Vehicle lane changing track prediction method, device and system

Abstract

The invention provides a vehicle lane changing track prediction method, device and system, and the method comprises the steps: obtaining a vertical coordinate, predicted based on a lane changing physical model, of a lane changing vehicle at a next moment through a calibrated lane changing physical model; through the trained residual learning network ResNet, obtaining the predicted abscissa and predicted value residual of the lane changing vehicle at the next moment; obtaining a space-time correlation feature vector through a hybrid neural network structure; performing bidirectional modulation on the predicted value residual error output by the ResNet to obtain a modulated residual error; on the basis of the modulated residual error, the corrected ordinate of the lane changing vehicle at the next moment is obtained; and taking the predicted abscissa of the lane-changing vehicle at the next moment and the corrected ordinate of the lane-changing vehicle at the next moment as the predicted coordinate of the lane-changing vehicle at the next moment. According to the scheme, the neural network algorithm and the parameter calibration-based vehicle lane changing physical model are fused, so that the accuracy and interpretability of vehicle lane changing track prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent transportation, in particular to a vehicle lane-changing trajectory prediction method, device and system. BACKGROUND

[0002] With the rapid development of expressway network and the continuous growth of traffic flow, traffic safety and efficiency problems are increasingly prominent, especially in the key sections such as expressway merging area, diverging area and construction area, frequent lane-changing behavior has become a high-risk inducement of traffic accidents. As an important node of different traffic flow intersection, due to high speed and dense traffic, drivers need to complete judgment and decision-making in a short time, resulting in frequent lane-changing operation, which is easy to cause scraping, rear-end and other accidents, and the merging area has become a traffic accident disaster area on the expressway. In the actual traffic environment, vehicles usually change lanes to merge into the main lane, avoid slow vehicles, enter the exit ramp or bypass the obstacle area, etc. However, due to the limitations of driver's reaction time, visual field and uncertainty of surrounding traffic environment perception, lane-changing behavior has significant risk, especially in the case of multiple vehicles changing lanes at the same time, high traffic density or blocked vision, which is more likely to cause chain collision.

[0003] Studies have shown that lane-changing behavior has become one of the main inducements of expressway traffic accidents. In addition, the current lane-changing efficiency on the expressway is generally low, and drivers often spend a long time in the process of observation, judgment and execution of lane-changing action, and there is a large individual difference, which further aggravates the volatility and instability of traffic flow. In special sections such as construction areas, due to lane reduction, unclear signs or sudden lane-changing requirements, the pressure on drivers to change lanes increases significantly, which has a great impact on the overall traffic order and efficiency, and even causes large-scale congestion and accident risk. It is of great significance to build a high-precision and high-robustness vehicle lane-changing trajectory prediction method to improve the predictability of lane-changing behavior, reduce accident risk and optimize traffic flow organization. By fusing vehicle historical motion characteristics, environmental information and lane-changing behavior model, the trajectory change of vehicle in the merging area and complex traffic situation can be accurately predicted, which not only provides more reliable decision basis for automatic driving system, but also can provide early warning for potential risks in assisted driving scenarios, thereby effectively improving traffic safety and overall operation efficiency.

[0004] Currently, lane-changing trajectory prediction models can be mainly divided into two categories: methods based on mathematical models and methods based on data driving. Among them, the lane-changing trajectory prediction method based on mathematical model generally describes the lane-changing behavior of the vehicle by constructing an explicit physical or geometric formula. Common models include safety distance model, polynomial trajectory model, Bezier curve trajectory model, etc. The safety distance model generally assumes that the surrounding vehicles travel at a constant speed, and by introducing mechanisms such as elliptical anti-collision rules and critical collision judgment conditions, the acceptable safety distance of the vehicle during lane-changing is calculated, thereby deriving a lane-changing trajectory that ensures driving safety. In addition, the polynomial model and the hyperbolic tangent function model use the motion characteristics of the vehicle to fit the trajectory, and the Bezier curve model generates a smooth and controllable curve in the geometric space, which is suitable for visualization and path planning. Although this kind of method has good interpretability and computational efficiency, it often relies on idealized assumptions and is difficult to accurately reflect complex and dynamic real traffic environments, so there may be large errors when dealing with complex traffic scenarios. The lane-changing trajectory prediction method based on data driving mines the vehicle behavior rules from historical data, thereby improving the adaptability of the model to complex scenarios. Common data-driven models include random forest, long short-term memory network (LSTM), graph neural network (GCN), etc. This kind of method can mine spatio-temporal features and multi-vehicle interaction relationships, and has strong modeling ability for high-dimensional data, but at the same time it also faces challenges such as poor model interpretability and strong dependence on high-quality training data.

[0005] In summary, the existing lane-changing trajectory prediction models still have a large trade-off between prediction accuracy and model interpretability, and there is still a lack of a lane-changing trajectory prediction model that combines data driving and physical models.

[0006] Therefore, there is an urgent need for a vehicle lane-changing trajectory prediction scheme to solve the above technical problems. SUMMARY

[0007] The present application provides a vehicle lane-changing trajectory prediction method, device and system, which combines neural network algorithm with parameter-calibrated vehicle lane-changing physical model, thereby improving the accuracy and interpretability of vehicle lane-changing trajectory prediction.

[0008] The present application provides a vehicle lane-changing trajectory prediction method, device and system, which combines neural network algorithm with parameter-calibrated vehicle lane-changing physical model, thereby improving the accuracy and interpretability of vehicle lane-changing trajectory prediction.

[0009] Based on the acquired state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the longitudinal coordinate of the lane-changing vehicle at the next time point predicted based on the lane-changing physical model is acquired through the calibrated lane-changing physical model;

[0010] Based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, a predicted horizontal coordinate of the lane-changing vehicle at the next moment and a predicted value residual are obtained through a trained residual learning network ResNet;

[0011] Based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, a spatio-temporal correlation feature vector is obtained through a hybrid neural network structure; wherein the hybrid neural network structure is composed of a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM and an attention mechanism Attention;

[0012] Based on the obtained spatio-temporal correlation feature vector, the predicted value residual output by the ResNet is bidirectionally modulated to obtain a modulated residual; wherein the bidirectional modulation process is: sequentially performing abnormal residual filtering and effective residual enhancement;

[0013] Based on the modulated residual, the longitudinal coordinate of the lane-changing vehicle at the next moment predicted by the lane-changing physical model is corrected to obtain a corrected longitudinal coordinate of the lane-changing vehicle at the next moment;

[0014] The predicted horizontal coordinate of the lane-changing vehicle at the next moment and the corrected longitudinal coordinate of the lane-changing vehicle at the next moment are taken as the predicted coordinates of the lane-changing vehicle at the next moment.

[0015] Optionally, based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the process of obtaining a spatio-temporal correlation feature vector through a hybrid neural network structure is:

[0016] Based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, spatial interaction features of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle are obtained through a convolutional neural network CNN;

[0017] Based on the spatial interaction features, time series features of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle are obtained through the BiLSTM;

[0018] The time series features output by the BiLSTM are weighted through the Attention to generate the spatio-temporal correlation feature vector.

[0019] Optionally, the process of abnormal residual filtering is:

[0020] A gating function is set: The expression form of the gating function is: ; wherein, is a spatio-temporal correlation feature vector; is a nonlinear activation function, i.e., a sigmoid function, whose output ranges between (0, 1), i.e., a real number greater than 0 and less than 1, for mapping the output value to a probability interval; and is a parameter of the gating function: is a weight matrix, for performing a linear transformation on to map to a new space; is a bias term, for adding an offset to after performing a linear transformation on ; and each element in and is a real number;

[0021] when , it is determined that the prediction value residual contains abnormal components; wherein, is a preset threshold, for determining whether the prediction value residual is abnormal;

[0022] based on obtaining the prediction value residual , an inhibited prediction value residual is obtained; = , is the prediction value residual; the value of is between (0, 1), i.e., a real number greater than 0 and less than 1.

[0023] Optionally, the process of effective residual enhancement is:

[0024] a cosine similarity between the and a historical motion trend vector extracted from the is calculated; wherein, the expression of is:

[0025] ; is a cosine similarity function, for measuring the similarity between two vectors, whose value ranges between [−1, 1], i.e., a real number greater than or equal to −1 and less than or equal to 1, the closer the value to 1, the more similar the two vectors; the closer the value to -1, the less similar the two vectors;

[0026] based on the cosine similarity , the , an enhanced prediction value residual is obtained; wherein, the is the inhibited prediction value residual, the and the historical motion trend vector extracted from the cosine similarity between ; is an enhancement coefficient, and is a real number greater than 0, used to control the intensity of enhancement.

