Vehicle longitudinal interaction intention estimation method, electronic device, and storage medium

Abstract

The application discloses a vehicle longitudinal interaction intention estimation method, an electronic device and a storage medium, and belongs to the technical field of intelligent driving. The method comprises the following steps: acquiring original observation data of surrounding traffic participants of a vehicle, and fusing state information of each traffic participant; screening a target traffic participant having a longitudinal interaction relationship with the vehicle; constructing a target intention prior distribution matched with a current traffic scene based on historical longitudinal behavior data of the target traffic participant; determining an observation likelihood for each type of longitudinal interaction intention based on the state information of the target traffic participant; determining a target posterior probability corresponding to each type of longitudinal interaction intention through a hierarchical Bayesian model based on the target intention prior distribution and the observation likelihood; and generating an intention confidence set based on the target posterior probability corresponding to each type of longitudinal interaction intention, so as to be used for vehicle behavior decision. The application improves the stability and reliability of the vehicle in estimating the intention of the surrounding traffic participants in the longitudinal interaction scene.

Description

Technical Field

[0001] This application belongs to the field of intelligent driving technology, and in particular relates to a method for estimating longitudinal interaction intent of a vehicle, an electronic device, and a storage medium. Background Technology

[0002] With the development of autonomous driving perception, prediction, and decision-making technologies, vehicles can acquire information about the surrounding traffic environment through onboard sensors, map data, and vehicle-to-everything (V2X) communication information, and generate corresponding driving decisions based on the movement status of surrounding traffic participants. In traffic scenarios such as lane merging, ramp merging, lane changing, cutting in, following, and overtaking, there is usually a longitudinal interaction problem between the vehicle and surrounding vehicles. The subsequent behavior of surrounding vehicles will directly affect the vehicle's acceleration / deceleration, yielding, following, or passage decisions.

[0003] In traditional methods, vehicles typically assess the motion trends of surrounding vehicles based on their relative distance, relative speed, collision time, time interval, acceleration changes, or short-term trajectory predictions, and then provide the assessment results to the vehicle's decision-making and planning module. Some solutions also employ behavioral classification models to predict the behavior types of surrounding vehicles, such as lane changing, lane keeping, lane insertion, or lane avoidance, to assist the vehicle in driving control in complex road environments.

[0004] However, in real-world traffic environments, vehicle interactions are often influenced by factors such as road structure, traffic flow conditions, differences in driving behavior, and instantaneous observation errors. Traditional methods, when dealing with highly interactive scenarios such as merging, lane changing, and cutting in, are prone to problems such as unstable intent judgment, difficulty in quantifying the reliability of results, and susceptibility of vehicle decisions to short-term observation fluctuations. These issues, in turn, affect the smoothness, safety, and predictability of vehicle behavior decisions. Summary of the Invention

[0005] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a method, electronic device, and storage medium for estimating the longitudinal interaction intention of a vehicle, in order to improve the stability and reliability of estimating the intentions of surrounding traffic participants in longitudinal interaction scenarios.

[0006] Firstly, this application provides a method for estimating longitudinal interaction intent in a vehicle, the method comprising: The original observation data of traffic participants around the vehicle are acquired, and the original observation data are fused to obtain the status information of each traffic participant. Target traffic participants with longitudinal interaction relationships with the vehicle are selected from the traffic participants around the vehicle. Based on the historical longitudinal behavior data of the target traffic participants, a prior distribution of target intent matching the current traffic scenario is constructed. Based on the state information of the target traffic participants, the observation likelihood is determined for various longitudinal interaction intentions; the longitudinal interaction intentions include at least the intention to cut in, the intention to give way, and the intention to keep. Based on the prior distribution of the target intent and the observation likelihood, the target posterior probability corresponding to various vertical interaction intents is determined by a hierarchical Bayesian model. Based on the target posterior probability corresponding to various vertical interaction intentions, an intention confidence set is generated for vehicle behavior decision-making.

[0007] Secondly, this application provides a vehicle longitudinal interaction intent estimation device, the device comprising: The data fusion module is used to acquire raw observation data of traffic participants around the vehicle and to fuse the raw observation data to obtain the status information of each traffic participant. The target filtering module is used to filter target traffic participants that have a longitudinal interaction relationship with the vehicle from traffic participants around the vehicle; The prior construction module is used to construct a prior distribution of target intent that matches the current traffic scenario based on the historical longitudinal behavior data of the target traffic participants. The likelihood determination module is used to determine the observation likelihood for various longitudinal interaction intentions based on the state information of the target traffic participant; the longitudinal interaction intentions include at least the intention to cut in, the intention to give way, and the intention to hold on. The posterior determination module is used to determine the target posterior probability corresponding to various vertical interaction intentions based on the target intention prior distribution and the observation likelihood, using a hierarchical Bayesian model. The confidence set generation module is used to generate intent confidence sets based on the target posterior probabilities corresponding to various vertical interaction intents for vehicle behavior decision-making.

[0008] Thirdly, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the vehicle longitudinal interaction intent estimation method as described in the first aspect above.

[0009] Fourthly, this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the vehicle longitudinal interaction intent estimation method as described in the first aspect above.

[0010] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the vehicle longitudinal interaction intent estimation method as described in the first aspect.

[0011] In a sixth aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the vehicle longitudinal interaction intent estimation method as described in the first aspect above.

[0012] The above-described one or more technical solutions in the embodiments of this application have at least one of the following technical effects: By acquiring and fusing raw observation data of traffic participants surrounding the vehicle, the state information of each participant can be obtained, providing a unified data foundation for subsequent target traffic participant selection and observation likelihood determination. By selecting target traffic participants with longitudinal interaction relationships with the vehicle from among those surrounding it, subsequent intent estimation can focus on objects that may influence the vehicle's longitudinal behavior decisions, reducing the influence of irrelevant traffic participants on the estimation process. Furthermore, by constructing a prior distribution of target intent that matches the current traffic scenario based on the historical longitudinal behavior data of target traffic participants, the prior distribution can simultaneously reflect the historical longitudinal behavior characteristics of target traffic participants and the current traffic scenario's influence on longitudinal interaction intent. The system has several advantages. First, by determining the observation likelihood of various longitudinal interaction intentions based on the state information of target traffic participants, the current state of these participants can be incorporated into the intention estimation process. Second, by determining the target posterior probability corresponding to various longitudinal interaction intentions based on the prior distribution and observation likelihood of the target intentions, and using a hierarchical Bayesian model, historical longitudinal behavior, current traffic scenario, and current state information can be integrated within the same probability update framework, improving the estimation stability of overtaking, yielding, and holding intentions. Third, by generating an intention confidence set based on the target posterior probability, a candidate intention set can be retained when multiple longitudinal interaction intentions are possible, providing a probabilistic basis for vehicle behavior decision-making. Therefore, the accuracy, stability, and reliability of vehicle intention estimation for target traffic participants in longitudinal interaction scenarios can be improved.

[0013] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0014] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is one of the flowcharts illustrating the vehicle longitudinal interaction intent estimation method provided in the embodiments of this application; Figure 2 This is a block diagram of the overall architecture of the vehicle interaction intent estimation system provided in the embodiments of this application; Figure 3 This is a flowchart illustrating the vehicle longitudinal interaction intent estimation method provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the hierarchical Bayesian filtering model provided in the embodiments of this application; Figure 5 This is a typical longitudinal interaction timing diagram of merging provided in the embodiments of this application; Figure 6 This is a schematic diagram of the active disambiguation logic provided in the embodiments of this application; Figure 7 This is a schematic diagram of multi-agent joint intent inference in an interaction scenario provided in an embodiment of this application; Figure 8 This is a schematic diagram of conformal prediction calibration and intention confidence set provided in an embodiment of this application; Figure 9 This is a schematic diagram of the vehicle longitudinal interaction intent estimation device provided in the embodiments of this application; Figure 10 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments that can be directly obtained based on the embodiments of this application are within the scope of protection of this application.

[0016] Interactions between vehicles typically revolve around longitudinal priority relationships, meaning different traffic participants need to determine the order of "who goes first, who goes last" near potential conflict zones. For example, one vehicle might accelerate to occupy a target space, thus initiating a right-of-way maneuver, while another vehicle might slow down or maintain a lower speed to create space, thus yielding. Alternatively, they could maintain their current relative formation without actively changing their priority relationship. Therefore, the longitudinal interaction intentions of the target traffic participants can include at least the intention to overtake, the intention to yield, and the intention to remain in their position.

[0017] In actual driving, the longitudinal interaction intentions of a target vehicle are typically influenced by factors such as relative distance, relative speed, changes in acceleration, turn signals, brake lights, road geometry, lane changes, and ramp connections. In different scenarios, the same kinematic behavior may correspond to different tendencies to cut in, yield, or maintain course.

[0018] In some implementations, vehicle longitudinal interaction intent estimation can employ methods such as rule-based thresholding, kinematic extrapolation, learning model classification, game-theoretic programming, or belief space programming. For example, the tendency of a target traffic participant to cut in or yield can be determined based on parameters such as time distance, collision time, and time difference to reach the potential conflict area; it can also be assumed that the target traffic participant maintains a constant speed, constant acceleration, or continues to move along the current trajectory for a short period of time, and the extrapolation results are used by the vehicle planning module for avoidance or following control; neural network models can be used to predict the behavior category of the target vehicle based on its historical trajectory, the status of surrounding vehicles, and road environment information; or game-theoretic programming, belief space programming, or active detection methods can be used to handle vehicle interaction problems.

[0019] However, the above methods still have certain limitations in complex longitudinal interaction scenarios. For example, rule-based thresholding methods are prone to judgment jumps under critical conditions or sensor measurement jitter; kinematic extrapolation methods tend to ignore the impact of the vehicle's actions on the subsequent behavior of the target traffic participants; learning models may suffer from generalization instability when road scenarios, urban traffic habits, or driving styles change, and they are usually unable to express the degree of uncertainty in intent judgment; game-theoretic programming or belief space programming methods have relatively complex calculation processes, and some methods focus more on continuous trajectory prediction or planning strategy solving, making it difficult to directly output discrete longitudinal interaction intent probabilities for vehicle behavior decision-making.

[0020] Therefore, in the process of estimating the longitudinal interaction intention of vehicles, it is still necessary to address issues such as differences in intention tendencies under different traffic scenarios, fluctuations in intention caused by short-term observation noise, social misunderstandings such as yielding or robbing each other, the chain interaction effects among multiple traffic participants, insufficient utilization of explicit signals and interaction responses, and the lack of statistical calibration and reliability quantification of intention probabilities.

[0021] Based on this, this application provides a method for estimating the longitudinal interaction intention of vehicles, which can be applied to traffic scenarios with longitudinal interaction relationships, such as lane merging, entering ramps, changing lanes, cutting in or being cut in, following others, and overtaking. It is used to estimate the probability of the intentions of target traffic participants to cut in, yield, and keep their distance, and to express the uncertainty of the intention estimation results.

[0022] The following description, in conjunction with the accompanying drawings, details the vehicle longitudinal interaction intent estimation method provided in this application through embodiments and application scenarios.

[0023] The vehicle longitudinal interaction intent estimation method provided in this application can be executed by an electronic device deployed in the vehicle, or a functional module or entity within the electronic device capable of implementing the method. This electronic device includes, but is not limited to, an in-vehicle controller, an autonomous driving domain controller, a central computing unit, a vehicle controller, a driver assistance controller, or other computing devices with data processing capabilities; this application does not limit the specific type of device.

