A target vehicle trajectory prediction method based on double-layer lane intention inference

Abstract

The present application relates to the technical field of unmanned vehicle under structured road intersection area, and particularly relates to a target vehicle trajectory prediction method based on double-layer lane intention inference, which comprises the following steps: obtaining a short-term prediction trajectory of a target vehicle based on an extended Kalman filter; analyzing the short-term prediction trajectory based on a path similarity scoring model to obtain a road prediction value selected by the target vehicle when entering an intersection; constructing a lane selection intention prediction model based on an interactive multi-model based on lateral deviation of the target vehicle, the lane selection intention prediction model obtaining a lane prediction value based on the road prediction value; generating a prediction trajectory of the target vehicle by using a quintic polynomial curve method based on the lane prediction value; and the present application can predict the driving trajectory of other driving vehicles with high precision under intersection scenarios.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle technology in structured road intersection areas, specifically to a target vehicle trajectory prediction method based on dual-lane intent inference. Background Technology

[0002] The realization of autonomous driving has a significant impact on improving traffic safety, increasing traffic efficiency, and optimizing the travel experience. Among these advancements, accurate prediction of the trajectories of surrounding vehicles is crucial for autonomous driving systems to promptly identify potential traffic conflicts and effectively plan safe routes. This is a key aspect of the commercialization and widespread application of autonomous driving technology, and is of great significance for accelerating the development of intelligent connected vehicles and promoting the industrial application of autonomous driving technology. Target vehicle trajectory prediction algorithms aim to predict the trajectories of other human-driven vehicles in structured road scenarios, especially urban intersections, providing accurate predicted trajectory information for autonomous vehicle traffic decisions and obstacle avoidance technologies. Similar technologies are commonly found in vehicle trajectory prediction and vehicle intention prediction technologies in scenarios such as urban road intersections, highways, and closed parks.

[0003] Chinese invention patent applications CN116933005A, CN116758727A, and CN117408406A disclose a method for predicting vehicle trajectories based on graph convolutional neural networks in a connected vehicle environment. CN116758727A discloses a method for predicting target vehicle trajectories in multi-lane scenarios. CN117408406A discloses a planning-coupled method and system for predicting multi-target vehicle trajectories. These solutions share the common feature of employing deep learning methods to utilize historical vehicle trajectory information to address the problem of vehicle trajectory prediction in dynamic scenarios. Whether using graph convolutional neural networks, social convolutional pooling networks, or long short-term memory neural networks (LSTM), all require feature encoding of historical vehicle trajectories, planned trajectories, and interactions between vehicles. These features are preprocessed and then trained on a neural network model using a large amount of data. Because these methods learn from historical trajectory data using neural networks, the real-time prediction results in real-world traffic environments may differ from experimental results. They are more suitable for specific traffic scenarios or road conditions and may have limitations in complex and ever-changing real-world traffic environments. Summary of the Invention

[0004] In view of the above problems, the present invention provides a target vehicle trajectory prediction method based on dual-lane intent inference, which solves the technical problem that the real-time prediction effect of the prior art is limited in real traffic environment.

[0005] This invention provides a target vehicle trajectory prediction method based on dual-lane intent inference, comprising the following steps:

[0006] Step S1: Based on extended Kalman filter, obtain the short-term predicted trajectory of the target vehicle; based on the path similarity scoring model, analyze the short-term predicted trajectory to obtain the predicted road value selected by the target vehicle when entering the intersection.

[0007] Step S2: Based on the lateral offset of the target vehicle, construct an interactive multi-model-based lane selection intention prediction model. The lane selection intention prediction model obtains lane prediction values ​​based on the road prediction values.

[0008] Step S3: Based on the lane prediction value, generate the predicted trajectory of the target vehicle using the fifth-order polynomial curve method.

[0009] Preferably, step S1 specifically includes:

[0010] Step S1-1: Search for a set of candidate lanes that the target vehicle can reach;

[0011] Step S1-2: Predict the short-time local state variables of the target vehicle and predict the short-term trajectory of the target vehicle based on the extended Kalman filter.

[0012] Steps S1-3: Establish a path similarity scoring model to evaluate the similarity between the short-term predicted trajectory of the target vehicle and each lane in the candidate lane set;

[0013] Steps S1-4: Determine the road that the target vehicle is most likely to choose based on the similarity, and use it as the predicted road value.

[0014] Preferably, step S1-1 includes: based on the real-time status information of the target vehicle, a depth-first search algorithm is used to search for a set of candidate lanes for the target vehicle from the high-definition map. The set of candidate lanes contains multiple candidate lanes, representing all possible choices for the target vehicle to drive to lanes in other roads at the intersection.

