CN118269968B - Prediction method of automatic driving collision risk fused with online map uncertainty - Google Patents

Abstract

The present invention belongs to the field of traffic control and relates to a method for predicting collision risks of autonomous driving by integrating uncertainty of online maps. The method mainly comprises generation of uncertainty of online maps, trajectory prediction integrating uncertainty of online maps, and collision risk prediction. Generation of uncertainty of online maps is based on online data collected by on-board cameras and radars to generate online maps and uncertainty parameters of online maps, and integrate them into a downstream trajectory prediction module for feature extraction, endpoint prediction and trajectory generation. The collision risk prediction module outputs trajectory points of target vehicles and surrounding vehicles according to trajectory prediction, outputs the probability of collision in future time steps, and calculates the remaining time of collision probability at the same time. The method uses online maps and takes uncertainty of online maps into consideration, significantly improves the accuracy and scalability of downstream risk prediction modules, and provides effective and accurate safety assessment signals for the takeover system of autonomous driving vehicles under human-machine co-driving.

Description

Prediction method of collision risk for autonomous driving integrating online map uncertainty

Technical Field

The present invention belongs to the field of traffic control and relates to the prediction of collision risk of an autonomous driving vehicle, and specifically, to a method for predicting collision risk of an autonomous driving vehicle by integrating uncertainty of an online map.

Background Art

Risk prediction is an important part of the autonomous driving stack. It affects the downstream planning and control modules of the autonomous driving vehicle. A good risk prediction method can provide effective and accurate safety assessment signals for the takeover system under human-machine co-driving.

Most of the existing risk predictions are based on data provided by high-precision maps. However, high-precision maps are expensive to label and maintain, can only be used in specific areas, and have poor scalability. Some scholars have turned their work to online estimation of high-precision maps based on sensor data, but these online estimation methods do not take into account the uncertainty of the estimated online maps. Therefore, it is necessary to design a new risk prediction method to generate the uncertainty of online maps and incorporate the generated uncertainty into risk prediction, so as to improve the accuracy of risk prediction and provide valuable reference for the research on early warning of autonomous driving takeover systems under human-machine co-driving.

Summary of the invention

In view of the above-mentioned technical problems and defects, the purpose of the present invention is to provide a method for predicting collision risks of autonomous driving by integrating the uncertainty of online maps. The method generates online maps and uncertainty parameters of online maps based on online data collected by on-board cameras and radars, and integrates them into the downstream trajectory prediction module for feature extraction, endpoint prediction and trajectory generation. After that, the collision risk prediction module outputs the trajectory points of the target vehicle and the surrounding vehicles according to the trajectory prediction, outputs the probability of collision in the future time step, and calculates the remaining time of the collision probability. The method takes into account the uncertainty of the online map, significantly improves the accuracy of the downstream risk prediction module, and provides an effective and accurate safety assessment signal for the takeover system of the autonomous driving vehicle under human-machine co-driving.

To achieve the above object, the present invention adopts the following technical solution:

A method for predicting collision risk of autonomous driving by integrating online map uncertainty, the method comprising the following steps:

Step 1: Generation of online map uncertainty;

Step 1.1: Collect the information of surrounding roads and agents obtained by the on-board camera and radar, and perform corresponding noise reduction processing;

Step 1.2: Input the denoised data into the BEV encoder to convert the acquired data into a common BEV spatial feature;

Step 1.3: Establish a Gaussian regression uncertainty model and a classification uncertainty model, take the BEV spatial features as input, use the online map regression model to generate map element vertices, the map element vertices include the position and type of the element vertices, and calculate the online map uncertainty related parameters through the Gaussian regression uncertainty model and the classification uncertainty model; wherein the Gaussian regression uncertainty model calculates the predicted position of each map element vertex and the parameters of the uncertainty related to the position, and the classification uncertainty model calculates the confidence of the type of each map element vertex;

Step 2: Trajectory prediction integrating online map uncertainty;

Step 2.1: Use multi-layer perceptron encoding to integrate the predicted positions of the obtained map element vertices, the parameters of the uncertainty associated with the positions, and the confidence of the types of the map element vertices into the online map;

Step 2.2: For the online map with fused uncertainty parameters, use VectorNet to encode multi-segment lines and perform feature extraction. When encoding lane line features, use VectorNet to encode a multi-segment line subgraph for each scene element, and use two independent subgraphs for the agent and lane to obtain the feature vector through encoding. ;