[0024] The following uses an electronic device as the execution subject to illustrate the vehicle longitudinal interaction intent estimation method provided in the embodiments of this application.

[0025] Figure 1 This is a flowchart illustrating the vehicle longitudinal interaction intent estimation method provided in some embodiments of this application. For example... Figure 1 As shown, the method includes steps 110 to 160.

[0026] Step 110: Obtain the raw observation data of traffic participants around the vehicle, and perform fusion processing on the raw observation data to obtain the status information of each traffic participant.

[0027] Among them, the self-vehicle can be the vehicle that performs the vehicle longitudinal interaction intention estimation method; traffic participants can be road participants located around the self-vehicle and that may affect the self-vehicle's driving decision, such as vehicles, two-wheeled vehicles or pedestrians.

[0028] Raw observation data can be data that is directly obtained by the vehicle through onboard sensing devices, onboard communication devices, or map data sources without unified fusion processing.

[0029] Status information can include information describing the position, motion status, and location of traffic participants relative to their own vehicles and on the road.

[0030] In one implementation, the electronic device can acquire raw observation data of traffic participants around the vehicle using at least one of a camera, millimeter-wave radar, lidar, inertial navigation sensor, wheel speed sensor, high-definition map, and V2X communication module. Since raw observation data from different sources may differ in sampling time, coordinate reference, data format, and measurement accuracy, the electronic device can perform time synchronization, coordinate transformation, and data association processing on the raw observation data, unifying the data from different sources into the vehicle's coordinate system. The vehicle coordinate system can be a coordinate system established with the vehicle as the reference object, used to represent the position and motion state of surrounding traffic participants relative to the vehicle. Subsequently, the electronic device can perform multi-sensor fusion and target tracking processing on the processed raw observation data to obtain the state vector of each traffic participant in the vehicle's coordinate system.

[0031] For example, for any traffic participant at time t, its state vector can be represented as:

[0032] in, This refers to the longitudinal relative distance between the traffic participant and the vehicle. The relative speed between the traffic participant and the vehicle. This is an estimate of the longitudinal acceleration of the traffic participant. The longitudinal acceleration of the vehicle. This represents the lateral offset of the traffic participant relative to the vehicle or the target lane. This is the lane number of the lane in which the traffic participant is located. The lateral offset can be used to subsequently determine whether the traffic participant is in the process of changing lanes, cutting in, or approaching the target lane.

[0033] For example, if there is a vehicle in the adjacent lane to the right front of the vehicle, the electronic device can identify the positional relationship between the vehicle and the lane lines based on camera data, determine the longitudinal relative distance and relative speed between the vehicle and the vehicle based on millimeter-wave radar data, supplement the spatial position of the vehicle based on lidar data, determine the longitudinal acceleration of the vehicle based on inertial navigation sensors and wheel speed sensors, and determine the lane number of the vehicle based on a high-definition map. After fusing the above data, the electronic device can obtain the state vector of the vehicle in the vehicle's coordinate system.

[0034] Step 120: Select target traffic participants from the traffic participants around the vehicle who have a longitudinal interaction relationship with the vehicle.

[0035] Electronic devices can filter target traffic participants with longitudinal interaction relationships with their own vehicles from among the traffic participants around the vehicle based on the status information of each traffic participant.

[0036] Among them, the target traffic participants can be traffic participants who may influence the longitudinal behavior decisions of the vehicle in the current or future period of time.

[0037] Longitudinal interaction relationships can be the interaction relationships formed between vehicles and traffic participants around the longitudinal sequence, such as who passes first, who passes last, whether to yield, whether to cut in front, or whether to maintain the current relative formation.

[0038] In one implementation, the electronic device can determine whether each traffic participant might affect the vehicle's longitudinal traffic order based on their position relative to the vehicle, direction of movement, relative speed, and relationship with the vehicle's travel path. Traffic participants that are far away, whose direction of movement is irrelevant, or who will not affect the vehicle's longitudinal decision-making can be excluded from the subsequent intent estimation process; traffic participants located in front of the vehicle, in adjacent lanes, merging areas, potentially cutting into the vehicle's travel path, or potentially forming a sequential traffic relationship with the vehicle can be identified as target traffic participants.

[0039] For example, if a vehicle is in the adjacent lane ahead of the vehicle and gradually approaches the vehicle's lane, this vehicle may influence whether the vehicle slows down to yield, maintains its speed, or adjusts its following distance. Electronic devices can identify this vehicle as a target traffic participant. Similarly, if a vehicle is in the same lane ahead of the vehicle and the longitudinal distance between them is relatively short, this vehicle can also be identified as a target traffic participant to subsequently estimate its intention to cut in, yield, or maintain its position.

[0040] Step 130: Based on the historical longitudinal behavior data of the target traffic participants, construct a prior distribution of the target intent that matches the current traffic scenario.

[0041] Historical longitudinal behavior data can be the behavioral data of the target traffic participant along the longitudinal direction of the road before the current moment, such as changes in speed, acceleration, longitudinal spacing, and tendencies to move forward or yield. The prior distribution of target intent can be the probability distribution of the target traffic participant's various longitudinal interaction intents before combining the current observation information.

[0042] In one implementation, the electronic device can determine the initial tendency of a target traffic participant regarding their intention to cut in, yield, or remain in the right-of-way based on changes in their longitudinal behavior over a past period. If the target traffic participant continuously approaches the space in front of the vehicle or exhibits a tendency to accelerate forward over a period of time, the prior probability of their intention to cut in can be relatively high; if the target traffic participant continuously decelerates or gradually increases the distance from the vehicle, the prior probability of their intention to yield can be relatively high; if the relative speed and longitudinal distance between the target traffic participant and the vehicle are relatively stable, the prior probability of their intention to remain in the right-of-way can be relatively high.

[0043] Electronic devices can also perform matching processing on the aforementioned prior distribution in conjunction with the current traffic scenario, so that the prior distribution of target intent can reflect the intent tendency in the current road interaction environment. For example, the same acceleration and moving forward behavior may have different meanings in ordinary following scenarios and merging scenarios. Therefore, electronic devices can adjust the prior probabilities of various longitudinal interaction intents according to the current traffic scenario to obtain a prior distribution of target intent that matches the current traffic scenario.

[0044] Step 140: Based on the state information of the target traffic participants, determine the observation likelihood for various types of longitudinal interaction intentions.

[0045] Observational likelihood can be defined as the probability of the occurrence of current state information or observational features extracted from current state information, given that the target traffic participant has a certain type of longitudinal interaction intention. Various longitudinal interaction intentions may include, but are not limited to, the intention to cut in, the intention to yield, and the intention to hold on.

[0046] In one implementation, the electronic device can assess the degree of matching between the current state information and different longitudinal interaction intentions based on state information such as the relative position, relative speed, acceleration changes, and road position changes between the target traffic participant and the vehicle. For example, if the target traffic participant gradually moves forward relative to the vehicle, and its motion indicates that it may take up space in front of the vehicle, the observation likelihood corresponding to the intention to cut in front can be relatively high; if the target traffic participant gradually decelerates and increases the space in front of the vehicle, the observation likelihood corresponding to the intention to yield can be relatively high; if the changes in the relative position and relative speed between the target traffic participant and the vehicle are small, the observation likelihood corresponding to the intention to hold on can be relatively high.

[0047] The electronic device can obtain the observation likelihoods corresponding to the intention to cut in, the intention to yield, and the intention to hold back. This observation likelihood is used to represent the degree to which the current state information of the target traffic participant supports different longitudinal interaction intentions, and serves as the input for subsequent probability updates in the hierarchical Bayesian model.

[0048] Step 150: Based on the prior distribution of target intent and observation likelihood, determine the target posterior probability corresponding to various vertical interaction intents through a hierarchical Bayesian model.

[0049] Among them, the target posterior probability can be the updated probability of the target traffic participant belonging to various types of vertical interaction intentions after combining historical longitudinal behavior data, current traffic scenario and current state information.

[0050] In one implementation, the electronic device can use the prior distribution of the target intent as the prior probability input for various vertical interaction intents, and the observation likelihood as the observational support input for the current state information regarding various vertical interaction intents. A hierarchical Bayesian model can update the aforementioned prior probabilities and observation likelihoods to obtain the target posterior probabilities corresponding to the preemptive intent, yielding intent, and hold-and-hold intent, respectively.

[0051] For example, if the prior distribution of the target intent indicates that the target traffic participant has a certain tendency to cut in line, and the current observation likelihood also supports this intention, then the electronic device can improve the posterior probability of the target intent through a hierarchical Bayesian model. If there is a difference between the prior distribution of the target intent and the current observation likelihood, for example, historical behavior shows that the target traffic participant tends to remain still, but current state information shows that they are slowing down to give way, then the electronic device can combine both to determine the posterior probability of the target for various longitudinal interaction intents, rather than making a judgment based solely on a single observation result.

[0052] Step 160: Based on the target posterior probability corresponding to various vertical interaction intentions, generate an intention confidence set for use in vehicle behavior decision-making.

[0053] An intent confidence set can be a collection of one or more longitudinal interaction intents, representing longitudinal interaction intents that may be valid under current observation conditions and statistical calibration constraints. Unlike methods that only output the intent with the highest probability, an intent confidence set can retain multiple candidate longitudinal interaction intents simultaneously, even when the posterior probabilities of the targets are close, the intent judgment is not sufficiently certain, or there are multiple reasonable interpretations of the intent.

[0054] In one implementation, the electronic device can determine the longitudinal interaction intentions that need to be included in the intention confidence set based on the target posterior probabilities corresponding to the intention to cut in, the intention to yield, and the intention to hold. If the target posterior probability of a certain type of longitudinal interaction intention is significantly higher than that of other longitudinal interaction intentions, the intention confidence set may only include that longitudinal interaction intention; if the target posterior probabilities of multiple longitudinal interaction intentions are relatively close, or if the current state information is insufficient to exclude multiple longitudinal interaction intentions, the intention confidence set may include multiple longitudinal interaction intentions.

[0055] In one implementation, the electronic device can determine the vertical interaction intents that need to be included in the intent confidence set based on the target posterior probabilities corresponding to various vertical interaction intents. If the target posterior probability of a certain type of vertical interaction intent is significantly higher than that of other vertical interaction intents, the intent confidence set may only include that vertical interaction intent; if the target posterior probabilities of multiple vertical interaction intents are relatively close, or if the current state information is insufficient to exclude multiple vertical interaction intents, the intent confidence set may include multiple vertical interaction intents.

[0056] Electronic devices can output the intent confidence set along with the target posterior probabilities corresponding to various longitudinal interaction intents, serving as the basis for vehicle behavior decisions such as acceleration, deceleration, maintaining speed, following, yielding, passing, or adjusting following distance. For example, when the intent confidence set only contains the intent to cut in front, the vehicle can decelerate in advance or increase the following distance; when the intent confidence set contains multiple longitudinal interaction intents, the vehicle can adopt a more conservative longitudinal control strategy.