[0015] Preferably, step S1-2 includes: step S1-2-1, predicting the short-time local state variables of the target vehicle; step S1-2-2, predicting the short-time trajectory based on the extended Kalman filter, using the following methods: (1) establishing a process model, (2) establishing a measurement model, (3) performing prediction using the extended Kalman filter, and (4) updating the extended Kalman filter.

[0016] Preferably, steps S1-3 specifically include: calculating the similarity score between the short-term predicted trajectory of the target vehicle and each lane in the candidate lane set in the following manner:

[0017]

[0018] in, It is the i-th lane and the short-term predicted trajectory The similarity score between them, where k1 and k2 are weighted coefficients for positional similarity and yaw angle similarity. and These are positional similarity and yaw angle similarity, respectively.

[0019] Location similarity The calculation method is as follows:

[0020]

[0021] in, x is the mean of the Euclidean distance, α is the index of the short-term predicted trajectory point, and x is the mean of the Euclidean distance. α y α Let x′ be the coordinates of the short-term predicted trajectory point. α y′ α is the coordinates of the nearest point in the short-term predicted trajectory, m is the width of the short-term future time window, and L is the mapping threshold parameter;

[0022] Yaw angle similarity The calculation method is as follows:

[0023]

[0024] in, The mean difference in yaw angles, α is the index of short-term predicted trajectory points, and θ is the mean difference in yaw angles. α For the short-term predicted trajectory point yaw angle, θ′ α This is for the heading in the lane corresponding to the short-term predicted trajectory point.

[0025] Preferably, steps S1-4 include: using the softmax function to process the candidate lane set. similarity score Normalization is performed to obtain the probability that the target vehicle selects different lanes.

[0026]

[0027] Where n3 is the total number of all candidate lanes;

[0028] The probability of a vehicle choosing each road is obtained by summing the probabilities of each lane in a lane group on the same road, and the road with the highest probability is taken as the predicted road value.

[0029] Preferably, step S2 specifically includes:

[0030] Step S2-1: Construct a lane selection intent prediction model based on interactive multi-model. Based on the lateral offset and lateral offset speed of the target vehicle, obtain the lane with the highest probability within the predicted lane value as the target lane.

[0031] Step S2-2: Calculate the aiming distance, and obtain the anchor point of the centerline of the target lane based on the aiming distance, as the lane prediction value.

[0032] Preferably, step S2-1 specifically includes: (1) constructing a state transition probability matrix, (2) performing interactive part calculation of the interactive multi-model, (3) performing filtering part calculation of the interactive multi-model, (4) performing lane model probability update of the interactive multi-model, and (5) selecting the lane with the highest posterior lane model probability as the target lane from the lanes within the road prediction value.

[0033] Preferably, step S2-2 specifically includes:

[0034] Calculate the aiming distance:

[0035] dis=v k T predict

[0036] Where T predict The aiming time is used as a reference to query the centerline anchor point on the target lane and obtain the status P of the anchor point. T (x T y T ,θ T ) is the lane prediction value, where x T g T θ T These are the horizontal and vertical coordinates of the anchor point, and the yaw angle, respectively.

[0037] Preferably, the following fifth-order polynomial trajectory curve interpolation formula is established:

[0038] x(t) = a5t 5 +a4t 4 +a3t 3 +a2t 2 +a1t+a0

[0039] y(t)=b5t 5 +b4t 4 +b3t 3 +b2t 2 +b1t+b0

[0040] Where t is time, a0, ..., a5, b0, ..., b5 are polynomial trajectory parameters, and x(t) and y(t) are the coordinates of the trajectory curve; the starting and ending constraints of the fifth-order polynomial trajectory curve are:

[0041] x(0)=x k y(0)=y k

[0042] x′(0)=v k cosθ k y′(0)=v k sinθ k

[0043] x″(0)=a k cosθ k y″(0)=a k sinθ k

[0044] x(T predict )=x T y(T predict )=y T

[0045] x′(T predict ) = v T cosθ T y′(T predict ) = v T cosθ T

[0046] x″(T predict )=0 y″(T predict ) = 0

[0047] Where, x k y k θ k v k a k The values ​​of x represent the target vehicle's position coordinates, yaw angle, longitudinal velocity, and longitudinal acceleration at time k. T y T θ T v T a T These are the position coordinates, yaw angle, longitudinal velocity, and longitudinal acceleration of the lane prediction value, respectively, and v T =v k a T =a k x′(·) and x″(·) are the first and second derivatives of the abscissa of the trajectory curve, respectively, and y′(·) and y″(·) are the first and second derivatives of the ordinate of the trajectory curve, respectively.

