Intelligent vehicle insurance state path decision-making method for two-wheeled vehicle collision protection

Abstract

The invention relates to the technical field of intelligent driving, in particular to an intelligent vehicle insurance state path decision-making method for two-wheeled vehicle collision protection, which comprises the following steps: S1, collecting left and right visual real-time images, and performing two-wheeled vehicle trajectory prediction according to the left and right visual real-time images to generate a two-wheeled vehicle trajectory prediction result; s2, performing personnel injury prediction according to the two-wheeled vehicle track prediction result, and generating a personnel injury prediction result; and S3, performing intelligent vehicle insurance state path decision according to a personnel injury prediction result. According to the scheme, intelligent vehicle insurance state path decision can be carried out, so that the problem of how to reduce injury to personnel under the condition of an unavoidable insurance state of collision is solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent driving technology, and in particular to an intelligent vehicle dangerous path decision method for collision protection of two-wheeled vehicles. Background Art

[0002] With the development of autonomous driving technology, smart cars are able to perceive dangerous traffic scenarios in advance and avoid them, reducing the occurrence of accidents. However, even high-level autonomous driving still has the possibility of collision accidents when facing dangerous scenarios such as the sudden protrusion of high-freedom traffic participants such as two-wheeled vehicles. Therefore, how smart cars make path decisions and implement collision strategies that minimize personal injuries to reduce the incidence of two-wheeled vehicle accidents and the casualties caused by them has become a key direction in the field of traffic safety research.

[0003] At present, in the field of intelligent vehicle path decision-making, most of the research focuses on avoiding collisions, improving vehicle safety, traffic efficiency and passenger comfort, but cannot solve the problem of how to reduce the harm to people in the dangerous situation where collisions are inevitable;

[0004] Therefore, there is an urgent need for an intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection, which can perform intelligent vehicle dangerous path decision to solve the problem of how to reduce damage to personnel in dangerous situations where collision is inevitable. Summary of the invention

[0005] The present invention aims to provide an intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection, which can perform intelligent vehicle dangerous path decision to solve the problem of how to reduce damage to personnel in dangerous situations where collision is inevitable.

[0006] The present invention provides the following basic solution: an intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection, including the following contents:

[0007] S1, collecting left and right visual real-time images, and constructing a trajectory prediction module based on the left and right visual real-time images, performing two-wheeled vehicle trajectory prediction, and generating two-wheeled vehicle trajectory prediction results;

[0008] S2. Performing personnel injury prediction based on the two-wheeled vehicle trajectory prediction result to generate a personnel injury prediction result;

[0009] S3. Make intelligent vehicle dangerous path decision based on the personnel injury prediction results.

[0010] Further, S101, extracting traffic flow scene data including two-wheeled vehicles based on a public autonomous driving data set; wherein the characteristics of the traffic flow scene data include: the position, speed, and acceleration of the two-wheeled vehicle and its surrounding vehicles;

[0011] S102, preprocessing the traffic flow scene data to obtain preprocessed data;

[0012] S103, constructing an environment perception model based on deep learning, for extracting traffic flow scene data containing features according to video images, and realizing real-time detection and tracking of traffic participants;

[0013] S104, using an optical video sensor, integrating the vehicle positioning and high-precision map information, and combining an environmental perception model to obtain historical trajectory data of the vehicle and surrounding traffic participants;

[0014] S105, using an optical video sensor and an environment perception model to extract key features of a specific target through a binocular stereo matching algorithm, wherein the key features include: a relative position, speed, and heading of the specific target;

[0015] S106, using a K-nearest neighbor algorithm to pre-process and classify the historical trajectory data of traffic participants to obtain historical trajectory data of different traffic participants;

[0016] S107. Using long short-term memory network, self-attention network, graph attention network and convolutional neural network, an encoding-decoding trajectory prediction network framework is constructed as a trajectory prediction model. According to the classified historical trajectory data, the trajectory of two-wheeled vehicles is predicted based on the multi-dimensional interactive characteristics of environmental status, participant behavior and rider driving intention.

[0017] Furthermore, the long short-term memory network, self-attention network, graph attention network and convolutional neural network are used to construct an encoding-decoding trajectory prediction network framework as a trajectory prediction model, including:

[0018] Use the encoder of the long short-term memory network and the self-attention network to process the classified historical trajectory data and obtain the hidden state of the historical trajectory data;

[0019] According to the hidden state of historical trajectory data, a graph attention network is used to capture the social interaction information between two-wheeled vehicles and their surrounding vehicles;

[0020] Using convolutional neural networks, high-latitude scene features of static maps are extracted from static map information;

[0021] Vector cascade of high-dimensional scene features and social interaction information;

[0022] The self-attention model decoder of the self-attention network is used to decode the vector cascade result and obtain the future trajectory coordinates as the two-wheeled vehicle trajectory prediction result.