[0057] According to the vehicle longitudinal interaction intent estimation method provided in this application, by acquiring and fusing the original observation data of traffic participants around the vehicle, the state information of each traffic participant is obtained, which can provide a unified data foundation for subsequent target traffic participant screening and observation likelihood determination; by screening target traffic participants with longitudinal interaction relationships with the vehicle from the traffic participants around the vehicle, the subsequent intent estimation can be focused on objects that may affect the vehicle's longitudinal behavior decision, reducing the influence of irrelevant traffic participants on the estimation process; by constructing a target intent prior distribution that matches the current traffic scenario based on the historical longitudinal behavior data of the target traffic participants, the prior distribution can simultaneously reflect the historical longitudinal behavior characteristics of the target traffic participants and the current traffic scenario. The impact of the preceding traffic scenario on longitudinal interaction intentions is investigated. By determining the observation likelihood for various longitudinal interaction intentions based on the state information of target traffic participants, the current state of the target traffic participants can be incorporated into the intention estimation process. By determining the target posterior probability corresponding to various longitudinal interaction intentions based on the prior distribution and observation likelihood of the target intentions and using a hierarchical Bayesian model, historical longitudinal behavior, current traffic scenario, and current state information can be integrated within the same probability update framework, improving the estimation stability of overtaking intentions, yielding intentions, and holding intentions. By generating an intention confidence set based on the target posterior probability, a candidate intention set can be retained when multiple longitudinal interaction intentions are possible, providing a probabilistic basis for vehicle behavior decision-making. Therefore, the accuracy, stability, and reliability of vehicle intention estimation for target traffic participants in longitudinal interaction scenarios can be improved.

[0058] In some embodiments, selecting target traffic participants with longitudinal interaction relationships with the vehicle from traffic participants surrounding the vehicle includes: identifying one or more candidate traffic participants from traffic participants surrounding the vehicle based on road topology information and potential conflict areas; determining an interaction intensity index between each candidate traffic participant and the vehicle based on the state information of each candidate traffic participant; wherein the interaction intensity index is determined based on at least one of collision time, time difference to potential conflict areas, and degree of overlap of potential conflict areas; and identifying target traffic participants with longitudinal interaction relationships with the vehicle from the candidate traffic participants according to the interaction intensity index corresponding to each candidate traffic participant.

[0059] Among them, road topology information can be used to characterize the connection relationship between the road where the vehicle is located and adjacent roads, the direction of lane extension, the lane merging relationship, the lane termination position, the ramp access position, and the spatial relationship between passable paths.

[0060] Potential conflict areas can be areas where a vehicle may intersect with other traffic participants in the future, have spatial overlaps, or compete for longitudinal priority during their journeys. Examples include ramp merging areas, merging points, lane narrowing areas, adjacent lane cutting-in areas, lane line overlapping areas, or areas where the vehicle's planned path overlaps with the possible paths of other traffic participants.

[0061] Electronic devices can perform preliminary screening of traffic participants around the vehicle based on road topology information and potential conflict zones to obtain candidate traffic participants. Potential conflict zones can be areas where the vehicle and other traffic participants may intersect, overlap spatially, or compete for longitudinal priority. Subsequently, the electronic device can calculate an interaction intensity index based on the state information of each candidate traffic participant. This interaction intensity index can be determined by one or more of the following: collision time, time difference of arrival at the potential conflict zone, and degree of overlap of the potential conflict zone. A higher interaction intensity index indicates that the candidate traffic participant is more likely to influence the vehicle's longitudinal traffic order. The electronic device can identify candidate traffic participants whose interaction intensity index meets preset conditions or who are ranked higher as target traffic participants.

[0062] In one implementation, the electronic device can determine the interaction intensity index of each candidate traffic participant based on one or more of the following: collision time, time difference of arrival at the potential conflict area, and degree of overlap of the potential conflict areas.

[0063] Collision time is used to characterize the time required for a vehicle to reach a conflict state with a candidate traffic participant under the current relative speed and relative distance conditions. The smaller the collision time, the higher the interaction intensity.

[0064] The time difference in arrival at the potential conflict zone is used to characterize the degree to which vehicles and candidate traffic participants may enter the conflict zone simultaneously in time. The smaller the time difference, the higher the interaction intensity.

[0065] The degree of overlap of potential conflict areas is used to characterize the spatial overlap between the planned path of the vehicle and the possible paths of the candidate traffic participants. The higher the degree of overlap, the stronger the interaction.

[0066] Electronic devices can filter target traffic participants based on interaction intensity indicators, such as selecting candidate traffic participants whose interaction intensity is greater than a threshold or whose ranking is high as target traffic participants, and further estimate the probability of their intention to cut in, yield, or hold back.

[0067] For example, in the scenario of merging at an entrance ramp, if a candidate vehicle approaches the merging area of ​​the main road and has a small time difference with the arriving vehicle, and their paths overlap significantly, then the candidate vehicle has a high interaction intensity and can be selected as a target traffic participant. Similarly, in the scenario of lane narrowing, a candidate vehicle that may merge into the lane of the arriving vehicle, with a small arrival time difference and high path overlap, can also be identified as a target traffic participant.

[0068] In the above embodiments, by determining the interaction intensity index based on road topology information, potential conflict areas, and the state information of candidate traffic participants, and then filtering target traffic participants accordingly, it is possible to identify objects that are more likely to affect the longitudinal passage order of the vehicle from multiple traffic participants around the vehicle, reduce irrelevant traffic participants from entering the subsequent intent estimation process, thereby improving the targeting of target filtering and the efficiency of intent estimation.

[0069] In some embodiments, a prior distribution of target intent matching the current traffic scenario is constructed based on the historical longitudinal behavior data of the target traffic participants. This includes: determining the current traffic scenario variables based on the road scenario information of the road where the vehicle is located; generating an initial prior distribution of intent corresponding to various longitudinal interaction intents based on the historical longitudinal behavior data of the target traffic participants; and conditionally adjusting the initial prior distribution of intent based on the current traffic scenario variables to obtain a prior distribution of target intent matching the current traffic scenario. The prior distribution of target intent is updated online as the longitudinal behavior data of the target traffic participants is continuously acquired during vehicle operation.

[0070] Among them, road scene information is used to characterize the scene features of the road where the vehicle is currently located and the road area where the target traffic participants are located. It can include high-definition map topology information, the vehicle's planned route, road markings, and traffic signs.

[0071] In one implementation, the electronic device can identify current traffic scene variables by combining the state information of the target traffic participant and road scene information. The current traffic scene variables can be denoted as... It is used to characterize the longitudinal interaction scenario between the target traffic participant and the vehicle, such as ramp merging scenario, lane narrowing scenario, lane changing scenario, cutting in scenario, following scenario, or overtaking scenario.

[0072] For example, when a target traffic participant is located on an entrance ramp and the distance between the target traffic participant and the conflict zone on the main road is less than a preset distance threshold, the electronic device can detect the current traffic scenario variables. This is identified as a ramp merging scenario, i.e. When the number of lanes ahead decreases and the target traffic participant is located in a lane that is about to disappear, electronic devices can determine the current traffic scenario variable as a lane narrowing scenario, i.e. When the lateral velocity of the target traffic participant is not zero, and the target traffic participant is crossing a lane line or approaching a target lane, the electronic device can determine the current traffic scenario variables as a lane-changing or cutting-in scenario, i.e. For other traffic scenarios, electronic devices can also be further defined according to engineering requirements.

[0073] Electronic devices can also construct the longitudinal interaction intentions of target traffic participants as intention random variables. ,in, , Indicating the intention to steal, Indicates the intention to give way. It indicates the intention to maintain.

[0074] Electronic devices can generate initial intent prior distributions corresponding to various longitudinal interaction intents based on the historical longitudinal behavior data of target traffic participants. Historical longitudinal behavior data can be used to characterize the behavioral changes exhibited by target traffic participants along the longitudinal direction of the road before the current moment, such as changes in longitudinal speed, changes in longitudinal acceleration, changes in longitudinal distance from their own vehicle, and tendencies to move forward or yield.

[0075] In one implementation, the electronic device can generate a historical behavior summary based on historical longitudinal behavior data. And based on the historical behavior summary, it generates the initial intent prior distribution corresponding to various vertical interaction intents. This historical behavior summary Information may include longitudinal acceleration distribution, longitudinal speed changes, whether there were multiple attempts to change lanes, whether there was continuous acceleration to move ahead, or whether there was continuous deceleration to yield, etc., within a preset time range in the past.

[0076] After obtaining the initial intent prior distribution, the electronic device can then adjust its settings based on the current traffic scenario variables. The initial intent prior distribution is conditionally adjusted to obtain a target intent prior distribution that matches the current traffic scenario. In different traffic scenarios, the same historical longitudinal behavioral data may correspond to different longitudinal interaction intent tendencies. Therefore, scenario-conditional adjustment can make the target intent prior distribution more consistent with the current road interaction environment.

[0077] In one implementation, the prior distribution of the target intent can be parameterized using a Dirichlet distribution, for example:

[0078] in, To match the variables of the current traffic scenario and driving style type Relevant prior parameters, These can be latent variables or online estimators representing the driving style type of the target traffic participant.

[0079] In another implementation, the prior distribution of the target intent can be represented using feature mapping and a normalized exponential function, for example:

[0080] in, For the current traffic scenario variables Summary of historical behavior The constructed eigenvectors These are learnable weight parameters.

[0081] In the above embodiments, by determining the current traffic scene variables based on road scene information and using the current traffic scene variables to conditionally adjust the initial intent prior distribution generated from historical longitudinal behavior data, the target intent prior distribution can simultaneously reflect historical longitudinal behavior and the current road interaction scene, thereby improving the scene adaptability and accuracy of subsequent longitudinal interaction intent estimation.

[0082] In some embodiments, the method further includes: obtaining initial prior parameters; wherein the initial prior parameters are obtained by meta-learning training of prior parameters using a labeled longitudinal behavior dataset containing multiple road scenes in an offline phase; during vehicle operation, updating the driving style information of the target traffic participant based on continuously acquired longitudinal behavior data of the target traffic participant; updating the initial prior parameters online based on the driving style information to obtain target prior parameters matching the driving style of the target traffic participant; and updating the target intent prior distribution online based on the target prior parameters to obtain the updated target intent prior distribution.

[0083] Electronic devices can further incorporate meta-learning mechanisms and online driving style update mechanisms to adaptively correct the prior distribution of target intent.

[0084] In the offline phase, electronic devices can use an annotated longitudinal behavior dataset containing multiple road scenarios to perform meta-learning training on prior parameters to obtain initial prior parameters. The annotated longitudinal behavior dataset can include sample data from multiple cities, multiple road types, or multiple traffic scenarios. Each sample can include historical longitudinal behavior data of the sample traffic participants, road scene information, and annotated longitudinal interaction intentions. Through meta-learning training, the electronic device can obtain initial prior parameters suitable as a starting point for online adaptation.

[0085] Through the above meta-learning training, the electronic device can obtain initial prior parameters that are suitable as the starting point for online adaptation, so that when updating the prior parameters in new road scenarios or new target traffic participants, it is not necessary to completely retrain.

[0086] In one implementation, when the target intent prior distribution is parameterized using a Dirichlet distribution, the initial prior parameters can be initialization parameters in the Dirichlet prior parameter space. Meta-learning training can be implemented using model-independent meta-learning algorithms, Reptile-like algorithms, or other meta-learning algorithms for learning parameter initialization. For example, the initial prior parameters can be learned using the following objective function:

[0087] in, These are the initial prior parameters obtained through meta-learning. Let the initial values ​​of the prior parameters to be optimized be... For the first One training task, For training mission The corresponding training loss, For learning rate, For training mission The corresponding verification data, To verify the loss.

