Automatic driving cooperative motion planning method fusing interactive trajectory prediction

Abstract

The invention relates to a collaborative motion planning method fusing interactive trajectory prediction, and belongs to the field of automatic driving automobiles. The method comprises the steps that S1, a polynomial parameterization trajectory generation model meeting vehicle kinematics constraints is constructed, and a future driving action set is generated based on environment information and the current position state of a vehicle; s2, constructing an interactive prediction-decision planning model based on deep learning, extracting interactive hidden features based on bidirectional interactive behaviors of the autonomous vehicle and surrounding vehicles, and predicting tracks of the surrounding vehicles; and S3, constructing an adaptive trajectory generation model, building a learnable trajectory selection module, scoring the driving action set by using a trajectory prediction result in the step S2, and finally generating an optimal trajectory by combining a deviation generation network optimization result. According to the invention, interactive modeling can be carried out on future states of the vehicle and the surrounding agents, and future trajectories of the surrounding vehicles under different planning interactions are predicted, so that the prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving vehicles and relates to an autonomous driving collaborative motion planning method integrating interactive trajectory prediction. Background Art

[0002] Thanks to the progress in computer vision, deep learning, sensor fusion and motion planning, autonomous driving technology has developed rapidly in recent years. This technology gradually realizes the transition from assisted driving with limited human intervention to fully autonomous driving without human intervention through autonomous vehicle operation. The autonomous driving system generally includes four modules: environmental perception fusion, obstacle trajectory prediction, joint decision planning and trajectory tracking control. The modules work together to achieve driving tasks. Among them, joint decision planning plays a key role in ensuring safety and is also the core embodiment of the system's intelligence.

[0003] At present, autonomous driving technology is mainly divided into three types: modular autonomous driving, modular end-to-end autonomous driving, and fully end-to-end autonomous driving. Modular autonomous driving is a traditional and relatively mature autonomous driving technology. This technology decomposes the autonomous driving task into multiple independent functional modules, including perception, prediction, decision planning and control, so that each module can independently perform corresponding functions according to different task requirements. Its advantage is that the output results of each module can be clearly displayed, which is convenient for flexible adjustment and optimization. However, due to the relative independence of each sub-module, modular autonomous driving may cause delays in data transmission and accumulated errors, which in turn cause safety hazards. Therefore, it is usually applied to L2-level assisted driving systems. However, in complex traffic environments, the two-way interaction between the ego vehicle and surrounding intelligent agents (such as other vehicles and pedestrians) makes the traffic situation more complicated. The future trajectory of the ego vehicle will not only be affected by the surrounding vehicles, but the driving intention of the ego vehicle will also change the future state of the surrounding vehicles. These factors increase the risk of prediction, decision-making and planning failure caused by accumulated errors and interactive behaviors in modular autonomous driving, which may lead to safety accidents.

[0004] Therefore, there is an urgent need for a safe motion planning method that can effectively deal with the interaction between vehicles to ensure that the autonomous driving system is safer and more reliable in dynamic and complex environments. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide an autonomous driving collaborative motion planning method that integrates interactive trajectory prediction, which can interactively model the future states of the self-vehicle and surrounding intelligent agents, and predict the future trajectories of the surrounding vehicles under different planning interactions, thereby improving the prediction accuracy. At the same time, the self-vehicle can imitate the behavioral decisions of human drivers to produce a safer, more efficient, comfortable and human-like motion planning trajectory, and the entire process of this method is differentiable.

[0006] A collaborative motion planning method for autonomous driving integrating interactive trajectory prediction includes the following steps:

[0007] S1: Construct a polynomial parameterized trajectory generation model that satisfies the vehicle kinematic constraints and generates a set of future driving actions based on environmental information and the vehicle's current position state;

[0008] S2: Build an interactive prediction-decision planning model based on deep learning. Based on the two-way interactive behavior between the autonomous driving vehicle and surrounding vehicles, extract the interactive hidden features and predict the trajectories of surrounding vehicles.