Step 2.3: Get the feature vector It is passed to the interactive modeling module composed of the agent-to-lane module, lane-to-lane module, lane-to-agent module and agent-to-agent module stack to update the features. The updated features are ; Among them, the agent-to-lane module focuses on the relationship between the agent and the lane, the lane-to-lane module focuses on the relationship between the lane and the lane, the lane-to-agent module focuses on the relationship between the lane and the agent, the agent-to-agent module focuses on the relationship between the agent and the agent, the agent-to-lane module, the lane-to-lane module, the lane-to-agent module, and the agent-to-agent module are all multi-head attention blocks;

Step 2.4: Based on meta-information f and features The concatenated feature vector , use fixed-point predictors and environment-adaptive predictors to predict the endpoints of the target vehicle and surrounding agent trajectories, and use multi-layer perceptrons (MLPs) to generate trajectories;

Step 3: Collision risk prediction based on predicted trajectory;

Step 3.1: Obtain the coordinates of the target vehicle and the surrounding agents in the future time step according to the trajectory generated in step 2.4, input the predicted coordinates of the target vehicle and the surrounding agents in the future time step into the collision probability prediction module, and predict the probability of collision between the surrounding agents in the total future time step;

Step 3.2: Input the collision probability obtained in step 3.1 into the critical collision probability remaining time calculation module, calculate the critical collision probability remaining time, and transmit it to the human-machine co-driving system.

As a preferred embodiment of the present invention, in step 1.1, the surrounding road information includes: road type, speed limit, whether it is at an intersection, traffic lights, and road slope; wherein the road type includes intersections, left-turn lanes, right-turn lanes, and through lanes; the surrounding agent information includes: agent type, speed, and acceleration, and the agent type includes pedestrians and vehicles.

As a preferred embodiment of the present invention, in step 1.3, the established Gaussian regression uncertainty model is expressed as:

;

Where V is the number of vertices of the map element M, and the coordinates of the i- th vertex are , is the mean, which represents the expected value of the vertex position, is the standard deviation, a measure of the uncertainty in the predicted position, is the variance; i and j are used to represent the jth dimension of the i- th vertex of the map element M, is the output of the Gaussian regression uncertainty model;

The classification uncertainty model uses a multi-layer perceptron (MLP) to output the confidence of the type of each map element vertex.

As a preferred embodiment of the present invention, in step 2.1, the model for integrating the predicted position of the map element vertex and the parameters of the uncertainty related to the position and the confidence of the type of the map element vertex into the online map is:

In the formula, is the Gaussian distribution mean vector of the i-th vertex, is the standard deviation vector of the uncertainty of the i-th vertex output by the Gaussian regression uncertainty model, is the category probability vector composed of the confidence of the i-th vertex, [;;] represents the connection of three vectors, It is a multi-layer perceptron MLP.

As a preferred embodiment of the present invention, in step 2.3, the interaction modeling module uses a self-attention encoder followed by a feed-forward network (FFN) to update the agent-to-agent module and the lane-to-lane module, and uses a cross-attention encoder followed by a feed-forward network to update the lane-to-agent module and the agent-to-lane module, and each encoder uses a multi-head attention mechanism (MHA):

;

= K = V = ;

;

In the formula, , , is the learned projection, norm regularized; Q, K, and V are the query vector, key vector, and value vector of the attention mechanism, respectively. Activation function, are all eigenvectors;

The multi-head attention block is defined as follows:

;

Stack agent to lane module, lane to lane module, lane to agent module, agent to agent module to update features , get the updated features .

As a preferred embodiment of the present invention, in step 2.4, the meta information f and the feature The spliced model is:

= （ , f);

In the formula, is a multi-layer perceptron MLP, Meta information and features The concatenated vector,meta-information f includes the direction and position information of the agent at the prediction moment.

As a preferred embodiment of the present invention, in step 2.4, the fixed-point predictor is used to predict the endpoint of the target agent. The fixed-point predictor predicts the endpoint with the help of a multi-layer perceptron, and the formula is as follows:

= （ );

In the formula, is the predicted endpoint, Multilayer Perceptron MLP, is the meta-information f and feature The concatenated feature vector.