[0088] During vehicle operation, electronic devices can update the driving style information of target traffic participants based on continuously acquired longitudinal behavioral data. Driving style information can characterize the behavioral tendencies of target traffic participants during longitudinal interactions, such as being more aggressive, conservative, or maintaining a stable approach. Based on the driving style information, the electronic devices can update the initial prior parameters online to obtain target prior parameters matching the driving style of the target traffic participants, and then update the target intent prior distribution online based on the target prior parameters to obtain the updated target intent prior distribution.

[0089] In this way, meta-learning can be used to train initial prior parameters that have the ability to initialize across road scenarios. During vehicle operation, the prior parameters are updated online according to the driving style information of the target traffic participants, so that the prior distribution of target intent can be further adapted to the differences in road scenarios and the individual behavioral tendencies of the target traffic participants, thereby improving the cross-scenario adaptability and individual adaptability of longitudinal interaction intent estimation.

[0090] In some embodiments, determining current traffic scene variables based on road scene information of the road where the vehicle is located includes: determining the positional relationship between the target traffic participant and the potential conflict area, the lane change state of the lane where the target traffic participant is located, and the lateral movement state of the target traffic participant relative to the lane line based on the road scene information of the road where the vehicle is located; determining current traffic scene variables based on positional relationship, lane change state, and lateral movement state; wherein, the current traffic scene variables are at least used to distinguish between ramp merging scenarios, lane narrowing scenarios, and lane changing or cutting-in scenarios.

[0091] Electronic devices can determine the positional relationship between a target traffic participant and a potential conflict area, the lane change status of the target traffic participant's lane, and the lateral movement status of the target traffic participant relative to the lane lines based on road scene information of the road where the vehicle is located, and determine the current traffic scene variables accordingly. Among them, the positional relationship can be used to characterize whether the target traffic participant is approaching the main road conflict area, the distance between it and the potential conflict area, or the estimated arrival time; the lane change status can be used to characterize whether there is a reduction in the number of lanes, lane termination, or lane merging in the target traffic participant's lane; and the lateral movement status can be used to characterize whether the target traffic participant is crossing lane lines or approaching the target lane.

[0092] For example, if the target traffic participant is located on an entrance ramp and the distance between it and the conflict area on the main road is less than a preset distance threshold, the electronic device can determine the current traffic scene variable as a ramp merging scenario; if the number of lanes ahead decreases and the target traffic participant is located in a lane that is about to disappear, the electronic device can determine the current traffic scene variable as a lane narrowing scenario; if the lateral speed of the target traffic participant is not zero and it is crossing a lane line or approaching the lane where its own vehicle is located, the electronic device can determine the current traffic scene variable as a lane changing or cutting in.

[0093] In the above embodiments, the current traffic scenario variables are determined by the positional relationship between the target traffic participant and the potential conflict area, the lane change state of the lane, and the lateral movement state relative to the lane line. This can distinguish between ramp merging, lane narrowing, and lane changing or cutting-in scenarios from aspects such as the degree of conflict proximity, lane structure changes, and lateral cutting-in trends, thereby improving the scenario adaptability of the prior distribution of target intentions.

[0094] In some embodiments, the observation likelihood for various longitudinal interaction intentions is determined based on the state information of the target traffic participants, including: extracting longitudinal motion features, lateral motion features, explicit signal features, and interaction response features based on the state information of the target traffic participants; constructing observation conditional probability models corresponding to various longitudinal interaction intentions based on the longitudinal motion features, lateral motion features, explicit signal features, and interaction response features; and determining the observation likelihood for various longitudinal interaction intentions based on the observation conditional probability models.

[0095] Electronic devices can extract multi-source observation features for constructing observation likelihood based on the state information of target traffic participants. These multi-source observation features can form the multi-source observation vector at the current time. This is used to characterize the motion state, lateral behavior, explicit intent expression, and response to vehicle actions of a target traffic participant in the current traffic scenario. The state information of the target traffic participant can be denoted as... The current traffic scenario variables can be denoted as: Vertical interaction intent can be denoted as .

[0096] Multi-source observation features can include longitudinal motion features, lateral motion features, dominant signal features, and interactive response features. Among them, longitudinal motion features can include longitudinal relative distance. Relative velocity Target acceleration Collision time and time interval At least one of the following: lateral motion features may include at least one of the lateral distance between the target traffic participant and the lane line, lateral velocity components, and vehicle attitude angle change rate; explicit signal features may include at least one of the turn signal status, brake light brightness, brake light flashing frequency, and V2X broadcast intent messages or path messages; interactive response features may be used to characterize the target traffic participant's response to the vehicle's recent actions, such as whether the target traffic participant decelerates after the vehicle accelerates slightly, or whether the target traffic participant continues to approach the target lane after the vehicle decelerates to make room.

[0097] In one implementation, the electronic device can construct an observation likelihood function. This represents the probability of a multi-source observation vector occurring given that the target traffic participant has a specific longitudinal interaction intention, possesses current state information, and is within the current traffic scenario. The observation conditional probability model can be implemented by combining interpretable physical feature functions with learnable weight parameters, for example:

[0098] in, For the first One characteristic function, For vertical interaction intent and current traffic scenario variables The corresponding weight parameters, The corresponding bias term is denoted as . Feature functions can include relative velocity sign features, target acceleration direction features, collision time interval features, consistency features between turn signal status and lane-changing phases, consistency features between brake light status and deceleration behavior, and interactive response features, etc. In addition to the above forms, observation conditional probability models can also be implemented using Gaussian mixture models, logistic regression models, or Naive Bayes models.

[0099] In some embodiments, the electronic device may also introduce counterfactual trajectory verification as an additional likelihood factor. For each type of candidate longitudinal interaction intent, the electronic device can generate a corresponding counterfactual predicted trajectory based on the current state information and kinematic model of the target traffic participant, and calculate the trajectory deviation between the actual observed trajectory and each counterfactual predicted trajectory. Counterfactual likelihood factor It can be determined in the following form:

[0100] in, The actual observed trajectories of the target traffic participants. Candidate vertical interaction intent The corresponding counterfactual prediction trajectory, This represents the scale parameter corresponding to the counterfactual trajectory bias. Electronic devices can fuse feature-based likelihood factors with counterfactual likelihood factors to generate observed likelihoods corresponding to various longitudinal interaction intentions.

[0101] In the above embodiments, by extracting the longitudinal motion features, lateral motion features, explicit signal features, and interactive response features of the target traffic participants, and combining them with physical feature functions, learnable weight parameters, and counterfactual trajectory verification to construct observation likelihood, it is possible to distinguish different longitudinal interaction intentions from the two levels of multi-source features and trajectory consistency, thereby improving the accuracy of observation likelihood and the reliability of intention estimation.

[0102] In some embodiments, based on the prior distribution of the target intent and the observation likelihood, the target posterior probability corresponding to various vertical interaction intents is determined by a hierarchical Bayesian model, including: in the coarse-grained layer, various vertical interaction intents are taken as coarse-grained intent states, and an intent transition probability matrix between each coarse-grained intent state is constructed based on a Markov chain; based on the prior distribution of the target intent, the observation likelihood, and the intent transition probability matrix, Bayesian filtering is performed on each coarse-grained intent state to update it, thereby obtaining the target posterior probability corresponding to various vertical interaction intents.

[0103] Electronic devices can use hierarchical Bayesian models to update the probabilistics of various vertical interaction intentions. Hierarchical Bayesian models can be divided into coarse-grained layers and fine-grained layers. The coarse-grained layer can be used to represent the strategic-level vertical interaction intentions of target traffic participants, while the fine-grained layer can be used to represent the tactical-level vertical movement behavior states of target traffic participants.

[0104] In the coarse-grained layer, electronic devices can view the longitudinal interaction intentions of target traffic participants as Markov states that evolve over time.

[0105] Electronic devices can be based on current traffic scenario variables Construct the intent transition probability matrix between each coarse-grained intent state:

[0106] in, Represents variables in the current traffic scenario. Below, the longitudinal interaction intention of the target traffic participant changes from the intention state of the previous moment. Transition to the intentional state at the current moment The probability of intention transition can be set to have an intention persistence constraint, that is, the self-transition probability of each coarse-grained intention state is greater than the transition probability between different intention states, so as to reduce frequent intention switching caused by short-term observation fluctuations.

[0107] Electronic devices can update each coarse-grained intent state using Bayesian filtering based on the prior distribution of the target intent, the observation likelihood, and the intent transition probability matrix.

[0108] in, This represents the observation information at the current moment. This represents the sequence of observation information from the initial time to the current time. This indicates that the current moment corresponds to the intention state. The observational likelihood, Indicates the probability of intention transition. This represents the posterior probability at the previous time step.

[0109] After normalization, the electronic device can obtain a posterior probability vector:

[0110] in, , and These represent the target posterior probabilities that the target traffic participant currently has the intention to cut in, the intention to yield, and the intention to remain in the right-of-way, respectively.

[0111] In some embodiments, in the fine-grained layer, the electronic device divides the longitudinal motion behavior of the target traffic participant into multiple fine-grained behavior states; based on the coarse-grained posterior probability and the current traffic scene variables, the posterior probability of each fine-grained behavior state is updated to obtain the fine-grained posterior probability; the longitudinal motion behavior prediction result of the target traffic participant is determined based on the fine-grained posterior probability, and the longitudinal motion behavior prediction result is used for vehicle behavior decision-making.

[0112] Electronic devices can construct fine-grained behavioral states. ,in It can include states such as aggressive acceleration, gentle acceleration, constant speed, gentle deceleration, and emergency braking.

[0113] Electronic devices can update the fine-grained behavioral states based on coarse-grained posterior probabilities and current traffic scenario variables.

[0114] in, Indicates vertical interaction intent and current traffic scenario variables Below, the target traffic participants are in a fine-grained behavioral state. The probability, This represents the posterior probability of the target obtained from the coarse-grained layer.

[0115] For example, in a ramp merging scenario, if the target traffic participant has a high probability of previously intending to cut in, and current observations show that they are accelerating towards the main road conflict zone, the electronic device can maintain the continuity of the cutting-in intention through the intention transfer probability matrix. This results in a higher posterior probability of cutting in at the coarse-grained level and an increased posterior probability of aggressive or moderate acceleration at the fine-grained level. Similarly, in a lane narrowing scenario, if the target traffic participant continuously decelerates and increases longitudinal spacing, the posterior probability of cutting in at the coarse-grained level decreases, while the posterior probability of yielding increases, and the posterior probability of moderate deceleration or sudden braking at the fine-grained level also increases.

[0116] By modeling the temporal evolution of longitudinal interaction intentions using an intention transition probability matrix, frequent intention jumps caused by single-frame classification can be avoided. The coarse-grained layer outputs the posterior probabilities of various longitudinal interaction intentions, providing a basis for vehicle behavior decisions; the fine-grained layer describes the specific longitudinal motion state, providing a reference for setting longitudinal control parameters.

