A collision risk prediction method for autonomous vehicles considering occluded areas

Abstract

The present invention belongs to the field of road vehicle control and relates to a method for predicting collision risks of an autonomous driving vehicle taking into account an obscured area. The method first obtains historical trajectories of surrounding vehicles and constructs an online map. Then, the surrounding vehicle trajectory information and the online map are processed to obtain a scene embedding vector. For a non-obstructed area, a multi-layer perceptron is used to generate #imgabs0# anchor points. For an obstructed area, a uniform distribution algorithm is used to generate #imgabs1# anchor points. The anchor points and the scene embedding vectors are subjected to a series of processes to obtain anchor point embeddings #imgabs2#. The surrounding vehicle trajectory and the anchor point occupancy probability are predicted by a multi-layer perceptron. Finally, the overall collision risk is calculated based on parameters such as the predicted surrounding vehicle trajectory and the anchor point occupancy probability. The method makes up for the deficiency of traditional risk prediction methods that rely only on sensor data by inferring the occupancy of the obscured area, and effectively addresses the problem of limited field of view caused by the obstruction of adjacent vehicles.

Description

Technical Field

[0001] The present invention belongs to the field of road vehicle control and relates to predicting or avoiding possible or imminent collisions, and in particular, to a method for predicting collision risks of an autonomous driving vehicle taking into account occluded areas. Background Art

[0002] Collision risk prediction is usually based on the output of the trajectory module. However, most existing trajectory predictions rely solely on data obtained by sensors. This trajectory prediction method has a series of problems, such as limited perception of the field of view caused by obstructions by nearby vehicles, incomplete perception of the scene, and sensor failure in extreme weather. The obscured parts often pose major safety hazards. Therefore, it is necessary to design a new collision risk prediction method that fully considers the obscured parts when making collision risk predictions, thereby achieving more accurate risk predictions. Summary of the invention

[0003] In view of the above-mentioned technical problems and defects, the purpose of the present invention is to provide a collision risk prediction method for an autonomous driving vehicle taking into account the occluded area. The method relies on the data collected by existing sensors to not only predict the trajectory of the agent (intelligent body), but also infer the occupancy of the occluded part. By combining the occlusion inference result with the expected trajectory information of the intelligent body, the accuracy and reliability of the collision risk prediction are comprehensively improved, the decision-making and planning capabilities of the autonomous driving vehicle in complex environments are significantly enhanced, and solid support is provided for the safe implementation of autonomous driving technology.

[0004] To achieve the above object, the present invention adopts the following technical solution:

[0005] A method for predicting collision risk of an autonomous driving vehicle considering an obscured area, the method comprising the following steps:

[0006] Step 1: Generation and encoding of historical trajectories and online maps of weekly vehicles;

[0007] Step 1.1: Collect surrounding agent information and weather environment information obtained from vehicle-mounted sensors;

[0008] Step 1.2: Use the noise reduction module to perform noise reduction on the data from the vehicle-mounted sensor; and use the online map generation model to build an online map;

[0009] Step 1.3: vectorize and encode the processed agent information and the constructed online map;

[0010] Step 2: Occlusion inference and trajectory prediction;

[0011] Step 2.1: Encode the trajectory information of the surrounding agents using a one-dimensional convolutional neural network to obtain a feature vector ; At the same time, the online map information is encoded using PointNet to obtain the feature vector ;

[0012] Step 2.2: Transform the feature vectors of the surrounding agents and the feature vector of the online map Concatenate and get the feature vector ;

[0013] Step 2.3: Use transformer encoder to encode the feature vector Encode and obtain the scene embedding vector ;

[0014] Step 2.4: Determine the occluded area and the non-occluded area based on the information obtained by the vehicle-mounted sensor. For the non-occluded area, that is, the area observed by the vehicle-mounted sensor, use a multi-layer perceptron to generate Anchor points represent the presence of agents at the anchor point position, that is, the position of the agent observed by the vehicle-mounted sensor; for the occluded area, a uniform distribution algorithm is used to generate Anchor points, define anchor points ;in, Represents the total number of anchor points in the occluded area and the non-occluded area, N a =N1+N0, R represents the dimensional space;

[0015] Step 2.5: Anchor Points Encode it so that its dimension is the same as the scene embedding vector Align and get anchor code ;

[0016] Step 2.6: Encode the anchor points and the scene embedding vector Perform multi-head attention operation to obtain the feature vector ;