[0023] Furthermore, the process of training the encoding-decoding trajectory prediction network framework is as follows:

[0024] The classified historical trajectory data is divided into the preset nodes, range historical trajectory and future trajectory, the historical trajectory data before the node is taken as the historical trajectory, and the historical trajectory data after the node is taken as the future trajectory;

[0025] The historical trajectory is used as the input of the long short-term memory network and the first self-attention model encoder. The output of the first self-attention model encoder is input into the graph neural network and the second self-attention model encoder. At the same time, the output of the first self-attention model encoder and the future trajectory are used as the input of the second self-attention model encoder.

[0026] The static map information is used as the input of the convolutional neural network, and the output of the convolutional neural network is input into the first multi-layer perceptron;

[0027] Input the high-dimensional scene features into the first multi-layer perceptron, perform vector cascade on the output results and social interaction information, input the vector cascade results into the multi-layer perceptron, and perform Gaussian distribution processing;

[0028] The output of the graph neural network and the output of the first multi-layer perceptron are vector-concatenated, and the vector concatenation result is input into the second multi-layer perceptron and subjected to Gaussian distribution processing. At the same time, the vector concatenation result is convolved with the output of the second self-attention model encoder, and the convolution result is input into the third multi-layer perceptron and subjected to Gaussian distribution processing;

[0029] The results of the two Gaussian distributions are convolved and input into the fourth multi-layer perceptron. The output of the fourth multi-layer perceptron is input into the self-attention model decoder for decoding to obtain the predicted trajectory.

[0030] The predicted trajectory and the future trajectory are carefully compared and analyzed. If the difference is within the preset difference range, it means that the trajectory prediction model training is completed. If the difference does not meet the preset difference range, the trajectory prediction model training continues.

[0031] Furthermore, the self-attention model encoder:

[0032]

[0033] Among them, x t are input features, is the hidden state of the encoder at time step t;

[0034] The self-attention model decoder:

[0035]

[0036] Where c is the context vector from the encoder, is the hidden state of the decoder at time step t, yt is the prediction output, which is the future trajectory coordinates of the traffic participants (vehicles).

[0037] Further, the S2 includes:

[0038] S201. Analyze the damage characteristics of different collision forms based on the two-wheeled vehicle history database, and establish the inline characteristics between different collision forms and damage mechanisms;

[0039] S202, reconstructing the collision accident and restoring the human body dynamic response information at the moment of the collision, so as to accurately capture the kinematic mechanism of the rider's injury; in this embodiment, the collision accident reconstruction is performed using Madymo software;

[0040] S203, establishing a simulation matrix, and building a two-wheeler rider injury database based on a large number of collision simulation experiments;

[0041] S204, screening the two-wheeled vehicle rider injury database and performing data preprocessing operations; wherein the two-wheeled vehicle rider injury database includes collision speed, collision angle, collision position and corresponding injury values; wherein the data preprocessing operation is performed, and the data is the injury value included in the injury database;

[0042] S205, building a rider collision injury prediction model based on the decision tree and the neural network, and training and verifying the injury prediction model based on the two-wheeled vehicle rider injury database established in step S2-4;

[0043] S206, obtaining a potential collision speed, collision position, and collision angle between the two-wheeled vehicle and the autonomous driving vehicle based on the trajectory prediction model;

[0044] S207: using the injury prediction model trained in S205, taking the collision speed, collision position and collision angle obtained in S206 as input parameters of the injury prediction model, obtaining the potential injury value of the rider as a personal injury prediction result.

[0045] Further, the S202 includes:

[0046] S20201. Based on the Madymo modeling environment, build a vehicle model and use force-deformation curves to characterize the contact characteristics of the vehicle;

[0047] S20202. Build a two-wheeled vehicle model and fix different structural parts of the two-wheeled vehicle through different types of hinge connections;

[0048] S20203. Use GEBOD module to scale the human model in Madymo environment;

[0049] S20204. Determine initial design variables and their value ranges that affect accident reconstruction, wherein the initial design variables include: collision speed, collision angle, and collision position of the vehicle;

[0050] S20205. Reconstruction objectives, constraints, and initial variables are set respectively. The platform automatically reads the results and uses them to calculate the objective function and penalty function to determine the feasibility of the current solution, and then decides the direction of algorithm optimization.

[0051] S20206, when the position error is controlled within 20% and the simulation result is consistent with the relevant information of the accident record, the accident reconstruction can be terminated; if not, the parameters such as the number of initial variables, the value range, the maximum number of optimizations, etc. are changed, and the next automatic simulation and optimization cycle is entered until the result meets the termination condition;

[0052] S20207. When the number of iterations reaches a specified upper limit, analyze the cyclist's kinematics and injury response in the global optimal solution.