As a preferred embodiment of the present invention, in step 2.4, the environment adaptive predictor is used to predict the endpoints of multiple surrounding agents, and the environment adaptive predictor adopts a dynamic weight learning method to adapt to the specific situation of each agent:

;

;

;

;

In the formula, , are trainable parameters, , It is the weight dynamic adjustment matrix in the endpoint prediction module, norm is layer normalization, and ReLU is the activation function; is a dynamic weight matrix and Features The matrix after regularization and nonlinear fitting, is the predicted endpoint.

As a preferred embodiment of the present invention, in step 3.1, the collision probability prediction module establishes a collision intersection function, a target vehicle and surrounding agents in the prediction time step The collision probability prediction function within the time range of ; Among them, the collision intersection function is used to determine whether the matrix of the target vehicle and the surrounding vehicle has overlapping parts in each time step of the future prediction. If so, it outputs 1, otherwise it outputs 0;

The collision intersection function is:

;

In the formula, To predict the time horizon of future trajectories, For the target vehicle, For surrounding agents, is the matrix of the target vehicle, is the matrix of surrounding agents;

The target vehicle and the surrounding agents at the prediction time step The collision probability prediction function within the time range is:

;

In the formula, is the event of collision between the target vehicle and surrounding agents, = ; The target vehicle and the surrounding agents at the prediction time step The integral probability of a collision occurring; It refers to the time between the target vehicle and the surrounding agents [ ] is the distribution probability obtained through interactive prediction;

Assume that there are n surrounding agents and define the target vehicle at T There is at least one collision with a surrounding agent within the range. （ ), in the future total time steps The probability of a collision occurring within （ )) is calculated as:

;

In the formula, p（ （ )) Calculates the total time steps from s to n weekly vehicles The collision probability.

As a preferred embodiment of the present invention, in step 3.2, the calculation formula of the critical collision probability remaining time calculation module is:

;

In the formula, means at all possible prediction time steps Find the smallest one so that The collision probability Exceeds the critical value CCP.

Advantages and beneficial effects of the present invention:

(1) The method provided by the present invention uses online maps as the input of the collision risk prediction model, avoiding the high maintenance cost of using high-precision maps and improving the scalability of the collision risk prediction model.

(2) The present invention establishes a generative model for online map uncertainty, incorporates the uncertainty of online maps into downstream prediction tasks, and effectively improves the accuracy of collision risk prediction.

(3) The endpoint prediction in the vehicle trajectory prediction module of the present invention uses a fixed-point predictor and an environment-adaptive predictor. When predicting the endpoints of multiple agents, the environment-adaptive predictor can learn dynamic weights according to the environment in which each agent is located, which can effectively avoid the waste of computing resources and improve the response speed of the risk prediction module.

(4) The present invention adopts a method that combines the cutting-edge technology of deep learning with traditional risk prediction, fully considers the problems encountered in the upstream module of the risk prediction module, and proposes corresponding improvement plans, providing a new idea and perspective for the establishment of a risk early warning system in the human-machine co-driving takeover system.

(5) In the method provided by the present invention, TTCCP provides a specific time value, indicating how long the system estimates that the vehicle will face a potential collision risk value that reaches or exceeds the critical level defined by the system; this value can be used to activate a safety warning, providing an effective and accurate safety assessment signal for the takeover system of the autonomous driving vehicle under human-machine co-driving.

BRIEF DESCRIPTION OF THE DRAWINGS

By referring to the following description in conjunction with the accompanying drawings, and with a more comprehensive understanding of the present invention, other objects and results of the present invention will become more apparent and easy to understand. In the accompanying drawings:

FIG1 is a flow chart of a method for predicting collision risk of autonomous driving by integrating online map uncertainty provided by the present invention;

FIG2 is a flow chart of an environment adaptive predictor predicting an endpoint of the present invention;

FIG3 is a flowchart of collision risk prediction based on predicted trajectory of the present invention.

DETAILED DESCRIPTION

In order to enable those skilled in the art to better understand the technical solution and advantages of the present invention, the present application is described in detail below in conjunction with the accompanying drawings, but it is not intended to limit the protection scope of the present invention.

