A collision risk prediction method for autonomous vehicles targeting the long-tail phenomenon

Abstract

The present invention belongs to the field of road vehicle control and relates to a method for predicting collision risks of autonomous vehicles targeting the long-tail phenomenon. The method screens high-risk trajectory samples and adopts a trajectory generation method that integrates a diffusion probability model and a Transformer structure to achieve high-fidelity modeling and synthesis of rare high-risk driving behaviors, thereby enhancing the model's learning ability for long-tail samples. In addition, considering that in long-tail scenarios, the kinematic characteristics and traffic flow density characteristics of vehicles usually show significant fluctuations, the kinematic characteristics and traffic flow density characteristics of surrounding traffic participants are constructed and extracted, and the future trajectories of surrounding vehicles are predicted through multimodal feature fusion. On this basis, a probabilistic risk assessment mechanism for different risk scenarios is constructed to dynamically assess potential collision risks or the probability of abnormal behavior, ultimately achieving refined predictions of potential high-risk events in complex traffic environments.

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. Specifically, it relates to a collision risk prediction method for autonomous driving vehicles targeting the long-tail phenomenon. Background Art

[0002] With the rapid development of autonomous driving technology, vehicle safety has become a key prerequisite for its implementation. Autonomous vehicles rely on real-time perception and prediction of the surrounding traffic environment to proactively identify and avoid potential collision risks, thereby ensuring safe driving. However, existing autonomous vehicle risk prediction methods still have significant shortcomings, particularly when dealing with the "long tail phenomenon" in real-world traffic scenarios. The long tail phenomenon refers to a small number of extremely dangerous traffic scenarios (such as abrupt lane changes, sudden braking, and sudden collisions). Due to the extremely scarce data for these extremely dangerous scenarios and the relatively abundant data for most scenarios (such as normal, stable driving), this imbalance in data distribution causes existing prediction models to overlook or underestimate the probability and severity of these extremely dangerous scenarios, resulting in large prediction errors for high-risk trajectories, delayed response, and misidentification of risks. Therefore, there is an urgent need to design a new collision risk prediction method that can accurately predict the long tail phenomenon to improve the autonomous driving system's ability to identify potential risks and improve its prediction accuracy in complex scenarios. Summary of the Invention

[0003] In view of the above-mentioned technical problems and shortcomings, the present invention proposes a collision risk prediction method for autonomous vehicles targeting the long-tail phenomenon. The method first processes the operating data collected by the autonomous vehicle, focusing on screening high-risk trajectory samples. On this basis, a trajectory generation method that fuses a diffusion probability model with a Transformer structure is adopted to achieve high-fidelity modeling and synthesis of rare high-risk driving behaviors, thereby compensating for the lack of data on rare dangerous events in long-tail scenarios and enhancing the model's learning ability for long-tail samples. In addition, considering that in long-tail scenarios, the vehicle's kinematic characteristics (such as speed and acceleration) and traffic flow density characteristics (such as local traffic density and lane change frequency) often show significant fluctuations, the present invention further constructs and extracts the kinematic characteristics and traffic flow density characteristics of surrounding traffic participants, and predicts the future trajectories of surrounding vehicles through multimodal feature fusion. On this basis, a probabilistic risk assessment mechanism is constructed for different risk scenarios to dynamically assess potential collision risks or the probability of abnormal behavior, ultimately achieving refined prediction of potential high-risk events in complex traffic environments.

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

[0005] A collision risk prediction method for autonomous vehicles targeting the long-tail phenomenon is proposed. This method uses a multimodal enhanced trajectory dataset to train a collision risk prediction model. The trained collision risk prediction model is then used to perform risk prediction. The specific risk prediction steps are as follows:

[0006] Step 1. Construct the dynamic feature vector of surrounding vehicles;

[0007] Step 2. Output the dynamic fluctuation type based on the dynamic characteristic vectors of the surrounding vehicles;

[0008] Step 3. Construct traffic flow feature vector;

[0009] Step 4. Output the traffic flow state according to the traffic flow feature vector;

[0010] Step 5. Trajectory prediction based on multimodal feature fusion:

[0011] Encode the historical trajectory information of surrounding vehicles to obtain historical trajectory features , encode high-precision map information to obtain map features ; Perform feature mapping on the dynamic feature vectors of surrounding vehicles and traffic flow feature vectors to obtain dynamic feature and traffic flow characteristics ; The historical trajectory characteristics and dynamic characteristics Splicing to get features , map features Traffic flow characteristics Splicing to get features ; The feature and features Perform interaction modeling to obtain features ; for features Process and obtain features ; for features Decode the future trajectory, speed, and acceleration of surrounding vehicles;

