A Deep Learning Vehicle Trajectory Prediction Method Incorporating a Kinematic Model

Abstract

The present invention discloses a deep learning vehicle trajectory prediction method integrating a kinematic model, which includes the following steps: obtaining the surrounding scene information of an autonomous vehicle and converting it into the coordinate system of the target vehicle; using a deep learning model to establish the historical motion state information and interaction relationship of the vehicle, and predicting the centroid acceleration and front wheel steering angle of the target vehicle; classifying vehicle types according to the vehicle length and constraining the predicted values of the centroid acceleration and front wheel steering angle of the vehicle, and calculating the trajectory coordinates of the target vehicle from the predicted values using a vehicle two-degree-of-freedom model. The present invention integrates the vehicle two-degree-of-freedom motion model into the deep learning prediction model, taking into account both the high-precision long-time-domain prediction performance of deep learning and the kinematic properties of the vehicle, so that the predicted trajectory results are more in line with the real vehicle motion law. The present invention can improve the feasibility and accuracy of the predicted trajectory, reduce the situation that the main vehicle planning is too conservative due to infeasible prediction results, and improve the driving safety and smoothness of autonomous vehicles.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving, and in particular to a deep learning vehicle trajectory prediction method integrating a kinematic model. Background Art

[0002] In recent years, autonomous vehicles have become an important development direction for the intelligentization of modern automobiles. Autonomous vehicles not only need to perceive the surrounding scene information in real time, but also need to understand the scene information and predict the changing trend of the surrounding traffic scene in the future. Currently, autonomous vehicles predict the future trend of traffic scenes by predicting the most likely trajectory of surrounding vehicles in the future.

[0003] The research methods of trajectory prediction can be divided into two categories:

[0004] (1) Trajectory prediction method based on physical model: This method first assumes that the model conforms to a certain vehicle kinematic model or vehicle dynamics model, and predicts the trajectory based on the predicted motion state of the target vehicle at a historical moment. It is generally used for short-term prediction tasks within 2 seconds, such as collision warning in assisted driving systems. Based on the physical model, it can only predict the trajectory in specific simple scenarios, and does not model the interaction relationship between intelligent agents; the prediction trajectory error accumulates and increases over time, which cannot meet the long-term prediction requirements for complex interaction scenarios in autonomous driving.

[0005] (2) Trajectory prediction method based on machine learning: This method learns the motion patterns of vehicles in traffic scenes and predicts the most likely trajectory pattern based on historical motion status and surrounding scenes. Among them, the trajectory prediction method based on deep learning can extract scene information and potential interactions, and can perform trajectory prediction of multiple motion modes for a long time in complex scenes, which can basically meet the prediction time and accuracy requirements of autonomous driving.

[0006] At present, the mainstream trajectory prediction method based on deep learning directly predicts the future trajectory coordinates of the target vehicle without considering the kinematic and dynamic properties of the target vehicle and does not distinguish between different types of vehicles. As a result, some predicted trajectories do not conform to the physical motion laws of the actual vehicle and have poor feasibility. Summary of the invention

[0007] The purpose of the present invention is to provide a deep learning vehicle trajectory prediction method that integrates a kinematic model to solve the problem in the prior art that it is impossible to take into account both long-term trajectory prediction and the kinematic properties of different types of vehicles, and to improve the accuracy and feasibility of the predicted trajectories of autonomous driving vehicles for different surrounding vehicles in traffic scenarios.

[0008] To achieve the above object, the technical solution of the present invention is as follows: A deep learning vehicle trajectory prediction method integrating a kinematic model, comprising the following steps:

[0009] A. Obtain the surrounding scene information of the autonomous vehicle and convert it to the target vehicle coordinate system

[0010] A1. Obtain the surrounding road vector information

[0011] Obtain the road structure information around the autonomous vehicle vehicle ego from the high-precision map, where the road structure information includes the position coordinates of the lane center line, lane boundary line, and crosswalk area;

[0012] Divide the road structure information into multiple equidistant road segments, and extract the starting point coordinate a, ending point coordinate b, and direction c from each road segment to jointly form the surrounding road vector information roadvector nbrs ={a, b, c};