[0117] In some embodiments, an intent confidence set is generated based on the target posterior probabilities corresponding to various vertical interaction intents, including: obtaining a statistical quantile threshold determined based on a calibration dataset during the offline phase; the statistical quantile threshold is determined based on a preset statistical coverage rate and the inconsistency score corresponding to each sample in the calibration dataset, the inconsistency score being used to characterize the degree of inconsistency between the posterior probability of the vertical interaction intent corresponding to the sample and the actual vertical interaction intent; during the online deployment phase, the current inconsistency score corresponding to each type of vertical interaction intent is determined based on the target posterior probabilities corresponding to various vertical interaction intents; vertical interaction intents with a current inconsistency score less than or equal to the statistical quantile threshold are determined as target vertical interaction intents, and the target vertical interaction intents constitute the intent confidence set; wherein, the statistical quantile threshold is configured to ensure that the coverage probability of the intent confidence set to the actual vertical interaction intent is not less than the preset statistical coverage rate.

[0118] Electronic devices can further superimpose a conformal prediction calibration layer after obtaining the target posterior probabilities corresponding to various vertical interaction intentions through a hierarchical Bayesian model, in order to generate an intention confidence set based on the target posterior probabilities. The conformal prediction calibration layer can be used to statistically calibrate the Bayesian posterior output, so that the generated intention confidence set meets the preset statistical coverage requirements under limited sample conditions.

[0119] During the offline calibration phase, the electronic device can determine the statistical quantile threshold using an independent calibration dataset. For each sample in the calibration dataset, the electronic device can calculate a non-consistency score. The inconsistency score characterizes the degree of inconsistency between the posterior probability of the longitudinal interaction intent corresponding to a sample and the true longitudinal interaction intent. In one implementation, the inconsistency score can be determined based on the ranking loss or Hinge loss of the posterior probability of the true longitudinal interaction intent among each candidate longitudinal interaction intent. Electronic devices can utilize the inconsistency scores of multiple samples in the calibration dataset and a preset statistical coverage. Determine the statistical quantile threshold .

[0120] During the online deployment phase, electronic devices can determine the current inconsistency score corresponding to each type of longitudinal interaction intent based on the target posterior probability corresponding to the current target traffic participants. Longitudinal interaction intents with a current inconsistency score less than or equal to the statistical quantile threshold are identified as target longitudinal interaction intents, and these target longitudinal interaction intents constitute the intent confidence set.

[0121] in, This represents the intent confidence set corresponding to the current observation information sequence.

[0122] Using the above conformal prediction calibration method, the intended confidence set can satisfy the marginal coverage constraint, that is:

[0123] in, This indicates the true intent of vertical interaction. The marginal coverage constraint means that, within the preset statistical coverage rate... In this case, the probability that a genuine vertical interaction intent is included in the intent confidence set is no less than the preset statistical coverage. When the intent confidence set contains only a single vertical interaction intent, the electronic device can output a relatively clear intent judgment result; when the intent confidence set contains multiple vertical interaction intents, the electronic device can determine that there is multiple candidate uncertainty in the current intent judgment.

[0124] Therefore, by obtaining the statistical quantile threshold through offline calibration and generating an intent confidence set based on the current inconsistency score during the online deployment phase, the intent confidence set can meet the preset statistical coverage requirements. This allows for the retention of potentially valid candidate intents when the intent is uncertain, providing a more reliable intent estimation basis for vehicle behavior decisions.

[0125] In some embodiments, the method further includes: determining an intent uncertainty index for target traffic participants based on the target posterior probabilities corresponding to various types of longitudinal interaction intents; determining the probability distribution of longitudinal interaction intents corresponding to the vehicle based on the vehicle's current planning behavior; determining mutual yielding risk, mutual snatching risk, and chain interaction risk based on the target posterior probabilities corresponding to target traffic participants, the probability distribution of longitudinal interaction intents corresponding to the vehicle, and the interaction consistency factor; and triggering an active disambiguation process when the intent confidence set contains multiple longitudinal interaction intents, the mutual yielding risk meets a first preset risk condition, the mutual snatching risk meets a second preset risk condition, the chain interaction risk meets a third preset risk condition, or the intent uncertainty index meets a preset uncertainty condition.

[0126] After obtaining the target posterior probability corresponding to the target traffic participant, electronic devices can perform cooperative consistency detection on the longitudinal interaction relationship between the vehicle and the target traffic participant to identify social misunderstanding risks such as both parties yielding, both parties cutting in, or uncertain intentions.

[0127] Electronic devices can determine an indicator of the intent uncertainty of a target traffic participant based on the target posterior probability. The intent uncertainty indicator can be based on the target posterior probability vector. posterior entropy Confirmed. The larger the posterior entropy, the higher the degree of uncertainty in the current intention judgment.

[0128] Electronic devices can also adjust their behavior based on the vehicle's current plan. Determine the probability distribution of the longitudinal interaction intent corresponding to the vehicle. For example, when the vehicle's current planned behavior is to accelerate through a potential conflict area, the probability of the vehicle's intent to cut in line can be relatively high; when the vehicle's current planned behavior is to slow down and give way, the probability of the vehicle's intent to give way can be relatively high.

[0129] In dual-vehicle interaction scenarios, electronic devices can construct a conditional joint intent distribution based on the target posterior probability of the target traffic participant, the longitudinal interaction intent probability distribution of the vehicle, and the interaction consistency factor:

[0130] in, Indicates the vertical interaction intent of the target traffic participants. This indicates the vehicle's intention to interact longitudinally. Indicates observation information, Represents the variables of the current traffic scenario. This indicates the current planned behavior of the vehicle. The interaction consistency factor is used to characterize the rationality of different combinations of longitudinal interaction intentions between the vehicle and the target traffic participants.

[0131] For example, in the current traffic scenario, if one party intends to cut in front while the other intends to yield, this combination may be highly reasonable; if both parties intend to cut in front, this combination may be less reasonable; if both parties intend to yield, this combination may result in decreased traffic efficiency or unnecessary speed reduction.

[0132] Based on the aforementioned joint intent distribution, electronic devices can determine yielding risk and collision risk. Yielding risk characterizes the probability that both the vehicle and the target traffic participant tend to yield, resulting in unnecessary deceleration or sudden braking. Collision risk characterizes the probability that both the vehicle and the target traffic participant tend to overtake, leading to an increased risk of collision.

[0133] Electronic devices can determine mutual concession risk and mutual theft risk based on conditional joint intent distribution:

[0134]

[0135] Where M(Y) represents the risk of both parties giving way, resulting in unnecessary deceleration or sudden braking, and M(R) represents the risk of both parties rushing to overtake, resulting in an increased risk of collision.

[0136] When the intent confidence set contains multiple vertical interaction intents, the mutual concession risk meets the first preset risk condition, the mutual snatching risk meets the second preset risk condition, the chain interaction risk meets the third preset risk condition, or the intent uncertainty index meets the preset uncertainty condition, the electronic device can trigger the active disambiguation process.

[0137] Therefore, by combining the target posterior probability of the target traffic participants, the probability distribution of the longitudinal interaction intention of the vehicle, and the interaction consistency factor, it is possible to identify risks such as uncertain intentions, both parties yielding, both parties rushing to pass, and the transmission of interactions among multiple traffic participants. When the triggering conditions are met, an active disambiguation process is initiated, thereby reducing the impact of social misunderstandings on vehicle behavior decisions and improving the safety and coordination of vehicle decisions in longitudinal interaction scenarios.

[0138] In some embodiments, based on the target posterior probability corresponding to the target traffic participant, the probability distribution of the longitudinal interaction intention corresponding to the vehicle, and the interaction consistency factor, the mutual yielding risk, mutual snatching risk, and cascading interaction risk are determined, including: constructing an interaction scenario graph; wherein the interaction scenario graph includes multiple nodes and directed edges connecting the nodes, each node corresponds to a target traffic participant and carries the target posterior probability of the target traffic participant for various longitudinal interaction intentions, and each directed edge corresponds to the interaction relationship between two target traffic participants and carries the interaction consistency factor; based on the interaction scenario graph, the target posterior probabilities of various longitudinal interaction intentions are transmitted between multiple nodes to update the joint intention posterior distribution among multiple target traffic participants; based on the updated joint intention posterior distribution, the mutual yielding risk and mutual snatching risk between the vehicle and each target traffic participant, and between different target traffic participants, are determined, and the cascading interaction risk among multiple traffic participants is determined.

[0139] When there are multiple target traffic participants with longitudinal interaction relationships with the vehicle, electronic devices can construct an interaction scenario graph. Perform joint intent inference among multiple agents. Represents a set of nodes. This represents a set of directed edges. Each node corresponds to a traffic participant, which can include vehicles and multiple target traffic participants. Each node can carry the probability of various vertical interaction intentions corresponding to that traffic participant. Each directed edge connects pairs of traffic participants with interaction relationships and carries an interaction consistency factor. , used to characterize the The traffic participant and the first Variables of traffic participants in the current traffic scenario The rationality of different combinations of intentions.

[0140] Electronic devices can transmit various vertical interaction intent probabilities between multiple nodes based on an interaction scenario graph to update the joint posterior distribution of intent among multiple traffic participants. In one implementation, the electronic device can use belief propagation or cyclic belief propagation to pass messages in the interaction scenario graph. During message passing, the vertical interaction intent probability of a node can be propagated along directed edges to adjacent nodes, and the intent probabilities of adjacent nodes are updated by combining the interaction consistency factor on the directed edges.

[0141] Based on the updated joint intention posterior distribution, electronic devices can determine the mutual yielding and competing risks between the vehicle and each target traffic participant, and between different target traffic participants, as well as the cascading interaction risks among multiple traffic participants. Cascading interaction risks can be used to characterize the multi-hop interaction effects formed after a traffic participant's longitudinal interaction intention changes are transmitted to other traffic participants through interaction relationships. For example, a vehicle yielding leads to a following vehicle adjusting its course, and the following vehicle's adjustment further affects vehicles behind it.

[0142] Furthermore, electronic devices can also determine the cascading interaction risks among multiple traffic participants based on the interaction scenario graph. Cascading interaction risks can be used to characterize the multi-hop interaction impact formed after a change in the longitudinal interaction intention of one traffic participant is transmitted to other traffic participants through interaction relationships. For example, in multi-vehicle following or merging scenarios, the yielding of the vehicle in front may cause the following vehicle to need to adjust its following distance, and the adjustment of the following vehicle may affect the acceleration and deceleration behavior of vehicles further behind; another example is that vehicles rushing on the entrance ramp may cause vehicles on the main road to decelerate, and the deceleration of vehicles on the main road further affects the following strategy of vehicles behind them. Through the interaction scenario graph, electronic devices can analyze the above multi-hop interaction impacts along directed edges and determine the cascading interaction risks among multiple traffic participants.

[0143] In this way, electronic devices can not only identify the risk of social misunderstanding between their own vehicle and a single target traffic participant in a two-vehicle scenario, but also jointly infer the longitudinal interaction intentions of multiple traffic participants in a multi-target interaction scenario, thereby identifying the risks of mutual yielding, mutual snatching and chain interaction caused by multi-vehicle interaction.

[0144] Therefore, by constructing an interaction scenario graph in a multi-target interaction scenario and transmitting the longitudinal interaction intention probability of each traffic participant in the graph structure, it is possible to infer the joint intention relationship between the vehicle and multiple target traffic participants, thereby identifying the mutual yielding risk, mutual snatching risk and chain interaction risk among multiple vehicles, improving the completeness of social misunderstanding detection and the synergy of vehicle behavior decision-making in complex traffic scenarios.