[0053] Further, the data preprocessing operation described in S204 is specifically as follows:

[0054] X stand =(X-mean(X))std(X)

[0055] Among them, X is the original data, X stand is the preprocessed data, mean(X) is the mean, and std(X) is the standard deviation.

[0056] Further, the specific calculation process of the decision tree model described in S205 is:

[0057]

[0058] in, represents the predicted damage value, K is the total number of trees, F represents the tree model, x i is the i-th input sample of the model;

[0059] The specific calculation process of the neural network model described in S205 is:

[0060] z (l+1) =W (l) a (l) +b (l)

[0061] Among them, z (l+1) is the output value of the model layer l, W (l) is the model weight value, b (l) is the model bias value, a (l) is the input value of the model layer l.

[0062] Further, the S3 includes:

[0063] S301, obtaining several candidate vehicle trajectories;

[0064] S302, taking the vehicle dynamics constraint as a hard constraint;

[0065] S303, building a collision detection module and using it as a hard constraint to avoid collisions between the vehicle and the two-wheeled vehicle to the greatest extent;

[0066] S304, traversing all candidate vehicle trajectories, first filtering out trajectories that do not satisfy vehicle dynamics constraints and that collide;

[0067] S305, sorting the screened vehicle trajectories according to the personnel injury prediction results obtained in S2 above, and taking the vehicle trajectory corresponding to the minimum personnel injury prediction result as the final planned trajectory of the autonomous driving vehicle;

[0068] S306: When all vehicle trajectories may collide with the two-wheeled vehicle, based on the personnel injury prediction result, the vehicle trajectory that causes the least damage to the two-wheeled vehicle rider is preferentially selected.

[0069] Beneficial effects of this solution: This solution introduces the injury indicators of vulnerable traffic participants into the path planning system of autonomous driving vehicles by building an injury prediction model for two-wheeled vehicle riders, thereby realizing the active protection of vulnerable road users by autonomous driving vehicles and improving the safety and advancement of autonomous driving systems;

[0070] Specifically, this solution builds an encoding-decoding trajectory prediction network framework based on existing data and combined with a variety of analysis models to achieve two-wheeled vehicle trajectory prediction that takes into account the multi-dimensional interactive characteristics of environmental conditions, participant behaviors and rider driving intentions; based on the predicted two-wheeled vehicle trajectory prediction results, combined with two-wheeled vehicle epidemiological analysis, machine learning and neural network methods are used to predict personal injuries, and personal injury predictions corresponding to various two-wheeled vehicle trajectory prediction results are obtained; finally, based on the personal injury prediction, intelligent vehicle dangerous path decision-making is carried out to select the path with the least damage, thereby solving the problem of how to reduce injuries to personnel in dangerous situations where collisions are inevitable.

[0071] In summary, this solution can make intelligent vehicle dangerous path decisions to solve the problem of how to reduce damage to personnel in dangerous situations where collisions are inevitable. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic diagram of a flow chart of an embodiment of an intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to the present invention;

[0073] Figure 2 A schematic diagram of an encoding-decoding trajectory prediction network framework of an embodiment of an intelligent vehicle dangerous path decision method for collision protection of two-wheeled vehicles according to the present invention;

[0074] Figure 3 The present invention is a schematic diagram of the encoding-decoding trajectory prediction network framework training process of an embodiment of an intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection. DETAILED DESCRIPTION

[0075] The following is further described in detail through specific implementation methods:

[0076] The embodiment is basically as shown in the attached Figure 1 As shown: A smart vehicle dangerous path decision method for two-wheeled vehicle collision protection includes the following contents:

[0077] S1, collecting left and right visual real-time images, and constructing a trajectory prediction model based on the left and right visual real-time images, performing two-wheeled vehicle trajectory prediction, and generating two-wheeled vehicle trajectory prediction results;

[0078] The two-wheeled vehicle trajectory prediction results include: collision position, collision speed, and collision direction;

[0079] Specifically, the S1 includes:

[0080] S101. Extracting traffic flow scene data including two-wheeled vehicles from a public autonomous driving data set; wherein the features of the traffic flow scene data include: information such as the position, speed, acceleration, etc. of the two-wheeled vehicle and its surrounding vehicles; the autonomous driving data set is a public autonomous driving video image set;

[0081] S102, preprocessing the traffic flow scene data to obtain preprocessed data; wherein the preprocessing includes: cleaning, denoising and standardization to ensure data quality;

[0082] S103, constructing an environment perception model based on deep learning, for extracting traffic flow scene data containing features according to video images, and realizing real-time detection and tracking of traffic participants;