[0013] A2. Obtain the historical motion state information of the autonomous vehicle

[0014] Obtain the historical motion state information of the autonomous vehicle vehicle ego at H time steps in the past T h seconds according to the sampling frequency f:

[0015]

[0016] The historical state information at the h-th time step of the historical motion state information is where the value range of h is from -H + 1 to 0, L ego and B ego are the longitude and latitude of the autonomous vehicle in the geodetic coordinate system respectively, ψ ego is the yaw angle of the autonomous vehicle, v ego is the centroid velocity of the autonomous vehicle, l ego , w ego are the length and width of the autonomous vehicle respectively;

[0017] A3. Obtain the historical motion state information of the surrounding vehicles of the autonomous vehicle

[0018] Real-time detect M surrounding vehicles vehicle nbrs ={vehicle1, vehicle2,.., vehicle M} through the target detection module of the autonomous vehicle environment perception system, and obtain the past T of each surrounding vehicle through the tracking moduleh Historical motion state information state within seconds nbrs ={state1, state2,..., state M}, go to step A4;

[0019] The historical motion state information of each surrounding vehicle is obtained at H time steps in the past T seconds according to the sampling frequency f, where the value range of j is from 1 to M, and the historical motion state information of the jth surrounding vehicle is h The motion state information of the jth surrounding vehicle at the hth time step is where L and B j are the longitude and latitude of the jth surrounding vehicle in the geodetic coordinate system respectively, ψ j is the yaw angle of the jth surrounding vehicle, v j is the centroid velocity of the jth surrounding vehicle, l j and w j are the length and width of the jth surrounding vehicle respectively, and id j is the identification code of the jth surrounding vehicle; j

[0020] A4. Select the target vehicle and coordinate transformation

[0021] Select a vehicle from the surrounding vehicles vehicle nbrs as the target vehicle vehicle target , merge the surrounding vehicles other than the target vehicle and the autonomous driving vehicle into other vehicles vehicle other , and establish a body rectangular coordinate system XOY with the position of the target vehicle at the current moment as the origin, the vehicle head orientation as the positive direction of the X-axis, and the left side of the vehicle as the positive direction of the Y-axis;

[0022] Transform the surrounding road vector information roadvector nbrs , the historical state information state of the autonomous driving vehicle ego and the historical motion state information state of the surrounding vehicles nbrs obtained in steps A1, A2, and A3 to the body rectangular coordinate system of the target vehicle. After coordinate transformation, filter out the historical motion state information of other vehicles vehicle nbrs from the historical motion state information state of the surrounding vehicles ego and the historical state information state of the autonomous driving vehicle other as the historical motion state information state other ;

[0023] ​B. Establish the historical motion state information and interaction relationships of the vehicle using a deep learning model to predict the centroid acceleration and front wheel angle of the target vehicle

[0024] B1. Encode the historical motion state information of the vehicle

[0025] Input the historical motion state information state of the target vehicle obtained in step A3 target Into the first long short-term memory network LSTM1, and LSTM1 outputs the encoded embedding of the historical motion state of the target vehicle target ;

[0026] Input the historical motion state information state of other vehicles obtained in step A3 other Into the second long short-term memory network LSTM2, and LSTM2 outputs the encoded embedding of the historical motion state of other vehicles other ;

[0027] B2. Encode various interaction relationships

[0028] Input the encoded embedding of the historical motion state of the target vehicle target And the encoded embedding of the historical motion state of other vehicles other Into the first multi-level scene gating module MCG1 to obtain the interaction feature feature between the target vehicle and other vehicles t2o ;

[0029] Input the encoded embedding of the historical motion state of the target vehicle other And the surrounding road vector information roadvector nbrs Into the second multi-level scene gating module MCG2 to obtain the interaction feature feature between the target vehicle and the surrounding road t2r ;

[0030] Input the interaction feature feature between the target vehicle and other vehicles t2o And the interaction feature feature between the target vehicle and the surrounding road t2r Into the third multi-level scene gating module MCG3 to obtain the global interaction feature feature global ;