[0027] Optionally, based on the modulated residual error, the longitudinal coordinate of the next time lane-changing vehicle predicted by the lane-changing physical model is corrected to obtain a corrected longitudinal coordinate of the next time lane-changing vehicle, and the process is:

[0028] The corrected longitudinal coordinate of the next time lane-changing vehicle is obtained based on the following formula ; ; wherein, is the longitudinal coordinate of the next time lane-changing vehicle predicted based on the lane-changing physical model, is the current time, indicating the current time point, is the time interval, is the enhanced predicted value residual error; is an adaptive coefficient based on the spatiotemporal feature, calculated by a multi-layer perception (MLP) layer: , is a spatiotemporal correlation feature vector, and are parameters of the multi-layer perception (MLP) layer: is a weight matrix, used for linear transformation of to map to a new space; is a bias term, used to add an offset to after linear transformation of ; and each element in and is a real number; is a nonlinear activation function, i.e., a sigmoid function, whose output range is between (0, 1), i.e., a real number greater than 0 and less than 1, used to map the output value to a probability interval.

[0029] Optionally, the training process of the residual learning network is:

[0030] obtain a first residual error of the longitudinal coordinate at the next time ; wherein, the is the difference between the longitudinal coordinate of the next time lane-changing vehicle predicted based on the lane-changing physical model and the real longitudinal coordinate of the next time lane-changing vehicle;

[0031] The residual learning network based on one or more layers of LSTM is constructed, and the lane-changing vehicle and the state information of the target vehicle related to the lane-changing vehicle are input into the residual learning network to obtain the horizontal coordinate of the lane-changing vehicle at the next moment and the vertical coordinate at the next moment predicted by the residual learning network ;

[0032] Based on the 、 , the first error is obtained; and the second error is obtained based on the predicted horizontal coordinate of the lane-changing vehicle at the next moment and the real horizontal coordinate of the lane-changing vehicle at the next moment;

[0033] The residual learning network is trained until the absolute values of the first error and the second error tend to 0 and reach the preset accuracy.

[0034] Optionally, a mapping function of the residual and the feature value is established: ; wherein, is a parameter of the mapping function, is the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle; ;

[0035] The function obtains the target parameter through training, so that the absolute value of the first error tends to 0 and reaches the preset accuracy.

[0036] Optionally, a mapping function of the vehicle horizontal coordinate and the feature value is established: ; wherein, is a mapping function, is a parameter of the mapping function, is the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, ;

[0037] The function obtains the target parameter through training, so that the absolute value of the second error tends to 0 and reaches the preset accuracy.

[0038] The present specification also provides a vehicle lane-changing trajectory prediction device, comprising: a first prediction module, a second prediction module, a modulation module, a correction module, and a determination module.

[0039] The first prediction module is configured to obtain the vertical coordinate of the lane-changing vehicle at the next moment predicted based on a lane-changing physical model based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle through the lane-changing physical model.

[0040] The second prediction module is configured to obtain a predicted horizontal coordinate of the lane-changing vehicle at the next moment and a predicted value residual based on the acquired state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, and through a trained residual learning network ResNet.

[0041] The modulation module is configured to obtain a space-time correlation feature vector based on the acquired state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, and through a hybrid neural network structure composed of a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM and an attention mechanism Attention; and to perform bidirectional modulation on the predicted value residual output by the ResNet based on the obtained space-time correlation feature vector, and obtain a modulated residual; wherein the bidirectional modulation process comprises sequentially performing abnormal residual filtering and effective residual enhancement.

[0042] The correction module is configured to correct the longitudinal coordinate of the lane-changing vehicle at the next moment predicted by the lane-changing physical model based on the modulated residual, and obtain a corrected longitudinal coordinate of the lane-changing vehicle at the next moment.

[0043] The determination module is configured to take the predicted horizontal coordinate of the lane-changing vehicle at the next moment and the corrected longitudinal coordinate of the lane-changing vehicle at the next moment as the predicted coordinates of the lane-changing vehicle at the next moment.

[0044] The present specification also provides a vehicle lane-changing trajectory prediction system comprising a processor and a memory, wherein the memory stores a program, and the processor executes the program to implement the vehicle lane-changing trajectory prediction method of any one of the technical solutions of the above method.

[0045] The present application provides a vehicle lane-changing trajectory prediction method, device and system, which combines neural network algorithm with parameter calibration based vehicle lane-changing physical model, thereby improving the accuracy and interpretability of vehicle lane-changing trajectory prediction.

[0046] Further, the present application combines physical model and residual learning network, provides prior knowledge through physical model, and residual learning model is responsible for correcting the prediction error of the physical model. The residual of the physical model is the difference between the observation data and the predicted value, which usually changes little, so the prediction residual is relatively simple, and the dependence on a large amount of data is greatly reduced. This makes the scheme can effectively run in the case of data shortage, significantly reduces the demand for large-scale data. In this way, the scheme can still maintain high prediction accuracy and stability in the scene of insufficient or poor quality data.

[0047] Furthermore, the present invention performs bidirectional modulation on the predicted value residual output by the residual learning network based on the spatiotemporal correlation feature vector output by the hybrid neural network structure to obtain the modulated residual. On the one hand, it solves the limitations of fixed threshold filtering in complex scenarios, and on the other hand, it strengthens the residual component consistent with the historical motion trend to avoid the correction result violating the laws of vehicle dynamics. At the same time, the spatiotemporal correlation feature vector can be traced back to specific historical states (such as the target vehicle deceleration event), which makes the residual correction process have physical meaning and greatly enhances the interpretability of the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0049] Fig. 1 A schematic flow chart of a vehicle lane change trajectory prediction method provided in Example 1 of this specification;

[0050] Fig. 2 A schematic flow chart of a vehicle lane change trajectory prediction method provided in Example 2 of this specification;

[0051] Fig. 3 This is a schematic diagram of an example scenario provided in Example 3 of this specification;

[0052] Fig. 4 This is a comparison chart of the predicted value and the actual value of the GIPPS+LSTM network in Example 4 of this specification;

[0053] Fig. 5 This is a comparison table of evaluation indicators of various prediction methods in Example 5 of this specification;

[0054] Fig. 6 A schematic diagram of a vehicle lane change trajectory prediction device provided in Example 6 of this specification;

[0055] Fig. 7 A schematic diagram of the structure of an electronic device provided in Example 7 of this specification;

[0056] Fig. 8 A schematic diagram of the principle of a computer-readable medium provided in Example 8 of this specification. DETAILED DESCRIPTION

[0057] The following description is presented to enable any person skilled in the art to practice the application as claimed. The preferred embodiments disclosed herein are only examples of the application and alternative embodiments can be devised by those skilled in the art without departing from the spirit and scope of the application. The disclosure of the application is presented to enable any person skilled in the art to practice the application as claimed.

[0058] The application will be described with reference to the accompanying drawings, of which: Figs. 1-8 Exemplary embodiments of the present application are described more fully hereinafter with reference to the accompanying drawings, in which exemplary embodiments of the application are shown. The exemplary embodiments, however, can be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these exemplary embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the application to those skilled in the art, with broad and enabling scope. In the drawings, the same reference numbers are used throughout the several views to indicate the same or similar components, parts, and sections of the application, thus, repeated description is omitted.

[0059] In the description of the specific embodiments of the application, certain features, structures, characteristics or other details of the application are described to enable a full and complete understanding of the embodiments of the application. However, it is not intended that the application be limited to the specific embodiments described. In some embodiments, certain features, structures, characteristics or other details of the application can be practiced without one or more of the specific features, structures, characteristics or other details. Therefore, the specific

[0060] In the description of the specific embodiments of the application, certain features, structures, characteristics or other details of the application are described to enable a full and complete understanding of the embodiments of the application. However, it is not intended that the application be limited to the specific embodiments described. In some embodiments, certain features, structures, characteristics or other details of the application can be practiced without one or more of the specific features, structures, characteristics or other details. Therefore, the specific

[0061] The flow charts shown in the drawings are merely illustrative and do not necessarily include all of the content and operations / steps, nor are they necessarily performed in the order described. For example, some operations / steps can be further broken down, and some operations / steps can be combined or partially combined, so the actual order of execution can be changed according to actual conditions.

[0062] The block diagrams shown in the drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0063] The term "and / or" or "and / or" includes all combinations of one or more of the associated listed items.

[0064] The application provides a highway merging area lane-changing vehicle trajectory prediction method based on physics enhanced residual learning.

[0065] Fig. 1 A vehicle lane-changing trajectory prediction method flowchart for embodiment 1 of the present specification comprises the following steps:

[0066] S101, based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the longitudinal coordinate of the lane-changing vehicle at the next moment predicted based on the lane-changing physical model is obtained through the calibrated lane-changing physical model.

[0067] Optionally, the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle at least comprises the obtained state information of the lane-changing vehicle, the front vehicle of the current lane, the front vehicle of the target lane and the rear vehicle of the target lane.