[0048] Based on the starting and ending point constraints, the fifth-order polynomial trajectory parameters are solved, and the interpolation result of the fifth-order polynomial trajectory curve is used as the predicted trajectory of the target vehicle.

[0049] Compared with the prior art, the present invention has at least the following beneficial effects:

[0050] (1) By combining the dynamic characteristics and real-time driving status of the target vehicle, this invention can more accurately predict the short-term trajectory and lane selection intention of the vehicle. The introduction of a prediction and path similarity scoring model based on vehicle dynamics can improve the prediction accuracy of the actual driving path of the vehicle, thereby providing more reliable data support for subsequent traffic management and autonomous driving system decision-making.

[0051] (2) The lane selection intent prediction model constructed by the present invention using the interactive multi-model method can effectively handle the variability and uncertainty of vehicle behavior. The interactive multi-model method adapts to the changes in target vehicle behavior by combining the estimates of multiple models, making the prediction results more flexible to adapt to the complex situations in actual driving.

[0052] (3) The present invention uses the fifth-order polynomial curve method to generate the predicted trajectory of the target vehicle, which not only ensures the smoothness of the trajectory and reduces the unsafe factors caused by sudden changes in the predicted trajectory, but also generates a trajectory that is more in line with the characteristics of actual vehicle driving and can adapt to different driving and road conditions. Attached Figure Description

[0053] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0054] Figure 1 This is a flowchart of the target vehicle trajectory prediction method based on dual-lane intent inference disclosed in an embodiment of the present invention;

[0055] Figure 2 This invention discloses a schematic diagram of obtaining candidate lanes and roads for a target vehicle in an embodiment of the invention.

[0056] Figure 3 This is a schematic diagram of the kinematics of the target vehicle disclosed in an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the target vehicle path similarity scoring results disclosed in an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram illustrating the lane selection intention prediction of a target vehicle on a road, as disclosed in an embodiment of the present invention. Detailed Implementation

[0059] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0060] This invention provides a target vehicle trajectory prediction algorithm based on dual-layer lane intention inference. It primarily relies on lane and vehicle pose information at structured road intersections, using a dynamically based accurate short-term predicted trajectory as input. Through a designed road similarity evaluation model, it infers the vehicle's road selection intention when entering the intersection. Subsequently, based on an interactive multi-model framework using a multi-lane model, it infers the vehicle's specific lane selection intention on different roads. Finally, a fifth-order polynomial method is used to quickly generate the trajectory for the lane with the highest prediction probability.

[0061] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment. For example, a specific implementation of this invention... Figure 1 As shown, a target vehicle trajectory prediction method based on dual-lane intent inference is disclosed, including the following steps:

[0062] Step S1: Based on extended Kalman filter, obtain the short-term predicted trajectory of the target vehicle; based on the path similarity scoring model, analyze the short-term predicted trajectory to obtain the predicted road value selected by the target vehicle when entering the intersection.

[0063] Step S1-1: Search for a set of candidate lanes that the target vehicle can reach. For example... Figure 2 As shown, when a vehicle approaches an intersection, going straight, turning right, or turning left will lead to the corresponding road, and each road has multiple lanes. In this step, based on the real-time status information of the target vehicle, a depth-first search algorithm searches for a set of candidate lanes for the target vehicle from a high-resolution map. The candidate lane set contains multiple candidate lanes, representing all possible lane choices that a target vehicle at an intersection can use to reach other roads. The road set is represented as follows:

[0064]

[0065] in, Let n be the set of all roads, and n1 be the total number of roads. For roads, for The set of all lanes on the road, n2 is The total number of lanes on the road, Let n be the set of candidate lanes, and n3 be the total number of all candidate lanes.

[0066] Then, based on the differences in traffic flow at the intersection, a candidate lane set is formed. It is divided into several clearly defined lane groups, corresponding to semantic behaviors such as left turn, right turn, and straight ahead.

[0067] Step S1-2: Predict the short-time trajectory of the target vehicle based on the kinematic model. This mainly includes short-time local state variable prediction and short-time trajectory prediction based on extended Kalman filter.