[0009] S3: Build an adaptive trajectory generation model and a learnable trajectory selection module. Use the trajectory prediction results of step S2 to score the driving action set, and combine the deviation to generate the network optimization result to finally generate the best trajectory.

[0010] Further, in step S1, a polynomial parameterized trajectory generation model is constructed, which specifically includes the following steps:

[0011] S11: Obtaining the local information of the vehicle: According to the current position of the vehicle and the high-precision map information provided, obtain the current motion state P of the vehicle c And the future drivable lane path point R, that is:

[0012] P c =[x c ,y c ,θ c ,v c ,a c ]

[0013] R=[L 1 ,L 2 ,...,L n ]

[0014]

[0015] Among them, x c ,y c ,θ c ,v c ,a cThey represent the horizontal coordinate, vertical coordinate, heading angle, speed and acceleration of the current position of the vehicle, respectively. n and They represent the nth future drivable lane and the hth path point information of the lane respectively. Respectively represent the abscissa, ordinate, navigation angle, curvature, and speed limit of the h-th path point in the n-th lane;

[0016] S12: Constructing a polynomial parameterized trajectory generation model: Based on the current position information of the vehicle and the future drivable path information provided in step S11, in order to adapt to the complex structure of the urban road scene, the trajectory models of lateral displacement and longitudinal displacement and time are respectively established using quintic polynomials and quartic polynomials in the Frenet coordinate system:

[0017] Δs(t)=b 0 +b 1 t+b 2 t 2 +b 3 t 3 +b 4 t 4

[0018] Δd(t)=r 0 +r 1 t+r 2 t 2 +r 3 t 3 +r 4 t 4 +r 5 t 5

[0019] Among them, Δd(t) and Δs(t) represent the horizontal and vertical displacements at time t, respectively. 0 ,b 1 ,...,b 4 ] and [r 0 ,r 1 ,...,r 4 ] are the parameters of the polynomial parameterized model, which can be obtained by the current position information of the vehicle P c and the terminal position information P e To solve, P e The calculation can be performed by assuming different driving speeds; finally, the obtained Δd(t) and Δs(t) are converted to the Cartesian coordinate system to obtain the final planning trajectory set:

[0020] Further, in step S2, an interactive prediction-decision planning model based on deep learning is constructed, which specifically includes the following steps:

[0021] S21: Interactive prediction-decision planning model input: The input of the interactive prediction-decision planning model mainly includes the historical trajectory characteristics of surrounding vehicles The local map information where the vehicle is currently located And the self-vehicle planning trajectory set obtained in step S12

[0022]

[0023] Among them, X i and They are respectively represented as the feature information and location features of the i-th intelligent agent, i.e., the vehicle, t h represents the historical time domain, l i ,w i Respectively represent the abscissa, ordinate, heading angle, lateral speed, longitudinal speed, vehicle length and width at time t, M n Represents the nth local lane information, They are respectively represented as the horizontal coordinate, vertical coordinate, navigation angle, lane type, speed limit and traffic light status from the first waypoint to the jth waypoint;

[0024] S22: Input feature encoding: Encode the agent features obtained in step S21 using the LSTM network to obtain hidden features Map feature information is encoded using the MLP network (multi-layer perceptron network) to obtain hidden features The planned trajectory set is encoded using another MLP network to obtain hidden features

[0025] S23: Global feature encoding: In order to obtain the global feature information of the agent and the local map, the encoding results of the agent and the map obtained in step S22 are encoded with global features using the self-attention mechanism of the two-layer Transformer;

[0026] G featrue =Attention(Q F ,K F ,V F ,mask)

[0027]

[0028] G feature = {G agent ,G map}

[0029] Among them, G featrue is a global feature, which can be decomposed into the agent global feature G agent and the global map feature Gmap , Q F ,K F ,V F ,mask are the query,key,value and corresponding mask of the attention mechanism, respectively. Attention and concatenate represent the attention mechanism and concatenation, respectively;