[0031] Input feature t2o 、featuer t2r And feature global These three types of interaction features are concatenated together using the concatenation function concat to form the implicit feature featuer hidden ;

[0032] B3. Decode the motion parameters of the target vehicle

[0033] Use a multi-layer perceptron (MLP) as the decoder to decode the latent feature obtained from step B2 hidden into the predicted values of the centripetal acceleration a pred containing N different motion modes and the predicted values of the front wheel steering angle γ pred for N different motion modes, and the calculation formula is as follows:

[0034] a pred , γ pred = MLP(feature hidden )

[0035]

[0036]

[0037]

[0038]

[0039] where MLP is the multi-layer perceptron method, a pred is the predicted value of the centripetal acceleration containing N different motion modes, and γ pred is the predicted value of the front wheel steering angle containing N different motion modes, represents the predicted value of the centripetal acceleration of the nth motion mode, represents the predicted value of the front wheel steering angle of the nth motion mode, and the value range of n is from 1 to N, represents the predicted value of the centripetal acceleration of the nth motion mode at the i-th future time step, represents the predicted value of the front wheel steering angle of the nth motion mode at the i-th future time step, and the value range of i is from 1 to F;

[0040] C. Divide the vehicle types according to the vehicle length and constrain the predicted values of the centripetal acceleration and front wheel steering angle of the vehicle. Use the vehicle two-degree-of-freedom model to calculate the trajectory coordinates of the target vehicle from the predicted values

[0041] C1. Divide the vehicle types and set the range of motion parameters

[0042] Divide the vehicle types according to the vehicle length and set the maximum acceleration of different types of vehicles:

[0043] Small vehicles: vehicle length ≤ 7m, maximum acceleration a 小型max is 5m / s 2 ;

[0044] Medium vehicles: 7m < vehicle length ≤ 10m, maximum acceleration a 中型max is 3.5m / s2 ;

[0045] Large vehicle: vehicle length > 10m, maximum acceleration a 大型max is 2.5 m / s 2 ;

[0046] Set the maximum front wheel steering angle γ of all vehicle types max to 36°;

[0047] C2. Predicted values of constrained motion parameters

[0048] The predicted centroid acceleration and front wheel steering angle of N motion modes of the predicted target vehicle at F time steps in the future are obtained from step B3, and are constrained according to the acceleration and maximum front wheel steering angle of different vehicle types: If the predicted value of the vehicle centroid acceleration a γ seconds at F time steps is greater than the maximum acceleration a pred of this type of vehicle, then the predicted value of the vehicle centroid acceleration takes the maximum acceleration a max of this type of vehicle; If the predicted value of the vehicle front wheel steering angle γ max is greater than the maximum front wheel steering angle γ pred , then the predicted value of the vehicle front wheel steering angle takes the maximum front wheel steering angle γ max of this type of vehicle; max ;

[0049] C3. Calculate the predicted trajectory from the vehicle two-degree-of-freedom motion model

[0050] Use the vehicle two-degree-of-freedom motion model for the predicted target vehicle, and input the predicted centroid acceleration value a pred of the target vehicle and the predicted front wheel steering angle γ pred of the target vehicle at N motion modes obtained in step C2 to calculate the trajectory coordinates of the target vehicle at F time steps in the future. The calculation formula is as follows:

[0051] Obtain the initial value of the motion state, with the motion state of the target vehicle at the current time step as the initial value:

[0052]

[0053]

[0054]

[0055]

[0056] Among them, is the initial speed value of the target vehicle, v 0 is the speed of the target vehicle at the current moment, is the initial lateral coordinate value of the target vehicle, is the initial longitudinal coordinate value of the target vehicle, is the initial yaw angle value of the target vehicle;

[0057] Calculate the predicted values of the trajectory coordinates for each future time step by the vehicle two-degree-of-freedom model and

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Finally, obtain the predicted values of the trajectory coordinates of the N motion modes of the target vehicle at F time steps in the future T f seconds, Trajectory pred is:

[0065]