As shown in FIGS. 1 to 3 , this embodiment provides a method for predicting collision risk of autonomous driving by integrating uncertainty of online maps, and the method includes the following steps:

Step 1: Generation of online map uncertainty;

Step 1.1: Collect the information of surrounding roads and agents obtained by the on-board camera and radar, and perform corresponding noise reduction processing;

In this embodiment, the surrounding road information includes: road type, speed limit, whether it is at an intersection, traffic lights, and road slope; wherein the road type includes intersections, left-turn lanes, right-turn lanes, and straight lanes; the surrounding agent information includes: agent type, speed, acceleration, and other information; the agent type includes intelligent agents such as pedestrians and vehicles, and in this invention, it mainly refers to vehicles, and the speed mainly refers to the vehicle speed;

Step 1.2: Input the denoised data into the BEV encoder to convert the acquired data into a common bird's eye view (BEV) spatial feature;

Step 1.3: Establish a Gaussian regression uncertainty model and a classification uncertainty model, take the BEV spatial features as input, use the online map regression model to generate map element vertices, the map element vertices include the position and type of the element vertices, and calculate the online map uncertainty related parameters through the Gaussian regression uncertainty model and the classification uncertainty model; wherein the Gaussian regression uncertainty model calculates the predicted position of each map element vertex and the parameters of the position-related uncertainty, and the classification uncertainty model calculates the confidence of the type of each map element vertex;

In this embodiment, the established Gaussian regression uncertainty model is expressed as:

;

Where V is the number of vertices of the map element M, and the coordinates of the i- th vertex are , is the mean, which represents the expected value of the vertex position, is the standard deviation, a measure of the uncertainty in the predicted position, is the variance; i and j are used to represent the jth dimension of the i- th vertex of the map element M, is the output of the Gaussian regression uncertainty model;

The classification uncertainty model uses a multi-layer perceptron (MLP) to output the confidence of the type of each map element vertex.

Step 2: Trajectory prediction integrating online map uncertainty;

Step 2.1: Use multi-layer perceptron encoding to integrate the predicted positions of the map element vertices, the parameters of the position-related uncertainty, and the confidence of the types of the map element vertices into the online map;

In this embodiment, the model for integrating the predicted position of the map element vertex and the parameters of the position-related uncertainty and the confidence of the type of the map element vertex into the online map is:

;

In the formula, is the Gaussian distribution mean vector of the i-th vertex, is the standard deviation vector of the uncertainty of the i-th vertex output by the Gaussian regression uncertainty model, is the category probability vector composed of the confidence of the i-th vertex, [;;] represents the connection of three vectors, It is a multi-layer perceptron MLP.

Step 2.2: For the online map that incorporates uncertainty parameters, use VectorNet to encode polyline segments and perform feature extraction. When encoding lane line features, use VectorNet to encode a polyline subgraph for each scene element (i.e., polyline). Use two independent subgraphs for the agent (vehicle) and lane, and obtain the feature vector by encoding. ;

Step 2.3: Get the feature vector It is passed to the interactive modeling module composed of the agent-to-lane module (AL), lane-to-lane module (LL), lane-to-agent module (LA) and agent-to-agent module (AA) stack to update the features. The updated features are ; Among them, the agent-to-lane module focuses on the relationship between the agent and the lane, the lane-to-lane module focuses on the relationship between the lane and the lane, the lane-to-agent module focuses on the relationship between the lane and the agent, and the agent-to-agent module focuses on the relationship between the agent and the agent;

Specifically, the interaction modeling module uses a self-attention encoder followed by a feedforward network (FFN) to update internal relations (AA, LL), and a cross-attention encoder followed by a feedforward network (FFN) to update cross-relations (LA, AL). To improve the expressiveness of the model, each encoder uses a multi-head attention mechanism (MHA):

;

= K = V = ;

;

In the formula, , , is the learned projection, norm regularized; Q, K, and V are the query vector, key vector, and value vector of the attention mechanism, respectively. Activation function, are all eigenvectors;

The multi-head attention block is defined as follows:

;

Stack the AA, LL, LA, and AL modules to update the features , get the updated features .

Step 2.4: Based on meta-information f and features The concatenated feature vector , use fixed-point predictors and environment-adaptive predictors to predict the endpoints of the target vehicle and surrounding agent (circumvehicle) trajectories, and use multi-layer perceptron MLP to generate trajectories;

Specifically, the meta information f and features The spliced model is:

= （ , f);

In the formula, is a multi-layer perceptron MLP, Meta information and features The concatenated vector,meta-information f includes the direction and position information of the agent at the prediction moment.

In this embodiment, the fixed-point predictor is used to predict the endpoint of the target agent (target vehicle). The fixed-point predictor predicts the endpoint with the help of a multi-layer perceptron. The formula is as follows:

= （ );

In the formula, is the predicted endpoint, Multilayer Perceptron MLP, is the meta-information f and feature The concatenated feature vector.