[0012] Step 6. Collision risk prediction:

[0013] Based on historical data, the vehicle with the shortest distance from the autonomous vehicle is selected as the target vehicle, the maximum speed and maximum acceleration of the target vehicle are extracted, and the standard deviation of acceleration is calculated; a dynamic collision risk index is constructed based on the maximum speed, maximum acceleration and standard deviation of acceleration of the target vehicle; finally, the dynamic fluctuation coefficient is determined according to the dynamic fluctuation type, and the traffic flow state label is determined according to the traffic flow state, and the dynamic collision risk index is corrected using the dynamic fluctuation coefficient and traffic flow state label.

[0014] As a preferred embodiment of the present invention, the steps of constructing the multimodal enhanced trajectory dataset are as follows:

[0015] Step A. Data collection and feature construction:

[0016] Step A1. Filter trajectory data and construct tail trajectory data samples;

[0017] Step A2: normalize the filtered tail trajectory data;

[0018] Step A3. Trajectory feature extraction;

[0019] Extracting multi-dimensional features of the target vehicle from the standardized tail trajectory data ; Wherein, P represents the position sequence of the target vehicle, V represents the velocity sequence of the target vehicle, A represents the acceleration sequence of the target vehicle, and K represents the curvature sequence of the target vehicle;

[0020] Step B. Data enhancement based on diffusion model:

[0021] Step B1. Based on the constructed multi-dimensional features of the target vehicle , add Gaussian noise at each time step , get the features ;

[0022] Step B2. Use the Transformer neural network to predict the Gaussian noise added at each time step , then the features Each time step feature Denoise and get features ;

[0023] Step B3. De-noised features Input into the multi-layer perceptron to generate trajectory data;

[0024] Step B4. Sampling different Gaussian noises multiple times , repeat the process of steps B2 to B3 to generate multiple different trajectory samples, thereby constructing a multimodal enhanced trajectory dataset.

[0025] As a preferred embodiment of the present invention, step 1 calculates the speed mean based on the historical data of the surrounding vehicles i. , mean acceleration , speed standard deviation , peak speed , velocity skewness , speed variation coefficient , construct the dynamic characteristic vector of the surrounding vehicle i ;

[0026] Step 2: The dynamic characteristic vector of the surrounding vehicle i Cluster centers representing high dynamic fluctuations, medium dynamic fluctuations, and low dynamic fluctuations Compare and output the dynamic fluctuation type of the surrounding vehicle i.

[0027] As a preferred embodiment of the present invention, step 3 calculates the traffic density based on the sensing range radius R of the sensor on the autonomous vehicle and the number of vehicles N within the sensing range. , traffic flow , the average speed of surrounding vehicles , construct traffic flow feature vector ={ };

[0028] Step 4: Transform the traffic flow feature vector around the autonomous vehicle and cluster centers representing high density, medium density, and low density Compare and output the traffic flow status.

[0029] As a preferred embodiment of the present invention, in step 5, the historical trajectory information is encoded using the LSTM encoder to obtain the historical trajectory features. , use the point Net-based encoder to encode the high-precision map information to obtain map features ; Use a multi-layer perceptron to perform feature mapping on the dynamic feature vectors of surrounding vehicles and traffic flow feature vectors to obtain the dynamic features of surrounding vehicles and traffic flow characteristics .

[0030] As a preferred embodiment of the present invention, in step 5, the characteristic and features Input into the multi-head cross attention module for interactive modeling to obtain features ; Use feedforward neural network to train features Process and obtain features ; Use multi-layer perceptron to classify features to decode.

[0031] As a preferred embodiment of the present invention, the specific steps of the collision risk prediction in step 6 are:

[0032] Step 6.1. Calculate the The distances from the N surrounding vehicles to the autonomous vehicle at the moment, and the vehicle with the smallest distance to the autonomous vehicle is selected as the target vehicle;

[0033] Step 6.2. Extract the maximum speed of the target vehicle based on its future speed and acceleration , maximum acceleration ;

[0034] Step 6.3. Calculate the acceleration standard deviation based on the future acceleration of the target vehicle;

[0035] Step 6.4. Use the target vehicle's maximum speed , maximum acceleration and acceleration standard deviation Constructing a dynamic collision risk index , the expression is:

[0036] ；

[0037] Where, , is the weight coefficient of velocity, acceleration and fluctuation, Indicates the maximum speed limit of the road. For safe acceleration, is the expected safety fluctuation value;

[0038] Step 6.5. Setting the Kinetic Fluctuation Coefficient Corrected dynamic collision risk indicator , and obtain the dynamic collision risk index after considering dynamic fluctuations , the expression is:

[0039] ;

[0040] Step 6.6. Set traffic flow status label Dynamic collision risk index after considering dynamic fluctuations Make corrections to obtain the final dynamic collision risk index , the expression is:

[0041] ;

[0042] Step 6.7. The final dynamic risk indicator Input to the risk judgment module for interpretation and output of risk level .