[0068] Specifically:

[0069] The state information of the lane-changing vehicle of the current lane, the front vehicle of the current lane, the front vehicle of the target lane and the rear vehicle of the target lane at least comprises:

[0070] The vehicle length of the lane-changing vehicle of the current lane, the current time horizontal and vertical direction instantaneous speed and horizontal and vertical direction instantaneous acceleration of the lane-changing vehicle, the current time horizontal and vertical coordinates of the lane-changing vehicle; the vehicle length of the front vehicle of the current lane, the current time horizontal and vertical direction instantaneous speed and horizontal and vertical direction instantaneous acceleration of the front vehicle of the current lane, the current time horizontal and vertical coordinates of the front vehicle of the current lane; the vehicle length of the front vehicle of the target lane, the current time horizontal and vertical direction instantaneous speed and horizontal and vertical direction instantaneous acceleration of the front vehicle of the target lane, the current time horizontal and vertical coordinates of the front vehicle of the target lane; the current time horizontal and vertical coordinates of the rear vehicle of the target lane, the current time horizontal and vertical direction instantaneous speed and horizontal and vertical direction instantaneous acceleration of the rear vehicle of the target lane.

[0071] Optionally, the lane-changing physical model is built, and the genetic algorithm is used to calibrate the parameters of the lane-changing physical model; wherein the lane-changing physical model is a vehicle lane-changing model applying Gipps general collision avoidance algorithm.

[0072] Specifically:

[0073] The parameter calibration process of the Gipps lane-changing model is as follows:

[0074] The Gipps lane-changing model adopts a genetic algorithm to calibrate the parameters of the lane-changing physical model, and the specific steps are as follows:

[0075] (1) Determine the fitness function

[0076]

[0077] The average value (MSE) of the error square sum between the longitudinal coordinate of the lane-changing vehicle predicted by the lane-changing physical model at the next time and the actual longitudinal coordinate of the lane-changing vehicle at the next time is taken as the fitness function, which is used to measure the pros and cons of the model prediction accuracy; is the number of samples.

[0078] (2) Initialize the population, the population size is 100, and each individual is the parameter to be calibrated , , , , , , the value range of the parameters is (-3.5, 0), (-3.5, 0), (-3.5, 0), (-3.5, 0), (0, 1), (0, 1);

[0079] (3) Selection operation, adopt the method of tournament selection, the specific operation is to randomly select two or more individuals from the population, compare their fitness, and select the individual with the best fitness;

[0080] (4) Cross operation, for the selected parent individual, weighted fusion is performed on each parameter according to the crossover rate;

[0081] (5) Mutation operation, add Gaussian noise to each parameter according to the mutation rate, and the mutation amplitude is adaptive to the parameter range;

[0082] (6) Elite reservation, the individual with the best fitness of each generation is directly entered into the next generation.

[0083] First, initialize the population, calculate the fitness (MSE) of each generation individual, record the current optimal solution, then execute the selection, crossover and mutation operations to generate a new population, then merge the elite individuals, output the optimal parameters and the minimum MSE of the final generation.

[0084] Optionally, based on the acquired lane-changing vehicle, the state information of the target vehicle related to the lane-changing vehicle, through the calibrated lane-changing physical model, the process of acquiring the longitudinal coordinate of the lane-changing vehicle at the next moment predicted based on the lane-changing physical model is:

[0085] The lane-changing vehicle in the merging area of the expressway and the front vehicle in the current lane and the front and rear vehicles in the target lane are selected as the research objects.

[0086] The historical state information of the lane-changing vehicle in the current lane, the front vehicle in the current lane, the front vehicle in the target lane, and the rear vehicle in the target lane is extracted, and the time window of data collection is 5s before lane changing to 5s after lane changing, and the time interval is 0.1s. Specifically, the following parameters are included: the vehicle length of the lane-changing vehicle in the current lane, the current moment instantaneous speed and instantaneous acceleration in the transverse and longitudinal directions of the lane-changing vehicle, the current moment transverse and longitudinal coordinates of the lane-changing vehicle; the vehicle length of the front vehicle in the current lane, the current moment instantaneous speed and instantaneous acceleration in the transverse and longitudinal directions of the front vehicle in the current lane, the current moment transverse and longitudinal coordinates of the front vehicle in the current lane; the vehicle length of the front vehicle in the target lane, the current moment instantaneous speed and instantaneous acceleration in the transverse and longitudinal directions of the front vehicle in the target lane, the current moment transverse and longitudinal coordinates of the front vehicle in the target lane; the current moment transverse and longitudinal coordinates of the rear vehicle in the target lane, the current moment instantaneous speed and instantaneous acceleration in the transverse and longitudinal directions of the rear vehicle in the target lane.

[0087] The collision avoidance algorithm with the front vehicle in the current lane is considered, and the specific calculation formula is as follows:

[0088] wherein, is the instantaneous speed of the lane-changing vehicle at the next moment to avoid collision with the front vehicle in the current lane, is the vehicle length of the front vehicle in the current lane, is the current moment x coordinate of the front vehicle in the current lane, is the current moment x coordinate of the lane-changing vehicle, is the current moment instantaneous speed of the lane-changing vehicle, is the current moment instantaneous speed of the front vehicle in the current lane, and is the maximum deceleration of the lane-changing vehicle and the front vehicle in the current lane, : current moment, indicating the current time point; is the time interval.

[0089] The collision avoidance algorithm with the front vehicle in the target lane is considered, and the calculation formula is as follows:

[0090] wherein the instantaneous speed of the vehicle at the next moment in order to avoid collision with the front vehicle in the target lane, the current moment x coordinate of the front vehicle in the target lane, the vehicle length of the front vehicle in the target lane, the maximum deceleration of the front vehicle in the target lane, the time interval. Other parameter meanings refer to the above description.

[0091] Considering the collision avoidance algorithm with the rear vehicle in the target lane, the calculation formula is as follows:

[0092] wherein the instantaneous speed of the vehicle at the next moment in order to avoid collision with the rear vehicle in the target lane, the vehicle length of the vehicle changing lane, the current moment x coordinate of the rear vehicle in the target lane, the current moment instantaneous speed of the rear vehicle in the target lane, the maximum deceleration of the rear vehicle in the target lane, the time interval. Other parameter meanings refer to the above description.

[0093] the longitudinal instantaneous speed of the vehicle changing lane at the next moment is:

[0094]

[0095] wherein, represents the influence of the front vehicle in the target lane on the speed of the vehicle changing lane, and its value range is a constant greater than or equal to 0 and less than or equal to 1, represents the influence proportion of the current lane and the target lane vehicle on the speed of the vehicle changing lane, and its value range is a constant greater than or equal to 0 and less than or equal to 1, the time interval. Other parameter meanings refer to the above description.

[0096] Using kinematic formula, the longitudinal coordinate of the vehicle changing lane at the next moment is obtained, and the calculation formula is as follows:

[0097]

[0098] wherein, represents the longitudinal coordinate of the vehicle changing lane at the next moment calculated by the lane changing physical model, the current moment longitudinal coordinate of the vehicle changing lane the current moment longitudinal instantaneous speed of the vehicle changing lane; the longitudinal instantaneous speed of the vehicle changing lane at the next moment; : the current moment, indicating the current time point; the time interval.

[0099] S102, based on the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the trained residual learning network ResNet is used to obtain the predicted horizontal coordinate of the lane-changing vehicle at the next moment and the predicted value residual;

[0100] S103, based on the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, a spatio-temporal correlation feature vector is obtained through a hybrid neural network structure; wherein the hybrid neural network structure is composed of a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM and an attention mechanism Attention;

[0101] Optionally, based on the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the process of obtaining the spatio-temporal correlation feature vector through the hybrid neural network structure is:

[0102] Based on the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the spatial interaction feature of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle is obtained through the convolutional neural network CNN;

[0103] Based on the spatial interaction feature, the time series feature of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle is obtained through the BiLSTM;

[0104] The time series feature output by the BiLSTM is weighted through the Attention to generate the spatio-temporal correlation feature vector.

[0105] Optionally, the spatial interaction feature includes at least one of the following: the distribution of the relative position between the lane-changing vehicle and the target vehicle related to the lane-changing vehicle in space; the distribution of the speed difference between the lane-changing vehicle and the target vehicle related to the lane-changing vehicle in space; the time series feature includes at least one of the following: the change trend of the lateral speed of the lane-changing vehicle, and the deceleration trend of the target vehicle related to the lane-changing vehicle.

[0106] S104, based on the obtained spatio-temporal correlation feature vector, the predicted value residual output by the ResNet is bidirectionally modulated to obtain a modulated residual; wherein the process of bidirectional modulation is: sequentially performing abnormal residual filtering and effective residual enhancement;

[0107] Optionally, the process of abnormal residual filtering is:

[0108] The gating function is set as: The expression form of the gating function is: ; wherein, is the spatio-temporal correlation feature vector; is a nonlinear activation function, i.e., sigmoid function, whose output ranges between (0, 1) i.e., real numbers greater than 0 and less than 1, for mapping the output value to a probability interval; and are parameters of the gating function: is a weight matrix, for performing linear transformation on to map to a new space; is a bias term, for adding an offset to after performing linear transformation on ; and each element in

[0109] when , it is determined that the prediction value residual contains abnormal components; wherein, is a preset threshold, for judging whether the prediction value residual is abnormal;

[0110] based on obtaining the prediction value residual , the suppressed prediction value residual is obtained; wherein, = , is the prediction value residual; the value of ranges between (0, 1) i.e., real numbers greater than 0 and less than 1.