[0145] When an electronic device detects that the intent confidence set contains multiple longitudinal interaction intents, the intent uncertainty index is high, the risk of yielding to each other is high, the risk of competing for something is high, or the risk of chained interaction is high, it can trigger an active disambiguation process. The active disambiguation process can be used to change the interaction state between the vehicle and the target traffic participant through low-risk probing actions, so that the target traffic participant exhibits a more explicit longitudinal behavioral response.

[0146] In some embodiments, the proactive disambiguation process includes: determining multiple candidate probing actions; the candidate probing actions include low-risk longitudinal actions used to change the interaction state between the vehicle and the target traffic participant; determining the information gain and conditional risk value corresponding to each candidate probing action based on the target posterior probability corresponding to various longitudinal interaction intentions; selecting candidate probing actions from the multiple candidate probing actions whose conditional risk value is less than or equal to a preset safety threshold, and determining the candidate probing action with the largest information gain from the selected candidate probing actions as the target probing action; controlling the vehicle to execute the target probing action, and acquiring new state information after executing the target probing action; redetermining the target posterior probability corresponding to various longitudinal interaction intentions based on the new state information, and generating an updated intention confidence set based on the redetermined target posterior probability.

[0147] Candidate probing actions can be low-risk longitudinal movements used to alter the interaction between the vehicle and the target traffic participant, such as slight acceleration, slight deceleration, maintaining the lane and adjusting the following distance, briefly maintaining the current speed, or gently releasing longitudinal space. Candidate probing actions are not intended to directly seize passage or force yielding, but rather to elicit the behavioral response of the target traffic participant within safety constraints.

[0148] Electronic devices can determine the information gain and conditional risk value for each candidate probing action based on the target posterior probability corresponding to various longitudinal interaction intentions. The information gain characterizes the degree to which the uncertainty of the longitudinal interaction intention is reduced after executing the candidate probing action; the conditional risk value characterizes the tail risk corresponding to executing the candidate probing action under the longitudinal interaction intention distribution of the target traffic participants.

[0149] In one implementation, the target probing action can be determined according to the following constraint optimization method:

[0150] And it satisfies:

[0151] in, Probing actions towards the target For the set of candidate trial actions, To provide information gain on the longitudinal interaction intent based on the observation information at the next time step after executing the candidate probing action. To execute candidate probing actions based on the longitudinal interaction intent distribution of target traffic participants. The corresponding collision risk or emergency risk, Confidence level Conditional Value at Risk (VaR) This is a preset safety threshold.

[0152] The electronic device can filter candidate trial actions from multiple candidates, selecting those with a conditional risk value less than or equal to a preset safety threshold. From these selected trial actions, it determines the one with the highest information gain as the target trial action. The electronic device then controls the vehicle to execute the target trial action and acquires new state information afterward. Based on this new state information, the electronic device can redetermine the target posterior probability corresponding to various longitudinal interaction intentions and generate an updated intention confidence set based on the redefined posterior probability, thus forming a closed loop of "disambiguation action - observation feedback - posterior update - confidence set update".

[0153] Therefore, by triggering an active disambiguation process when the intent is uncertain or the risk of misunderstanding is high, and selecting actions that meet the conditional risk value constraint from multiple candidate probing actions, it is possible to avoid probing actions being executed when the tail risk is high. Furthermore, by selecting the target probing action with the largest information gain from the candidate probing actions that meet the safety constraints, it is possible to encourage the target traffic participants to exhibit clearer behavioral responses while ensuring safety. The target posterior probability and intent confidence set are updated with new state information, thereby reducing the uncertainty of longitudinal interaction intent and improving the safety and reliability of vehicle behavior decisions.

[0154] In some embodiments, this application also includes online calibration and self-evaluation steps. Specifically: during vehicle operation, calibration statistics of the intent estimation result are obtained based on a preset sliding window; wherein, the calibration statistics include at least one of the following: the deviation between the target posterior probability and the actual longitudinal interaction intent for various types of longitudinal interaction intents, the intent estimation accuracy, and the intent confidence set coverage; calibration drift is detected based on the calibration statistics; if the calibration drift meets preset drift conditions, the prior parameters of the target intent prior distribution are recalibrated online, and / or the statistical quantile threshold used to generate the intent confidence set is updated; the calibration statistics and the corresponding parameter update results are recorded for maintenance or parameter updating of the vehicle longitudinal interaction intent estimation method.

[0155] Electronic devices can obtain calibration statistics of intent estimation results based on a preset sliding window. The calibration statistics can include at least one of the following: the deviation between the target posterior probability of various vertical interaction intents and the actual vertical interaction intents, the intent estimation accuracy, and the intent confidence set coverage.

[0156] Electronic devices can detect calibration drift based on calibration statistics. For example, when the average deviation between the target posterior probability and the actual longitudinal interaction intent within a preset sliding window exceeds a preset deviation threshold, or when the intent confidence set coverage is lower than a preset coverage threshold, it can be determined that the calibration drift meets a preset drift condition. Upon detecting that the calibration drift meets the preset drift condition, the electronic device can perform online recalibration of the prior parameters of the target intent prior distribution and / or update the statistical quantile thresholds used to generate the intent confidence set.

[0157] The electronic device can also record calibration statistics and corresponding parameter update results for subsequent system maintenance, model performance evaluation, parameter updates, or offline retraining.

[0158] Therefore, by continuously acquiring calibration statistics through a sliding window and detecting the deviation between the target posterior probability and the actual longitudinal behavior, estimation deviations can be detected in a timely manner. When the calibration drift exceeds a threshold, the prior parameters are recalibrated online or the statistical quantile threshold is updated to maintain the reliability of the posterior probability and the intention confidence set. At the same time, the statistics and parameter update results are recorded to provide data support for system maintenance and optimization.

[0159] In some embodiments, this application also provides a vehicle longitudinal interaction intent estimation system for performing the aforementioned method. The system can be deployed on an onboard computing platform (such as an autonomous driving domain controller, central computing unit, or driver assistance controller) as an independent functional service in the autonomous driving software stack, interacting with perception, mapping, and planning modules via onboard Ethernet, shared memory, or ROS2 interface.

[0160] The system includes the following main modules: Sensor and V2X interface module: Acquires data from cameras, millimeter-wave radar, lidar, IMU / wheel speed, and V2X communication messages.

[0161] Target detection and tracking fusion module: performs multi-sensor fusion and target tracking on raw data, and outputs the trajectory and state vector X of traffic participants. t (Including longitudinal relative distance, relative speed, longitudinal acceleration, lateral offset, lane number, etc.).

[0162] Scene recognition module: Based on map topology, vehicle route planning, road markings, and traffic sign information, outputs the current traffic scene variables. Information on conflict areas (such as ramp merging, lane narrowing, lane changing, etc.).

[0163] Prior generation module (including meta-learning adaptor): based on and target historical behavior summary Generate scene conditional priors And through meta-learning, it achieves rapid cross-domain adaptation.

[0164] Likelihood calculation module (including counterfactual verifier): based on multi-source observation features , and Calculate observational likelihood Furthermore, it combines counterfactual trajectory verification to improve intent differentiation.

[0165] Hierarchical Bayesian filtering module: outputs coarse-grained strategic intent posterior probability and fine-grained tactical behavior posterior probability (such as aggressive acceleration, gentle deceleration, etc.).

[0166] Conformal prediction calibration module: Performs calibration on the posterior probability to generate an intent confidence set with statistical coverage guarantee.

[0167] Joint Intent and Misunderstanding Detection Module: Combining the probability of the intention of the vehicle and the target, as well as the interaction consistency factor, it calculates the risk of mutual yielding, mutual snatching, and chain interaction, and performs joint intent inference in multi-target scenarios.

[0168] Active disambiguation and interface module (including CVaR constraint solver): When the triggering condition is met, select the action with the largest information gain from the low-risk candidate actions for execution, and output the final intent probability, confidence set and recommended interaction label.

[0169] Online calibration and self-evaluation module: Maintains calibration statistics based on a sliding window, detects drift and triggers prior parameter recalibration or confidence set threshold update.

[0170] For example, in an entrance ramp merging scenario, sensors collect data on entrance vehicles and main road vehicles; the target fusion module outputs the entrance vehicle state vector; the scene recognition module determines that the current traffic scenario is ramp merging; the prior generation module generates an intention prior distribution based on historical behavior; the likelihood module calculates the observation likelihood; the hierarchical Bayesian filtering module updates the coarse-grained posterior probability; the conformal calibration module generates a confidence set; the joint intention module assesses the risk of mutual preemption / yielding; and the proactive disambiguation module selects low-risk trial actions and outputs the final intention probability, confidence set, and recommended interaction tags.

[0171] Through the above system structure, the modules can work together to enable the vehicle to obtain probabilistic, calibrable, disambiguating and time-stable longitudinal interaction intent estimation results in complex longitudinal interaction scenarios.

[0172] To facilitate understanding of the vehicle longitudinal interaction intent estimation method, system composition, and collaborative relationships between functional modules provided in this application, the embodiments of this application are described below with reference to the accompanying drawings. It should be noted that the following drawings are only for illustrating the technical solutions of the embodiments of this application and do not constitute a limitation on the processing steps, the number of modules, the module connection method, or the execution order. In practical applications, modules can be merged, split, or deployed in parallel according to the software architecture of the in-vehicle computing platform, and each step can also be adaptively adjusted according to the vehicle operating scenario, sensor configuration, and decision-making and planning requirements.

[0173] like Figure 2 As shown, the vehicle interaction intent estimation system provided in this application embodiment may include ten functional modules from M1 to M10. Each module works together in accordance with the data flow relationship to complete longitudinal interaction intent estimation, confidence calibration, misunderstanding detection and active disambiguation processing.

[0174] The M1 sensor and V2X interface module is used to access vehicle sensor and V2X communication data, providing subsequent modules with raw observation data such as camera, radar, lidar, vehicle status, map, and communication messages. The M2 target detection and tracking fusion module performs target detection, target association, multi-sensor fusion, and target tracking on the data provided by M1, outputting the trajectory and status information of the target traffic participants. The M3 scene recognition module identifies current traffic scene variables based on road scene information and provides scene-related information to subsequent prior generation, misunderstanding detection, conformal prediction calibration, and online calibration modules. The M4 prior generation module may include meta-learning adaptation functionality, used to generate a scene-conditional target intent prior distribution based on current traffic scene variables and historical behavior summaries of target traffic participants. The M5 likelihood calculation module may include counterfactual verification functionality, used to calculate the observation likelihood corresponding to various longitudinal interaction intentions based on the multi-source observation characteristics and status information of target traffic participants, and enhance the distinguishability between different intentions through counterfactual trajectory verification. The target intent prior distribution output by M4 and the observation likelihood output by M5 are jointly input into the M6 ​​hierarchical Bayesian filtering module. The M6 ​​hierarchical Bayesian filtering module performs hierarchical Bayesian filtering updates to obtain the target posterior probability for the target traffic participant. The M9 conformal prediction calibration module performs conformal prediction calibration based on the Bayesian posterior probability, generating an intent confidence set. The M7 joint intent and misunderstanding detection module may include an interaction scene graph function, used to determine mutual yielding risk, mutual snatching risk, uncertainty indicators, or chain interaction risks based on the target posterior probability, the vehicle's intent probability, and interaction consistency factors. The M10 online calibration and self-evaluation module monitors the calibration performance during system operation and interacts with the conformal prediction calibration module and scene recognition module to support drift detection and parameter updates. The M8 active disambiguation and interface module may include CVaR risk constraint functionality, used to select a target probing action based on CVaR safety constraints and information gain when uncertainty or misunderstanding risks meet trigger conditions, and output the processing results to the decision planning, longitudinal control, risk assessment, or human-machine co-driving modules.