[0017] Step 2.7: Transform the feature vector and anchor code Perform residual normalization to obtain the eigenvector ;

[0018] ;

[0019] In the formula, represents the residual normalization layer;

[0020] Step 2.8: Transform the feature vector Use feedforward neural network to process and get the feature vector ;

[0021] ;

[0022] In the formula, It is a feed-forward neural network;

[0023] Step 2.9: Transform the feature vector and the eigenvector Perform residual normalization to get anchor point embedding ;

[0024] ；

[0025] In the formula, represents the residual normalization layer, Embed for anchor point;

[0026] Step 2.10: Embedding the Anchor Points The first multi-layer perceptron is used to predict K multimodal trajectories of the surrounding agents. , the second multi-layer perceptron is used to predict the probability of each trajectory of the agent , the third multi-layer perceptron is used to predict the probability of anchor point occupancy in the occluded area , that is, the probability that there is an agent at the anchor point in the occluded area; the fourth multi-layer perceptron is used to predict the speed of the surrounding agents and acceleration ;

[0027] Step 3: Risk prediction;

[0028] Step 3.1: Based on the predicted K trajectories of the surrounding agents, the predicted position of the agent at the future moment and the position of the anchor point in the occluded area Perform distance calculation to obtain the distance from the agent to the anchor point of the occluded area at time t ;

[0029] Step 3.2: Based on the distance from the agent to the anchor point of the occluded area at time t and critical collision distance Calculate real-time collision probability ;

[0030] Step 3.3: Correct the real-time collision probability according to weather conditions to obtain the real-time collision probability after adding weather conditions correction ;

[0031] Step 3.4: Calculate the risk impact factor of weather conditions on speed and acceleration weighted by the weather conditions ;

[0032] Step 3.5: Combine the probabilities of each trajectory , Occlusion area anchor point occupancy probability , Added real-time collision probability after weather conditions correction , risk factors of weather conditions on speed and acceleration weighted , calculate the final risk value ;

[0033] Step 3.6: Accumulate the risk values ​​of all time steps to obtain the overall collision risk of the vehicle at the future moment ;

[0034] Step 3.7: Assign overall collision risk Input into the risk assessment module for risk assessment.

[0035] As a preferred embodiment of the present invention, in step 1.1, the surrounding agent information includes: the type of surrounding agents, the trajectory information of surrounding agents, and the type of surrounding vehicles. The types of surrounding agents include pedestrians and vehicles. The trajectory information of surrounding agents includes the trajectory coordinates of the agents, the speed, acceleration, steering angle, etc. of the agents; the types of surrounding vehicles include bicycles, electric vehicles, cars, and trucks; the weather environment information includes surrounding road information and weather information, and the surrounding road information includes: the type of road, speed limit, whether it is at an intersection, traffic lights, and the slope of the road; wherein the types of roads include intersections, left-turn lanes, right-turn lanes, and through lanes.

[0036] As a preferred embodiment of the present invention, in step 1.2, a Kalman filter algorithm is used to perform noise reduction processing on the data from the vehicle-mounted sensor; and StreamMapNet is used to construct an online map;

[0037] In step 1.3, the processed agent information and the constructed online map are vectorized and encoded using the VectorNet encoder.

[0038] As a preferred embodiment of the present invention, in step 2.2, the feature vectors of the surrounding agents are and the feature vector of the online map The model for splicing is expressed as:

[0039] ,M);

[0040] In the formula, is the feature vector obtained after concatenation, is the feature vector of the surrounding agent, is the feature vector of the online map, To concatenate two feature vectors sequentially;

[0041] In step 2.3, the feature vector is encoded using the transformer The encoding model is expressed as:

[0042] ;

[0043] In the formula, represents the scene embedding vector, Represents the encoder of the Transformer, which consists of a multi-head self-attention head, a feed-forward neural network, and two residual normalization layers.

[0044] As a preferred embodiment of the present invention, in step 2.4, for the occluded area, a uniform distribution algorithm is used to generate The specific steps of anchor points are: divide the occluded area into Anchor points are generated at the center of each grid or at a fixed position. By dividing the grid, the anchor points are evenly distributed in the entire occluded area.