[0043] As a further preferred embodiment of the present invention, the peak speed in step 1 is , velocity skewness , speed variation coefficient The expression is shown as:

[0044] ;

[0045] ;

[0046] ;

[0047] in, represents the velocity at time step f, is the historical trajectory time step.

[0048] As a further preference of the present invention, in step 6.4 , Set to 1.0, 1.2, 1.0 respectively; Set as , Set as ;

[0049] Step 6.5: High Dynamics Fluctuation , mesodynamic fluctuations Low dynamic fluctuations ;

[0050] Step 6.6: Medium to high density , medium density , at low density .

[0051] Advantages and beneficial effects of the present invention:

[0052] (1) By screening high-risk trajectory data and combining it with a trajectory generation method based on a diffusion probability model and a Transformer structure, the present invention can synthesize rare high-risk scenarios with high fidelity, effectively expand long-tail data samples, and improve the model's coverage and learning effect for extremely dangerous scenarios.

[0053] (2) The present invention constructs a high-dimensional input space that integrates historical trajectories, kinematic information, and traffic flow density characteristics, enabling the model to more comprehensively understand the dynamic evolution process in a complex traffic environment, thereby having strong predictive robustness and generalization capabilities in a variety of potential risk scenarios.

[0054] (3) The present invention not only predicts the future trajectory of the target vehicle, but also introduces the behavior modeling and density perception of surrounding traffic participants. Combining the trajectory generation and risk identification modules, it constructs a probabilistic risk assessment mechanism for high-risk scenarios, which can more precisely identify and quantify potential collision risks.

[0055] (4) The method provided by the present invention is applicable to a variety of scenarios, including urban congested roads, high-speed driving scenarios, and even when sensor perception decreases in severe weather. It can stably perform risk prediction on traditional models or improved autonomous driving platforms, providing good support for the implementation of autonomous driving in diverse environments.

[0056] (5) Compared with traditional rule-based or static model-based methods, the present invention can achieve earlier and more accurate risk warnings through intensive training of deep learning models on long-tail data, improve the system's response efficiency and decision-making quality in sudden and dangerous scenarios, and enhance overall driving safety.

[0057] (6) The method provided by the present invention is compatible with existing hardware conditions. By analyzing the environmental perception data collected by the vehicle-mounted sensors and the limited historical driving data, it does not require the installation of additional special sensing equipment or expensive hardware modules. It has high economy and feasibility and is easy to integrate and deploy with the existing autonomous driving system. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] By referring to the following description in conjunction with the accompanying drawings, and with a more complete 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:

[0059] Figure 1 A flow chart of a method for predicting collision risk of an autonomous vehicle for the long-tail phenomenon provided by the present invention;

[0060] Figure 2 Construct a flow chart for the multimodal enhanced trajectory dataset of the present invention;

[0061] Figure 3 Constructing a flow chart for the vehicle dynamics features surrounding the present invention;

[0062] Figure 4 A flow chart constructed for the traffic flow characteristics of the present invention;

[0063] Figure 5 This is a flow chart of trajectory prediction based on multimodal feature fusion of the present invention;

[0064] Figure 6 This is a collision risk prediction flow chart of the present invention. DETAILED DESCRIPTION

[0065] In order to enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application is described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0066] like Figure 1 As shown, this embodiment provides a method for predicting collision risks of autonomous vehicles for the long-tail phenomenon. The method uses a multimodal enhanced trajectory dataset to train a collision risk prediction model, and then uses the trained collision risk prediction model to perform risk prediction. The specific risk prediction steps are as follows:

[0067] Step 1. Calculate the average speed based on the historical data of surrounding vehicles i , mean acceleration , speed standard deviation , peak speed , velocity skewness , speed variation coefficient , construct the dynamic characteristic vector of the surrounding vehicle i ;