[0066] where the superscript of the variable is the i-th future time step, the value range of i is from 1 to F, pred n is the n-th predicted motion mode, the value range of n is from 1 to N, v is the predicted value of the centroid velocity of the target vehicle, x is the lateral coordinate of the centroid of the target vehicle in the body right-angle coordinate system, y is the longitudinal coordinate of the centroid of the target vehicle in the body right-angle coordinate system, γ is the front wheel angle of the target vehicle, a is the centroid acceleration of the target vehicle, t is the duration of each time step, β is the centroid side slip angle of the target vehicle, l f is the distance from the front axle of the target vehicle to the centroid, l r is the distance from the rear axle of the target vehicle to the centroid, ψ is the yaw angle of the target vehicle, and yaw is the heading angle of the target vehicle.

[0067] Furthermore, the length of the equidistant road section is 2 - 3m.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] 1. The present invention integrates the vehicle two-degree-of-freedom motion model into the deep learning prediction model, taking into account both the high-precision long-time-domain prediction performance of deep learning and the vehicle kinematic properties, making the predicted trajectory results more in line with the motion law of the real vehicle.

[0070] 2. By differentiating the kinematic laws of different types of vehicles, the smoothness and uniformity of the predicted trajectory can be improved, making the predicted trajectory conform to the kinematic laws of the vehicle, enhancing the feasibility and accuracy of the predicted trajectory, reducing the situation where the ego-vehicle planning is too conservative due to infeasible prediction results, and improving the driving safety and ride comfort of autonomous vehicles.

[0071] 3. The present invention uses a multi-level scenario gating mechanism module MCG to explicitly model different types of interactions:

[0072] 1) The interaction between the target vehicle vehicle target and other vehicles vehicle other ;

[0073] 2) The interaction between the target vehicle vehicle target and the surrounding road vehicle other ;

[0074] 3) The global interaction among the target vehicle vehicle target , other vehicles vehicle other , and the surrounding road vehicle other ;

[0075] Compared with previous prediction models, it can extract the interaction relationships between different types of scenario elements in the scenario more comprehensively, improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a schematic flow chart of the present invention.

[0077] Figure 2 is a detailed schematic flow chart of the present invention.

[0078] Figure 3 is a schematic flow chart of the deep learning prediction model of the present invention.

[0079] Figure 4 is a schematic flow chart of the context gating module (CG).

[0080] Figure 5 is a schematic flow chart of the multiple context gating module (MCG).

[0081] Figure 6 is a schematic diagram of the vehicle two-degree-of-freedom motion model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0082] The following specific embodiments illustrate the implementation manners of the present invention, and those skilled in the art can easily understand the advantages of the present invention from the content disclosed in this specification. The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts fall within the protection scope of the present invention. In order to provide a deep understanding of the present invention, many specific details will be included in the following description. The present invention can also be implemented without these details. In addition, some specific details will be omitted in the description to avoid obscuring or confusing the focus of the present invention.

[0083] The terms "first", "second", "third", etc. in this application are used to distinguish different objects, rather than to describe a specific order.

[0084] To make the objectives, technical solutions and advantages of the present invention clearer, the implementation manners of the present invention will be further described in detail below with reference to the accompanying drawings:

