, straight, and right turn at intersections; 2) All indicators are standardized to a percentage score of 0-100, with higher values ​​indicating better performance; 3) Bold indicates the highest score among the corresponding indicators, and underline indicates the second highest score for that indicator.

[0221] (4) Interactive Closed-Loop Simulation Verification: To further verify the adaptability of the method of the present invention under the condition that surrounding vehicles can dynamically respond, an interactive closed-loop simulation was conducted. The results are shown in Table 2. In the interactive closed-loop simulation, the comprehensive score of the method of the present invention reached 92.64, which is the highest value among the comparison methods. Compared with the non-interactive closed-loop simulation, in the interactive closed-loop simulation, surrounding vehicles will dynamically respond according to the behavior of the master vehicle, thus better reflecting the interactive game characteristics in the real intersection scenario. The method of the present invention still achieved the highest comprehensive score under this condition, indicating that it can not only generate a safe trajectory, but also guide the planning process through the interaction intensity, reduce the disturbance of the master vehicle to surrounding vehicles, and improve the interactive coordination of the trajectory.

[0222] Table 2 Comparison of Interactive Closed-Loop Simulation Results

[0223]

[0224] Note: 1) The simulation covers three scenarios: left turn, straight, and right turn at intersections; 2) All indicators are standardized to a percentage score of 0-100, with higher values ​​indicating better performance; 3) Bold indicates the highest score among the corresponding indicators, and underline indicates the second highest score for that indicator.

[0225] (5) Contribution effect of each module in the invention: Under the same intersection closed-loop simulation conditions, the ablation verification of each module of the invention was performed. The results are as follows: Figure 5 As shown, when using only the traditional MCTS without constraints, the collision-free rate is 57.14% and the safety margin is 46.87%, indicating that the traditional search strategy is difficult to effectively handle the risks of multi-vehicle interactions at complex intersections. Introducing dynamic priors improves the safety margin and comfort by 25.6% and 8.5%, respectively, demonstrating that prior guidance can effectively compress the search space and improve the stability of trajectory generation. Further adding interaction intensity adjustment increases the safety margin to 91.36%, indicating that interaction intensity can effectively guide vehicles to avoid high-risk interaction areas. Adding a driving aggression module improves the collision-free rate by 5.8%, indicating that this module can enhance the driver's response to aggressive vehicles. Continuing to add trajectory continuity constraints improves the collision-free rate and task progress by 6.1% and 7.9%, respectively, indicating that this constraint can improve trajectory executability and traffic efficiency.

[0226] In summary, dynamic priors, interaction intensity adjustment, driving aggression, and trajectory continuity constraints can all effectively improve the safety, stability, and traffic efficiency of the method of the present invention in intersection trajectory planning.

[0227] (6) Case Visualization: Three cases were selected for closed-loop trajectory planning visualization. Each scenario was displayed with a trajectory snapshot at a 3-second interval (see Figure 6 ). Figure 6 (a) In the left-turn scenario, before entering the intersection, the planned trajectory exhibits a reasonable pre-offset in the turning direction, adjusting the driving direction in advance to match the expected trajectory curvature. During the crossing of the intersection, a potential conflict arises between the vehicle and an oncoming right-turning vehicle. At this point, the IISP (Interaction-Intensity-guided Stage-wise Planner) model of this invention automatically adjusts the left-turn path to a larger radius and selects a yielding strategy to ensure safety. Upon exiting the intersection, the vehicle avoids the same exit lane as the oncoming vehicle, choosing an empty lane to reduce following pressure and improve overall traffic efficiency. Although this behavior differs slightly from the expert trajectory, it is more practical from a traffic efficiency perspective. Figure 6(b) For right turns, maintain a smooth curvature at the beginning of the turn, consistent with the expert trajectory, and then switch to a lane with fewer vehicles to reduce the risk of interacting with other vehicles. Figure 6 (c) Demonstrating the straight-ahead process, the model consistently maintains its trajectory along the lane centerline. When a slow-moving vehicle is detected ahead, it automatically decelerates to maintain a safe following distance, demonstrating excellent longitudinal control and following strategy. Overall, IISP exhibits trajectory generation capabilities close to human driving logic in multiple typical intersection scenarios, effectively predicting, proactively avoiding, and balancing traffic efficiency and safety.