[0068] Step 2. Transform the dynamic feature vector of the surrounding vehicle i into Cluster centers representing high dynamic fluctuations, medium dynamic fluctuations, and low dynamic fluctuations Compare and output the dynamic fluctuation type of surrounding vehicle i;

[0069] Step 3. Calculate the traffic density based on the sensing range radius R of the sensor on the autonomous vehicle and the number of vehicles N within the sensing range , traffic flow , the average speed of surrounding vehicles , construct traffic flow feature vector ={ };

[0070] Step 4. Transform the traffic flow feature vector around the autonomous vehicle and cluster centers representing high density, medium density, and low density Make comparisons and output traffic flow status;

[0071] Step 5. Trajectory prediction based on multimodal feature fusion:

[0072] Step 5.1. Vectorize and encode the historical trajectory information and high-precision map of surrounding vehicles;

[0073] Step 5.2. Encode the historical trajectory information of surrounding vehicles to obtain historical trajectory features , encode high-precision map information to obtain map features ;

[0074] Step 5.3. Dynamic characteristic vectors of N surrounding vehicles and traffic flow characteristic vector ={ } Perform feature mapping to obtain the dynamic characteristics of surrounding vehicles and traffic flow characteristics ;

[0075] Step 5.4. Characterize the historical trajectory and vehicle dynamics Splicing to get features , map features Traffic flow characteristics Splicing to get features ;

[0076] Step 5.5. Features and features Input into the multi-head cross attention module for interactive modeling to obtain features ;

[0077] Step 5.6. Use feedforward neural network to train features Process and obtain features ;

[0078] Step 5.7. Features Decode the future trajectories of N surrounding vehicles }, the speed of the future },Future acceleration ;

[0079] Step 6. Collision risk prediction:

[0080] According to historical data, the vehicle with the shortest distance from the autonomous vehicle is selected as the target vehicle. Based on the future speed and acceleration of the target vehicle, the maximum speed of the target vehicle is extracted. , maximum acceleration , calculate the standard deviation of acceleration , and based on the maximum speed of the target vehicle , maximum acceleration and acceleration standard deviation Constructing a dynamic collision risk index Finally, the dynamic fluctuation coefficient is determined according to the dynamic fluctuation type, the traffic flow state label is determined according to the traffic flow state, and the dynamic fluctuation coefficient is used , Traffic flow status label Corrected dynamic collision risk indicator .

[0081] like Figure 2 As shown, in this embodiment, the steps for constructing the multimodal enhanced trajectory dataset are as follows:

[0082] Step A. Data collection and feature construction:

[0083] Step A1. Screening of high-risk trajectory data;

[0084] Collect driving data of autonomous vehicles in actual road traffic environments, and obtain historical information of surrounding vehicles through on-board sensors (such as lidar, millimeter-wave radar, camera, GPS sensor, etc.); then preprocess the raw data and filter out trajectory segments containing obvious abnormal driving behavior or high collision risk events, which are used as tail trajectory data samples.

[0085] In this embodiment, the historical information of surrounding vehicles includes: surrounding vehicle trajectory information, surrounding vehicle types, road types, and other information; the surrounding vehicle trajectory information includes the coordinates of the vehicle's historical trajectory, the vehicle's speed, acceleration, steering angle, etc.; the surrounding vehicle types include bicycles, electric vehicles, cars, trucks, etc. Obviously abnormal driving behavior or high collision risk events include, but are not limited to: vehicle emergency braking events, vehicle violent steering or sharp turns, vehicle rapid lane changes accompanied by significant deceleration or acceleration events, rapid reduction in vehicle spacing, and collision risk indicators (such as time to collision (TTC)) less than a set threshold.

[0086] Step A2. Standardize the filtered tail trajectory data, which includes time synchronization and coordinate system synchronization. Specifically, the pulse signal provided by GPS is used to synchronize the time of multiple vehicle-mounted sensors, and the coordinate system synchronization can refer to the existing method for coordinate system synchronization.

[0087] Step A3. Trajectory feature extraction;

[0088] Extracting multi-dimensional features of the target vehicle from the standardized tail trajectory data ]; where P represents the position sequence of the target vehicle, V represents the velocity sequence of the target vehicle, A represents the acceleration sequence of the target vehicle, and K represents the curvature sequence of the target vehicle. represents the total historical time steps, Represents the multidimensional features of the t-th time step;

[0089] Position sequence of target vehicles ;in, represents the position at time step t, ;

[0090] Speed sequence of target vehicle ;in, represents the velocity at time step t;

[0091] Acceleration sequence of the target vehicle ;in, represents the acceleration at the t-th time step and can be calculated using the following formula:

[0092]

[0093] in, , are the velocities at two adjacent moments (time t+1 and time t); is the sampling time interval, set to 0.1s;

[0094] Direction angle sequence of target vehicle ;in, Represents the direction angle at time step t;

[0095] The curvature sequence of the target vehicle is ;in, represents the curvature at the t-th time step and can be calculated using the following formula:

[0096]

[0097] in, are the direction angles at two adjacent moments (time t+1 and time t), is the position coordinate of two adjacent moments (time t+1 and time t).