[0111] because the value of ranges between (0, 1), when it is less than the threshold , its value will be smaller, after multiplying , the value of the residual will be reduced, thereby achieving the effect of suppressing abnormal values, thereby improving the robustness and accuracy of the model.

[0112] Optionally, the process of effective residual enhancement is:

[0113] calculate the cosine similarity between the and the historical motion trend vector extracted from the ; wherein, the expression of

[0114] ; is a cosine similarity function, for measuring the similarity between two vectors, whose value ranges between [−1, 1] i.e., real numbers greater than or equal to −1 and less than or equal to 1, the closer the value to 1, the more similar the two vectors; the closer the value to -1, the less similar the two vectors;

[0115] Based on cosine similarity 、 , and obtain the enhanced predicted value residual ; , wherein the is the residual of the predicted value after suppression, For the With The historical movement trend vector extracted from The cosine similarity between is the enhancement coefficient, which is a real number greater than 0 and is used to control the intensity of the enhancement.

[0116] This method enhances the effective residual components by similarity with historical trajectory trends, thereby improving the model's ability to capture effective information in the data.

[0117] S105. Based on the modulated residual, correct the vertical coordinate of the lane-changing vehicle at the next moment predicted by the lane-changing physical model to obtain a corrected vertical coordinate of the lane-changing vehicle at the next moment;

[0118] Optionally, based on the modulated residual, the vertical coordinate of the lane-changing vehicle at the next moment predicted by the lane-changing physical model is corrected, and the process of obtaining the corrected vertical coordinate of the lane-changing vehicle at the next moment is:

[0119] Based on the following formula, the corrected vertical coordinate of the vehicle changing lanes at the next moment is obtained: ; ;in, is the vertical coordinate of the vehicle that will change lanes at the next moment predicted based on the lane-changing physical model, is the current moment, indicating the current time point, is the time interval, is the residual of the predicted value after enhancement; is the adaptive coefficient based on spatiotemporal features, which is calculated through a multi-layer perceptron MLP layer: , is the spatiotemporal correlation eigenvector, and Parameters of the multi-layer perceptron MLP layer: is the weight matrix used to Perform a linear transformation to Map to a new space; is the bias term used to In the pair After the linear transformation, an offset is added; and Each element in is a real number; is a nonlinear activation function, i.e., the sigmoid function, whose output ranges between (0, 1), i.e., a real number greater than 0 and less than 1, used to map the output value to a probability interval.

[0120] The implementation scheme: fuse the predicted value of the lane-changing physical model and the modulated residual. This process adjusts the influence of the residual through an adaptive coefficient to ensure the stability and accuracy of the model under different conditions.

[0121] The behavior of the adaptive coefficient is as follows:

[0122] When the historical trajectory is stable: When the historical trajectory is stable, tends to 1. This means that the contribution of the residual is close to the original value, and the combination of the predicted value of the lane-changing physical model and the residual is more direct.

[0123] When the scene mutates: When the scene mutates, automatically decreases. This means that the contribution of the residual is weakened to maintain the stability of the model. This helps to avoid excessive influence of the residual on the prediction result in the mutation case.

[0124] S106: Take the predicted horizontal coordinate of the next time lane-changing vehicle and the corrected vertical coordinate of the next time lane-changing vehicle as the predicted coordinates of the next time lane-changing vehicle.

[0125] The following is an example:

[0126] Take a certain highway lane-changing scene as an example (the host vehicle (lane-changing vehicle) changes from the left lane to the right lane, and the target vehicle is the front vehicle on the right lane), and the specific implementation steps are as follows:

[0127] 1. Historical data collection

[0128] Collect the state data of the past 5 times (t-0.4s to t), as shown in the following table:

[0129] Time Host vehicle y coordinate (unit: meter, m) Host vehicle vertical velocity Vy (unit: meter / second, m / s) Target vehicle y coordinate (unit: meter, m) Target vehicle vertical velocity Vy (unit: meter / second, m / s) Relative distance d (m) t-0.4s 10.2 1.2 15.6 28.0 5.4 t-0.3s 10.4 1.3 15.9 28.1 5.5 t-0.2s 10.7 1.5 16.2 28.0 5.5 t-0.1s 11.0 1.6 16.5 27.5 5.5 t 11.3 1.7 16.8 27.0 5.5

[0130] 1. Spatio-temporal feature extraction

[0131] The CNN layer captures the spatial interaction feature of “the target vehicle and the host vehicle are longitudinally stable, but the target vehicle speed decreases from 28.0m / s to 27.0m / s”;

[0132] The BiLSTM layer learns the time sequence rule of “the host vehicle Vy shows an increasing trend (1.2→1.7m / s), and the target vehicle Vy shows a decreasing trend”.

[0133] The attention mechanism gives higher weights to the features from t-0.1s to t (because the target vehicle starts to decelerate), generating a spatiotemporal correlation feature vector .

[0134] 2. Residual modulation

[0135] If the residual learning network outputs an original residual = +0.8m (indicating that the lane-changing physical model may underestimate the lane-changing speed of the vehicle); where "+0.8m" indicates that the residual value is 0.8 meters. The "+" sign here indicates the positive direction of the residual. Specifically, it indicates that the network-predicted residual is a positive value, i.e., the output value is 0.8 meters greater than the input value.

[0136] If the gating mechanism calculates = 0.85 (> =0.3), it is determined that the residual is not abnormal and does not need to be filtered. Here, is a preset threshold value used to determine whether the residual is abnormal.

[0137] If =0.92 (indicating that the direction of the residual is consistent with the increasing trend of Vy of the vehicle), is 0.5, then the enhanced = 0.8 × (1 + 0.5×0.92) = 1.17m.

[0138] 3. Final correction

[0139] If the lane-changing physical model predicts the longitudinal coordinate of the next moment = 11.6m, and the adaptive coefficient = 0.9, the corrected longitudinal coordinate = 11.6 + 1.17×0.9 = 12.65m, which is more consistent with the actual trend of the vehicle accelerating to change lanes.

[0140] Optionally, the training process of the residual learning network is: obtaining a first residual of the longitudinal coordinate of the next moment; wherein the is the difference between the longitudinal coordinate of the lane-changing vehicle predicted by the lane-changing physical model at the next moment and the actual longitudinal coordinate of the lane-changing vehicle at the next moment;

[0141] Specifically:

[0142] ; wherein is the longitudinal coordinate of the lane-changing vehicle predicted by the lane-changing physical model at the next moment; is the actual longitudinal coordinate of the lane-changing vehicle at the next moment. ; wherein a longitudinal coordinate of the lane-changing vehicle at a next moment calculated according to a lane-changing physical model, a longitudinal coordinate of the lane-changing vehicle at a current moment a longitudinal instantaneous speed of the lane-changing vehicle at a current moment; a longitudinal instantaneous speed of the lane-changing vehicle at a next moment, a time interval; : a current moment, representing a current time point.

[0143] a residual learning network based on one or more layers of LSTM is constructed, and state information of the lane-changing vehicle and a target vehicle related to the lane-changing vehicle is input into the residual learning network, so as to obtain a horizontal coordinate of the lane-changing vehicle at a next moment and a second residual error of a longitudinal coordinate at the next moment predicted by the residual learning network ; wherein, LSTM is a long short-term memory network.

[0144] based on the , , a first error is obtained; and based on the predicted horizontal coordinate of the lane-changing vehicle at the next moment and a real horizontal coordinate of the lane-changing vehicle at the next moment, a second error is obtained;

[0145] The residual learning network is trained until the absolute values of the first error and the second error tend to 0 and reach a preset precision.

[0146] Optionally, a mapping function of the residual error and the feature value is established: ; wherein, is a parameter of the mapping function, is the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle; ;

[0147] The function is trained to obtain a target parameter , so that the absolute value of the first error tends to 0 and reaches a preset precision.

[0148] Optionally, a mapping function of the vehicle horizontal coordinate and the feature value is established: ; wherein, is a mapping function, is a parameter of the mapping function, is the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, ;

[0149] The function is trained to obtain a target parameter , so that the absolute value of the second error tends to 0 and reaches a preset precision.

[0150] Compared with the prior art, the advantages of the present application are:

[0151] The embodiment scheme improves the accuracy and interpretability of vehicle lane-changing trajectory prediction by fusing a neural network algorithm with a vehicle lane-changing physical model based on parameter calibration.

[0152] Secondly, the application combines a physical model and a residual learning network, provides prior knowledge through the physical model, and the residual learning model is responsible for correcting the prediction error of the physical model. The residual of the physical model is the difference between the observed data and the predicted value, which usually changes little, so the prediction residual is relatively simple and greatly reduces the dependence on a large amount of data. This makes the scheme effective in the case of data scarcity, significantly reducing the demand for large-scale data. In this way, the scheme can maintain high prediction accuracy and stability in the case of insufficient or poor quality data.