[0045] As a preferred embodiment of the present invention, in step 2.5, a multi-layer perceptron is used to align the anchor points. Encode it so that its dimension is the same as the scene embedding vector Alignment, expressed as:

[0046] ；

[0047] In the formula, represents a multi-layer perceptron;

[0048] In step 2.6, encode the anchor point and the scene embedding vector The model for multi-head attention operation is:

[0049] ;

[0050] ;

[0051] In the formula, They are query matrix, key matrix, and value matrix respectively. They are query vector, key vector, and value vector respectively. is the feature vector output by the multi-head attention mechanism, is the activation function, is the dimension of the potential vector, which takes a value of 128, and T represents the transposition of the matrix.

[0052] As a preferred embodiment of the present invention, in step 3.1, the predicted position of the agent at the future time and the position of the anchor point in the occluded area The distance calculation model is:

[0053] ;

[0054] In the formula, represents the distance from the agent to the anchor point of the occluded area at time t;

[0055] In step 3.2, the real-time collision probability The calculation model is:

[0056] ;

[0057] In the formula, is a parameter that controls the speed at which the collision probability changes. , It is the critical collision distance, which is controlled at 0.5-2m.

[0058] As a preferred embodiment of the present invention, in step 3.3, the real-time collision probability calculation model after adding weather condition correction is:

[0059] ;

[0060] In the formula, represents the real-time collision probability, Indicates the real-time collision probability after adding weather conditions correction, is the weather condition influencing factor, which is set to 1.0 on sunny days, 1.5 on rainy days, 2.0 on snowy days, and 1.8 on foggy days;

[0061] Step 3.4: The risk impact factor calculation model after weather conditions weight speed and acceleration is:

[0062] ;

[0063] Where W(t) is the weather condition influencing factor, is the speed risk weighting factor, , is the speed of the agent at time t, is the acceleration risk weighting factor, is the acceleration of the agent at time t, represents the absolute value, It is the risk impact factor of weather conditions weighted by speed and acceleration.

[0064] As a preferred embodiment of the present invention, in step 3.5, the final risk value The calculation model is:

[0065] ;

[0066] In the formula, represents the final risk value at time step t;

[0067] In step 3.6, the calculation model of the overall collision risk of the vehicle at the future moment is:

[0068] ;

[0069] Where T1 represents the number of future time steps.

[0070] As a preferred embodiment of the present invention, in step 3.7, two thresholds are set in the risk judgment module, and the overall collision risk is Compare with the threshold to determine the collision risk. The specific judgment rules are:

[0071] Low risk: ;

[0072] Medium risk: ;

[0073] High risk: ;

[0074] In the formula, and are the predefined low-risk threshold and medium-risk threshold, which are 0.2 and 0.5 respectively.

[0075] Advantages and beneficial effects of the present invention:

[0076] (1) The method provided by the present invention integrates trajectory prediction and occlusion inference into one model, which greatly reduces the waste of computing resources and is conducive to improving the real-time performance of the model.

[0077] (2) The method provided by the present invention makes up for the deficiency of traditional risk prediction methods that only rely on sensor data by inferring the occupancy of the blocked area, and effectively addresses the problem of limited field of view caused by blockage by nearby vehicles.

[0078] (3) The method provided by the present invention combines the intelligent agent trajectory prediction and occlusion inference results, so that it can still achieve more accurate risk prediction in complex environments and effectively reduce the safety hazards caused by perception blind spots.

[0079] (4) The method provided by the present invention reduces the complete reliance on sensors by inferring the occluded area. This can continue to provide reliable and effective risk prediction when sensor data is incomplete, noisy or faulty. This can reduce the workload of the sensor, extend its service life, and reduce safety issues caused by sensor failure.

[0080] (5) The method provided by the present invention introduces weather conditions, so that the system can not only operate effectively in normal driving environments, but also make reliable risk predictions in extreme weather conditions such as rainy days, snowy days, and foggy days; the addition of weather factors improves the safety of the system, ensuring that more accurate driving decisions can be made under low visibility or low friction conditions, significantly reducing the risk of accidents. At the same time, the model prediction results are dynamically adjusted in combination with weather data, which enhances the robustness of the system in the face of sensor performance degradation. It relies on existing sensor data without adding additional hardware equipment, has high economy and feasibility, and is easy to integrate and apply in existing systems.