In this embodiment, the environment adaptive predictor is used to predict the endpoints of multiple surrounding agents (circumferential vehicles). As shown in FIG2 , the environment adaptive predictor uses dynamic weight learning to adapt to the specific situation of each agent (vehicle):

;

;

;

;

In the formula, , are trainable parameters, , It is the weight dynamic adjustment matrix in the endpoint prediction module, norm is layer normalization, and ReLU is the activation function; is a dynamic weight matrix and Features The matrix after regularization and nonlinear fitting, is the predicted endpoint.

Step 3: Collision risk prediction based on predicted trajectory;

Step 3.1: Obtain the coordinates of the target vehicle and surrounding vehicles in the future time step according to the trajectory generated in step 2.4, input the predicted coordinates of the target vehicle and surrounding vehicles in the future time step into the collision probability prediction module, and predict the probability of collision between the surrounding agents in the total future time step; as shown in Figure 3, it specifically includes the following steps:

Step 3.1.1: Get the coordinates of the four vertices of the target vehicle and surrounding vehicles;

Step 3.1.2: Construct the four vertices of the target vehicle and the surrounding vehicle into matrices respectively;

Step 3.1.3: Input the collision probability prediction module (collision judgment function) to determine whether a collision occurs;

Step 3.1.4: Input the probability calculation function (collision probability prediction function) of the target vehicle and one of the surrounding vehicles to calculate the probability;

Step 3.1.5: Calculate the collision probability between the target vehicle and all surrounding vehicles.

In this embodiment, the collision probability prediction module establishes a collision intersection function, a target vehicle and surrounding agents in the prediction time step. The collision probability prediction function within the time range of ; Among them, the collision intersection function is used to determine whether the matrix of the target vehicle and the surrounding vehicle has overlapping parts in each time step of the future prediction. If there is an output 1, it means that a collision has occurred; if there is no output 0, it means that no collision has occurred;

The collision intersection function is:

;

In the formula, To predict the time horizon of future trajectories, For the target vehicle, For surrounding agents, is the matrix of the target vehicle, is the matrix of surrounding agents;

The target vehicle and the surrounding agents (surrounding vehicles) at the prediction time step The collision probability prediction function within the time range is:

;

In the formula, is the event of the target vehicle colliding with surrounding vehicles. = ; The target vehicle and surrounding vehicles at the prediction time step The integral probability of a collision occurring; It refers to the time between the target vehicle and the surrounding agents [ ] is the distribution probability obtained through interactive prediction;

Assume that there are n vehicles around and define the target vehicle at T There is at least one collision with a surrounding vehicle within the range. （ ), in the future total time steps The probability of a collision occurring within （ )) is calculated as:

;

In the formula, p（ （ )) Calculates the total time steps from s to n weekly vehicles The collision probability.

Step 3.2: Input the collision probability obtained in step 3.1 into the critical collision probability remaining time (TTCCP) calculation module, calculate the critical collision probability remaining time, and transmit it to the human-machine co-driving system; the calculation formula of the critical collision probability remaining time (TTCCP) calculation module is:

;

In the formula, means at all possible prediction time steps Find the smallest one so that The collision probability Exceeds the critical value CCP.

In this embodiment, TTCCP provides a specific time value, indicating how long the system estimates that the vehicle will face a potential collision risk value that reaches or exceeds a critical level defined by the system; this value can be used to activate a safety warning, providing an effective and accurate safety assessment signal for the takeover system of the autonomous driving vehicle under human-machine co-driving.

The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method for predicting collision risk of autonomous driving that integrates uncertainty of online maps.

The present invention also provides a computer-readable medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for predicting collision risk of autonomous driving by integrating uncertainty of online maps as described above is implemented.

Those skilled in the art will appreciate that all or part of the functions of the various methods/modules in the above embodiments may be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program may be stored in a computer-readable storage medium, which may include: a read-only memory, a random access memory, a disk, an optical disk, a hard disk, etc. The program is executed by a computer to implement the above functions. For example, the program is stored in the memory of the device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

In addition, when all or part of the functions in the above-mentioned embodiments are implemented by means of a computer program, the program can also be stored in a storage medium such as a server, another computer, a disk, an optical disk, a flash drive or a mobile hard disk, and saved to the memory of a local device by downloading or copying, or the system of the local device is updated. When the program in the memory is executed by the processor, all or part of the functions in the above-mentioned embodiments can be implemented.