[0153] Finally, based on the spatio-temporal correlation feature vector output by the hybrid neural network structure, the prediction value residual output by the residual learning network is bidirectionally modulated to obtain a modulated residual, which on the one hand solves the limitations of fixed threshold filtering in complex scenarios, and on the other hand strengthens the residual component consistent with the historical motion trend, avoiding the violation of vehicle dynamics rules in the correction result; at the same time, the spatio-temporal correlation feature vector can be traced back to a specific historical state (such as a target vehicle deceleration event), so that the residual correction process has physical meaning, greatly enhancing the interpretability of the result.

[0154] Fig. 2 The vehicle lane-changing trajectory prediction method flowchart provided for Embodiment 2 of the present specification is specifically described as follows:

[0155] The embodiment proposes a highway lane-changing vehicle trajectory prediction method based on enhanced physical information residual learning. The core of this method is to combine physical models with data-driven methods, and fully consider the traffic flow characteristics of lane-changing vehicles in the merging area. Through in-depth analysis of vehicle historical data, a physical model suitable for lane-changing behavior in the merging area is identified, and an optimized deep learning network architecture is constructed based on this. This method aims to combine physical models with data-driven prediction to improve the accuracy and interpretability of highway lane-changing vehicle trajectory prediction.

[0156] Specifically, the method flow of the embodiment is as shown in Fig. 2 , and specifically includes the following steps:

[0157] Step 201: Obtain historical state information of the lane-changing vehicle and surrounding vehicles;

[0158] Select the lane-changing vehicle in the merging area of the highway and the front vehicle in the current lane and the front and rear vehicles in the target lane as the research object.

[0159] Extract the historical state information of the lane-changing vehicle in the current lane, the front vehicle in the current lane, the front vehicle in the target lane, and the rear vehicle in the target lane, and the time window for data collection is 5s before lane changing to 5s after lane changing, and the time interval is 0.1s.

[0160] Specifically, the following parameters are included: the vehicle length of the lane-changing vehicle in the current lane, the current time instant lateral and longitudinal instantaneous speed and acceleration of the lane-changing vehicle, the current time instant lateral and longitudinal coordinates of the lane-changing vehicle; the vehicle length of the front vehicle in the current lane, the current time instant lateral and longitudinal instantaneous speed and acceleration of the front vehicle in the current lane, the current time instant lateral and longitudinal coordinates of the front vehicle in the current lane; the vehicle length of the front vehicle in the target lane, the current time instant lateral and longitudinal instantaneous speed and acceleration of the front vehicle in the target lane, the current time instant lateral and longitudinal coordinates of the front vehicle in the target lane; the current time instant lateral and longitudinal coordinates of the rear vehicle in the target lane, the current time instant lateral and longitudinal instantaneous speed and acceleration of the rear vehicle in the target lane.

[0161] Step 202: Build a lane-changing physical model;

[0162] Build a lane-changing physical model, and select a vehicle lane-changing model applying Gipps general collision avoidance algorithm.

[0163] Step 203: Calibrate the parameters of the lane-changing physical model using genetic algorithm;

[0164] The specific steps are as follows:

[0165] (1) Determine the fitness function

[0166]

[0167] The average value (MSE) of the error square sum between the next time instant longitudinal coordinate of the lane-changing vehicle predicted by the lane-changing physical model and the actual next time instant longitudinal coordinate of the lane-changing vehicle is used as the fitness function, which is used to measure the pros and cons of the prediction accuracy of the model; is the time interval, is the sample number.

[0168] (2) Initialize the population, the population size is 100, and each individual is the parameter to be calibrated ( , , , , , ), the value range of the parameter is (-3.5, 0), (-3.5, 0), (-3.5, 0), (-3.5, 0), (0, 1), (0, 1);

[0169] (3) Selection operation, using the method of tournament selection, the specific operation is to randomly select two or more individuals from the population, compare their fitness, and select the individual with the best fitness;

[0170] (4) Cross operation, for the selected parent individual, each parameter is weighted and fused according to the crossover rate;

[0171] (5) Mutation operation, add Gaussian noise to each parameter according to the mutation rate, and the mutation range is adaptive to the parameter range;

[0172] (6) Elite reservation, the individual with the best fitness of each generation is directly entered into the next generation.

[0173] The specific execution steps are as follows: first, initialize the population for each generation individual: calculate the fitness (MSE), record the current optimal solution, then execute the selection, crossover and mutation operations to generate a new population, then merge the elite individuals, output the optimal parameters and minimum MSE of the final generation.

[0174] Step 204: Calculate the residual error and the longitudinal coordinate of the next time vehicle of the lane-changing vehicle using the calibrated lane-changing physical model;

[0175] The collision avoidance algorithm with the front vehicle in the current lane is considered, and the specific calculation formula is as follows:

[0176] wherein, is the instantaneous speed of the lane-changing vehicle at the next time to avoid collision with the front vehicle in the current lane, is the vehicle length of the front vehicle in the current lane, is the x-coordinate of the front vehicle in the current lane at the current time, is the x-coordinate of the lane-changing vehicle at the current time, is the instantaneous speed of the lane-changing vehicle at the current time, is the instantaneous speed of the front vehicle in the current lane at the current time, and are the maximum deceleration of the lane-changing vehicle and the front vehicle in the current lane, : current time, indicating the current time point, is the time interval.

[0177] The collision avoidance algorithm with the front vehicle in the target lane is considered, and the calculation formula is as follows:

[0178] wherein the instantaneous speed of the vehicle at the next moment in order to avoid collision with the front vehicle in the target lane, the current moment x coordinate of the front vehicle in the target lane, the vehicle length of the front vehicle in the target lane, the maximum deceleration of the front vehicle in the target lane, the time interval, and other parameters, please refer to the description above.

[0179] Considering the collision avoidance algorithm with the rear vehicle in the target lane, the calculation formula is as follows:

[0180] wherein the instantaneous speed of the vehicle at the next moment in order to avoid collision with the rear vehicle in the target lane, the vehicle length of the vehicle, the current moment x coordinate of the rear vehicle in the target lane, the current moment instantaneous speed of the rear vehicle in the target lane, the maximum deceleration of the rear vehicle in the target lane, the time interval, and other parameters, please refer to the description above.

[0181] The longitudinal instantaneous speed of the vehicle at the next moment is:

[0182]

[0183] wherein, represents the influence of the front vehicle in the target lane on the speed of the vehicle, and its value range is a constant greater than or equal to 0 and less than or equal to 1, represents the proportion of the influence of the current lane and the target lane on the speed of the vehicle, and its value range is a constant greater than or equal to 0 and less than or equal to 1, the time interval, and other parameters, please refer to the description above.

[0184] Using kinematic formula, the longitudinal coordinate of the vehicle at the next moment is obtained, and the calculation formula is as follows:

[0185]

[0186] wherein, represents the longitudinal coordinate of the vehicle at the next moment calculated by the lane changing physical model, the current moment longitudinal coordinate of the vehicle, the current moment longitudinal instantaneous speed of the vehicle, the longitudinal instantaneous speed of the vehicle at the next moment, the time interval, : the current moment, indicating the current time point.

[0187] residual error is the difference between the predicted longitudinal coordinate of the lane-changing vehicle at the next moment based on the lane-changing physical model and the true longitudinal coordinate of the lane-changing vehicle at the next moment:

[0188] Step 205: building a residual learning network and training;

[0189] A residual learning network based on one or more layers of LSTM is built and trained, and the input feature values are the state information of the vehicle , at least including: the state information of the lane-changing vehicle in the current lane, the front vehicle in the current lane, the front vehicle in the target lane, and the rear vehicle in the target lane.

[0190] The state information of the lane-changing vehicle in the current lane, the front vehicle in the current lane, the front vehicle in the target lane, and the rear vehicle in the target lane at least includes: the vehicle length of the lane-changing vehicle in the current lane, the current moment instantaneous speed and instantaneous acceleration in the horizontal and vertical directions of the lane-changing vehicle, the current moment horizontal and vertical coordinates of the lane-changing vehicle; the vehicle length of the front vehicle in the current lane, the current moment instantaneous speed and instantaneous acceleration in the horizontal and vertical directions of the front vehicle in the current lane, the current moment horizontal and vertical coordinates of the front vehicle in the current lane; the vehicle length of the front vehicle in the target lane, the current moment instantaneous speed and instantaneous acceleration in the horizontal and vertical directions of the front vehicle in the target lane, the current moment horizontal and vertical coordinates of the front vehicle in the target lane; the current moment horizontal and vertical coordinates of the rear vehicle in the target lane, the current moment instantaneous speed and instantaneous acceleration in the horizontal and vertical directions of the rear vehicle in the target lane.

[0191] The output target value of the residual learning network is the horizontal coordinate of the lane-changing vehicle at the next moment and the residual error at the next moment .