[0081] (6) The method provided by the present invention is applicable to a variety of driving scenarios, including low-speed driving in cities and on highways, and has adaptability to all scenarios, further expanding the application scope of autonomous driving technology in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] 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 clear and easy to understand. In the accompanying drawings:

[0083] Figure 1 A flow chart of a method for predicting collision risk of an autonomous vehicle taking into account an obscured area provided by the present invention;

[0084] Figure 2 For the present invention Figure 1 The data processing flow chart of the decoder provided in;

[0085] Figure 3 A conceptual diagram of calculating the collision distance for the risk prediction part provided by the present invention;

[0086] Figure 4 The present invention is a flow chart of risk prediction. DETAILED DESCRIPTION

[0087] 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.

[0088] like Figures 1 to 4 As shown, this embodiment provides a method for predicting collision risk of an autonomous driving vehicle taking into account an obscured area, the method comprising the following steps:

[0089] Step 1: Generation and encoding of historical trajectories and online maps of weekly vehicles;

[0090] Step 1.1: Collect surrounding agent information and weather environment information obtained from vehicle-mounted sensors (cameras, radars, etc.); wherein the surrounding agent information includes: the type of surrounding agents, the trajectory information of surrounding agents, and the type of surrounding vehicles. The types of surrounding agents include pedestrians and vehicles. The trajectory information of surrounding agents includes the trajectory coordinates of the agents, the speed, acceleration, steering angle, etc. of the agents; the types of surrounding vehicles include bicycles, electric vehicles, cars, and trucks; the weather environment information includes surrounding road information and weather information. The surrounding road information includes: the type of road, speed limit, whether it is at an intersection, traffic lights, and the slope of the road; wherein the types of roads include intersections, left-turn lanes, right-turn lanes, and through lanes.

[0091] Step 1.2: Use the noise reduction module to perform noise reduction on the data from the vehicle-mounted sensor; and use the online map generation model to build an online map;

[0092] Specifically, in this embodiment, the Kalman filter algorithm is used to perform noise reduction processing on the data from the vehicle-mounted sensor; and StreamMapNet is used to build an online map.

[0093] Step 1.3: vectorize and encode the processed agent information and the constructed online map;

[0094] Specifically, in this embodiment, the processed agent information and the constructed online map are vectorized and encoded using the VectorNet encoder.

[0095] Step 2: Occlusion inference and trajectory prediction;

[0096] Step 2.1: Encode the trajectory information of the surrounding agents using a one-dimensional convolutional neural network to obtain a feature vector ; At the same time, the online map information is encoded using PointNet to obtain the feature vector ;

[0097] Step 2.2: Transform the feature vectors of the surrounding agents and the feature vector of the online map Concatenate and get the feature vector ;

[0098] Specifically, in this embodiment, the feature vectors of the surrounding agents are and the feature vector of the online map The model for splicing is expressed as:

[0099] ,M);

[0100] In the formula, is the feature vector obtained after concatenation, is the feature vector of the surrounding agent, is the feature vector of the online map, To concatenate two feature vectors sequentially.

[0101] Step 2.3: Use transformer encoder to encode the feature vector Encode and obtain the scene embedding vector ;

[0102] Specifically, in this embodiment, a transformer encoder is used to transform the feature vector The encoding model is expressed as:

[0103] ;

[0104] In the formula, represents the scene embedding vector, Represents the encoder of the Transformer, which consists of a multi-head self-attention head, a feed-forward neural network, and two residual normalization layers.

[0105] Step 2.4: Determine the occluded area and the non-occluded area based on the information obtained by the on-board sensor. For the non-occluded area, that is, the area observed by the on-board sensor, use a multi-layer perceptron (MLP) to generate Anchor points represent the location (anchor point location) where an agent exists, that is, the location of the agent observed by the vehicle-mounted sensor; for the occluded area, a uniform distribution algorithm is used to generate Anchor points, define anchor points ;in, Represents the total number of anchor points in the occluded area and the non-occluded area, N a =N1+N0, R represents the dimensional space;

[0106] In this embodiment, a uniform distribution algorithm is used to generate The specific steps of anchor points are: divide the occluded area into Anchor points are generated at the center of each grid or at a fixed position. By dividing the grid, the anchor points are evenly distributed in the entire occluded area.

[0107] Step 2.5: Anchor Points Encode it so that its dimension is the same as the scene embedding vector Align and get anchor code ;

[0108] Specifically, in this embodiment, a multi-layer perceptron is used to identify the anchor points. Encode it so that its dimension is the same as the scene embedding vector Alignment, expressed as:

[0109] ；

[0110] In the formula, Represents a multilayer perceptron.