[0175] like Figure 3 As shown, the vehicle longitudinal interaction intent estimation method provided in this application embodiment may include a complete processing flow from S1 to S10.

[0176] In the S1 data acquisition and fusion phase, the electronic device acquires raw observation data of traffic participants around the vehicle from data sources such as cameras, radar, lidar, and V2X communication, and then performs fusion processing. Next, in the S2 interaction candidate selection phase, the electronic device can combine conflict areas and interaction intensity indicators to select target traffic participants with longitudinal interaction relationships with the vehicle from among the traffic participants around it. In the S3 scene recognition and context construction phase, the electronic device can identify the variables of the current traffic scene. This traffic scenario variable can be used to characterize traffic scenarios such as merging, narrowing, and lane cutting. After S3, the method flow can be divided into two paths: one is to enter the S4 intention hypothesis and prior adaptation stage, which combines meta-learning cross-domain adaptation mechanism to generate the scenario-conditional target intention prior distribution; the other is to enter the S5 multi-source observation feature extraction stage, which extracts observation features such as kinematic signals and interaction responses. The prior information output from S4 and the observation features output from S5 jointly enter the S6 likelihood model construction stage, where the electronic device can construct the observation likelihood by combining counterfactual trajectory verification factors. In the S7 hierarchical Bayesian filtering update stage, the electronic device can perform probability updates based on the coarse-grained intention layer and the fine-grained behavior layer to obtain the target posterior probability corresponding to various vertical interaction intentions. Then, in the S7+ conformal prediction calibration stage, the electronic device can generate an intention confidence set based on the intention confidence set and coverage guarantee. In the S8 joint intention and misunderstanding detection stage, the electronic device can identify the risk of social misunderstanding based on the interaction scenario graph or multi-agent inference. In the S9 CVaR constrained active disambiguation stage, the electronic device can determine the target probing action by combining information gain and tail risk constraints. Finally, during the S10 online calibration and self-evaluation phase, electronic devices can undergo drift detection and prior recalibration to maintain calibration performance during long-term system operation.

[0177] Figure 3 The solid arrow on the right indicates the feedback update process. Active disambiguation, online calibration, or new observation results can be fed back to the preceding processing stage, so that the system forms a closed loop of "observation-estimation-calibration-disambiguation-re-observation-re-update".

[0178] like Figure 4 As shown, the hierarchical Bayesian filtering model can include a Level-1 strategic intent layer and a Level-2 tactical behavior layer. The Level-1 strategic intent layer is used to represent coarse-grained intent states. The coarse-grained intent state can include the intent to preempt, the intent to yield, and the intent to hold, represented as R / Y / K in the diagram. The Level-2 tactical behavior layer is used to represent the fine-grained behavior state. This fine-grained behavioral state can include specific longitudinal motion behaviors such as acceleration, deceleration, and constant speed.

[0179] In the time dimension, Figure 4 It shows from , arrive The evolutionary relationship. In the coarse-grained layer, the previous time step... Probability of intent transfer Transfer to the current moment The current moment It can also be done through Transfer to the next moment This time transition relationship is used to express the persistence and hysteresis characteristics of vertical interaction intent.

[0180] At any given moment, the coarse-grained intention state can constrain the corresponding fine-grained behavioral state. For example, correspond , correspond , correspond Fine-grained behavioral states are further linked to multi-source observations. , , Related to the current traffic scenario , , It can serve as a scenario condition that influences the interpretation of observations and the updating of behavioral states. Figure 4 The conformal prediction calibration layer on the right is set after the coarse-grained posterior update and is used to generate an intent confidence set that meets the statistical coverage requirements based on Bayesian posterior probability.

[0181] like Figure 5 As shown, the typical longitudinal interaction time series of merging can be represented by three time series: relative distance, posterior probability of intent, and posterior entropy. Figure 5 The horizontal axis represents time, in seconds. The upper curve represents the vertical relative distance. It also provides a safe distance threshold. The longitudinal relative distance can first decrease and then increase over time, used to represent the approach and distance relationship between the target traffic participant and the vehicle during the merging process.

[0182] Figure 5 The central region represents the change in the posterior probability of intent over time, including , and In the example, initially, the intention to hold and the intention to cut in have certain probabilities; as the target traffic participant approaches the merging area, the probability of cutting in increases, while the probability of holding in decreases; near the intention transition point, the posterior probability of the intention changes significantly, and then the probability of yielding increases and tends to stabilize. The intention transition points marked by dashed lines in the figure are used to indicate the locations where the target traffic participant's main intention judgment changes.

[0183] Figure 5 The lower region represents the posterior entropy. The posterior entropy changes over time, and a disambiguation trigger threshold is set. When the posterior entropy is higher than the disambiguation trigger threshold, it can be considered that the current intent judgment has high uncertainty, and the corresponding area in the figure can be used as a candidate area to trigger active disambiguation. As new observation information is obtained or disambiguation actions are performed, the posterior entropy decreases, indicating that the intent judgment tends to be clearer. Figure 5It can be seen that electronic devices can comprehensively judge the changes in the longitudinal interaction intentions of target traffic participants during the merging process by considering relative distance, posterior probability of intent, and posterior entropy, and trigger active disambiguation in the high uncertainty interval.

[0184] like Figure 6 As shown, the active disambiguation logic can be executed based on Bayesian posterior probability, intent confidence set, misunderstanding index, and CVaR safety constraints. The electronic device can first obtain the Bayesian posterior probability. ,in , , Let represent the target posterior probabilities of the intention to cut in, the intention to yield, and the intention to hold back, respectively.

[0185] Electronic devices can determine posterior entropy Whether it exceeds the threshold, misinterpretation index The system checks whether the threshold is exceeded or whether the intent confidence set contains multiple candidate intents. If the result is no, the electronic device can output the confirmed intent and recommended tags. If the result is yes, the electronic device enters the active disambiguation process to generate a set of candidate trial actions. .

[0186] For each candidate trial action, the electronic device can perform an information gain evaluation, which can be expressed as: This is used to characterize the degree to which the uncertainty of intent is reduced after performing a candidate probing action. Simultaneously, electronic devices can also perform CVaR safety constraint judgments, i.e., determine... Whether it is valid or not.

[0187] After the electronic device selects a target and initiates a probe, it can control the vehicle to execute the probe. Obtain new observations And update the Bayesian posterior based on the new observations. Figure 6 The feedback arrows in the diagram indicate that the observations after active disambiguation can be fed back to the Bayesian posterior estimation process, thus forming a closed-loop update.

[0188] like Figure 7 As shown, in a multi-target interaction scenario, electronic devices can construct an interaction scenario graph to perform joint intent inference among multiple agents. The interaction scenario graph can include a vehicle node (Ego) and multiple target traffic participant nodes, such as target vehicle 1, target vehicle 2, target vehicle 3, and target vehicle 4. Each target node can carry the vertical interaction intent probability of that target traffic participant; for example, target vehicle 1 carries... , Target vehicle 2 carries , Target vehicle 3 carries Target vehicle 4 carries , Each vehicle node can carry the probability of its current planned behavior.

[0189] Figure 7 The edges in the diagram represent pairs of traffic participants with interactive relationships. The labels on the edges are... , , , as well as , It can represent the interaction consistency factor, used to characterize the variables of two traffic participants in the current traffic scenario. The rationality of different combinations of intentions. Different edges in the diagram can correspond to interaction relationships of different intensities or types.

[0190] Electronic devices can perform joint intent inference on the graph using BeliefPropagation. Specifically, each node can transmit intent probability messages to neighboring nodes along the edges and update the posterior distribution of the joint intent by combining the interaction consistency factors on the edges. In this way, electronic devices can detect mutual yielding patterns, mutual cutting-off patterns, and multi-vehicle chain interaction risks. For example, if target vehicle 1 has a high probability of cutting off the other vehicle, and the other vehicle also has a high probability of cutting off the other vehicle, a mutual cutting-off risk may occur; if multiple vehicles have a high probability of yielding, a mutual yielding risk may occur; if a vehicle in front yields, causing a vehicle behind to need to adjust, and further affecting vehicles further behind, a chain interaction risk may occur.

[0191] like Figure 8 As shown, conformal prediction calibration and intention confidence set can be illustrated using reliability plots and examples of intention confidence set outputs. Figure 8 Figure (a) shows the reliability graphs before and after calibration. The horizontal axis represents the predicted probability, the vertical axis represents the actual frequency, and the dashed line represents the ideal calibration relationship. The red bars represent the reliability results before calibration, and the green bars represent the reliability results after conformal prediction calibration. It can be seen that the probability output before calibration deviates from the ideal calibration line, while after conformal prediction calibration, the predicted probability is closer to the actual frequency, indicating that the calibrability of the intended probability output has been improved.

[0192] Figure 8 Figure (b) shows an example of the intent confidence set output. The horizontal axis represents different scenarios, the vertical axis represents the posterior probability, and the different colored bars correspond to the intent to rush ahead. Intention to yield and maintain intention In scenario 1, the probability of an intention to cut in line is significantly higher than that of an intention to yield or to hold on, and the electronic device can output an intention confidence set. This scenario corresponds to determining the intent. In scenario 2, both the intent to cut in and the intent to yield are relatively high, and the electronic device can output an intent confidence set { This scenario corresponds to conflicting or uncertain intentions. In scenario 3, the posterior probabilities of the intentions to cut in, yield, and hold are relatively close, and the electronic device can output an intention confidence set { This scenario corresponds to a highly uncertain state.

[0193] pass Figure 8 As can be seen, the conformal prediction calibration module can generate an intent confidence set based on the Bayesian posterior probability, so that the output not only includes the single intent with the highest probability, but also retains multiple candidate intents under uncertain conditions, thereby providing intent estimation results with statistical coverage guarantee for vehicle behavior decision-making.

[0194] The vehicle longitudinal interaction intent estimation method provided in this application can be executed by a vehicle longitudinal interaction intent estimation device. This application uses the vehicle longitudinal interaction intent estimation device executing the method as an example to illustrate the vehicle longitudinal interaction intent estimation device provided in this application.

[0195] like Figure 9 As shown, the vehicle longitudinal interaction intent estimation device 900 includes: a data fusion module 901, a target screening module 902, a priori construction module 903, a likelihood determination module 904, a posterior determination module 905, and a confidence set generation module 906.

[0196] The data fusion module 901 is used to acquire the raw observation data of traffic participants around the vehicle and to fuse the raw observation data to obtain the status information of each traffic participant. The target filtering module 902 is used to filter target traffic participants with longitudinal interaction relationships with the vehicle from traffic participants around the vehicle. Prior construction module 903 is used to construct a prior distribution of target intent that matches the current traffic scenario based on the historical longitudinal behavior data of the target traffic participants; The likelihood determination module 904 is used to determine the observation likelihood for various longitudinal interaction intentions based on the state information of the target traffic participants; longitudinal interaction intentions include at least the intention to cut in, the intention to give way, and the intention to hold on. The posterior determination module 905 is used to determine the target posterior probability corresponding to various vertical interaction intentions based on the target intention prior distribution and observation likelihood, using a hierarchical Bayesian model. The confidence set generation module 906 is used to generate an intent confidence set based on the target posterior probability corresponding to various vertical interaction intents for vehicle behavior decision-making.