[0030] S24: interactive feature information modeling: combining hidden features of the planning trajectory set obtained in step S22 The agent global feature G obtained in step S23 agent and the global map feature G map , using the cross attention mechanism to model the interaction of each predicted agent i:

[0031]

[0032]

[0033] Among them, F i,am and F i,ae They are respectively represented as the interaction features between the ith agent and the map and the planned trajectory, Q i,F , mask m , mask e Respectively represent the query, key, value and corresponding mask required to calculate the attention. Represents the global characteristics of the i-th agent;

[0034] S25: Prediction trajectory decoding: Combine the agent-map interaction feature F obtained in step S24 i,am and the interactive features of the agent-planning trajectory F i,ae , using an MLP network to decode the future trajectory of the i-th agent

[0035]

[0036] Among them, t f Represents the future time domain.

[0037] Further, in step S3, an adaptive trajectory generation model is constructed, which specifically includes the following steps:

[0038] S31: Learnable cost weight: Combined with the hidden features of the vehicle's historical trajectory obtained in step S22 and the vehicle-map interactive hidden feature F obtained in step S24 e,am Concatenate and use an MLP network to decode the weight of the cost function

[0039]

[0040] Among them, τ s ,τ e ,τ c They are expressed as the weights of safety, efficiency and comfort respectively;

[0041] S32: Trajectory scoring modeling: Combine the weights obtained in step S31 Construct trajectory scoring model;

[0042] U o =τ s U s +τ e U e +τ c U c

[0043]

[0044] Among them, U o represents the trajectory score, U s ,U e ,U c denote the cost functions of safety, efficiency and comfort respectively, o is the index of the predicted agent in the safety cost function, S safety is the safety threshold, They represent the horizontal and vertical coordinates of the planned trajectory of the ego-vehicle and the predicted trajectory of the agent o at the future time, They represent the speed, lateral and longitudinal acceleration of the planned trajectory of the vehicle in the future, v limit Indicates lane speed limit;

[0045] S33: Trajectory deviation optimization network: Combine the planned trajectory scores obtained in step S32 and select the trajectory with the highest score As the basic trajectory of the trajectory deviation optimization network, find the corresponding hidden features of the ego-vehicle planning trajectory Combined with the hidden features of the vehicle's historical trajectory obtained in step S22 and the vehicle-map interactive hidden feature F obtained in step S24 e,am , use the MLP network to decode the coordinate deviation of the trajectory and add it to the trajectory to generate the optimal trajectory;

[0046] S34: Loss function formulation: In order to make trajectory prediction more accurate and the driving behavior of the vehicle more human-like, a loss function is designed

[0047]

[0048] in, denote the prediction trajectory imitation loss, planning trajectory imitation loss and planning selection score loss respectively, and λ 1 ,λ 2 ,λ 3 Represents the weight corresponding to each loss, G i ,G e They represent the true trajectories of the predicted agent i and the ego vehicle, respectively. represents the planning trajectory obtained by the model, smoothhl1 and softmax are smooth L1 loss and activation function respectively, y z is the planning trajectory score, a and Z are the number of predicted agents and the number of planning trajectories respectively.

[0049] The beneficial effects of the present invention are:

[0050] 1) The present invention designs a polynomial parameterized model, which uses polynomial curves to generate a set of planned trajectories that can be exercised in the future based on the current position of the vehicle and the characteristics of the local environment, and combines the kinematic model to generate candidate trajectories that are more in line with the actual driving conditions.

[0051] 2) The present invention designs an interactive prediction-decision-making planning model based on deep learning. The model can extract the historical trajectory characteristics, environmental characteristics and self-vehicle planning trajectory characteristics of the intelligent agent, use the Transformer network to establish the interactive characteristics of the vehicle-map and vehicle-planning trajectory, and generate a predicted trajectory under interactive behavior.