[0085] Figure 1 It is a schematic flow chart of a deep learning vehicle trajectory prediction method integrating a vehicle kinematic model designed by the present invention. Figure 2 It is a detailed schematic flow chart of the present invention. First, obtain the historical motion state information of the autonomous vehicle, the historical motion state information of surrounding vehicles, and the vector information of the surrounding roads, which can comprehensively consider various dynamic and static scene elements affecting the prediction result in the traffic scene, and select the target vehicle to transform the coordinate system with the target vehicle as the center, which can achieve input data normalization and improve the accuracy of the model; then use a deep learning model adopting an encoder-decoder architecture. The structural diagram of the deep learning prediction model is as Figure 3As shown, the historical motion state information of the target vehicle and other vehicles is encoded using an LSTM network, and the interaction relationships among the target vehicle, other vehicles, and the surrounding roads are encoded using an MCG module. By explicitly modeling different types of interaction relationships, the extraction of various interaction features is achieved. An MLP is used to decode the centripetal acceleration and front wheel steering angle of the target vehicle in N different motion modes within the future time domain. Using a deep learning model saves complex and cumbersome feature engineering compared to machine learning, and can represent real traffic scenarios in the form of latent features to achieve a more compact information representation and a more excellent information extraction and fusion ability. Considering the differences in the dynamic performance of different types of vehicles, the target vehicle is divided into three categories according to its length, and different maximum centripetal accelerations are set respectively. The predicted values of the deep learning model are constrained by the maximum centripetal acceleration and maximum front wheel steering angle of different types of vehicles, which can prevent the trajectory prediction model from predicting trajectories that do not conform to the vehicle kinematic laws, improve the feasibility and accuracy of the predicted trajectories, and use a vehicle two-degree-of-freedom motion model to model the motion law of the vehicle. The constrained maximum centripetal acceleration and maximum front wheel steering angle of the vehicle are integrated over time to obtain the future predicted trajectory.

[0086] Figure 4 It is a schematic diagram of the process of the Context Gating (CG) module, and its calculation formula is as follows:

[0087] c k = MLP(a k ) ⊙ MLP(b k )

[0088] d k = Pooling(MLP(a k ) ⊙ MLP(b k ))

[0089] Among them, a k , b k are the inputs of the CG module, ⊙ is the use of vector dot product, MLP is the use of the multi-layer perceptron method, c k , d k are the outputs of the CG module, Pooling is the use of the max pooling method, and k is the number of levels of the CG module.

[0090] Figure 5 It is a schematic diagram of the process of the Multiple Context Gating (MCG) module, which is composed of K stacked CG modules, and its calculation formula is as follows:

[0091] a k+1 = μ(c k , a k )

[0092] b k+1 = μ(d k ,b k )

[0093] c k ,d k = CG k (a k ,b k )

[0094] where k is the number of stages of the CG module, and its value range is from 1 to K. a k 、b k are the inputs of the k-th stage CG module, c k 、d k are the outputs of the k-th stage CG module, μ is for calculating the mean value, CG k is for using the k-th stage CG module, a k 、b k are the inputs of the MCG module, a k+1 、b k+1 are the outputs of the MCG module. Specifically, K takes the value of 5.

[0095] Figure 6 is a schematic diagram of the two-degree-of-freedom motion model of the vehicle. The vehicle is simplified into a two-degree-of-freedom motion model composed of the front wheels, rear wheels, and the vehicle body. In the ground coordinate system XOY, the distance from the front wheel axle to the vehicle's center of mass is l f , and the distance from the rear wheel axle to the vehicle's center of mass is l r . v is the velocity vector of the vehicle's center of mass, the angle between the front wheel orientation and the vehicle body orientation is the front wheel steering angle γ, the angle between the vehicle body orientation and the X-axis of the coordinate system is the yaw angle ψ, the angle between the vehicle's center of mass velocity and the vehicle body orientation is the center of mass side slip angle β, and the angle between the vehicle's center of mass velocity and the X-axis of the coordinate system is the heading angle yaw.

[0096] The above describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed.