[0083] Specifically, in this embodiment, the environment perception model adopts a BP neural network, and a training set is constructed through preprocessed traffic flow scene data and corresponding automatic driving data. The training set is used to train the environment perception model, so that the environment perception model can analyze and process the input video image and extract the traffic flow scene data containing features;

[0084] S104, using an optical video sensor, integrating the vehicle positioning and high-precision map information, and combining an environmental perception model to obtain historical trajectory data of the vehicle and surrounding traffic participants;

[0085] Specifically, an optical video sensor is used to collect video images; wherein the video images are real-time left and right visual images;

[0086] The environment perception model extracts traffic flow scene data containing features based on video images;

[0087] According to the traffic flow scene data, the positioning of the vehicle itself and the high-precision map information are integrated to analyze the trajectory data of the vehicle itself and the surrounding traffic participants within a preset time period as historical trajectory data; the preset time period is the time that has passed, so the trajectory data is historical trajectory data, and the length of the preset time period is set according to demand; specifically, according to the position, speed, and acceleration of the two-wheeled vehicle and the surrounding vehicles, combined with the positioning of the vehicle itself and the high-precision map information, the form trajectory of all traffic participants within the preset time is updated in the map, thereby obtaining historical trajectory data;

[0088] S105, using an optical video sensor and an environment perception model to extract key features of a specific target through a binocular stereo matching algorithm, wherein the key features include: a relative position, speed, and heading of the specific target;

[0089] Specifically, an optical video sensor is used to collect video images; wherein the video images are real-time left and right visual images;

[0090] The environment perception model extracts traffic flow scene data containing key features of specific targets based on video images;

[0091] S106, using a K-nearest neighbor (KNN) algorithm to pre-process and classify the historical trajectory data of the traffic participants, and obtaining the historical trajectory data of different traffic participants to distinguish the traffic participants corresponding to the historical trajectory data;

[0092] S107. Using long short-term memory network, self-attention network, graph attention network and convolutional neural network, an encoding-decoding trajectory prediction network framework is constructed as a trajectory prediction model, and the trajectory of two-wheeled vehicles is predicted based on the multi-dimensional interactive characteristics of environmental state, participant behavior and rider driving intention according to the classified historical trajectory data;

[0093] Specifically, the long short-term memory network, self-attention network, graph attention network and convolutional neural network are used to construct the encoding-decoding trajectory prediction network framework as the trajectory prediction model. Figure 2 As shown: Use the long short-term memory network (LSTM) and the self-attention network (Transformer) to process the classified historical trajectory data and obtain the hidden state of the historical trajectory data;

[0094] The specific structure of the long short-term memory network is:

[0095] Forget Gate:

[0096] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0097] Among them, f t Represents the activation vector of the forget gate, W f and b f are weights and biases respectively, σ is the sigmoid activation function, h t-1 is the hidden state of the previous time step, x t is the current input, and is the classified historical trajectory data;

[0098] Input Gate:

[0099] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0100]

[0101] Among them, i t is the activation vector of the input gate, is the candidate cell state;

[0102] Update cell status:

[0103]

[0104] Output Gate:

[0105] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0106] h t =o t *tanh(C t );

[0107] The specific structure of the self-attention network is:

[0108]

[0109] Where Q, K, and V are query, key, and value matrices respectively, and d k is the dimension of the key;

[0110] Output the hidden state of historical trajectory data with higher dimensions;

[0111] Specifically, the self-attention model encoder:

[0112]

[0113] Among them, x t is the input feature, which is the historical trajectory data; is the hidden state of the encoder at time step t;

[0114] According to the hidden state of historical trajectory data, a graph attention network (GAT) is used to capture the social interaction information between two-wheeled vehicles and their surrounding vehicles; the social interaction information refers to the mutual influence relationship between different traffic participants (cars, two-wheeled vehicles, etc.), that is, the hidden state of historical trajectory data after attention mechanism processing;

[0115] The input of the graph attention network is the hidden state of the historical trajectory data, and the output is the hidden state of the historical trajectory data after being processed by the attention mechanism;

[0116] The specific structure of the graph attention network is:

[0117] Calculation of attention coefficient:

[0118]

[0119] Among them, e ij is the initial value of the attention coefficient between node i and node j; is a learnable weight vector used to calculate the attention coefficient; W is a weight matrix used to linearly transform the features of the node; and are the feature vectors of node i and node j respectively; LeakyReLU is the activation function used to increase nonlinearity;

[0120] Normalized attention coefficient:

[0121]

[0122] Among them, α ij is the normalized attention coefficient, indicating the importance of node j to node i; is the set of neighbor nodes of node i;

[0123] Node feature updates:

[0124]