[0192] The data is divided into a training set, a validation set and a test set, the residual learning network is trained by the training set, and MSE is used as the loss function to find an optimal , so that the error with the true x is minimized, and an optimal is found, so that the error with the true value is minimized.

[0193]

[0194]

[0195] Specifically:

[0196] A mapping function of the residual error and the feature value is established: ; wherein, is the parameter of the mapping function, ​​state information of the lane-changing vehicle and a target vehicle related to the lane-changing vehicle; ;

[0197] The function obtains target parameters through training , so that the absolute value of the first error , the error of the true value , approaches 0 and reaches a preset accuracy.

[0198] A mapping function of the vehicle horizontal coordinate and the characteristic value is established: ; wherein, is the mapping function, is a parameter of the mapping function, is state information of the lane-changing vehicle and a target vehicle related to the lane-changing vehicle, ;

[0199] The function obtains target parameters through training , so that the absolute value of the second error , the error of the true x, approaches 0 and reaches a preset accuracy.

[0200] Step 206: using the trained residual learning network to predict the residual and the horizontal coordinate of the lane-changing vehicle at the next moment;

[0201] Using the trained model, the residual is predicted, the characteristic values of the vehicle in the previous 5s are input, and the predicted values of the vehicle in the next 5s are obtained and the horizontal coordinate at the next moment.

[0202] Step 207: based on the obtained state information of the lane-changing vehicle and a target vehicle related to the lane-changing vehicle, a spatio-temporal correlation feature vector is obtained through a hybrid neural network structure; wherein the hybrid neural network structure is composed of a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM and an attention mechanism Attention;

[0203] Step 208: based on the obtained spatio-temporal correlation feature vector, the predicted residual output by the residual learning network is bidirectionally modulated to obtain a modulated residual; wherein the process of bidirectional modulation is: sequentially performing abnormal residual filtering and effective residual enhancement.

[0204] Optionally, the process of abnormal residual filtering is:

[0205] A gating function is set: ; the expression form of the gating function is: ; wherein, is the spatio-temporal correlation feature vector; It is a nonlinear activation function, namely the sigmoid function, whose output range is between (0,1), that is, a real number greater than 0 and less than 1, and is used to map the output value to a probability interval; and Parameters of the gating function: is the weight matrix used to Perform a linear transformation to Map to a new space; is the bias term used to In the pair After the linear transformation, an offset is added; and Each element in is a real number;

[0206] when , then it is determined that the predicted value residual contains abnormal components; among them, When it is a preset threshold, it is used to determine whether the residual of the predicted value is abnormal;

[0207] Based on obtaining the predicted value residual, , obtain the predicted value residual after suppression ;in, = , is the residual of the predicted value; The value of is between (0,1), that is, a real number greater than 0 and less than 1.

[0208] because The value of is between (0,1), when it is less than the threshold When , its value will be relatively small, multiplied by After that, the residual value will be reduced, thereby achieving the effect of suppressing outliers and improving the robustness and accuracy of the model.

[0209] Optionally, the process of effective residual enhancement is:

[0210] Calculate the With from the The historical movement trend vector extracted from The cosine similarity between ;in, The expression is:

[0211] ; It is the cosine similarity function, which is used to measure the similarity between two vectors. Its value range is between [−1,1], that is, a real number greater than or equal to −1 and less than or equal to 1. The closer the value is to 1, the more similar the two vectors are; the closer the value is to -1, the less similar the two vectors are.

[0212] based on cosine similarity , the obtaining an enhanced predicted value residual ; wherein the is the suppressed predicted value residual, is the cosine similarity between the historical motion trend vector extracted from the historical trajectory trend; is an enhancement coefficient, and is a real number greater than 0, used to control the intensity of enhancement.

[0213] This method enhances the effective residual component by similarity with the historical trajectory trend, thereby improving the model's ability to capture effective information in the data.

[0214] Step 209: Based on the modulated residual, correcting the longitudinal coordinate of the next time lane-changing vehicle predicted by the lane-changing physical model to obtain a corrected longitudinal coordinate of the next time lane-changing vehicle;

[0215] Optionally, based on the modulated residual, the longitudinal coordinate of the next time lane-changing vehicle predicted by the lane-changing physical model is corrected to obtain a corrected longitudinal coordinate of the next time lane-changing vehicle.

[0216] Based on the following formula, the corrected longitudinal coordinate of the next time lane-changing vehicle is obtained ; ; wherein, is the longitudinal coordinate of the next time lane-changing vehicle predicted based on the lane-changing physical model, is the current time, representing the current time point, is the time interval, is the enhanced predicted value residual; is an adaptive coefficient based on spatiotemporal features, calculated by a multi-layer perceptron (MLP) layer: , is a spatiotemporal correlation feature vector, and are parameters of the multi-layer perceptron (MLP) layer: is a weight matrix, used for linear transformation of to map to a new space; is a bias term, used to add an offset to after linear transformation of ; and each element is a real number; It is a nonlinear activation function, namely the sigmoid function, whose output range is between (0,1), that is, a real number greater than 0 and less than 1, and is used to map the output value to a probability interval.

[0217] The implementation scheme is to fuse the predicted value of the lane change physical model and the modulated residual. This process is achieved through an adaptive coefficient To adjust the impact of residuals to ensure the stability and accuracy of the model under different circumstances.

[0218] Adaptive coefficient The behavior is as follows:

[0219] When the historical trajectory is stable: When the historical trajectory is stable, It approaches 1. This means that the contribution of the residual is close to the original value, and the combination of the predicted value of the lane change physics model and the residual is more direct.

[0220] When the scene suddenly changes: When the scene suddenly changes, Automatically reduced. This means that the contribution of the residuals is weakened to maintain the stability of the model. This helps to avoid excessive influence of the residuals on the prediction results in the case of sudden changes.

[0221] Step 210: The predicted horizontal coordinate of the vehicle that will change lanes at the next moment and the corrected vertical coordinate of the vehicle that will change lanes at the next moment are used as the predicted coordinates of the vehicle that will change lanes at the next moment.

[0222] Finally, the predicted coordinates of the lane-changing vehicle are calculated as follows ( (t+ , ); here: = ;in, = + × .

[0223] The effect of the method of this embodiment is verified by specific experiments below.

[0224] To verify the predictive performance of the prediction method proposed in this invention, the NGSIM dataset was selected as the experimental data source. To ensure that the model has sufficient training samples and fully learns the data characteristics, the dataset was divided into training, validation, and test sets in a ratio of 6:2:2.

[0225] The model input is the features of a target vehicle and four surrounding vehicles (such as Fig. 3 shown).

[0226] 6 features for each vehicle (including: the lateral and longitudinal coordinates of the target vehicle, the target lane vehicles, and the current lane vehicles, the lateral and longitudinal instantaneous velocities, and the lateral and longitudinal instantaneous accelerations, i.e., (x, y, ), and a total of 30 dimensions of feature vectors. To balance the computational complexity and prediction accuracy of the model, 100 time steps (10 s) were used as a sample, with the first 50 time steps (5 s) as the historical time steps and the last 50 time steps (5 s) as the prediction time steps, and the prediction step was 0.1 s.

[0227] Combined with the lane-changing physical model, the residual learning network based on single-layer LSTM (Long Short-Term Memory Network), and the hybrid neural network structure, the predicted coordinates of the lane-changing vehicle were obtained. Finally, after the model training and prediction, the evaluation indicators were used to compare the prediction accuracy of the model.

[0228] The overall root mean square error (RMSE), the lateral position root mean square error (LaP_RMSE), the longitudinal position root mean square error (LoP_RMSE), the lateral position mean absolute percentage error (LaP_MAPE), and the longitudinal position mean absolute percentage error (LoP_MAPE) within the next 5 s were selected as the evaluation indicators. The calculation formula is as follows: where, is the time interval, and other parameters are described above.

[0229]

[0230] where N is the number of samples;

[0231] T is the length of the time series, i.e., how many data points there are in the time dimension. Here T−τ represents the range from time point t to time point t+τ in the time series, and how many time points are used to calculate the error; here I is a positive real number (such as 1) that adjusts the proportion. If it is 1, then this formula is the average of the RMSE of all samples; at time point , the squared error between the x value predicted by the LSTM model: and the true x value: ; at time point , the squared error between the y value predicted by the PERL model: and the true y value: .

[0232] ; the parameter descriptions involved in the formula are described above.

[0233] ; the parameter descriptions involved in the formula are described above.

[0234] ​ 100%; the parameter descriptions involved in the formula are described above.

[0235] ; wherein, ; The parameter descriptions involved in the formula are described above.

[0236]

[0237] ;

[0238] ;

[0239] The parameter descriptions involved in the formula are described above.

[0240]

[0241] wherein is the real horizontal coordinate of the lane-changing vehicle at the next moment, is the real longitudinal coordinate of the lane-changing vehicle at the next moment.