The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art of the present invention, according to the idea of the present invention, some simple deductions, deformations or substitutions can be made. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims (10)

1. A method for predicting collision risk of autonomous driving by integrating online map uncertainty, characterized in that the method comprises the following steps: Step 1: Generation of online map uncertainty; Step 1.1: Collect the information of surrounding roads and agents obtained by the on-board camera and radar, and perform corresponding noise reduction processing; Step 1.2: Input the denoised data into the BEV encoder to convert the acquired data into a common BEV spatial feature; Step 1.3: Establish a Gaussian regression uncertainty model and a classification uncertainty model, take the BEV spatial features as input, use the online map regression model to generate map element vertices, the map element vertices include the position and type of the element vertices, and calculate the online map uncertainty related parameters through the Gaussian regression uncertainty model and the classification uncertainty model; wherein the Gaussian regression uncertainty model calculates the predicted position of each map element vertex and the parameters of the uncertainty related to the position, and the classification uncertainty model calculates the confidence of the type of each map element vertex; Step 2: Trajectory prediction integrating online map uncertainty; Step 2.1: Use multi-layer perceptron encoding to integrate the predicted positions of the obtained map element vertices, the parameters of the uncertainty associated with the positions, and the confidence of the types of the map element vertices into the online map; Step 2.2: For the online map with fused uncertainty parameters, use VectorNet to encode multi-segment lines and perform feature extraction. When encoding lane line features, use VectorNet to encode a multi-segment line subgraph for each scene element, and use two independent subgraphs for the agent and lane to obtain the feature vector through encoding. ; Step 2.3: Get the feature vector It is passed to the interactive modeling module composed of the agent-to-lane module, lane-to-lane module, lane-to-agent module and agent-to-agent module stack to update the features. The updated features are ; Among them, the agent-to-lane module focuses on the relationship between the agent and the lane, the lane-to-lane module focuses on the relationship between the lane and the lane, the lane-to-agent module focuses on the relationship between the lane and the agent, the agent-to-agent module focuses on the relationship between the agent and the agent, the agent-to-lane module, the lane-to-lane module, the lane-to-agent module, and the agent-to-agent module are all multi-head attention blocks; Step 2.4: Based on meta-information f and features The concatenated feature vector , use fixed-point predictors and environment-adaptive predictors to predict the endpoints of the target vehicle and surrounding agent trajectories, and use multi-layer perceptrons (MLPs) to generate trajectories; Step 3: Collision risk prediction based on predicted trajectory; Step 3.1: Obtain the coordinates of the target vehicle and the surrounding agents in the future time step according to the trajectory generated in step 2.4, input the predicted coordinates of the target vehicle and the surrounding agents in the future time step into the collision probability prediction module, and predict the probability of collision between the surrounding agents in the total future time step; Step 3.2: Input the collision probability obtained in step 3.1 into the critical collision probability remaining time calculation module, calculate the critical collision probability remaining time, and transmit it to the human-machine co-driving system. 2. According to claim 1, a method for predicting collision risk of autonomous driving integrating online map uncertainty is characterized in that, in step 1.1, the surrounding road information includes: road type, speed limit, whether it is at an intersection, traffic lights, and road slope; wherein the road type includes intersections, left-turn lanes, right-turn lanes, and through lanes; the surrounding agent information includes: agent type, speed, and acceleration, and the agent type includes pedestrians and vehicles. 3. The method for predicting collision risk of autonomous driving integrating online map uncertainty according to claim 1, characterized in that in step 1.3, the established Gaussian regression uncertainty model is expressed as: ; Where V is the number of vertices of the map element M, and the coordinates of the i- th vertex are , is the mean, which represents the expected value of the vertex position, is the standard deviation, a measure of the uncertainty in the predicted position, is the variance; i and j are used to represent the jth dimension of the i- th vertex of the map element M, is the output of the Gaussian regression uncertainty model; The classification uncertainty model uses a multi-layer perceptron (MLP) to output the confidence of the type of each map element vertex. 