[0125] in, is the updated feature vector of node i; σ is a nonlinear activation function, such as sigmoid or ReLU function is usually used; ReLU function is used in this embodiment;

[0126] Using convolutional neural network (CNN), high-dimensional scene features of static maps are extracted from static map information;

[0127] Specifically, the input of the convolutional neural network is static map information, and the output is high-dimensional scene features after convolution processing;

[0128] The static map information can be a high-precision map or a simple map containing only lane line information;

[0129] The specific structure of the convolutional neural network is:

[0130]

[0131] Among them, W m is the convolution kernel; b is the bias; f is the nonlinear activation function, * indicates the convolution operation; x m is the input static map information, C(x) is the output high-latitude scene feature;

[0132] Vector cascade of high-dimensional scene features and social interaction information;

[0133] The self-attention model decoder is used to decode the vector cascade result and obtain the future trajectory coordinates, i.e. the predicted trajectory, as the trajectory prediction result of the two-wheeled vehicle;

[0134] Self-Attention Model Decoder:

[0135]

[0136] Where c is the context vector from the encoder, is the hidden state of the decoder at time step t, y t is the prediction output, which is the future trajectory coordinates of the traffic participants (vehicles).

[0137] The process of training the encoding-decoding trajectory prediction network framework, such as Figure 3 As shown:

[0138] The classified historical trajectory data is divided into the preset nodes, range historical trajectory and future trajectory, the historical trajectory data before the node is taken as the historical trajectory, and the historical trajectory data after the node is taken as the future trajectory;

[0139] The historical trajectory is used as the input of the long short-term memory network and the first self-attention model encoder. The output of the first self-attention model encoder (the hidden state of the historical trajectory data) is input into the graph neural network and the second self-attention model encoder. At the same time, the output of the first self-attention model encoder and the future trajectory are used as the input of the second self-attention model encoder.

[0140] The static map information is used as the input of the convolutional neural network, and the output of the convolutional neural network (high-dimensional scene features) is input into the first multi-layer perceptron;

[0141] Input the high-dimensional scene features into the first multi-layer perceptron, perform vector cascade on the output results and social interaction information, input the vector cascade results into the multi-layer perceptron, and perform Gaussian distribution processing;

[0142] The output of the graph neural network (social interaction information) and the output of the first multi-layer perceptron are vector-concatenated, and the vector concatenation result is input into the second multi-layer perceptron and subjected to Gaussian distribution processing. At the same time, the vector concatenation result is convolved with the output of the second self-attention model encoder (the hidden state of the future trajectory), and the convolution result is input into the third multi-layer perceptron and subjected to Gaussian distribution processing;

[0143] The results of the two Gaussian distributions are convolved and input into the fourth multi-layer perceptron. The output of the fourth multi-layer perceptron is input into the self-attention model decoder for decoding to obtain the predicted trajectory.

[0144] The predicted trajectory and the future trajectory are carefully compared and analyzed. If the difference is within the preset difference range, it means that the trajectory prediction model training is completed. If the difference does not meet the preset difference range, the trajectory prediction model training continues.

[0145] S2. Performing personnel injury prediction based on the two-wheeled vehicle trajectory prediction result to generate a personnel injury prediction result;

[0146] The S2 prediction of personal injury includes epidemiological analysis of two-wheeled vehicles, establishment of an intelligent accident reconstruction platform, and prediction of personal injury using machine learning and neural network methods. The overall principle includes:

[0147] Specifically, S2 includes:

[0148] S201. Analyze the damage characteristics of different collision forms according to the two-wheeled vehicle historical database, and establish the inline characteristics between different collision forms and damage mechanisms; the two-wheeled vehicle historical database may adopt a public database, or a database may be built according to demand; the inline characteristics refer to the correlation characteristics between different collision forms between vehicles and two-wheeled vehicles and different injury forms of two-wheeled vehicle riders;

[0149] S202, reconstructing the collision accident and restoring the human body dynamic response information at the moment of collision, so as to accurately capture the kinematic mechanism of the rider's injury; in this embodiment, the collision accident is reconstructed by using Madymo software, and the collision speed, collision position and collision angle at the moment of collision between the two-wheeled vehicle and the car are input into Madymo software to obtain the human body dynamic response information;

[0150] Specifically, S202 includes:

[0151] S20201. Based on the Madymo modeling environment, build a vehicle model and use force-deformation curves to characterize the contact characteristics of the vehicle; the vehicle model is a model of the vehicle that will collide with the two-wheeled vehicle;

[0152] S20202. Build a two-wheeled vehicle model, and fix different structural parts of the two-wheeled vehicle through different types of hinge connections; wherein the two-wheeled vehicle model is a model of a two-wheeled vehicle that is to be collided;