[0242] As Fig. 4 shown is a comparison diagram of 50-step predicted values and real values of the highway merging area lane-changing trajectory prediction model (Gipps-LSTM) based on the enhanced physical information residual learning (PERL) of the application, and the prediction result shows that the predicted trajectory is highly consistent with the real trajectory. In the complex vehicle lane-changing scene, the model can accurately predict the motion trend of the vehicle, and has strong robustness. In addition, the predicted trajectory shows good smoothness without obvious jumps or deviations, indicating that the model has successfully captured the continuity feature of the vehicle motion.

[0243] In order to compare the prediction performance of the method, seven prediction models are selected for comparative analysis. Specifically, they include: recurrent neural network (RNN), long short-term memory network (LSTM), graph convolutional network (GCN), trajectory prediction method considering dynamic coupling of space-time features (Spatial-Cross Attention Mechanism-LSTM), and Gipps-RNN, Gipps-GCN, Gipps-Spatial-Cross Attention Mechanism-LSTM which combine enhanced physical information residual learning with RNN, GCN, Spatial-Cross Attention Mechanism-LSTM.

[0244] The performance evaluation results of each model on the test set are shown in Table 1. Fig. 5 As shown in Table 1, the RMSE of the model of the present application on 5s prediction is 0.4411m, the LaP-RMSE is 0.6187m, and the LoP-RMSE is 0.0795m, which are significantly better than the pure data-driven model. For example, compared with LSTM, the error of the model of the present application on RMSE and LoP-RMSE is reduced by 89.8% and 98.7%, respectively. These results show that the model of the present application not only has higher advantages in overall trajectory prediction accuracy, but also can better capture the local and global features of the trajectory. At the same time, compared with other prediction models, the MAPE of the model of the present application is 0.0253, and the R² is 0.9404, which are better than other models. The high R² value indicates that the model can better fit the data and explain most of the variability, showing strong prediction ability; while the low MAPE value means that the prediction error is small, and the accuracy of the model at each data point is high. Therefore, the model is better than other models in fitting degree and actual prediction accuracy, and has higher practical application value.

[0245] The scheme of the present embodiment improves the accuracy and interpretability of vehicle lane-changing trajectory prediction by fusing neural network algorithm with vehicle lane-changing physical model based on parameter calibration.

[0246] Secondly, the application combines a physical model and a residual learning network, provides prior knowledge through the physical model, and the residual learning model is responsible for correcting the prediction error of the physical model. The residual of the physical model is the difference between the observation data and the predicted value, which usually changes little, so the prediction residual is relatively simple and greatly reduces the dependence on a large amount of data. This makes the scheme effective in the case of data scarcity, significantly reducing the demand for large-scale data. In this way, the scheme can still maintain high prediction accuracy and stability in the scene of insufficient or poor quality data.

[0247] Finally, based on the spatio-temporal correlation feature vector output by the hybrid neural network structure, the residual learning network outputs the prediction value residual, and the modulated residual is obtained, which on the one hand solves the limitations of fixed threshold filtering in complex scenes, and on the other hand strengthens the residual component consistent with the historical motion trend, avoids the violation of the vehicle dynamics law of the correction result; at the same time, the spatio-temporal correlation feature vector can be traced back to a specific historical state (such as a target vehicle deceleration event), so that the residual correction process has physical meaning, greatly enhancing the explainability of the result.

[0248] Fig. 6 The vehicle lane change trajectory prediction device schematic diagram provided for embodiment 6 of the present application comprises: a first prediction module 601, a second prediction module 602, a modulation module 603, a correction module 604, and a determination module 605.

[0249] The first prediction module 601 is configured to acquire the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, and acquire the longitudinal coordinate of the lane-changing vehicle at the next moment based on the lane-changing physical model.

[0250] The second prediction module 602 is configured to acquire the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, and obtain the predicted horizontal coordinate of the lane-changing vehicle at the next moment and the prediction value residual through the trained residual learning network ResNet.

[0251] The modulation module 603 is configured to acquire the state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, and obtain the spatio-temporal correlation feature vector through the hybrid neural network structure; wherein the hybrid neural network structure is composed of a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM, and an attention mechanism Attention; and is further configured to perform bidirectional modulation on the prediction value residual output by the ResNet based on the obtained spatio-temporal correlation feature vector, and obtain the modulated residual; wherein the process of bidirectional modulation is: sequentially performing abnormal residual filtering and effective residual enhancement.

[0252] The correction module 604 is configured to correct the longitudinal coordinate of the next-time lane-changing vehicle predicted by the lane-changing physical model based on the modulated residual error, to obtain a corrected longitudinal coordinate of the next-time lane-changing vehicle.

[0253] The determination module 605 is configured to take the predicted longitudinal coordinate of the next-time lane-changing vehicle and the corrected longitudinal coordinate of the next-time lane-changing vehicle as the predicted coordinates of the next-time lane-changing vehicle.

[0254] The scheme of the embodiment improves the accuracy and interpretability of vehicle lane-changing trajectory prediction by fusing the neural network algorithm with the vehicle lane-changing physical model based on parameter calibration.

[0255] Secondly, the present application combines the physical model and the residual error learning network, the physical model provides prior knowledge, and the residual error learning model is responsible for correcting the prediction error of the physical model. The residual error of the physical model is the difference between the observed data and the predicted value, which usually changes little, so the prediction residual error is relatively simple, and the dependence on a large amount of data is greatly reduced. This makes the scheme can effectively run in the case of data shortage, significantly reducing the demand for large-scale data. In this way, the scheme can still maintain high prediction accuracy and stability in the scene of insufficient or poor quality data.

[0256] Finally, the present application modulates the prediction residual error output by the residual error learning network based on the spatio-temporal correlation feature vector output by the hybrid neural network structure, to obtain a modulated residual error, which on the one hand solves the limitations of fixed threshold filtering in complex scenes, and on the other hand strengthens the residual error component consistent with the historical motion trend, avoiding the violation of vehicle dynamics law in the correction result; at the same time, the spatio-temporal correlation feature vector can be traced back to a specific historical state (such as a target vehicle deceleration event), so that the residual error correction process has physical meaning, greatly enhancing the interpretability of the result.

[0257] The present specification also provides a vehicle lane-changing trajectory prediction system, comprising a processor and a memory, wherein the memory stores a program, and the processor executes the program to realize the vehicle lane-changing trajectory prediction method described in Embodiment 1 or the vehicle lane-changing trajectory prediction method described in Embodiment 2.

[0258] Fig. 7 An electronic device according to the embodiment of the present application is provided. Fig. 7 The electronic device 300 according to the embodiment of the present application is described below with reference to Fig. 7 The electronic device 300 shown is merely an example and should not impose any limitation on the functions and use range of the embodiment of the present application.

[0259] As Fig. 7As shown, the electronic device 300 is in the form of a general computing device. The components of the electronic device 300 can include, but are not limited to, at least one processing unit 310, at least one storage unit 320, a bus 330 that connects the various system components, including the storage unit 320 and the processing unit 310, a display unit 340, and the like.

[0260] The storage unit stores program code that can be executed by the processing unit 310 such that the processing unit 310 performs the steps described in the above method section of this specification in accordance with the various exemplary embodiments of the present application. For example, the processing unit 310 can execute the steps shown in Fig. 1

[0261] The storage unit 320 can include a readable medium in the form of volatile storage such as a random access memory (RAM) 3201 and / or cache 3202, and can further include a read-only memory (ROM) 3203.

[0262] The storage unit 320 can also include program / utility 3204 having a set of programs / modules 3205, including an operating system, one or more application programs, other programs, and programmatic data, each or some combination thereof, which can include implementation of a network environment.

[0263] The bus 330 can represent one or more of several types of bus structures, including a storage bus or bus controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of a variety of bus architectures.

[0264] The electronic device 300 can also communicate with one or more external devices 400 such as a keyboard or pointing device, a Bluetooth device, etc.; other devices that enable a user to interact with the electronic device 300; and / or one or more devices that enable the electronic device 300 to communicate with one or more other computing devices. Such communication can be via an input / output (I / O) interface 350. Further, the electronic device 300 can communicate with one or more networks, such as a local area network (LAN), a wide area network (WAN), and / or the public network, such as the Internet, via a network adapter 360. The network adapter 360 can communicate with the other components of the electronic device 300 via the bus 330.

[0265] ​Through the description of the above embodiments, it is easy for those skilled in the art to understand that the exemplary embodiments described in the present invention can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the embodiment of the present invention can be embodied in the form of a software product, which can be stored in a computer-readable storage medium (which can be a CD-ROM, USB flash drive, mobile hard disk, etc.) or on a network, and includes a number of instructions to enable a computing device (which can be a personal computer, server, or network device, etc.) to execute the above method according to the present invention. When the computer program is executed by a data processing device, the computer-readable medium is enabled to implement the above method of the present invention, that is: Fig. 1 The method shown.

[0266] Fig. 8 A schematic diagram of a computer-readable medium provided in accordance with an embodiment of this specification.