[0068] Step S1-2-1: Prediction of short-time local state variables. For example... Figure 3 When the target vehicle moves, its position coordinates are (x... k y k The yaw angle is the angle θ of counterclockwise rotation with the positive x-axis as the reference. k The longitudinal velocity is v k The longitudinal acceleration is a k The yaw rate is γ k Yaw angular acceleration is

[0069] Based on the longitudinal velocity and yaw rate of the target vehicle in a short-term historical time window, the moving average method is used to predict the longitudinal velocity and yaw rate of the target vehicle in a short-term future time window:

[0070]

[0071] in, Let the longitudinal velocity and yaw rate be the values ​​at time k+m in the future time window. Let m be the longitudinal velocity and yaw rate at time k+1 in the future time window, m be the width of the short-term future time window, n be the width of the short-term historical time window, a and b be the parameter indices, and w be the yaw rate. a v is the weighting coefficient. k-a+1 γ k-a+1 The longitudinal velocity and yaw rate are given at time k-a+1 within the short-term historical time window.

[0072] In some embodiments, the short-term historical time window is set to 1 second (n=10), the short-term future time window is set to 1 second (m=10), and the sampling period T is 0.1 seconds.

[0073] Step S1-2-2: Predict short-time trajectory based on extended Kalman filter. Construct a state estimation model, using the longitudinal velocity and yaw rate of the short-term future time window as observations, and perform extended Kalman filtering on other state variables to obtain the other state variables (x) of the short-term future time window.k y k θ k v k γ k ), where based on (x) k y k θ k This can represent the short-term future trajectory of the target vehicle. Extended Kalman filtering consists of four parts: establishing a process model, establishing a measurement model, prediction, and updating.

[0074] (1) Establish a process model:

[0075] x k+1 =f(x) k )+q k

[0076] = [f1, f2, f3, f4, f5] + q k

[0077] Where, x k =(x k y k θ k v k γ k ) is the state vector of the target vehicle, q k ~N(0,Q) k The noise is a process noise that has a zero mean and a power spectral density equal to the process noise matrix Q. k The normal distribution, where Q k =diag{0, 0, q dθ ,q dv q dγ}, q dθ q represents the noise intensity during the yaw angle process. dv q represents the noise intensity during the longitudinal velocity process. dγ This indicates the noise intensity during the yaw rate process.

[0078] f(x k The state transition equation is based on vehicle kinematics and is calculated as follows:

[0079] f1:x k+1 =x k +(v k cosθ k )T

[0080] f2:y k+1 =y k +(v k sinθ k )T

[0081] f3: θk+1 =θ k +γ k T

[0082] f4:v k+1 =v k

[0083] f5:γ k+1 =γ k

[0084] Where T is the sampling period.

[0085] (2) Establish a measurement model:

[0086]

[0087] Where H is the measurement matrix, r k ~N(0, R) k The noise is a measurement noise that has a mean of zero and a noise covariance matrix of R. k The normal distribution, R k =diag{r v r γ}, r v The noise intensity of the longitudinal velocity measurement, r γ This indicates the noise intensity of the yaw rate measurement.

[0088] The measurement model takes into account that the estimates of speed and yaw rate become increasingly inaccurate over time, and therefore sets the variance of speed and yaw rate to increase linearly.

[0089] (3) Perform the prediction step f of the extended Kalman filter. ekf，predict The calculation process is as follows:

[0090]

[0091] in, It is the prior estimated state at time k. It is the posterior estimated state at time k-1, and f(·) is the state transition equation. It is the prior covariance matrix at time k. It is the posterior covariance matrix at time k-1. It is the Jacobian matrix at time k-1, Q k-1 It is the process noise matrix at time k-1.

[0092] (4) Perform the update step f of the extended Kalman filter. ekf，update The calculation process is as follows:

[0093]

[0094] Among them, K k It is the Kalman gain at time k.

[0095] Repeat the update steps of the extended Kalman filter to finally obtain the short-term predicted trajectory.

[0096] Steps S1-3: Establish a path similarity scoring model. To determine the semantic behavior of the target vehicle, such as turning left, turning right, and going straight, this invention infers a candidate lane set based on the lane similarity scoring model. Similarity score for all lanes:

[0097]

[0098] in, It is the i-th lane and the short-term predicted trajectory The similarity score between them, k1 and k2 are the weighting coefficients for positional similarity and yaw angle similarity, k1+k2=1; and These are positional similarity and yaw angle similarity, respectively.