[0153] S20203. Use the GEBOD module in Madymo to scale the human body model in the Madymo environment to meet the human body characteristics of cyclists in real collision accidents;

[0154] S20204. Determine initial design variables and their value ranges that affect accident reconstruction, wherein the initial design variables include: collision speed, collision angle, and collision position of the vehicle; in this embodiment, the initial design variables and their value ranges that affect accident reconstruction are determined in the intelligent accident reconstruction platform;

[0155] S20205. Reconstruction objectives, constraints, and initial variables are set respectively. The intelligent accident reconstruction platform automatically reads the results and calculates the objective function and penalty function to determine the feasibility of the current solution, and then determines the direction of algorithm optimization. In this embodiment, the reconstruction is carried out in the intelligent accident reconstruction platform. The reconstruction objective is to restore the real collision scene. The constraints are the value range of the variables to be solved. The initial variables are the collision speed, collision angle, and collision position. The objective function is a function that needs to be minimized or maximized during the optimization process. In accident reconstruction, the objective function is used to measure the difference between the reconstruction result and the actual accident situation. For example, the objective function is the error measure between the actual accident data and the reconstruction result, such as cyclist injury. The penalty function is used to penalize solutions that do not meet the constraints to guide the optimization process towards solutions that meet the constraints. In accident reconstruction, the best solution is usually determined by minimizing the objective function, that is, the optimization algorithm will look for a solution that minimizes the objective function value. In accident reconstruction, commonly used optimization algorithms include: genetic algorithm, particle swarm optimization algorithm, simulated annealing algorithm, and gradient descent algorithm.

[0156] S20206, when the position error is controlled within 20% and the simulation result is consistent with the relevant information of the accident record, the accident reconstruction can be terminated; if not, the parameters such as the number of initial variables, the value range, the maximum number of optimizations, etc. are changed, and the next automatic simulation and optimization cycle is entered until the result meets the termination condition;

[0157] S20207. When the number of iterations reaches a specified upper limit, analyzing the kinematics and injury response of the cyclist in the global optimal solution; specifically, analyzing the kinematics and injury response of the cyclist in the global optimal solution by comparing the injury of the cyclist in the simulation with the injury of the cyclist in the actual collision accident;

[0158] S203, establishing a simulation matrix, and building a two-wheeled vehicle rider injury database based on a large number of collision simulation experiments; wherein the simulation matrix includes different collision speeds, collision positions and collision angles, and is used to obtain a large amount of collision scene data and build a two-wheeled vehicle rider injury database;

[0159] S204, screening the two-wheeled vehicle rider injury database and performing data preprocessing operations; wherein the two-wheeled vehicle rider injury database includes collision speed, collision angle, collision position and corresponding injury values; wherein the data preprocessing operation is performed, and the data is the injury value included in the injury database;

[0160] The data preprocessing operation described in S204 is specifically as follows:

[0161] X stand =(X-mean(X))std(X)

[0162] Among them, X is the original data, X stand is the preprocessed data, mean(X) is the mean, and std(X) is the standard deviation;

[0163] S205, building a rider collision injury prediction model based on a decision tree and a neural network, and training and verifying the injury prediction model based on a two-wheeled vehicle rider injury database;

[0164] The specific calculation process of the decision tree model described in S205 is:

[0165]

[0166] in, represents the predicted damage value, i.e., the damage value, K is the total number of trees, F represents the tree model, x i is the i-th input sample of the model, i.e., the collision speed, collision angle, and collision position in the two-wheeler rider injury database;

[0167] The specific calculation process of the neural network model described in S205 is:

[0168] z (l+1) =W (l) a (l) +b (l)

[0169] Among them, z (l+1) is the output value of the model layer l, W (l) is the model weight value, b (l) is the model bias value, a (l) is the input value of the model layer l; the input value is the collision speed, collision angle, and collision position in the two-wheeler rider injury database; the output value is the damage value;

[0170] S206, obtaining a potential collision speed, collision position, and collision angle between the two-wheeled vehicle and the autonomous driving vehicle based on the trajectory prediction model;

[0171] S207: using the injury prediction model trained in S205, taking the collision speed, collision position and collision angle obtained in S206 as input parameters of the injury prediction model, obtaining the potential injury risk of the rider, that is, the injury value, as the personal injury prediction result.

[0172] S3. Make intelligent vehicle dangerous path decision based on personnel injury prediction results;

[0173] Specifically, S3 includes:

[0174] S301, obtaining several candidate vehicle trajectories. In this embodiment, a series of candidate vehicle trajectories are obtained by sampling, and sampling is performed within the vehicle's drivable area.