[0052] 3) The present invention designs an adaptive trajectory generation model, which can use the historical characteristics of the vehicle and the interaction characteristics with the map to decode the adaptive trajectory selection weight, and select the adaptive trajectory through safety, efficiency and comfort scores, combined with the deviation network model to generate the optimal trajectory. Compared with the traditional modular autonomous driving, it takes into account the impact of the two-way interaction between the vehicle and the surrounding vehicles and the cumulative error between models to ensure the safe driving of the vehicle.

[0053] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0055] Figure 1 This is a logical framework diagram for the overall implementation of the interactive motion planning method for autonomous driving based on deep learning of the present invention. DETAILED DESCRIPTION

[0056] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0057] See also Figure 1 , the present invention designs a collaborative motion planning method for autonomous driving that integrates interactive trajectory prediction. Considering the problem of two-way vehicle interaction in complex traffic environments, a polynomial parameterized model is constructed to generate a set of planning trajectories. Combining the current map information of the vehicle and the characteristics of surrounding intelligent agents, an interactive prediction-decision planning model is constructed, which can extract the two-way interaction characteristics between the vehicle and the surrounding intelligent agents, and predict the future trajectory of the surrounding intelligent agents. An adaptive trajectory generation model is constructed, which combines the influence of safety, efficiency and comfort, and can adaptively generate more human-like motion planning trajectories according to different environmental information to ensure driving safety.

[0058] The method specifically comprises the following steps:

[0059] S1: Construct a polynomial parameterized trajectory generation model that satisfies the vehicle kinematic constraints and generates a set of future driving actions based on environmental information and the vehicle's current position state;

[0060] Constructing a polynomial parameterized trajectory generation model specifically includes the following steps:

[0061] S11: Obtaining the local information of the vehicle: According to the current position of the vehicle and the high-precision map information provided, obtain the current motion state P of the vehicle c And the future drivable lane path point R, that is:

[0062] P c =[x c ,y c ,θ c ,v c ,a c ]

[0063] R=[L 1 ,L 2 ,...,L n ]

[0064]

[0065] Among them, x c ,y c ,θ c ,v c ,a c They represent the horizontal and vertical coordinates, heading angle, speed and acceleration of the current position of the vehicle, respectively. n and They represent the nth future drivable lane and the hth path point information of the lane respectively. They respectively represent the horizontal and vertical coordinates, navigation angle, curvature and speed limit of the h-th path point in the n-th lane.

[0066] S12: Constructing a polynomial parameterized trajectory generation model: Based on the current position information of the vehicle and the future drivable path information provided in step S11, in order to adapt to the complex structure of the urban road scene, the trajectory models of lateral displacement and longitudinal displacement and time are respectively established using quintic polynomials and quartic polynomials in the Frenet coordinate system:

[0067] Δs(t)=b 0 +b 1 t+b 2 t 2 +b 3 t 3 +b 4 t 4

[0068] Δd(t)=r 0 +r 1 t+r 2 t 2 +r 3 t 3 +r 4 t 4 +r 5 t 5

[0069] Among them, Δd(t) and Δs(t) represent the horizontal and vertical displacements at time t, respectively. 0 ,b 1 ,...,b 4 ] and [r 0 ,r 1 ,...,r 4 ] are the parameters of the polynomial parameterized model, which can be obtained by the current position information of the vehicle P c and the terminal position information P e To solve, P e The calculation can be performed by assuming different driving speeds. Finally, the obtained Δd(t) and Δs(t) are converted to the Cartesian coordinate system to obtain the final planning trajectory set

[0070] S2: Construct an interactive prediction-decision planning model based on deep learning. Based on the two-way interactive behavior between the autonomous driving vehicle and surrounding vehicles, extract the interactive hidden features and predict the trajectories of surrounding vehicles. The specific steps include:

[0071] S21: Interactive prediction-decision planning model input: The input of the interactive prediction-decision planning model mainly includes the historical trajectory characteristics of surrounding vehicles The local map information where the vehicle is currently located And the self-vehicle planning trajectory set obtained in step S12

[0072]

[0073] Among them, X i and They are respectively represented as the feature information and location features of the i-th intelligent agent, i.e., the vehicle, and t h represents the historical time domain, l i ,w i Respectively represent the horizontal and vertical coordinates, heading angle, horizontal and vertical speed, vehicle length and width at time t, M n Represents the nth local lane information, They are respectively represented by the horizontal and vertical coordinates, navigation angle, lane type, speed limit and traffic light status from the first point to the jth path point.

[0074] S22: Input feature encoding: Encode the agent features obtained in step S21 using the LSTM network to obtain hidden features Map feature information is encoded using the MLP network to obtain hidden features The planned trajectory set is encoded using another MLP network to obtain hidden features

[0075] S23: Global feature encoding: In order to obtain the global feature information of the agent and the local map, the encoding results of the agent and the map obtained in step S22 are encoded with global features using the self-attention mechanism of the two-layer Transformer:

[0076] G featrue =Attention(Q F ,K F ,V F ,mask)

[0077]

[0078] G feature = {Gagent ,G map}

[0079] Among them, G featrue is a global feature, which can be decomposed into the agent global feature G agent and the global map feature G map , Q F ,K F ,V F ,mask are the query,key,value and corresponding mask of the attention mechanism, respectively.,Attention,and,concatenate,represent the attention mechanism and concatenation, respectively.

[0080] S24: interactive feature information modeling: combining hidden features of the planning trajectory set obtained in step S22 The agent global feature G obtained in step S23 agent and the global map feature G map , using the cross attention mechanism to model the interaction of each predicted agent i:

[0081]

[0082] Among them, F i,am and F i,ae They are respectively represented as the interaction features between the ith agent and the map and the planned trajectory, Q i,F , mask m , mask e They represent the query, key, value and corresponding mask required to calculate the attention respectively. Represents the global characteristics of the i-th agent.

[0083] S25: Prediction trajectory decoding: Combine the agent-map interaction feature F obtained in step S24 i,am and the interactive features of the agent-planning trajectory F i,ae , use an MLP to decode the future trajectory of the i-th agent:

[0084]

[0085] Among them, t f Represents the future time domain.

[0086] S3: Build an adaptive trajectory generation model, construct a learnable trajectory selection module, use the trajectory prediction results of step S2 to score the driving action set, and generate network optimization results based on the deviation to finally generate the best trajectory; specifically, the following steps are included:

[0087] S31: Learnable cost weight: Combined with the hidden features of the vehicle's historical trajectory obtained in step S22 and the vehicle-map interactive hidden feature F obtained in step S24 e,am Splice and use an MLP network to decode the weight of the cost function

[0088]

[0089] Among them, τ s ,τ e ,τ c They are represented as the weights of safety, efficiency and comfort respectively.

[0090] S32: Trajectory scoring modeling: Combined with the weights obtained in S31 Build a trajectory scoring model:

[0091] U o =τ s U s +τ e U e +τ c U c

[0092]

[0093] Among them, U s ,U e ,U c denote the cost functions of safety, efficiency and comfort respectively, o is the index of the predicted agent in the safety cost function, S safety is the safety threshold, They represent the horizontal and vertical coordinates of the planned trajectory of the ego-vehicle and the predicted trajectory of the agent o in the future, They represent the speed and lateral and longitudinal acceleration of the planned trajectory of the vehicle in the future, v limit Indicates the speed limit of the lane.

[0094] S33: Trajectory deviation optimization network: Combine the planned trajectory scores obtained in step S32 and select the trajectory with the highest score As the basic trajectory of the trajectory deviation optimization network, find the corresponding hidden features of the ego-vehicle planning trajectory Combined with the hidden features of the vehicle's historical trajectory obtained in step S22 and the vehicle-map interactive hidden feature F obtained in step S24 e,am , use the MLP network to decode the coordinate deviation of the trajectory and add it to the trajectory to generate the optimal trajectory.