[0267] accomplish Fig. 1 The computer program of the illustrated method can be stored on one or more computer-readable media. The computer-readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0268] The computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave, wherein the readable program code is carried. The data signal propagated may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or component. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination thereof.

[0269] The program code may be implemented in any of various ways, including procedure-based, object-oriented languages, such as Java, C++, etc., and traditional process-based languages, such as the "C" programming language. The program code can execute entirely on the viewer's computing device, partly on the viewer's device, as a stand-alone software package, partly on the viewer's computing device and partly on a remote computing device or entirely on the remote computing device or server. In the latter scenario, the remote computing device can be connected to the viewer's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device such as through the Internet using an Internet Service Provider. The application is not limited to a particular programming language. The program code can be implemented in assembly or machine language, if desired.

[0270] In light of the above, the present application can be implemented in hardware, or as software modules running on one or more processors, or as a combination of both. Those ordinarily skilled in the art will appreciate that the various components of some or all of the embodiments according to the present application can be implemented in whole or in part in a general purpose data processing device such as a microprocessor or digital signal processor (DSP). The present application can also be implemented as a program for executing part or all of the methods described herein on a device or apparatus such as a computer or computer program product. Such program implementing the present application can be stored on a computer readable medium or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, or provided on a carrier medium, or in any other form.

[0271] The above specific embodiments are described in order to explain the purpose, technical solutions and beneficial effects of the present application. It should be understood that the present application is not inherently related to any specific computer, virtual device or electronic device, and various general purpose devices can also implement the present application. The above description is only for specific embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

[0272] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments.

[0273] The above description is only for specific embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various changes and modifications to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A vehicle lane-changing trajectory prediction method, characterized in that, The method comprises the following steps: Based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the longitudinal coordinate of the lane-changing vehicle at the next moment predicted based on the lane-changing physical model is obtained through the calibrated lane-changing physical model; Based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the predicted lateral coordinate of the lane-changing vehicle at the next moment and the predicted value residual are obtained through the trained residual learning network ResNet; Based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the spatio-temporal correlation feature vector is obtained through the hybrid neural network structure; wherein the hybrid neural network structure is composed of a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM and an attention mechanism Attention; Based on the obtained spatio-temporal correlation feature vector, the predicted value residual output by the ResNet is bidirectionally modulated to obtain a modulated residual; wherein the bidirectional modulation process is: sequentially performing abnormal residual filtering and effective residual enhancement; Based on the modulated residual, the longitudinal coordinate of the lane-changing vehicle at the next moment predicted by the lane-changing physical model is corrected to obtain the corrected longitudinal coordinate of the lane-changing vehicle at the next moment; The predicted lateral coordinate of the lane-changing vehicle at the next moment and the corrected longitudinal coordinate of the lane-changing vehicle at the next moment are taken as the predicted coordinates of the lane-changing vehicle at the next moment.

2. The method of claim 1, wherein, The process of obtaining the spatio-temporal correlation feature vector based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle through the hybrid neural network structure is as follows: Based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, the spatial interaction features of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle are obtained through the convolutional neural network CNN; Based on the spatial interaction features, the time sequence features of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle are obtained through the BiLSTM; The time sequence features output by the BiLSTM are weighted through the Attention to generate the spatio-temporal correlation feature vector.

3. The method of claim 1, wherein, The process of abnormal residual filtering is as follows: Setting the gating function: The expression form of the gating function is: Where, is the spatiotemporal correlation feature vector; is a nonlinear activation function, i.e., a sigmoid function, whose output range is between (0, 1), used to map the output value to a probability interval; and are parameters of the gating function: is a weight matrix, used to perform linear transformation on to map to a new space; is a bias term, used to add an offset to after linear transformation on ; and each element in and is a real number; When , then determining that the prediction residual contains abnormal components; wherein, is a preset threshold, for determining whether the prediction residual is abnormal. based on the obtained prediction value residual, , obtain the suppressed prediction value residual ; wherein, = , is the prediction value residual; the value of is between (0, 1).

4. The method of claim 3, wherein, The process of effective residual enhancement is as follows: computing the cosine similarity between the history motion trend vector extracted from the ; wherein is expressed as:​ ; is a cosine similarity function used to measure the similarity between two vectors, whose value ranges between [−1, 1], the closer the value to 1, the more similar the two vectors are; the closer the value to -1, the less similar the two vectors are; based on a cosine similarity , the , obtaining an enhanced prediction value residual ; wherein the is a suppressed prediction value residual, is a cosine similarity between the and a historical motion trend vector extracted from the ; is an enhancement coefficient, and is a real number greater than 0, for controlling the intensity of enhancement.

5. The method of claim 4, wherein, The process of correcting the longitudinal coordinate of the lane-changing vehicle at the next moment predicted by the lane-changing physical model based on the modulated residual to obtain the corrected longitudinal coordinate of the lane-changing vehicle at the next moment is as follows: The corrected longitudinal coordinate of the lane-changing vehicle at the next moment is obtained based on the following formula ; ; wherein, is the longitudinal coordinate of the lane-changing vehicle at the next moment predicted based on a lane-changing physical model, is the current moment, representing the current time point, is the time interval, is the enhanced predicted value residual; is an adaptive coefficient based on the spatiotemporal features, calculated by a multi-layer perception (MLP) layer: , is a spatiotemporal correlation feature vector, and are parameters of the multi-layer perception (MLP) layer: is a weight matrix used for linear transformation of to map to a new space; is a bias term used to add an offset to after linear transformation of ; and each element is a real number; is a nonlinear activation function, i.e., a sigmoid function, whose output range is between (0, 1), used to map the output value to a probability interval.

6. The method of claim 1, wherein, The training process of the residual learning network is as follows: acquiring a first residual error of the longitudinal coordinate at the next moment ; wherein the is a difference between the longitudinal coordinate of the lane-changing vehicle at the next moment predicted based on the lane-changing physical model and the real longitudinal coordinate of the lane-changing vehicle at the next moment; The residual learning network based on one or more layers of LSTM is constructed, and state information of a lane-changing vehicle and a target vehicle related to the lane-changing vehicle is input into the residual learning network, so as to obtain a horizontal coordinate of the lane-changing vehicle at a next moment and a second residual error of a vertical coordinate at the next moment predicted by the residual learning network ; based on the , , a first error is obtained; and a second error is obtained based on the predicted lateral coordinate of the next-time lane-changing vehicle and the real lateral coordinate of the next-time lane-changing vehicle; The residual learning network is trained until the absolute values of the first error and the second error tend to 0 and reach the preset precision.

7. The method of claim 6, wherein, establishing a mapping function of residual error and eigenvalue: ; wherein, is a parameter of the mapping function, is state information of the lane-changing vehicle, a target vehicle related to the lane-changing vehicle; ; The function obtains the target parameter through training So that the absolute value of the first error approaches 0 and reaches a preset precision.

8. The method of claim 6, wherein, A mapping function of vehicle lateral coordinate and characteristic value is established: ; wherein, is a mapping function, is a parameter of the mapping function, is state information of a lane-changing vehicle, a target vehicle related to the lane-changing vehicle, ; The function obtains the target parameter through training So that the absolute value of the second error approaches 0 and reaches a preset precision.

9. A vehicle lane-changing trajectory prediction device characterized by comprising: The method comprises the following steps: A first prediction module, a second prediction module, a modulation module, a correction module and a determination module are included. The first prediction module is used to obtain the longitudinal coordinate of the lane-changing vehicle at the next moment predicted based on the lane-changing physical model based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle through the calibrated lane-changing physical model. The second prediction module is configured to obtain a predicted horizontal coordinate of the lane-changing vehicle at a next moment and a prediction residual based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, and a trained residual learning network ResNet. The modulation module is configured to obtain a space-time correlation feature vector based on the obtained state information of the lane-changing vehicle and the target vehicle related to the lane-changing vehicle, and a hybrid neural network structure, wherein the hybrid neural network structure is composed of a convolutional neural network CNN, a bidirectional long short-term memory network BiLSTM, and an attention mechanism Attention; and to perform bidirectional modulation on the prediction residual output by the ResNet based on the obtained space-time correlation feature vector, to obtain a modulated residual, wherein the bidirectional modulation process includes sequentially performing abnormal residual filtering and effective residual enhancement. The correction module is configured to correct a longitudinal coordinate of the lane-changing vehicle at the next moment predicted by the lane-changing physical model based on the modulated residual, to obtain a corrected longitudinal coordinate of the lane-changing vehicle at the next moment. The determination module is configured to take the predicted horizontal coordinate of the lane-changing vehicle at the next moment and the corrected longitudinal coordinate of the lane-changing vehicle at the next moment as a predicted coordinate of the lane-changing vehicle at the next moment. 10.A vehicle lane-changing trajectory prediction system, comprising a processor and a memory, wherein the memory has stored therein a program, characterized in that, The processor implements the method in any one of claims 1 to 8 when executing the program.

Images