[0099] To determine the positional similarity between the path and the short-term predicted trajectory, m short-term predicted trajectories are calculated. The mean Euclidean distance between a point on the lane and the nearest point on the lane is calculated, and the result is mapped from [0, L] and [L, ∞] to the range [0, 1]. This yields the positional similarity, calculated as follows:

[0100]

[0101] in, x is the mean of the Euclidean distance, α is the index of the short-term predicted trajectory point, and x is the mean of the Euclidean distance. α y α Let x′ be the coordinates of the short-term predicted trajectory point. α y′ α is the coordinates of the nearest point in the short-term predicted trajectory, m is the width of the short-term future time window, and L is the mapping threshold parameter.

[0102] To determine the yaw similarity between the path and the short-term predicted trajectory, m short-term predicted trajectories are calculated. The average difference between a point on the track and its heading in the lane is used to map the result from [0, 2π) to the range [0, 1], thus obtaining the positional similarity. The calculation method is as follows:

[0103]

[0104] in, The average difference in yaw angle is given by α, where α is the index of the short-term predicted trajectory point, θα is the yaw angle of the short-term predicted trajectory point, and θ′α is the heading in the lane corresponding to the short-term predicted trajectory point.

[0105] Steps S1-4: Classify the candidate lane set into different roads and calculate the road with the highest probability. Use the softmax function to process the candidate lane set. similarity score Normalization is performed to obtain the probability that the target vehicle chooses different lanes.

[0106]

[0107] Finally, by summing the probabilities of each lane in the lane group of the same road, the probability of the vehicle choosing each road is obtained, and the road with the highest probability is taken as the predicted road value.

[0108] Through the above methods, the present invention searches and determines a set of candidate lanes that the target vehicle can reach, providing a possible range of choices for subsequent predictions. Based on vehicle dynamics, the short-term trajectory of the target vehicle is predicted. This process considers the vehicle's current state and possible actions to predict its short-term movement trend. A path similarity scoring model is established to evaluate the degree of matching between each lane and the predicted trajectory of the target vehicle by calculating the similarity score of all candidate lanes. Finally, the set of candidate lanes is classified into different roads, and the road with the highest probability is calculated to determine the road most likely to be chosen by the target vehicle. This provides a foundation for understanding the general driving intentions of the target vehicle.

[0109] Step S2: Based on the lateral offset of the target vehicle, construct an interactive multi-model-based lane selection intention prediction model. The lane selection intention prediction model obtains lane prediction values ​​based on the road prediction values.

[0110] Step S2-1: Considering the lateral deviation and lateral deviation velocity of the target vehicle, construct a lane selection intention prediction model based on IMM (Interactive Multi-Model). Based on the results of previous steps, the target vehicle's intention to choose different roads can be determined. However, there are usually multiple lanes within a road, each with different probabilities. Therefore, it is necessary to infer the target vehicle's lane intention more precisely. This invention selects the lateral deviation l of the target vehicle from the lane centerline at time k. k and lateral speed As state input, an interactive multi-model framework is constructed based on the target multi-lane model, which mainly includes the following steps: constructing the state transition probability matrix (TPM), the interactive part of the IMM, the filtering part of the IMM, and the lane model probability update of the IMM.

[0111] (1) Construct the state transition probability matrix (TPM):

[0112] Based on the Markov assumption, lane intention inference within a road is conceptualized as a lane change prediction problem. This is achieved by considering real-time measured lateral offset velocities. To derive the time-varying transition probability matrix (TPM):

[0113]

[0114] Where the subscripts i and j represent lane numbers. n′ is the total number of lanes, p ij The probability of lane i changing to lane j at time k-1 is represented by the lateral offset velocity. Related time-varying functions:

[0115]

[0116] in, It is determined by the probability density function The calculated prior transition probabilities. The prior transition probability of the prior TPM for each row. Normalization is required. This is the cumulative value of the probability density function for lane i transitioning to lane j. In this invention, left lane change, following lane change, and right lane change are all modeled as normal distributions:

[0117]

[0118] Where, ρ ij and These are the mean and covariance of the lateral offset speed of the target vehicle transitioning from lane i to lane j.

[0119] (2) The interactive part of IMM:

[0120] In the interactive part of the IMM algorithm, the posterior lane model probability at time k-1 is used. Lane lateral deviation and the covariance of the probability model for selecting lane j The mixed probability of lane i to lane j at time k. Prior lane model probability Mixed state of lateral offset in lane j and the mixed covariance of lateral offset Perform update calculations:

[0121]

[0122] in, and These are the posterior lane model probability, lane lateral offset, and covariance of the j-lane probability model, respectively, derived from time k-1. For mixed probabilities, For the prior lane model probability, The mixed state of lateral offset for lane j. Let be the mixed covariance of the lateral offset of lane j. This refers to the mixed state of lateral offset of lane i.