[0175] S302, taking vehicle dynamics constraints as hard constraints; the vehicle dynamics constraints include: speed constraints, acceleration constraints and steering constraints;

[0176] S303, building a collision detection module, which is used together with the vehicle dynamics constraint as a hard constraint to avoid the collision between the vehicle and the two-wheeled vehicle to the greatest extent;

[0177] S304, traversing all candidate vehicle trajectories, first filtering out trajectories that do not satisfy vehicle dynamics constraints and that collide;

[0178] S305, sorting the screened vehicle trajectories according to the personnel injury prediction results obtained in S2 above, and taking the vehicle trajectory corresponding to the minimum personnel injury prediction result as the final planned trajectory of the autonomous driving vehicle;

[0179] S306: When all vehicle trajectories may collide with the two-wheeled vehicle, based on the personnel injury prediction result, the vehicle trajectory that causes the least damage to the two-wheeled vehicle rider is preferentially selected.

[0180] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for intelligent vehicle dangerous path decision-making for two-wheeled vehicle collision protection, characterized in that: It includes the following: S1, collecting left and right visual real-time images, and constructing a trajectory prediction model based on the left and right visual real-time images, performing two-wheeled vehicle trajectory prediction, and generating two-wheeled vehicle trajectory prediction results; S2. Performing personnel injury prediction based on the two-wheeled vehicle trajectory prediction result to generate a personnel injury prediction result; S3. Make intelligent vehicle dangerous path decision based on the personnel injury prediction results.

2. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 1 is characterized in that: S101: extracting traffic flow scene data including two-wheeled vehicles based on a public autonomous driving data set; The characteristics of traffic flow scene data include: the position, speed, and acceleration of two-wheeled vehicles and their surrounding vehicles; S102, preprocessing the traffic flow scene data to obtain preprocessed data; S103, constructing an environment perception model based on deep learning, for extracting traffic flow scene data containing features according to video images, and realizing real-time detection and tracking of traffic participants; S104, using an optical video sensor, integrating the vehicle positioning and high-precision map information, and combining an environmental perception model to obtain historical trajectory data of the vehicle and surrounding traffic participants; S105, using an optical video sensor and an environment perception model to extract key features of a specific target through a binocular stereo matching algorithm, wherein the key features include: a relative position, speed, and heading of the specific target; S106, preprocessing and classifying the historical trajectory data of traffic participants to obtain historical trajectory data of different traffic participants; S107. Using long short-term memory network, self-attention network, graph attention network and convolutional neural network, an encoding-decoding trajectory prediction network framework is constructed as a trajectory prediction model. According to the classified historical trajectory data, the trajectory of two-wheeled vehicles is predicted based on the multi-dimensional interactive characteristics of environmental status, participant behavior and rider driving intention.

3. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 2 is characterized in that: The long short-term memory network, self-attention network, graph attention network and convolutional neural network are used to construct an encoding-decoding trajectory prediction network framework as a trajectory prediction model, including: Use the encoder of the long short-term memory network and the self-attention network to process the classified historical trajectory data and obtain the hidden state of the historical trajectory data; According to the hidden state of historical trajectory data, a graph attention network is used to capture the social interaction information between two-wheeled vehicles and their surrounding vehicles; Using convolutional neural networks, high-latitude scene features of static maps are extracted from static map information; Vector cascade of high-dimensional scene features and social interaction information; The self-attention model decoder of the self-attention network is used to decode the vector cascade result and obtain the future trajectory coordinates as the two-wheeled vehicle trajectory prediction result.

4. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 3 is characterized in that: The process of training the encoding-decoding trajectory prediction network framework is as follows: The classified historical trajectory data is divided into the preset nodes, range historical trajectory and future trajectory, the historical trajectory data before the node is taken as the historical trajectory, and the historical trajectory data after the node is taken as the future trajectory; The historical trajectory is used as the input of the long short-term memory network and the first self-attention model encoder. The output of the first self-attention model encoder is input into the graph neural network and the second self-attention model encoder. At the same time, the output of the first self-attention model encoder and the future trajectory are used as the input of the second self-attention model encoder. The static map information is used as the input of the convolutional neural network, and the output of the convolutional neural network is input into the first multi-layer perceptron; Input the high-dimensional scene features into the first multi-layer perceptron, perform vector cascade on the output results and social interaction information, input the vector cascade results into the multi-layer perceptron, and perform Gaussian distribution processing; The output of the graph neural network and the output of the first multi-layer perceptron are vector-concatenated, and the vector concatenation result is input into the second multi-layer perceptron and subjected to Gaussian distribution processing. At the same time, the vector concatenation result is convolved with the output of the second self-attention model encoder, and the convolution result is input into the third multi-layer perceptron and subjected to Gaussian distribution processing; The results of the two Gaussian distributions are convolved and input into the fourth multi-layer perceptron. The output of the fourth multi-layer perceptron is input into the self-attention model decoder for decoding to obtain the predicted trajectory. The predicted trajectory and the future trajectory are carefully compared and analyzed. If the difference is within the preset difference range, it means that the trajectory prediction model training is completed. If the difference does not meet the preset difference range, the trajectory prediction model training continues.

5. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 4 is characterized in that: The self-attention model encoder: Among them, x t are input features, is the hidden state of the encoder at time step t; The self-attention model decoder: Where c is the context vector from the encoder, is the hidden state of the decoder at time step t, y t is the prediction output, which is the future trajectory coordinates of the traffic participants.

6. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 5 is characterized in that: The S2 comprises: S201. Analyze the damage characteristics of different collision forms based on the two-wheeled vehicle history database, and establish the inline characteristics between different collision forms and damage mechanisms; S202, reconstructing the collision accident and restoring the human body dynamic response information at the moment of the collision, so as to accurately capture the kinematic mechanism of the cyclist's injury; S203, establishing a simulation matrix, and building a two-wheeler rider injury database based on a large number of collision simulation experiments; S204, screening the two-wheeled vehicle rider injury database and performing data preprocessing operations; wherein the two-wheeled vehicle rider injury database includes collision speed, collision angle, collision position and corresponding injury values; wherein the data preprocessing operation is performed, and the data is the injury value included in the injury database; S205, building a rider collision injury prediction model based on the decision tree and the neural network, and training and verifying the injury prediction model based on the two-wheeled vehicle rider injury database established in step S2-4; S206, obtaining a potential collision speed, collision position, and collision angle between the two-wheeled vehicle and the autonomous driving vehicle based on the trajectory prediction model; S207: using the injury prediction model trained in S205, taking the collision speed, collision position and collision angle obtained in S206 as input parameters of the injury prediction model, obtaining the potential injury value of the rider as a personal injury prediction result.

7. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 6 is characterized in that: The S202 includes: S20201. Based on the Madymo modeling environment, build a vehicle model and use force-deformation curves to characterize the contact characteristics of the vehicle; S20202. Build a two-wheeled vehicle model and fix different structural parts of the two-wheeled vehicle through different types of hinge connections; S20203. Use GEBOD module to scale the human model in Madymo environment; S20204. Determine initial design variables and their value ranges that affect accident reconstruction, wherein the initial design variables include: collision speed, collision angle, and collision position of the vehicle; S20205. Reconstruction objectives, constraints, and initial variables are set respectively. The platform automatically reads the results and uses them to calculate the objective function and penalty function to determine the feasibility of the current solution, and then decides the direction of algorithm optimization. S20206, when the position error is controlled within 20% and the simulation result is consistent with the relevant information of the accident record, the accident reconstruction can be terminated; if not, the parameters such as the number of initial variables, the value range, the maximum number of optimizations, etc. are changed, and the next automatic simulation and optimization cycle is entered until the result meets the termination condition; S20207. When the number of iterations reaches a specified upper limit, analyze the cyclist's kinematics and injury response in the global optimal solution.

8. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 7 is characterized in that: The data preprocessing operation described in S204 is specifically as follows: X stand =(X-mean(X)) / std(X) Among them, X is the original data, X stand is the preprocessed data, mean(X) is the mean, and std(X) is the standard deviation.

9. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 8 is characterized in that: The specific calculation process of the decision tree model described in S205 is: in, represents the predicted damage value, K is the total number of trees, F represents the tree model, x i is the i-th input sample of the model; The specific calculation process of the neural network model described in S205 is: z (l+1) =W (l) a (l) +b (l) Among them, z (l+1) is the output value of the model layer l, W (l) is the model weight value, b (l) is the model bias value, a (l) is the input value of the model layer l.

10. The intelligent vehicle dangerous path decision method for two-wheeled vehicle collision protection according to claim 9, characterized in that: The S3 includes: S301, obtaining several candidate vehicle trajectories; S302, taking the vehicle dynamics constraint as a hard constraint; S303, building a collision detection module and using it as a hard constraint to avoid collisions between the vehicle and the two-wheeled vehicle to the greatest extent; S304, traversing all candidate vehicle trajectories, first filtering out trajectories that do not satisfy vehicle dynamics constraints and that collide; S305, sorting the screened vehicle trajectories according to the personnel injury prediction results obtained in S2 above, and taking the vehicle trajectory corresponding to the minimum personnel injury prediction result as the final planned trajectory of the autonomous driving vehicle; S306: When all vehicle trajectories may collide with the two-wheeled vehicle, based on the personnel injury prediction result, the vehicle trajectory that causes the least damage to the two-wheeled vehicle rider is preferentially selected.

Images