Claims

1. A deep learning vehicle trajectory prediction method integrating a kinematic model, characterized in that: It includes the following steps: A. Obtain the surrounding scene information of the autonomous vehicle and convert it to the target vehicle coordinate system A1. Obtain the surrounding road vector information Obtain the autonomous driving vehicle vehicle from the high-precision map ego The surrounding road structure information, where the road structure information includes the position coordinates of the lane center line, lane boundary line, and crosswalk area; Divide the road structure information into multiple equidistant road segments, and extract the starting point coordinates a, the ending point coordinates b, and the direction c from each road segment to jointly form the surrounding road vector information roadvector nbrs ={a, b, c}; A2. Obtain the historical motion state information of the autonomous vehicle The autonomous driving vehicle vehicle is obtained through the positioning system of the autonomous driving vehicle ego According to the sampling frequency f, the historical motion state information at H time steps in the past T h seconds: The historical state information of the historical motion state information at the h-th time step is where the value range of h is from -H + 1 to 0, L ego and B ego are the longitude and latitude of the autonomous vehicle in the earth coordinate system, respectively, and ψ ego is the yaw angle of the autonomous vehicle, v ego is the centroid velocity of the autonomous vehicle, and l ego , w ego are the length and width of the autonomous vehicle, respectively; A3. Obtain the historical motion state information of the surrounding vehicles of the autonomous vehicle M surrounding vehicles vehicle are detected in real time by the target detection module of the environmental perception system of the autonomous vehicle nbrs ={vehicle1, vehicle2,.., vehicle M}, and the historical motion state information state of each surrounding vehicle within the past T h seconds is obtained through the tracking module nbrs ={state1, state2,..., state M}, and go to step A4; The historical motion state information of each surrounding vehicle is obtained at H time steps in the past T seconds according to the sampling frequency f. The value range of j is from 1 to M, and the historical motion state information of the j-th surrounding vehicle is h The motion state information of the j-th surrounding vehicle at the h-th time step is where L j and B j are the longitude and latitude of the j-th surrounding vehicle in the geodetic coordinate system, ψ j is the yaw angle of the j-th surrounding vehicle, v j is the centroid velocity of the j-th surrounding vehicle, l j , w j are the length and width of the j-th surrounding vehicle respectively, and id j is the identification code of the j-th surrounding vehicle; A4. Select the target vehicle and perform coordinate system conversion Select a vehicle from the surrounding vehicles nbrs as the target vehicle, target and combine the surrounding vehicles other than the target vehicle and the autonomous vehicle into other vehicles. other Taking the position of the target vehicle at the current moment as the origin, with the vehicle head facing as the positive direction of the X-axis and the left side of the vehicle as the positive direction of the Y-axis, establish a body rectangular coordinate system XOY; Convert the surrounding road vector information roadvector obtained in steps A1, A2, and A3 nbrs , the historical state information state of the autonomous vehicle ego and the historical motion state information state of the surrounding vehicles nbrs to the body right-angle coordinate system of the target vehicle. After coordinate transformation, filter out other vehicles vehicle nbrs from the historical motion state information state of the surrounding vehicles ego and the historical state information state of the autonomous vehicle other whose historical motion state information is the historical motion state information state other ; B. Use a deep learning model to establish the vehicle historical motion state information and interaction relationships, and predict the centroid acceleration and front wheel angle of the target vehicle B1. Encode the vehicle historical motion state information Input the historical motion state information state of the target vehicle obtained in step A3 target into the first long short-term memory network LSTM1, and LSTM1 outputs the encoded historical motion state embedding of the target vehicle target ; Input the historical motion state information state of other vehicles obtained in step A3 other into the second long short-term memory network LSTM2, and LSTM2 outputs the encoded historical motion state embedding of other vehicles other ; B2. Encode various interaction relationships Encode the historical motion state of the target vehicle into an embedding target and encode the historical motion states of other vehicles into embeddings other Input them into the first multi-level scene gating module MCG1 to obtain the interaction features between the target vehicle and other vehicles t2o ; Encode the historical motion state of the target vehicle into an embedding other and the surrounding road vector information roadvector nbrs Input them into the second multi-level scene gating module MCG2 to obtain the interaction feature feature between the target vehicle and the surrounding roads t2r ; The interaction features of the target vehicle with other vehicles feature t2o The interaction features of the target vehicle with the surrounding roads feature