[0123] (3) The filtering part of IMM:

[0124] Then, for each lane model All are subjected to standard linear Kalman filtering, and the calculation process is as follows:

[0125]

[0126] Among them, KF prediction (·) and KF update (·) represent the prediction and update steps of the standard Kalman filter, respectively. A is the transition matrix, and Q is the process noise matrix. Let be the mixed covariance of the lateral offset of lane i. and Let k be the prior and posterior lane lateral offsets, respectively. This represents the average lateral offset of existing lanes. and Let be the prior and posterior covariances of lane i, respectively.

[0127] (4) Lane model probability update of IMM:

[0128] Finally, based on the measurement residuals of the filter for the i-th lane model at time k... Covariance Update the posterior lane model probability The calculation method is as follows:

[0129]

[0130] in, It is the likelihood measured by the filter of the i-th lane model. Let be the measurement residual of the filter for the i-th lane model at time k. Let l be the covariance of the filter for the i-th lane model at time k. k Let R be the lateral offset of the target vehicle from the lane centerline at time k, and R be the noise matrix.

[0131] Finally, in step S1, the lane with the highest posterior lane model probability is selected from the lanes within the road prediction values ​​as the target lane.

[0132] Step S2-2: Calculate the target distance and obtain the anchor point of the lane centerline with the highest probability based on the assumption that the target vehicle is traveling at a constant speed. In step S1, the lane with the highest posterior lane model probability is selected from the lanes within the road prediction value as the target lane, assuming a longitudinal speed v. k The aiming distance (dis) is constant and calculated according to the formula:

[0133] dis=v k T predict

[0134] Where T predict This is the aiming time. Then, using this distance as input, the centerline anchor point on the target lane is queried, and the state P of the anchor point is obtained. T (x T y T θ T ), where x T g T θ T These are the horizontal and vertical coordinates of the anchor point, and the yaw angle, respectively.

[0135] In this way, after determining the possible roads that the target vehicle may choose, the present invention further refines the prediction to the lane level, predicting the vehicle's specific lane selection intention on that road. Considering the lateral deviation and lateral deviation speed of the target vehicle, a lane selection intention prediction model based on Interactive Multi-Model (IMM) is constructed. The IMM model can handle the diversity and uncertainty of vehicle behavior, improving the accuracy of the prediction. By obtaining the anchor point of the most probable lane centerline based on the assumption that the target vehicle is traveling at a constant speed, the lane most likely to be chosen by the target vehicle is further determined.

[0136] Step S3: Generate the predicted trajectory with the highest probability for the target vehicle based on the fifth-order polynomial curve method. Use the real-time state of the target vehicle at time k as the initial state (start). k (x k y k θ k v k a k The endpoint state target(x) is the state of the target vehicle's target point on the reference path, i.e., the target lane anchor point with the highest probability obtained in the previous steps. T g T θ T v T =v k a T =a k A fifth-order polynomial trajectory curve interpolation method is established as follows:

[0137] x(t) = a5t 5 +a4t4 +a3t 3 +a2t 2 +a1t+a0

[0138] y(t)=b5t 5 +b4t 4 +b3t 3 +b2t 2 +b1t+b0

[0139] Where t is time, a0, ..., a5, b0, ..., b5 are polynomial coefficients, and x(t) and y(t) are the coordinates of the trajectory curve. The starting and ending constraints of the fifth-order polynomial trajectory curve are:

[0140] x(0)=x k y(0)=y k

[0141] x′(0)=v k cosθ k y′(0)=v k sinθ k

[0142] x″(0)=a k cosθ k y″(0)=a k sinθ k

[0143] x(T predict )=x T y(T predict )=y T

[0144] x′(T predict ) = v T cosθ T y′(T predict ) = v T cosθ T

[0145] x″(T predict )=0 y″(T predict ) = 0

[0146] Where x′(·) and x″(·) are the first and second derivatives of the abscissa of the trajectory curve, respectively, and y′(·) and y″(·) are the first and second derivatives of the ordinate of the trajectory curve, respectively.

[0147] Therefore, the parameters of the fifth-order polynomial trajectory are solved, and the results are used for target vehicle trajectory prediction.

[0148] Based on the above methods, this invention uses a quintic polynomial curve method to generate the predicted trajectory with the highest probability for the target vehicle, building upon the first two steps. The quintic polynomial curve provides a smooth trajectory while considering the vehicle's current position, speed, acceleration, and other states, generating a predicted trajectory that conforms to the vehicle's dynamic characteristics. This provides accurate and practical information on the target vehicle's behavior and trajectory, offering crucial support for decision-making, path planning, and collision avoidance functions in autonomous driving systems.