[0095] S34: Loss function formulation: In order to make trajectory prediction more accurate and the driving behavior of the vehicle more human-like, a loss function is designed

[0096]

[0097] in, They are respectively represented as the prediction trajectory imitation loss, the planning trajectory imitation loss and the planning selection score loss, λ 1 ,λ 2 ,λ 3 Represent the weights corresponding to each loss, G i ,G e They represent the true trajectories of the predicted agent i and the ego vehicle, respectively. represents the planning trajectory obtained by the model, smoothhl1 and softmax are smooth L1 loss and activation function respectively, y z is the planning trajectory score, a and Z are the number of predicted agents and the number of planning trajectories respectively.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A collaborative motion planning method for autonomous driving integrating interactive trajectory prediction, characterized in that: The method specifically comprises the following steps: S1: Construct a polynomial parameterized trajectory generation model that satisfies the vehicle kinematic constraints and generates a set of future driving actions based on environmental information and the vehicle's current position state; S2: Build an interactive prediction-decision planning model based on deep learning. Based on the two-way interactive behavior between the autonomous driving vehicle and surrounding vehicles, extract the interactive hidden features and predict the trajectories of surrounding vehicles. S3: Build an adaptive trajectory generation model and a learnable trajectory selection module. Use the trajectory prediction results of step S2 to score the driving action set, and combine the deviation to generate the network optimization result to finally generate the best trajectory.

2. The autonomous driving collaborative motion planning method according to claim 1, characterized in that: In step S1, a polynomial parameterized trajectory generation model is constructed, which specifically includes the following steps: S11: Obtaining the local information of the vehicle: According to the current position of the vehicle and the high-precision map information provided, obtain the current motion state P of the vehicle c And the future drivable lane path point R, that is: P c =[x c ,y c ,i c ,v c ,a c ] R=[L1,L2,...,L n ] Among them, x c ,y c ,θ c ,v c ,a c They represent the horizontal coordinate, vertical coordinate, heading angle, speed and acceleration of the current position of the vehicle, respectively. n and They represent the nth future drivable lane and the hth path point information of the lane respectively. Respectively represent the abscissa, ordinate, navigation angle, curvature, and speed limit of the h-th path point in the n-th lane; S12: Constructing a polynomial parameterized trajectory generation model: Based on the current position information of the vehicle and the future drivable path information provided in step S11, a trajectory model of lateral displacement and longitudinal displacement versus time is established using a quintic polynomial and a quartic polynomial in the Frenet coordinate system: Δs(t)=b0+b1t+b2t 2 +b3t 3 +b4t 4 Δd(t)=r0+r1t+r2t 2 +r3t 3 +r4t 4 +r5t 5 Among them, Δd(t) and Δs(t) represent the lateral and longitudinal displacements at time t, respectively. [b0, b1, ..., b4] and [r0, r1, ..., r4] are the parameters of the polynomial parameterized model. The current position information P of the vehicle is c and the terminal position information P e To solve, P e The calculation is performed by assuming different driving speeds; finally, the obtained Δd(t) and Δs(t) are converted to the Cartesian coordinate system to obtain the final planning trajectory set 3. The autonomous driving collaborative motion planning method according to claim 2, characterized in that: In step S2, an interactive prediction-decision planning model based on deep learning is constructed, which specifically includes the following steps: S21: Interactive prediction-decision planning model input: The input of the interactive prediction-decision planning model includes the historical trajectory characteristics of surrounding vehicles The local map information where the vehicle is currently located And the self-vehicle planning trajectory set obtained in step S12 Among them, X i and They are respectively represented as the feature information and location features of the i-th intelligent agent, i.e., the vehicle, t h represents the historical time domain, l i ,w i Respectively represent the abscissa, ordinate, heading angle, lateral speed, longitudinal speed, vehicle length and width at time t, M n Represents the nth local lane information, They are respectively represented as the horizontal coordinate, vertical coordinate, navigation angle, lane type, speed limit and traffic light status from the first waypoint to the jth waypoint; S22: Input feature encoding: Encode the agent features obtained in step S21 using the LSTM network to obtain hidden features Map feature information is encoded using the MLP network to obtain hidden features The planned trajectory set is encoded using another MLP network to obtain hidden features S23: Global feature encoding: In order to obtain the global feature information of the agent and the local map, the encoding results of the agent and the map obtained in step S22 are encoded with global features using the self-attention mechanism of the two-layer Transformer; G featrue =Attention(Q F ,K F ,V F ,mask) G feature ={G agent ,G map } Among them, G featrue is a global feature, which can be decomposed into the agent global feature G agent and the global map feature G map , Q F ,K F ,V F ,mask are the query,key,value and corresponding mask of the attention mechanism, respectively. Attention and concatenate represent the attention mechanism and concatenation, respectively; S24: interactive feature information modeling: combining hidden features of the planning trajectory set obtained in step S22 The agent global feature G obtained in step S23 agent and the global map feature G map , using the cross attention mechanism to model the interaction of each predicted agent i: Among them, F i,am and F i,ae They are respectively represented as the interaction features between the ith agent and the map and the planned trajectory, Q i,F , mask m , mask e Respectively represent the query, key, value and corresponding mask required to calculate the attention. Represents the global characteristics of the i-th agent; S25: Prediction trajectory decoding: Combine the agent-map interaction feature F obtained in step S24 i,am and the interactive features of the agent-planning trajectory F i,ae , using an MLP network to decode the future trajectory of the i-th agent Among them, t f Represents the future time domain.