[0098] Step B. Data augmentation based on diffusion model (DDPM):

[0099] Step B1. Based on the constructed multi-dimensional features of the target vehicle , add Gaussian noise at each time step , so that the multidimensional features Gradually degenerate and obey the random Gaussian distribution, and obtain the characteristics , , the feature of the tth time step after adding noise The expression is shown as:

[0100]

[0101] Where, is Gaussian noise; is the noise scheduling coefficient of time step t, that is, when the time step t increases, , represents the cumulative retention signal rate from time step 1 to time step t, represents the signal retention rate at the s-th time step, Getting smaller and smaller, resulting in Becoming more and more random and losing original features ； is the multidimensional feature at the tth time step.

[0102] Step B2. Use the existing Transformer neural network to predict the Gaussian noise added at each time step , then the features Each time step feature Denoise and get features , , the feature of the tth time step after denoising The expression is:

[0103]

[0104]

[0105] Where, Represents each time step The "signal retention rate" of

[0106] Step B3. The above denoised features Input into MLP (Multi-layer Perceptron), which generates trajectory data with physical rationality and consistency of behavioral characteristics. The expression is shown as:

[0107]

[0108] In the formula, the generated trajectory , Represents the trajectory of the t-th time step. The MLP mainly consists of two linear change layers (fully connected layers) and a ReLU activation function.

[0109] Step B4. Sampling different Gaussian noises multiple times ,By repeating the process from step B2 to B3, multiple different trajectory samples can be generated, thereby,constructing a multimodal enhanced trajectory dataset.

[0110] like Figure 3 As shown, in this embodiment, step 1 specifically includes the following steps:

[0111] Step 1.1. Based on the historical trajectory time step of the surrounding vehicle i Speed sequence within Calculate the mean speed ;

[0112] Step 1.2. Based on the historical trajectory time step of the surrounding vehicle i Speed sequence within Calculate the mean acceleration , expressed as:

[0113]

[0114] Where, represents the time interval of the historical trajectory of sampling, 、 Represent the velocities of the fth time step and f+1th time step respectively;

[0115] Step 1.3. Based on the speed sequence of surrounding vehicles i and speed mean Calculate speed standard deviation , expressed as:

[0116]

[0117] in, represents the velocity at time step f;

[0118] Step 1.4. Based on the speed sequence of surrounding vehicles i , mean speed and speed standard deviation Calculate peak speed , expressed as:

[0119]

[0120] Step 1.5. Based on the speed sequence of surrounding vehicles i , mean speed and speed standard deviation Calculating velocity skewness , expressed as:

[0121]

[0122] Step 1.6. Based on the mean speed of vehicle i and speed standard deviation Calculate the coefficient of variation of speed , expressed as:

[0123]

[0124] Step 1.7. Combine the above calculated indicators to obtain the dynamic characteristic vector of the surrounding vehicle i .

[0125] In this embodiment, the dynamic characteristic vectors of all surrounding vehicles can be obtained by the above method. The dynamic characteristic vectors of all surrounding vehicles are expressed as , N represents the total number of vehicles; in actual operation, the vehicle dynamics feature vectors in all collected data are Used as a data set, K-means clustering is used to cluster three cluster centers , the three cluster centers correspond to: high dynamic fluctuation, medium dynamic fluctuation, and low dynamic fluctuation; subsequently, the vehicle dynamics feature vector of the surrounding vehicle i is With three cluster centers By comparing, the dynamic fluctuation type of the surrounding vehicle i can be directly output.

[0126] like Figure 4 As shown, in this embodiment, step 3 specifically includes the following steps:

[0127] Step 3.1. Calculate the traffic density based on the sensing range radius R of the autonomous vehicle’s sensor and the number of vehicles N in the sensing area at time t. , expressed as:

[0128]

[0129] Where, represents pi, and R represents the radius of the sensing range.

[0130] Step 3.2. According to Detect the number of vehicles within range at all times and the number of vehicles N in the sensing area at time t to calculate the traffic flow , expressed as:

[0131]

[0132] Where, Indicates absolute value.