[0111] Step 2.6: Encode the anchor points and the scene embedding vector Perform multi-head attention operation to obtain the feature vector ;

[0112] In this embodiment, the anchor point is encoded and the scene embedding vector The model for multi-head attention operation is:

[0113] ;

[0114] ;

[0115] In the formula, They are query matrix, key matrix, and value matrix respectively. They are query vector, key vector, and value vector respectively. is the feature vector output by the multi-head attention mechanism, is the activation function, is the dimension of the potential vector, which takes a value of 128, and T represents the transposition of the matrix.

[0116] Step 2.7: Transform the feature vector and anchor code Perform residual normalization to obtain the eigenvector ;

[0117] ;

[0118] In the formula, represents the residual normalization layer;

[0119] Step 2.8: Transform the feature vector Use feedforward neural network to process and get the feature vector ;

[0120] ;

[0121] In the formula, It is a feed-forward neural network;

[0122] Step 2.9: Transform the feature vector and the eigenvector Perform residual normalization to get anchor point embedding ;

[0123] ；

[0124] In the formula, represents the residual normalization layer, Embed for anchor point;

[0125] Step 2.10: Embedding the Anchor Points The first MLP is used to predict K multimodal trajectories of the surrounding agents. , the second multi-layer perceptron is used to predict the probability of each trajectory of the agent , the third multi-layer perceptron is used to predict the probability of anchor point occupancy in the occluded area , that is, the probability that there is an agent at the anchor point in the occluded area; the fourth multi-layer perceptron is used to predict the speed of the surrounding agents and acceleration ;

[0126] In step 2 described in this embodiment, the constructed neural network can be trained with historical data, and the trained model is used as the final trajectory prediction and anchor point occupancy probability prediction model. Regarding the specific training method, those skilled in the art can refer to the prior art, which is not the key point of this application and will not be described in detail in this application.

[0127] Step 3: Risk prediction;

[0128] Step 3.1: Based on the predicted K trajectories of the surrounding agents, the predicted position of the agent at the future moment and the position of the anchor point in the occluded area Perform distance calculation to obtain the distance from the agent to the anchor point of the occluded area at time t ;

[0129] ;

[0130] In the formula, Represents the distance from the agent to the anchor point of the occluded area at time t.

[0131] Step 3.2: Based on the distance from the agent to the anchor point of the occluded area at time t and critical collision distance Calculate real-time collision probability ;

[0132] ;

[0133] In the formula, is a parameter that controls the speed at which the collision probability changes. , The choice depends on the vehicle speed, road type and safety requirements. It is the critical collision distance, which is controlled at 0.5-2m according to vehicle speed and road type.

[0134] Step 3.3: Correct the real-time collision probability according to weather conditions to obtain the real-time collision probability after adding weather conditions correction ;

[0135] ;

[0136] In the formula, represents the real-time collision probability, Indicates the real-time collision probability after adding weather conditions correction, It is the weather condition influencing factor, which is set to 1.0 for sunny days, 1.5 for rainy days, 2.0 for snowy days, and 1.8 for foggy days.

[0137] Step 3.4: Calculate the risk impact factor of weather conditions on speed and acceleration weighted by the weather conditions ;

[0138] ;

[0139] Where W(t) is the weather condition influencing factor, is the speed risk weighting factor, , is the speed of the agent at time t, is the acceleration risk weighting factor, is the acceleration of the agent at time t, represents the absolute value, It is the risk impact factor of weather conditions weighted by speed and acceleration.

[0140] Step 3.5: Combine the probabilities of each trajectory , Occlusion area anchor point occupancy probability , Added real-time collision probability after weather conditions correction , risk factors of weather conditions on speed and acceleration weighted , calculate the final risk value ;

[0141] ;

[0142] In the formula, represents the final risk value at time step t;

[0143] Step 3.6: Accumulate the risk values ​​of all time steps to obtain the overall collision risk of the vehicle at the future moment ;

[0144] ;

[0145] Where T1 represents the number of future time steps.

[0146] Step 3.7: Assign overall collision risk Input into the risk assessment module for risk assessment.

[0147] In this embodiment, two thresholds are set in the risk judgment module. Compare with the threshold to determine the collision risk. The specific judgment rules are:

[0148] Low risk: ;

[0149] Medium risk: ;

[0150] High risk: ;

[0151] In the formula, and are the predefined low-risk threshold and medium-risk threshold, which are 0.2 and 0.5 respectively.