[0149] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the shown specific order or sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations. The above descriptions are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A target vehicle trajectory prediction method based on dual-lane intent inference, characterized in that, Includes the following steps: Step S1: Based on extended Kalman filter, obtain the short-term predicted trajectory of the target vehicle; based on the path similarity scoring model, analyze the short-term predicted trajectory to obtain the predicted road value selected by the target vehicle when entering the intersection. Step S2: Based on the lateral offset of the target vehicle, construct an interactive multi-model-based lane selection intention prediction model. The lane selection intention prediction model obtains lane prediction values ​​based on the road prediction values. Step S3: Based on the lane prediction value, generate the predicted trajectory of the target vehicle using the fifth-order polynomial curve method; Step S2 details include: Step S2-1: Construct a lane selection intent prediction model based on interactive multi-model. Based on the lateral offset and lateral offset speed of the target vehicle, obtain the lane with the highest probability within the predicted lane value as the target lane. Step S2-2: Calculate the aiming distance, and obtain the anchor point of the centerline of the target lane based on the aiming distance, as the lane prediction value; Step S2-1 specifically includes: (1) Construct the state transition probability matrix: Where the subscripts i and j represent lane numbers, The total number of lanes. The probability of lane i changing to lane j at time k-1 is represented by the lateral offset velocity. The relevant time-varying function is calculated as follows; in, It is determined by the probability density function The calculated prior transition probability, It is the cumulative value of the probability density function for lane i to lane j. and These are the mean and covariance of the lateral offset speed of the target vehicle transitioning from lane i to lane j; (2) For the interactive part of the interactive multi-model, the following update calculations are performed: in, , ,and These are the posterior lane model probability, lane lateral offset, and covariance of the j-lane probability model, respectively, derived from time k-1. For mixed probabilities, For the prior lane model probability, The mixed state of lateral offset for lane j. Let be the mixed covariance of the lateral offset of lane j. This represents a mixed state of lateral offset for lane i; (3) The filtering part of the interactive multi-model, for each lane model All are subjected to standard linear Kalman filtering, and the calculation process is as follows: in, and These are the prediction and update steps of the standard Kalman filter, respectively. A It is a transition matrix. Q It is the process noise matrix. Let be the mixed covariance of the lateral offset of lane i. and Let k be the prior and posterior lane lateral offsets at time k. This represents the average lateral offset of existing lanes. and Let be the prior and posterior covariances of lane i, respectively; (4) The lane model probability update for interactive multi-model is calculated as follows: in, It is the likelihood measured by the filter of the i-th lane model. Let be the measurement residual of the filter for the i-th lane model at time k. Let be the covariance of the filter for the i-th lane model at time k. Let R be the lateral offset of the target vehicle from the lane centerline at time k, and let R be the noise matrix. This represents the probability of updating the posterior lane model; (5) Select the lane with the highest posterior lane model probability as the target lane from the lanes within the predicted road value.

2. The target vehicle trajectory prediction method based on dual-lane intent inference according to claim 1, characterized in that, Step S1 specifically includes: Step S1-1: Search for a set of candidate lanes that the target vehicle can reach; Step S1-2: Predict the short-time local state variables of the target vehicle and predict the short-term trajectory of the target vehicle based on the extended Kalman filter. Steps S1-3: Establish a path similarity scoring model to evaluate the similarity between the short-term predicted trajectory of the target vehicle and each lane in the candidate lane set; Steps S1-4: Determine the road that the target vehicle is most likely to choose based on the similarity, and use it as the predicted road value.

3. The target vehicle trajectory prediction method based on dual-lane intent inference according to claim 2, characterized in that, Step S1-1 includes: Based on the real-time status information of the target vehicle, a depth-first search algorithm is used to search for a set of candidate lanes for the target vehicle from the high-definition map. The set of candidate lanes contains multiple candidate lanes, representing all possible choices that the target vehicle can make to drive into other lanes at the intersection.