[0133] Step 3.3. According to the speed of N vehicles in the internal perception area at time t }Calculate average speed , expressed as:

[0134]

[0135] Step 3.4. Splicing to get traffic flow feature vector .

[0136] In this embodiment, the traffic flow feature vectors in all collected data are Used as a data set, K-means clustering is used to cluster three cluster centers The traffic flow states corresponding to the three cluster centers are high density, medium density, and low density. ={ }and By comparing, the traffic flow status can be directly output.

[0137] like Figure 5 As shown, in this embodiment, the trajectory prediction of step 5 multimodal feature fusion specifically includes the following steps:

[0138] Step 5.1. Use Vector Net encoder to vectorize the historical trajectory information and HD map.

[0139] Step 5.2. Use LSTM encoder to encode historical trajectory information to obtain historical trajectory features , use the point Net-based encoder to encode the high-precision map information to obtain map features ;

[0140] Step 5.3. Use MLP (Multi-layer Perceptron) to analyze the dynamic feature vectors of the N surrounding vehicles and traffic flow characteristic vector ={ } Perform feature mapping to obtain the dynamic characteristics of surrounding vehicles and traffic flow characteristics , the expression is:

[0141]

[0142]

[0143] Where, Represents a simple MLP consisting of two linear transformation layers and a ReLU activation function.

[0144] Step 5.4. Characterize the historical trajectory and vehicle dynamics Splicing to get features , map features Traffic flow characteristics Splicing to get features , the expression is:

[0145] , )

[0146] , )

[0147] Where, and is the feature obtained after splicing, For the historical trajectory characteristics, is the kinematic characteristic of the vehicle, For map features, is the traffic flow characteristic, To concatenate two feature vectors sequentially.

[0148] Step 5.5. Features and features Input into the multi-head cross attention module for interactive modeling to obtain features , the expression is:

[0149]

[0150] Where, represents multi-head cross attention, the number of heads is 8, Q represents the query matrix, and the source is the feature , K is the key matrix, the source is the feature , V is the value matrix, the source is the feature ;

[0151] Step 5.6. Use feedforward neural network to train features Process and obtain features , the expression is:

[0152]

[0153] Where, It is a feed-forward neural network.

[0154] Step 5.7. Use MLP to analyze features Decode the future trajectories of N surrounding vehicles }, the speed of the future },Future acceleration , the expression is:

[0155]

[0156]

[0157]

[0158] Where, }、 }、 represents the future trajectory of N surrounding vehicles, ={ } means in the future The trajectory coordinates of the time step, ={ } means in the future The speed of the time step, ={ } means in the future The acceleration of a time step.

[0159] like Figure 6 As shown, in this embodiment, the specific steps of step 6 collision risk prediction are:

[0160] Step 6.1. Calculate the The distance from the N surrounding vehicles to the autonomous vehicle at the moment , and select the vehicle with the smallest distance from the autonomous vehicle as the target vehicle; among them, the surrounding vehicle i is in the first Distance to the self-driving car at that moment The expression is:

[0161]

[0162]

[0163] Where, Indicates the surrounding vehicle i in the The distance to the self-driving car at that moment, Indicates that self-driving cars The x, y coordinates of the moment, Indicates the surrounding vehicle i in the The x, y coordinates of the moment, , Represents the minimum distance between surrounding vehicles and the autonomous vehicle;

[0164] Step 6.2. Based on the future speed of the target vehicle ={ }, acceleration , extract the maximum speed of the target vehicle , maximum acceleration ;

[0165] Step 6.3. Based on the future acceleration of the target vehicle Calculating the standard deviation of acceleration , the expression is:

[0166]

[0167]

[0168] Where, represents the average acceleration, Represents the maximum future time step for prediction.

[0169] Step 6.4. Use the target vehicle's maximum speed , maximum acceleration and acceleration standard deviation Constructing a dynamic collision risk index , the expression is:

[0170]

[0171] Where, , are the weight coefficients of velocity, acceleration, and fluctuation, which are set to 1.0, 1.2, and 1.0 respectively; Indicates the maximum speed limit of the road. is a safe acceleration, set , is the expected safety fluctuation value, set .