[0228] Figure 7 The complete trajectories of different models in non-interactive closed-loop simulation cases were compared. Figure 7 (a) is UrbanDriver (green track). Figure 7 (b) is MBAPPE (red track). Figure 7 (c) shows the planning model of this invention (blue trajectory). As can be seen from the figure, at the start of the simulation, UrbanDriver can stably control the vehicle to travel within the lane. However, as the simulation progresses, the trajectory gradually deviates from the drivable area, which may be caused by accumulated errors or local errors in the model in complex environments. In contrast, the master vehicle planned by MBAPPE and IISP can steadily travel on the center line of the road when entering the intersection and successfully complete the left turn. However, when interacting with surrounding vehicles, the trajectory planned by MBAPPE is closer to the surrounding vehicles, especially in the middle of the intersection, where there is a certain risk of conflict between the master vehicle and right-turning vehicles. Furthermore, MBAPPE has low task progress efficiency and ultimately fails to reach the endpoint of the expert trajectory, resulting in insufficient task progress efficiency. In contrast, the trajectory planned by IISP enters the intersection and moves into a lane closer to the inside, successfully avoiding the risk of conflict with oncoming right-turning vehicles. This advantage is due to IISP's effective modeling of interaction risks during the planning process. IISP not only considers the relative motion between the master vehicle and the vehicle in front but also comprehensively evaluates potential traffic risk factors. Furthermore, the trajectory planned by IISP performs exceptionally well in terms of mission progress efficiency, with the endpoint even surpassing that of expert trajectories.

[0229] Without causing conflict, those skilled in the art can freely combine and use the above-mentioned additional technical features.

[0230] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for planning the intersection trajectory of autonomous vehicles based on interaction intensity, characterized in that, Includes the following steps: S1. Obtain the status information of the autonomous driving master vehicle, surrounding vehicles, and map environment in the intersection scenario; S2. Based on the historical trajectory information of surrounding vehicles and map environment information, generate multimodal predicted trajectories of surrounding vehicles in the future planning time domain; S3. Calculate the driving aggression weight of surrounding vehicles based on their recent driving behavior characteristics. S4. Construct a local influence field for vehicles based on the trajectory of the autonomous driving master vehicle and the multimodal predicted trajectories of surrounding vehicles; S5. Calculate the interaction intensity based on the degree of overlap of the local influence fields between the autonomous driving master vehicle and surrounding vehicles, the relative motion direction, and the driving aggression weight. S6. Embed the interaction intensity into the autonomous vehicle trajectory planning process to generate a candidate trajectory set; S7. The candidate trajectory set is comprehensively scored, and the candidate trajectory with the highest score is selected as the planned trajectory of the autonomous vehicle in the intersection scenario.

2. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 1, characterized in that: Step S1, which involves obtaining the state information of the autonomous driving vehicle, surrounding vehicles, and map environment in the intersection scenario, includes constructing the current environmental state. : ； in, Indicates the status of the autonomous driving master vehicle. Indicates the first The status of surrounding vehicles. Represents high-precision map information; The vehicle status for: ； in, Indicates the vehicle's location. Indicates speed, Indicates acceleration. Indicates the heading angle; The high-precision map information Structured modeling is performed using lane diagrams.

3. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 1, characterized in that: In step S2, the multimodal trajectory is predicted. for: ； in, Indicates the first The first of the surrounding vehicles Predicted trajectory, This indicates the probability corresponding to the predicted trajectory; ； 。 4. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 1, characterized in that: Step S3, calculating the driving aggression weights of surrounding vehicles, includes obtaining the driving behavior sequence of surrounding vehicles over a recent period and extracting speed, acceleration, jerk, and their fluctuation features; inputting these features into a pre-trained driving style recognition model to obtain the first... Driving style embedding vector of surrounding vehicles Based on the aforementioned first Driving style embedding vector of surrounding vehicles Calculate the first Weight of driving aggression of surrounding vehicles : ； in, The cluster centers representing aggressive driving styles; Indicates the temperature coefficient; This represents the cluster center of driving style obtained from offline training; when A larger value indicates that surrounding vehicles are more likely to exhibit aggressive cutting-edge, rapid acceleration, or strong interactive behaviors; when A smaller value indicates that the behavior of surrounding vehicles is relatively stable or conservative.

5. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 1, characterized in that: Step S4, constructing the local influence field of the vehicle, includes building a local coordinate system with reference to the vehicle's current position and its predicted trajectory direction. This local coordinate system is the Frenet coordinate system, and for any spatial point... In the local coordinate system, it is represented as: ； in, Indicates the longitudinal distance along the centerline of the lane. Indicates lateral offset relative to the lane centerline; Define the vehicle in the Frenet coordinate system. exist The influence field strength on spatial point P at time t is: ； in, The current vehicle speed, For the width of the driving lane, For the time window with the greatest impact; This is a speed amplification term, used to control the amplification of the influence of speed. This is the speed amplification factor; The distance attenuation term represents the natural decay of the vehicle's influence on a spatial point with distance. It is defined as a two-dimensional Gaussian function, and the specific formula is as follows: ； in, and These are the Gaussian decay scales in the longitudinal and transverse directions, respectively; This is a directional deviation penalty term, used to suppress misjudgments caused by the inconsistency between the vehicle's current heading and the orientation of a spatial point. The specific formula is as follows: ； ； in, It is the direction angle from the vehicle's current position to spatial point P. For the vehicle's yaw angle, This is the directional deviation adjustment coefficient, which controls the sensitivity to influence the direction.

6. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 1, characterized in that: Step S5, calculating the interaction strength, includes setting the autonomous driving master vehicle at time... The local influence field is , No. The surrounding vehicles were in the first The local influence field under the predicted trajectory is The spatial overlap between the two is: ； in, This represents the effective threshold of the influence field.

7. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 6, characterized in that: Step S5, which calculates the interaction strength, also includes calculating the direction adjustment factor between the host vehicle and surrounding vehicles: ； in, Indicates the direction of the vehicle's speed. Indicates the surrounding vehicles in the first Velocity direction under the predicted trajectory; Next, the autonomous driving master vehicle and the first The surrounding vehicles were in the first The instantaneous interaction intensity under the predicted trajectory is: ； in, This indicates the weight of driving aggression.

8. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 7, characterized in that: Step S5, calculating the interaction intensity, further includes performing a probability-weighted fusion of the instantaneous interaction intensity under different predicted trajectory modes: ； Next, the summation over all surrounding vehicles is used to obtain the overall instantaneous interaction intensity: ； Then, the overall instantaneous interaction intensity is normalized: ； in, Indicates the compression factor. Indicates the center parameter of the interaction strength; Finally, within the future planning time domain, candidate trajectories The cumulative interaction strength is: ； in, Representing candidate trajectories At any moment The corresponding normalized instantaneous interaction intensity.

9. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 1, characterized in that: Step S6, generating candidate trajectories, involves constructing the discrete action space of the autonomous vehicle, assuming the master vehicle's control actions are: ； in, Indicates longitudinal acceleration. Indicates the steering angle; Discretize the longitudinal acceleration and steering angle separately to obtain the motion space: ； Next, a Monte Carlo tree search guided by interaction intensity is used to generate candidate trajectories for the states in the search tree. The following action The selection criteria are as follows: ； in, Indicates the value of a state action. This represents the prior probability of an action. Indicates the number of times the state has been accessed. Indicates the number of times the state action is accessed. Indicates the exploration coefficient. This represents the interaction intensity penalty coefficient. Indicates the execution of an action The intensity of interaction in the subsequent corresponding state; The prior probability of the action can be obtained by fusing manual priors and learned priors: ； in, This represents the learning prior obtained from the multimodal prediction results. This represents the artificial prior knowledge obtained based on smooth driving rules. This represents the dynamic fusion coefficient.

10. The method for planning the intersection trajectory of autonomous vehicles based on interaction intensity according to claim 1, characterized in that: Step S7 involves comprehensively scoring the candidate trajectory set and selecting the candidate trajectory with the highest score as the planned trajectory for the autonomous vehicle in the intersection scenario. This includes constructing an instantaneous reward function for each candidate trajectory. ； in, Indicates a security reward. This indicates a reward for improving traffic flow. Indicates a comfort reward. These represent the corresponding weights; The security reward for: ； in, Indicates the collision penalty item. This indicates a penalty for driving out of the permitted driving area. Indicates the normalized instantaneous interaction strength; Traffic efficiency reward for: ； in, This indicates the distance the main vehicle travels along the target path. Indicates the road speed limit; The comfort reward for: ； in, This represents the change in acceleration. This indicates the change in steering angle. Indicates lateral offset. Indicates heading deviation; Next, a value assessment network is used to estimate the value of the latter part of the candidate trajectory. The overall score is: ； in, This represents the value of the latter part of the trajectory output by the value assessment network. This represents the cumulative interaction intensity penalty coefficient; Finally, the candidate trajectory with the highest score is selected as the planned trajectory for the autonomous vehicle: ； in, Represents the set of candidate trajectories. This represents the optimal trajectory.

Images