4. The target vehicle trajectory prediction method based on dual-lane intent inference according to claim 3, characterized in that, Step S1-2 includes: Step S1-2-1: Predict the short-time local state variables of the target vehicle using the following method: in, , Let the longitudinal velocity and yaw rate be the values ​​at time k+m in the future time window. , Let m be the longitudinal velocity and yaw rate at time k+1 in the future time window, m be the width of the short-term future time window, n be the width of the short-term historical time window, and a and b be the parameter indices. These are the weighting coefficients. , The longitudinal velocity and yaw rate at time k-a+1 within the short-term historical time window; Step S1-2-2: Predict short-time trajectories based on extended Kalman filtering, using the following method: (1) Establishing a process model: in, The state transition equations for vehicle kinematics are calculated as follows: in, It is the state vector of the target vehicle at time k. Here are the location coordinates of the target vehicle. Yaw angle For longitudinal velocity, Yaw angular velocity, It is process noise, which has a mean of zero and a power spectral density equal to the process noise matrix. The normal distribution , Indicates the noise intensity during the yaw angle process. Indicates the noise intensity during the longitudinal velocity process. The noise intensity during the yaw rate process is represented by T, where T is the sampling period. (2) Establish a measurement model: in, Let k be the measured state vector of the target vehicle at time k. It is a measurement matrix. It is measurement noise, which has a mean of zero and a noise covariance matrix of... The normal distribution , Indicates the noise intensity of longitudinal velocity measurement. Indicates the noise intensity of the yaw rate measurement; (3) Perform predictions for the extended Kalman filter. The calculation process is as follows: in, It is the prior estimated state at time k. It is the posterior estimated state at time k-1. The state transition equation is... It is the prior covariance matrix at time k. It is the posterior covariance matrix at time k-1. It is the Jacobian matrix at time k-1. It is the process noise matrix at time k-1; (4) Update the extended Kalman filter. The calculation process is as follows: in, It is the Kalman gain at time k; The extended Kalman filter is updated repeatedly until all trajectories in the future time window are completed, ultimately yielding the short-term predicted trajectory. .

5. The target vehicle trajectory prediction method based on dual-lane intent inference according to claim 4, characterized in that, Steps S1-3 specifically include: The similarity score between the short-term predicted trajectory of the target vehicle and each lane in the candidate lane set is calculated as follows: in, It is the i-th lane and the short-term predicted trajectory Similarity score between them , It is a weighted coefficient for positional similarity and yaw angle similarity. and These are positional similarity and yaw angle similarity, respectively. Location similarity The calculation method is as follows: in, The mean of the Euclidean distance. For short-term predicted trajectory point index, For the coordinates of the short-term predicted trajectory points, is the coordinates of the nearest point in the short-term predicted trajectory, m is the width of the short-term future time window, and L is the mapping threshold parameter; Yaw angle similarity The calculation method is as follows: in, The average difference in yaw angle. For short-term predicted trajectory point index, For short-term predicted trajectory point yaw angle, This is for the heading in the lane corresponding to the short-term predicted trajectory point.

6. The target vehicle trajectory prediction method based on dual-lane intent inference according to claim 5, characterized in that, Steps S1-4 include: Use the softmax function to evaluate the candidate lane set. similarity score Normalization is performed to obtain the probability that the target vehicle selects different lanes. : in, The total number of all candidate lanes; The probability of a vehicle choosing each road is obtained by summing the probabilities of each lane in a lane group on the same road, and the road with the highest probability is taken as the predicted road value.

7. The target vehicle trajectory prediction method based on dual-lane intent inference according to claim 6, characterized in that, Step S2-2 specifically includes: Calculate the aiming distance : in It refers to the pre-aiming time, using the pre-aiming distance as a reference to query the centerline anchor point on the target lane and obtain the status of the anchor point. As the lane prediction value, where These are the horizontal and vertical coordinates of the anchor point and the yaw angle, respectively.

8. The target vehicle trajectory prediction method based on dual-lane intent inference according to claim 7, characterized in that, Step S3 specifically includes: Establish the following interpolation formula for the trajectory curve of a fifth-order polynomial: Where t is time. , For polynomial trajectory parameters, , The coordinates of the trajectory curve are given; the constraints for the start and end points of the fifth-order polynomial trajectory curve are as follows: in, These represent the target vehicle's position coordinates, yaw angle, longitudinal velocity, and longitudinal acceleration at time k, respectively. These are the position coordinates, yaw angle, longitudinal velocity, and longitudinal acceleration of the lane prediction value, respectively. , , These are the first and second derivatives of the x-coordinate of the trajectory curve, respectively. , These are the first and second derivatives of the ordinate of the trajectory curve, respectively; Based on the starting and ending point constraints, the fifth-order polynomial trajectory parameters are solved, and the interpolation result of the fifth-order polynomial trajectory curve is used as the predicted trajectory of the target vehicle.

Images