[0172] Step 6.5. Setting the Kinetic Fluctuation Coefficient Corrected dynamic collision risk indicator , and obtain the dynamic collision risk index after considering dynamic fluctuations , the expression is:

[0173]

[0174] Where, Indicates the dynamic fluctuation coefficient, high dynamic fluctuation , mesodynamic fluctuations Low dynamic fluctuations ;

[0175] Step 6.6. Set traffic flow status label Dynamic collision risk index after considering dynamic fluctuations Make corrections to obtain the final dynamic collision risk index , the expression is:

[0176]

[0177] Where, Indicates traffic flow status label, when the density is high , medium density , at low density , The final dynamic risk indicator.

[0178] Step 6.7. The final dynamic risk indicator Input to the risk judgment module for interpretation and output of risk level , the expression is:

[0179] ；

[0180] The technical solution provided by the present invention can effectively enhance the recognition and prediction capabilities of autonomous driving systems in rare and high-risk events, significantly improving the problem of insufficient prediction accuracy of existing methods in long-tail scenarios. It has important safety value and broad engineering application prospects.

[0181] 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 autonomous driving vehicles for the long-tail phenomenon.

[0182] The present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for predicting collision risk of an autonomous driving vehicle for the long-tail phenomenon.

[0183] 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 via hardware or via computer programs. When all or part of the functions in the above embodiments are implemented via computer programs, the program may be stored in a computer-readable storage medium, which may include a read-only memory, random access memory, a magnetic disk, an optical disk, a hard disk, etc., and the program is executed by a computer to implement the above functions. For example, the program may be stored in a memory of a device, and when the program in the memory is executed by a processor, all or part of the above functions may be implemented.

[0184] 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.

[0185] The above description of the present invention using specific examples is intended only to facilitate understanding of the present invention and is not intended to limit the present invention. A person skilled in the art of the present invention may make several simple deductions, modifications, or substitutions based on the principles of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for predicting collision risk of autonomous vehicles targeting the long-tail phenomenon, characterized in that: This method uses a multimodal enhanced trajectory dataset to train a collision risk prediction model, and then uses the trained collision risk prediction model to perform risk prediction. The specific risk prediction steps are as follows: Step 1. Construct the dynamic feature vector of surrounding vehicles; Step 2. Output the dynamic fluctuation type based on the dynamic characteristic vectors of the surrounding vehicles; Step 3. Construct traffic flow feature vector; Step 4. Output the traffic flow state according to the traffic flow feature vector; Step 5. Trajectory prediction based on multimodal feature fusion: The historical trajectory information of the surrounding vehicles is encoded to obtain the historical trajectory feature f1, and the high-precision map information is encoded to obtain the map feature M; the dynamic feature vectors of the surrounding vehicles and the traffic flow feature vector are feature mapped to obtain the dynamic feature E v and traffic flow characteristics E env ; Combine the historical trajectory feature f1 and the dynamic feature E v Splicing to get feature f2, map feature M and traffic flow feature E env Concatenate feature M1; interactively model feature f2 and feature M1 to obtain feature f3; process feature f3 to obtain feature f4; decode feature f4 to decode the future trajectory, speed, and acceleration of surrounding vehicles; Step 6. Collision risk prediction: Based on historical data, the vehicle with the shortest distance from the autonomous vehicle is selected as the target vehicle, the maximum speed and maximum acceleration of the target vehicle are extracted, and the standard deviation of acceleration is calculated; a dynamic collision risk index is constructed based on the maximum speed, maximum acceleration and standard deviation of acceleration of the target vehicle; finally, the dynamic fluctuation coefficient is determined according to the dynamic fluctuation type, and the traffic flow state label is determined according to the traffic flow state, and the dynamic collision risk index is corrected using the dynamic fluctuation coefficient and traffic flow state label.

2. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 1, characterized in that: The steps for constructing the multimodal enhanced trajectory dataset are as follows: Step A. Data collection and feature construction: step A1. Filter trajectory data and construct tail trajectory data samples; Step A2: normalize the filtered tail trajectory data; Step A3. Trajectory feature extraction; Extract the target vehicle's multidimensional features F = [P, V, A, K] from the standardized tail trajectory data; where P represents the target vehicle's position sequence, V represents the target vehicle's velocity sequence, A represents the target vehicle's acceleration sequence, and K represents the target vehicle's curvature sequence; Step B. Data enhancement based on diffusion model: Step B1. Based on the constructed multi-dimensional feature F = [P, V, A, K] of the target vehicle, Gaussian noise is added at each time step to obtain the feature F 1 ; Step B2. Use the Transformer neural network to predict the Gaussian noise added at each time step, and then perform the feature F 1 Each time step feature Perform denoising to obtain feature F 2 ; Step B3. De-noised feature F 2 Input into the multi-layer perceptron to generate trajectory data; Step B4. By sampling different Gaussian noises multiple times and repeating the process from steps B2 to B3, multiple different trajectory samples are generated, thereby constructing a multimodal enhanced trajectory dataset.

3. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 1, characterized in that: Step 1 Calculate the mean speed m based on the historical data of surrounding vehicles i i , mean acceleration a i , speed standard deviation std i , peak speed kurt i , velocity skew i , speed variation coefficient cv i , construct the dynamic characteristic vector x of the surrounding vehicle i i ={m i , a i , std i , kurt i , skew i , cv i }; Step 2: The dynamic characteristic vector x of the surrounding vehicle i i ={m i , a i , std i , kurt i , skew i , cv i }Compare with the cluster centers μ0, μ1, μ2 representing high dynamic fluctuation, medium dynamic fluctuation, and low dynamic fluctuation, and output the dynamic fluctuation type of the surrounding vehicle i.

4. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 1, characterized in that: Step 3: Calculate the traffic density K based on the sensing range radius R of the sensor on the autonomous vehicle and the number of vehicles N within the sensing range. env , traffic flow Q env , the average speed of surrounding vehicles V env , construct traffic flow feature vector X env ={K env , Q env , V env }; Step 4: Transform the traffic flow feature vector X around the autonomous vehicle env Compare with the cluster centers μ3, μ4, and μ5 representing high density, medium density, and low density, and output the traffic flow status.

5. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 1, characterized in that: In step 5, the historical trajectory information is encoded using the LSTM encoder to obtain the historical trajectory feature f1, and the high-precision map information is encoded using the point Net-based encoder to obtain the map feature M; Use a multi-layer perceptron to perform feature mapping on the dynamic feature vectors of surrounding vehicles and traffic flow feature vectors to obtain the dynamic feature E of surrounding vehicles v and traffic flow characteristics E env .

6. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 1, characterized in that: In step 5, feature f2 and feature M1 are input into the multi-head cross attention module for interaction modeling to obtain feature f3; feature f3 is processed using a feedforward neural network to obtain feature f4; and feature f4 is decoded using a multi-layer perceptron.

7. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 1, characterized in that: The specific steps of step 6 collision risk prediction are: Step 6.

1. Calculate the t-th historical time step h The distances from the N surrounding vehicles to the autonomous vehicle at the moment, and the vehicle with the smallest distance to the autonomous vehicle is selected as the target vehicle; Step 6.

2. Extract the maximum speed V of the target vehicle based on the future speed and acceleration of the target vehicle max , maximum acceleration a max ; Step 6.

3. Calculate the acceleration standard deviation based on the future acceleration of the target vehicle; Step 6.

4. Use the maximum speed V of the target vehicle max , maximum acceleration a max and acceleration standard deviation std a Constructing a dynamic collision risk index R dyn , the expression is: Where ω1, ω2, ω3 are the weight coefficients of velocity, acceleration, and fluctuation, V limit Indicates the maximum speed limit of the road, a safe is the safe acceleration, σ ref is the expected safety fluctuation value; Step 6.

5. Set the dynamic fluctuation coefficient α s Modified dynamic collision risk index R dyn , the dynamic collision risk index R after considering dynamic fluctuations is obtained style , the expression is: R style =a s .R dyn ; Step 6.

6. Set traffic flow state label α f The dynamic collision risk index R after considering dynamic fluctuations style After correction, the final dynamic collision risk index R is obtained, which is expressed as: R=a f .R style ; Step 6.

7. Input the final dynamic risk indicator R into the risk judgment module for interpretation and output the risk level RiskLevel.

8. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 3, characterized in that: The peak speed in step 1 is kurt i , velocity skew i , speed variation coefficient cv i The expression is shown as: Among them, v f represents the velocity at time step f, t h is the historical trajectory time step.

9. The method for predicting collision risk of an autonomous vehicle targeting the long-tail phenomenon according to claim 7, characterized in that: In step 6.4, ω1, ω2, and ω3 are set to 1.0, 1.2, and 1.0 respectively; a safe Set to 0.9m / s 2 ,σ ref Set to 0.4m / s 2 ; When the kinetic fluctuation type is high kinetic fluctuation, α s =1.2, when the dynamic fluctuation type is medium dynamic fluctuation α s =1.0, when the dynamic fluctuation type is low dynamic fluctuation α s =0.8; When the traffic flow state is high density, α in step 6.6 f =1.3, when the traffic flow state is medium density f =1.0, when the traffic flow state is low density f =0.7.

Images