t2r Input to the third multi-level scene gating module MCG3 to obtain the global interaction feature feature global ; Concatenate the following three types of interaction features, namely feature t2o , feature t2r , and feature global , using the concatenation function concat to form the implicit feature feature hidden ; B3. Decode the target vehicle motion parameters Use a multi-layer perceptron (MLP) as the decoder to decode the latent feature feature obtained from step B2 hidden into the predicted centripetal acceleration values a for N different motion modes pred and the predicted front wheel steering angle values γ for N different motion modes pred , and the calculation formula is as follows: a pred , γ pred = MLP(feature hidden ) where MLP uses the multi-layer perceptron method, a pred is the predicted value of the centroid acceleration containing N different motion modes, γ pred is the predicted value of the front wheel steering angle containing N different motion modes, represents the predicted value of the centroid acceleration of the n-th motion mode, represents the predicted value of the front wheel steering angle of the n-th motion mode, and the value range of n is from 1 to N, represents the predicted value of the centroid acceleration of the n-th motion mode at the i-th future time step, represents the predicted value of the front wheel steering angle of the n-th motion mode at the i-th future time step, and the value range of i is from 1 to F; C. Divide the vehicle types according to the vehicle length and constrain the predicted values of the centroid acceleration and front wheel angle of the vehicle. Use the vehicle two-degree-of-freedom model to calculate the trajectory coordinates of the target vehicle from the predicted values C1. Divide the vehicle types and set the range of motion parameters Divide the vehicle types according to the vehicle length and set the maximum acceleration of different types of vehicles: Small vehicle: vehicle length ≤ 7m, maximum acceleration a 小型max is 5m / s 2 ; Medium-sized vehicle: 7m < vehicle length ≤ 10m, maximum acceleration a 中型max is 3.5 m / s 2 ; Large vehicle: vehicle length > 10 m, maximum acceleration a 大型max is 2.5 m / s 2 ; Set the maximum front wheel steering angle γ for all types of vehicles max to 36°; C2. Constrain the predicted values of the motion parameters From step B3, the predicted centroid acceleration and front wheel angle of the target vehicle at N motion modes in the next F time steps within the next T seconds are obtained, and they are constrained according to the acceleration and maximum front wheel angle of different types of vehicles: If the predicted value of the vehicle centroid acceleration a f is greater than the maximum acceleration a pred of this type of vehicle, then the predicted value of the vehicle centroid acceleration is taken as the maximum acceleration a max of this type of vehicle; If the predicted value of the vehicle front wheel angle γ max is greater than the maximum front wheel angle γ pred of this type of vehicle, then the predicted value of the vehicle front wheel angle is taken as the maximum front wheel angle γ max of this type of vehicle; max ; C3. Calculate the predicted trajectory by the vehicle two-degree-of-freedom motion model Using the vehicle two-degree-of-freedom motion model for the predicted target vehicle, input the predicted value of the centroid acceleration a of the target vehicle on the N motion modes obtained in step C2 pred and the predicted value of the front wheel angle γ of the vehicle pred , calculate the trajectory coordinates of the target vehicle in the next F time steps. The calculation formula is as follows: Obtain the initial value of the motion state, using the motion state of the target vehicle at the current time step as the initial value: Among them, is the initial speed value of the target vehicle, v 0 is the speed of the target vehicle at the current moment, is the initial lateral coordinate value of the target vehicle, is the initial longitudinal coordinate value of the target vehicle, is the initial yaw angle value of the target vehicle; Calculate the predicted values of the trajectory coordinates for each future time step using the vehicle two-degree-of-freedom model and Finally, the predicted values of the trajectory coordinates of N motion modes of the target vehicle at F time steps in the future T f seconds are obtained as follows: pred That is: where the superscript of the variable is the $i$-th future time step, and the value range of $i$ is from 1 to $F$, and pred n is the $n$-th predicted motion mode, the value range of $n$ is from 1 to $N$, $v$ is the predicted value of the centroid velocity of the target vehicle, $x$ is the lateral coordinate of the centroid of the target vehicle in the body right-angle coordinate system, $y$ is the longitudinal coordinate of the centroid of the target vehicle in the body right-angle coordinate system, $\gamma$ is the front wheel steering angle of the target vehicle, $a$ is the centroid acceleration of the target vehicle, $t$ is the duration of each time step, $\beta$ is the centroid side slip angle of the target vehicle, $l$ f is the distance from the front axle of the target vehicle to the centroid, $l$ r is the distance from the rear axle of the target vehicle to the centroid, $\psi$ is the yaw angle of the target vehicle, and yaw is the heading angle of the target vehicle.

2. The deep learning vehicle trajectory prediction method integrating a kinematic model according to claim 1, characterized in that: The length of the equidistant road segments is 2 - 3 m.

Images