4. The autonomous driving collaborative motion planning method according to claim 3, characterized in that: In step S3, an adaptive trajectory generation model is constructed, which specifically includes the following steps: S31: Learnable cost weight: Combined with the hidden features of the vehicle's historical trajectory obtained in step S22 and the vehicle-map interactive hidden feature F obtained in step S24 e,am Concatenate and use an MLP network to decode the weight of the cost function Among them, τ s ,τ e ,τ c They are expressed as the weights of safety, efficiency and comfort respectively; S32: Trajectory scoring modeling: Combine the weights obtained in step S31 Construct trajectory scoring model; U o =t s U s +t e U e +t c U c Among them, U o represents the trajectory score, U s ,U e ,U c denote the cost functions of safety, efficiency and comfort respectively, o is the index of the predicted agent in the safety cost function, S safety is the safety threshold, They represent the horizontal and vertical coordinates of the planned trajectory of the ego-vehicle and the predicted trajectory of the agent o at the future time, They represent the speed, lateral and longitudinal acceleration of the planned trajectory of the vehicle in the future, v limit Indicates lane speed limit; S33: Trajectory deviation optimization network: Combine the planned trajectory scores obtained in step S32 and select the trajectory with the highest score As the basic trajectory of the trajectory deviation optimization network, find the corresponding hidden features of the ego-vehicle planning trajectory Combined with the hidden features of the vehicle's historical trajectory obtained in step S22 and the vehicle-map interactive hidden feature F obtained in step S24 e,am , use the MLP network to decode the coordinate deviation of the trajectory and add it to the trajectory to generate the optimal trajectory; S34: Loss Function Formulation: Designing Loss Function in, They represent the prediction trajectory imitation loss, planning trajectory imitation loss and planning selection score loss respectively, λ1, λ2, λ3 represent the weights corresponding to each loss, G i ,G e They represent the true trajectories of the predicted agent i and the ego vehicle, respectively. represents the planning trajectory obtained by the model, smoothhl1 and softmax are smooth L1 loss and activation function respectively, y z is the planning trajectory score, a and Z are the number of predicted agents and the number of planning trajectories respectively.

Images