4. The method for predicting collision risk of autonomous driving by integrating uncertainty of online maps according to claim 1, characterized in that in step 2.1, the model for integrating the predicted position of the map element vertex and the parameters of the uncertainty related to the position and the confidence of the type of the map element vertex into the online map is: ; In the formula, is the Gaussian distribution mean vector of the i-th vertex, is the standard deviation vector of the uncertainty of the i-th vertex output by the Gaussian regression uncertainty model, is the category probability vector composed of the confidence of the i-th vertex, [ ; ; ] represents the connection of three vectors, It is a multi-layer perceptron MLP. 5. The method for predicting collision risk of autonomous driving integrating online map uncertainty according to claim 1, characterized in that in step 2.3, the interactive modeling module uses a self-attention encoder followed by FFN to update the agent-to-agent module and the lane-to-lane module, and uses a cross-attention encoder followed by FFN to update the lane-to-agent module and the agent-to-lane module, and each encoder uses a multi-head attention mechanism: ; = K = V = ; ; In the formula, , , is the learned projection, norm regularized; Q, K, and V are the query vector, key vector, and value vector of the attention mechanism, respectively. Activation function, are all eigenvectors; The multi-head attention block is defined as follows: ; Stack agent to lane module, lane to lane module, lane to agent module, agent to agent module to update features , get the updated features . 6. The method for predicting collision risk of autonomous driving by integrating online map uncertainty according to claim 1, characterized in that in step 2.4, the meta information f and the feature The spliced model is: = （ , f); In the formula, is a multi-layer perceptron MLP, Meta information and features The concatenated vector,meta-information f includes the direction and position information of the agent at the prediction moment. 7. The method for predicting collision risk of autonomous driving integrating online map uncertainty according to claim 1, characterized in that in step 2.4, the fixed-point predictor is used to predict the endpoint of the target agent, and the fixed-point predictor predicts the endpoint with the help of a multi-layer perceptron, and the formula is as follows: = （ ); In the formula, is the predicted endpoint, Multilayer Perceptron MLP, is the meta-information f and feature The concatenated feature vector. 8. The method for predicting collision risk of autonomous driving integrating online map uncertainty according to claim 1 is characterized in that an environment adaptive predictor is used to predict the endpoints of multiple surrounding agents, and the environment adaptive predictor adopts a dynamic weight learning method to adapt to the specific situation of each agent: ; ; ; ; In the formula, , are trainable parameters, , It is the weight dynamic adjustment matrix in the endpoint prediction module, norm is layer normalization, and ReLU is the activation function; is a dynamic weight matrix and Features The matrix after regularization and nonlinear fitting, is the predicted endpoint. 9. The method for predicting collision risk of autonomous driving integrating online map uncertainty according to claim 1, wherein step 3.1 specifically comprises the following steps: Step 3.1.1: Get the coordinates of the four vertices of the target vehicle and the surrounding vehicle; Step 3.1.2: Construct the four vertices of the target vehicle and the surrounding vehicle into matrices respectively; Step 3.1.3: Input the collision probability prediction module to determine whether a collision occurs; Step 3.1.4: Input the probability calculation function of the target vehicle and one of the surrounding vehicles to calculate the probability; Step 3.1.5: Calculate the collision probability between the target vehicle and all surrounding vehicles; In step 3.1, the collision probability prediction module establishes the collision intersection function, the target vehicle and the surrounding agents at the prediction time step. The collision probability prediction function within the time range of ; Among them, the collision intersection function is used to determine whether the matrix of the target vehicle and the surrounding vehicle has overlapping parts in each time step of the future prediction. If so, it outputs 1, otherwise it outputs 0; The collision intersection function is: ; In the formula, To predict the time horizon of future trajectories, For the target vehicle, For surrounding agents, is the matrix of the target vehicle, is the matrix of surrounding agents; The target vehicle and the surrounding agents at the prediction time step The collision probability prediction function within the time range is: ; In the formula, is the event of collision between the target vehicle and surrounding agents, = ; The target vehicle and the surrounding agents at the prediction time step The integral probability of a collision occurring; It refers to the time between the target vehicle and the surrounding agents [ ] is the distribution probability obtained through interactive prediction; Assume that there are n surrounding agents and define the target vehicle at T There is at least one collision with a surrounding agent within the range. （ ), in the future total time steps The probability of a collision occurring within （ )) is calculated as: ; In the formula, p（ （ )) Calculates the total time steps from s to n weekly vehicles The collision probability. 10. The method for predicting collision risk of autonomous driving integrating online map uncertainty according to claim 1, characterized in that in step 3.2, the calculation formula of the critical collision probability remaining time calculation module is: ; In the formula, means at all possible prediction time steps Find the smallest one so that The collision probability Exceeds the critical value CCP.