[0197] The vehicle longitudinal interaction intent estimation device provided in this application embodiment can implement all the processes implemented in the above-described vehicle longitudinal interaction intent estimation method embodiment and achieve the same technical effect. To avoid repetition, it will not be described again here. The vehicle longitudinal interaction intent estimation device in this application embodiment can be an electronic device or a component in an electronic device, such as an integrated circuit or a chip.

[0198] In some embodiments, such as Figure 10 As shown, this application embodiment also provides an electronic device 1000, including a processor 1001, a memory 1002, and a computer program stored in the memory 1002 and executable on the processor 1001. When the program is executed by the processor 1001, it implements the various processes of the above-described vehicle longitudinal interaction intention estimation method embodiment and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0199] This application provides a non-transitory computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described vehicle longitudinal interaction intent estimation method embodiment and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0200] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-storable media, such as computer read-only memory (ROM), random-access memory (RAM), magnetic disks, or optical disks.

[0201] The computer-readable storage medium may include: read-only memory (ROM), random-access memory (RAM), magnetic disk or optical disk, etc.

[0202] This application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described vehicle longitudinal interaction intent estimation method.

[0203] This application provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described vehicle longitudinal interaction intent estimation method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0204] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0205] Although embodiments of this application have been shown and described, it is understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for estimating longitudinal interaction intent in vehicles, characterized in that, include: The original observation data of traffic participants around the vehicle are acquired, and the original observation data are fused to obtain the status information of each traffic participant. Target traffic participants with longitudinal interaction relationships with the vehicle are selected from the traffic participants around the vehicle. Based on the historical longitudinal behavior data of the target traffic participants, a prior distribution of target intent matching the current traffic scenario is constructed. Based on the state information of the target traffic participants, the observation likelihood is determined for various longitudinal interaction intentions; the longitudinal interaction intentions include at least the intention to cut in, the intention to give way, and the intention to keep. Based on the prior distribution of the target intent and the observation likelihood, the target posterior probability corresponding to various vertical interaction intents is determined by a hierarchical Bayesian model. Based on the target posterior probability corresponding to various vertical interaction intentions, an intention confidence set is generated for vehicle behavior decision-making.

2. The vehicle longitudinal interaction intent estimation method according to claim 1, characterized in that, The step of selecting target traffic participants with longitudinal interaction relationships with the vehicle from among traffic participants around the vehicle includes: Based on road topology information and potential conflict areas, identify one or more candidate traffic participants from traffic participants around the vehicle. Based on the state information of each candidate traffic participant, the interaction intensity index between each candidate traffic participant and the vehicle is determined; wherein, the interaction intensity index is determined based on at least one of collision time, time difference to potential conflict area, and degree of overlap of potential conflict areas. Based on the interaction intensity index corresponding to each candidate traffic participant, target traffic participants with longitudinal interaction relationships with the vehicle are determined from the candidate traffic participants.

3. The vehicle longitudinal interaction intent estimation method according to claim 1, characterized in that, The step of constructing a prior distribution of target intent that matches the current traffic scenario based on the historical longitudinal behavior data of the target traffic participants includes: Based on the road scene information of the road where the vehicle is located, determine the current traffic scene variables; Based on the historical longitudinal behavior data of the target traffic participants, generate the initial intent prior distribution corresponding to various longitudinal interaction intents; Based on the current traffic scenario variables, the initial intention prior distribution is conditionally adjusted to obtain a target intention prior distribution that matches the current traffic scenario; wherein, the target intention prior distribution is updated online as the longitudinal behavior data of the target traffic participants are continuously acquired during vehicle operation.

4. The vehicle longitudinal interaction intent estimation method according to claim 3, characterized in that, The method further includes: Obtain initial prior parameters; wherein, the initial prior parameters are obtained by meta-learning training of prior parameters using an labeled longitudinal behavior dataset containing multiple road scenes in the offline stage; During vehicle operation, the driving style information of the target traffic participants is updated based on the continuously acquired longitudinal behavior data of the target traffic participants. Based on the driving style information, the initial prior parameters are updated online to obtain target prior parameters that match the driving style of the target traffic participant; The target intent prior distribution is updated online based on the target prior parameters to obtain the updated target intent prior distribution.

5. The vehicle longitudinal interaction intent estimation method according to claim 3 or 4, characterized in that, The determination of current traffic scene variables based on road scene information of the road where the vehicle is located includes: Based on the road scene information of the road where the vehicle is located, determine the positional relationship between the target traffic participant and the potential conflict area, the lane change state of the lane where the target traffic participant is located, and the lateral movement state of the target traffic participant relative to the lane line. Based on the positional relationship, the lane change state, and the lateral movement state, the current traffic scenario variables are determined; wherein, the current traffic scenario variables are used at least to distinguish between ramp merging scenarios, lane narrowing scenarios, and lane changing or cutting-in scenarios.

6. The vehicle longitudinal interaction intent estimation method according to claim 1, characterized in that, The determination of the observation likelihood for various longitudinal interaction intentions based on the state information of the target traffic participants includes: Based on the state information of the target traffic participants, longitudinal motion features, lateral motion features, explicit signal features, and interactive response features are extracted. Based on the longitudinal motion features, the lateral motion features, the explicit signal features, and the interactive response features, an observation conditional probability model corresponding to various longitudinal interactive intentions is constructed. Based on the observation conditional probability model, the observation likelihood corresponding to each type of longitudinal interaction intent is determined.

7. The vehicle longitudinal interaction intent estimation method according to claim 1, characterized in that, The hierarchical Bayesian model is divided into a coarse-grained layer and a fine-grained layer; the determination of the target posterior probability corresponding to various vertical interaction intentions based on the target intention prior distribution and the observation likelihood using the hierarchical Bayesian model includes: In the coarse-grained layer, various vertical interaction intentions are treated as coarse-grained intention states, and an intention transition probability matrix between each coarse-grained intention state is constructed based on a Markov chain. Based on the prior distribution of the target intent, the observation likelihood, and the intent transition probability matrix, Bayesian filtering is performed on each coarse-grained intent state to obtain the target posterior probability corresponding to each type of vertical interaction intent.

8. The vehicle longitudinal interaction intent estimation method according to claim 7, characterized in that, The method further includes: In the fine-grained layer, the longitudinal movement behavior of the target traffic participant is divided into multiple fine-grained behavioral states; Based on the coarse-grained posterior probability and the current traffic scenario variables, the posterior probability of each fine-grained behavioral state is updated to obtain the fine-grained posterior probability. The longitudinal motion behavior prediction result of the target traffic participant is determined based on the fine-grained posterior probability, and the longitudinal motion behavior prediction result is used for vehicle behavior decision-making.

9. The vehicle longitudinal interaction intent estimation method according to claim 1 or 7, characterized in that, The generation of intent confidence sets based on the target posterior probabilities corresponding to various vertical interaction intents includes: Obtain the statistical quantile threshold determined based on the calibration dataset during the offline phase; the statistical quantile threshold is determined based on a preset statistical coverage and the inconsistency score corresponding to each sample in the calibration dataset, and the inconsistency score is used to characterize the degree of inconsistency between the posterior probability of the longitudinal interaction intent corresponding to the sample and the true longitudinal interaction intent. During the online deployment phase, the current inconsistency score corresponding to each type of vertical interaction intent is determined based on the target posterior probability corresponding to each type of vertical interaction intent. Vertical interaction intentions with a current inconsistency score less than or equal to the statistical quantile threshold are identified as target vertical interaction intentions, and the target vertical interaction intentions constitute an intention confidence set; wherein, the statistical quantile threshold is configured such that the probability of the intention confidence set covering the true vertical interaction intention is not less than a preset statistical coverage rate.

10. The vehicle longitudinal interaction intent estimation method according to claim 1, characterized in that, The method further includes: Based on the target posterior probability corresponding to various vertical interaction intentions, the intention uncertainty index of the target traffic participant is determined; Determine the probability distribution of the vehicle's corresponding longitudinal interaction intent based on the vehicle's current planned behavior; Based on the target posterior probability of the target traffic participants, the probability distribution of the longitudinal interaction intention of the vehicle, and the interaction consistency factor, the mutual yielding risk, mutual robbery risk, and chain interaction risk are determined. When the intent confidence set contains multiple vertical interaction intents, the mutual concession risk meets the first preset risk condition, the mutual robbery risk meets the second preset risk condition, the chain interaction risk meets the third preset risk condition, or the intent uncertainty index meets the preset uncertainty condition, the active disambiguation process is triggered.

11. The vehicle longitudinal interaction intent estimation method according to claim 10, characterized in that, The determination of mutual yielding risk, mutual robbery risk, and chain interaction risk based on the target posterior probability corresponding to the target traffic participant, the longitudinal interaction intention probability distribution corresponding to the vehicle, and the interaction consistency factor includes: Construct an interaction scenario graph; wherein the interaction scenario graph includes multiple nodes and directed edges connecting the nodes, each node corresponds to a target traffic participant and carries the target posterior probability of the target traffic participant corresponding to various vertical interaction intentions, and each directed edge corresponds to the interaction relationship between two target traffic participants and carries an interaction consistency factor. Based on the interaction scenario graph, the target posterior probabilities of various vertical interaction intentions are transmitted between multiple nodes to update the joint intention posterior distribution among multiple target traffic participants. Based on the updated joint intent posterior distribution, the mutual yielding risk and mutual snatching risk between the vehicle and each target traffic participant, and between different target traffic participants, are determined, as well as the chain interaction risk between multiple traffic participants.

12. The vehicle longitudinal interaction intent estimation method according to claim 10 or 11, characterized in that, The active disambiguation process includes: Multiple candidate probing actions are identified; the candidate probing actions include low-risk longitudinal actions used to change the interaction state between the vehicle and the target traffic participant; Based on the target posterior probability corresponding to various vertical interaction intentions, the information gain and conditional risk value corresponding to each candidate probing action are determined respectively. From the multiple candidate probing actions, select candidate probing actions whose conditional risk value is less than or equal to a preset safety threshold, and from the selected candidate probing actions, determine the candidate probing action with the greatest information gain as the target probing action. The autonomous vehicle is controlled to perform the target probing action, and new state information is obtained after the target probing action is performed; Based on the new state information, the target posterior probabilities corresponding to various vertical interaction intentions are redefined, and an updated intention confidence set is generated based on the redefined target posterior probabilities.

13. The vehicle longitudinal interaction intent estimation method according to claim 1, characterized in that, The method further includes: During vehicle operation, calibration statistics of intent estimation results are obtained based on a preset sliding window; wherein, the calibration statistics include at least one of the following: the deviation between the target posterior probability of various longitudinal interaction intents and the actual longitudinal interaction intents, intent estimation accuracy, and intent confidence set coverage. Calibration drift is detected based on the aforementioned calibration statistics; If the calibration drift meets the preset drift conditions, the prior parameters of the target intent prior distribution are recalibrated online, and / or the statistical quantile threshold used to generate the intent confidence set is updated; The calibration statistics and corresponding parameter update results are recorded for use in the maintenance or parameter update of the vehicle longitudinal interaction intent estimation method.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the vehicle longitudinal interaction intent estimation method as described in any one of claims 1-13.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the vehicle longitudinal interaction intent estimation method as described in any one of claims 1-13.

Images