[0152] In this embodiment, after the risk judgment is made, corresponding measures are taken according to the risk level. The specific control method is: if the risk is low, the vehicle will drive normally without changing the current speed and path; and the driver does not need to receive any warnings or prompts; if the risk is medium, the vehicle will slow down appropriately and start planning possible avoidance paths; at the same time, a warning sound or flashing lights will be issued to remind the driver of the increased risk; if the risk is high, emergency avoidance measures such as braking or changing the route will be taken immediately to avoid a collision; at the same time, a strong warning will be issued to the driver, indicating that a collision is about to occur and an emergency response is required.

[0153] 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 risks of an autonomous driving vehicle taking into account obscured areas.

[0154] 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 an autonomous driving vehicle taking into account obscured areas as described above is implemented.

[0155] 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.

[0156] 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.

[0157] 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

1. A method for predicting collision risk of an autonomous vehicle considering an obscured area, characterized in that: The method comprises the following steps: Step 1: Generation and encoding of historical trajectories and online maps of weekly vehicles; Step 1.1: Collect surrounding agent information and weather environment information obtained from vehicle-mounted sensors; Step 1.2: Use the noise reduction module to perform noise reduction on the data from the vehicle-mounted sensor; and use the online map generation model to build an online map; Step 1.3: vectorize and encode the processed agent information and the constructed online map; Step 2: Occlusion inference and trajectory prediction; Step 2.1: Encode the trajectory information of the surrounding agents using a one-dimensional convolutional neural network to obtain a feature vector ; At the same time, the online map information is encoded using PointNet to obtain the feature vector ; Step 2.2: Transform the feature vectors of the surrounding agents and the feature vector of the online map Concatenate and get the feature vector ; Step 2.3: Use transformer encoder to encode the feature vector Encode and obtain the scene embedding vector ; Step 2.4: Determine the occluded area and the non-occluded area based on the information obtained by the vehicle-mounted sensor. For the non-occluded area, that is, the area observed by the vehicle-mounted sensor, use a multi-layer perceptron to generate Anchor points represent the presence of agents at the anchor point position, that is, the position of the agent observed by the vehicle-mounted sensor; for the occluded area, a uniform distribution algorithm is used to generate Anchor points, define anchor points ; where R represents the dimensional space, Represents the total number of anchor points in the occluded area and the non-occluded area, N a =N1+N0; Step 2.5: Anchor Points Encode it so that its dimension is the same as the scene embedding vector Align and get anchor code ; Step 2.6: Encode the anchor points and the scene embedding vector Perform multi-head attention operation to obtain the feature vector ; Step 2.7: Transform the feature vector and anchor code Perform residual normalization to obtain the eigenvector ; ； In the formula, represents the residual normalization layer; Step 2.8: Transform the feature vector Use feedforward neural network to process and get the feature vector ; ； In the formula, It is a feed-forward neural network; Step 2.9: Transform the feature vector and the eigenvector Perform residual normalization to get anchor point embedding ; ； In the formula, represents the residual normalization layer, Embed for anchor point; Step 2.10: Embedding the Anchor Points The first multi-layer perceptron is used to predict K multimodal trajectories of the surrounding agents. , the second multi-layer perceptron is used to predict the probability of each trajectory of the agent , the third multi-layer perceptron is used to predict the probability of anchor point occupancy in the occluded area , that is, the probability that there is an agent at the anchor point in the occluded area; the fourth multi-layer perceptron is used to predict the speed of the surrounding agents and acceleration ; Step 3: Risk prediction; Step 3.1: Based on the predicted K trajectories of the surrounding agents, the predicted position of the agent at the future moment and the position of the anchor point in the occluded area Perform distance calculation to obtain the distance from the agent to the anchor point of the occluded area at time t ; Step 3.2: Based on the distance from the agent to the anchor point of the occluded area at time t and critical collision distance Calculate real-time collision probability ; Step 3.3: Correct the real-time collision probability according to weather conditions to obtain the real-time collision probability after adding weather conditions correction ; Step 3.4: Calculate the risk impact factor of weather conditions on speed and acceleration weighted by the weather conditions ; Step 3.5: Combine the probabilities of each trajectory , Occlusion area anchor point occupancy probability , Added real-time collision probability after weather conditions correction , risk factors of weather conditions on speed and acceleration weighted , calculate the final risk value ; Step 3.6: Add up the risk values ​​of all time steps to get the overall collision risk of the vehicle at the future moment ; Step 3.7: Assign overall collision risk Input into the risk assessment module for risk assessment.

2. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 1.1, the surrounding agent information includes: the type of surrounding agents, the trajectory information of surrounding agents, and the type of surrounding vehicles. The types of surrounding agents include pedestrians and vehicles. The trajectory information of surrounding agents includes the trajectory coordinates of the agents, the speed, acceleration, steering angle, etc. of the agents; the types of surrounding vehicles include bicycles, electric vehicles, cars, and trucks; the weather environment information includes surrounding road information and weather information. The surrounding road information includes: the type of road, speed limit, whether it is at an intersection, traffic lights, and the slope of the road; wherein the types of roads include intersections, left-turn lanes, right-turn lanes, and through lanes.

3. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 1.2, the Kalman filter algorithm is used to reduce the noise of the data from the vehicle-mounted sensor; at the same time, StreamMapNet is used to build an online map; In step 1.3, the processed agent information and the constructed online map are vectorized and encoded using the VectorNet encoder.

4. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 2.2, the feature vectors of the surrounding agents are and the feature vector of the online map The model for splicing is expressed as: ,M)； In the formula, is the feature vector obtained after concatenation, is the feature vector of the surrounding agent, is the feature vector of the online map, To concatenate two feature vectors sequentially; In step 2.3, the feature vector is encoded using the transformer The encoding model is expressed as: ； In the formula, represents the scene embedding vector, Represents the encoder of the Transformer, which consists of a multi-head self-attention head, a feed-forward neural network, and two residual normalization layers.

5. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 2.4, for the occluded area, a uniform distribution algorithm is used to generate The specific steps of anchor points are: divide the occluded area into Anchor points are generated at the center of each grid or at a fixed position. By dividing the grid, the anchor points are evenly distributed in the entire occluded area.

6. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 2.5, use a multi-layer perceptron to align the anchor points Encode it so that its dimension is the same as the scene embedding vector Alignment, expressed as: ； In the formula, represents a multi-layer perceptron; In step 2.6, encode the anchor point and the scene embedding vector The model for multi-head attention operation is: ； ； In the formula, They are query matrix, key matrix, and value matrix respectively. They are query vector, key vector, and value vector respectively. is the feature vector output by the multi-head attention mechanism, is the activation function, is the dimension of the potential vector, which is 128, and T represents the transposition of the matrix.

7. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 3.1, the predicted position of the agent at the future time and the position of the anchor point in the occluded area The distance calculation model is: ； In the formula, represents the distance from the agent to the anchor point of the occluded area at time t; In step 3.2, the real-time collision probability The calculation model is: ； In the formula, is a parameter that controls the speed at which the collision probability changes. , It is the critical collision distance, which is controlled at 0.5-2m.

8. The method for predicting collision risk of an autonomous driving vehicle taking into account an obscured area according to claim 1, characterized in that: In step 3.3, the real-time collision probability calculation model after adding weather condition correction is: ； In the formula, represents the real-time collision probability, Indicates the real-time collision probability after adding weather conditions correction, is the weather condition influencing factor, which is set to 1.0 on sunny days, 1.5 on rainy days, 2.0 on snowy days, and 1.8 on foggy days; Step 3.4: The risk impact factor calculation model after weather conditions weight speed and acceleration is: ； Where W(t) is the weather condition influencing factor, is the speed risk weighting factor, , is the speed of the agent at time t, is the acceleration risk weighting factor, is the acceleration of the agent at time t, represents the absolute value, It is the risk impact factor of weather conditions weighted by speed and acceleration.

9. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 3.5, the final risk value The calculation model is: ; In the formula, represents the final risk value at time step t; In step 3.6, the calculation model of the overall collision risk of the vehicle at the future moment is: ； Where T1 represents the number of future time steps.

10. The method for predicting collision risk of an autonomous driving vehicle considering an obscured area according to claim 1, characterized in that: In step 3.7, two thresholds are set in the risk judgment module. Compare with the threshold to determine the collision risk. The specific judgment rules are: Low risk: ; Medium risk: ; High risk: ; In the formula, and are the predefined low-risk threshold and medium-risk threshold, which are 0.2 and 0.5 respectively.

Images