A space-time coupling intelligent automobile risk assessment system considering motion uncertainty

Abstract

The application relates to a space-time coupling intelligent automobile risk assessment system considering motion uncertainty, which comprises a traffic participant uncertainty kinematics information prediction module, a traffic participant anisotropic risk field module, a driving risk space distribution pseudo picture construction module and a space-time coupling driving risk assessment module; the anisotropic driving risk field is constructed to describe the risk generated by other traffic participants as risk sources in the traffic environment, the uncertainty prediction of the kinematics information of the other traffic participants in the future is introduced, the uncertainty interaction coupling relationship between the ego vehicle and the other traffic participants and the dynamic evolution mechanism of the traffic situation are considered, the pseudo picture covering the risk space distribution information of the ego vehicle and eight surrounding traffic participants is constructed based on six-dimensional risk characteristics, the historical and future driving risk space distribution pseudo picture sequences are taken as inputs, and the space-time coupling driving risk level is output, so that the technical problem that it is difficult to realize the space-time coupling accurate driving risk assessment is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of intelligent car driving risk assessment systems, in particular to a kind of spatiotemporal coupling intelligent car driving risk assessment systems considering motion uncertainty. BACKGROUND

[0002] In recent years, intelligent car has become the strategic development direction of global automobile industry, and a large number of related technologies and products emerge in an endless stream.However, traffic accidents caused by imperfect design of automatic driving system or driving assistance system and other factors are common in the world.Therefore, it is urgent to develop a reliable intelligent car risk assessment system to fully guarantee the driving safety of intelligent car.

[0003] At present, academia and industry have carried out a lot of research and practice around intelligent car driving risk assessment, but there are still some unsolved problems in the related technology of existing risk assessment, for example:

[0004] (1) The existing technology is not accurate in describing the objective risk in the traffic environment, and the collision time, the following time and the traditional driving risk field are difficult to effectively describe the mapping relationship between the definition attribute, motion state of risk source and the risk generated by it.

[0005] (2) The existing technology does not fully consider the motion uncertainty of other traffic participants in the traffic environment, and it is difficult to effectively express and describe the interactive coupling relationship of strong random uncertainty between the ego car and other traffic participants and the dynamic evolution mechanism of traffic situation.

[0006] (3) Driving risk not only has spatial distribution characteristics, but also has continuity and dynamic evolution in time, while the existing technology can only express the spatial distribution or time distribution of driving risk separately, and it is difficult to realize the risk assessment of time and space coupling, thereby restricting the accuracy of intelligent car driving risk assessment.

[0007] Therefore, it is necessary to develop a set of spatiotemporal coupling driving risk assessment system that can accurately describe the objective risk in the traffic environment and fully consider the motion uncertainty of traffic participants, which has become a core technical problem to be solved in the field of intelligent car technology. SUMMARY

[0008] In order to solve the above technical problems, the present application provides a kind of spatiotemporal coupling intelligent car risk assessment system considering motion uncertainty, including traffic participant uncertainty kinematics information prediction module, traffic participant anisotropic risk field module, driving risk space distribution pseudo picture construction module and spatiotemporal coupling driving risk assessment module;

[0009] (1) The traffic participant uncertainty kinematic information prediction module takes the historical kinematic information of each traffic participant and the ego vehicle as input, and outputs the future kinematic information of each traffic participant considering its motion uncertainty in the form of a probability distribution within a prediction time window;

[0010] (2) The traffic participant anisotropic risk field module takes the historical kinematic information of each traffic participant and the future kinematic information of each traffic participant considering motion uncertainty output by the traffic participant uncertainty kinematic information prediction module as input, calculates the risk generated by each traffic participant at each historical and future time through the construction of an anisotropic risk field, and outputs the risk field strength generated by each traffic participant to the ego vehicle at each historical and future time;

[0011] (3) The driving risk space distribution pseudo-picture construction module takes the risk field strength generated by each traffic participant to the ego vehicle at each historical and future time, the kinematic information of the ego vehicle at each historical and future time, and the kinematic information of each traffic participant at each historical and future time as input, uses the risk field strength, risk field strength derivative variables, and kinematic information to form a 6-dimensional risk feature;

[0012] At each time in the past and future, the driving risk space distribution pseudo-picture construction module divides a certain spatial region around the ego vehicle into a 3x3 grid region according to the position of the ego vehicle at that time, and uses the 6-dimensional risk feature at that time to describe the risk state of each traffic participant in each grid relative to the ego vehicle and the risk state of the ego vehicle, thereby constructing a 3x3x6 pseudo-picture to describe the driving risk space distribution at that time, and finally outputting a pseudo-picture sequence composed of driving risk space distribution pseudo-pictures at each historical and future time;

[0013] (4) The space-time coupled driving risk assessment module receives the historical and future driving risk space distribution pseudo-picture sequence, extracts the driving risk space distribution information contained in each historical and future pseudo-picture through a convolutional neural network (CNN), processes the temporal dependence relationship and dynamic evolution trend between risks at each time through a long-short term memory network (LSTM), and finally outputs a time and space coupled intelligent vehicle driving risk level.

[0014] Furthermore, the traffic participants described in this invention include eight traffic participants located at the front (Fr), rear (Rear, Re), left (Left, L), left front (Left Front, LF), left rear (Left Rear, LR), right (Right, Ri), right front (Right Front, RF), and right rear (Right Rear, RR) of the EV; the longitudinal range of the selected traffic participants is 50-150m in front of and behind the EV, and the lateral range of the selected traffic participants is the current lane of the EV and the adjacent lanes on the left and right; according to whether they are moving, the traffic participants are divided into dynamic traffic participants and static traffic participants.

[0015] Furthermore, the traffic participant uncertainty kinematic information prediction module consists of four parts: an LSTM-based encoder, an attention mechanism, an LSTM-based decoder, and a Mixture Density Network (MDN).

[0016] 1.1 The input to the LSTM-based encoder is a time series of historical kinematic information of the vehicle and traffic participants. At the current time t0, the historical kinematic information time series X is:

[0017]

[0018] Where H is the length of the historical time window; x e and y e These refer to the longitudinal and lateral positions of the EV, respectively; v ex and v ey These represent the longitudinal and lateral velocities of the EV, respectively; x j and y j These represent the longitudinal and lateral positions of traffic participant j, respectively; v jx and v jy These are the longitudinal and lateral velocities of traffic participant j, respectively.

[0019] At each encoding time step t h The LSTM unit of the LSTM-based encoder receives the previous encoding time step t. h -1 hidden state and the current encoding time step t h Input Output the current encoding time step t h Hidden state Until the entire input sequence X is encoded into a hidden state sequence

[0020] 1.2, at each decoding time step t fThe attention mechanism calculates the decoder's performance at the previous decoding time step t. f -1 hidden state and The correlation between them is used to calculate the current decoding time step t. f Below, encoder hidden state sequence The weight coefficients of each hidden state are then used to obtain the current decoding time step t through weighted summation. f semantic vector;

[0021] 1.3, at each decoding time step t f The LSTM unit of the LSTM-based decoder receives the previous decoding time step t. f -1 hidden state and the current decoding time step t f The semantic vector is used to output the current decoding time step t. f Hidden state Until decoding is completed for all decoding time steps, the positions of relevant traffic participants around the EV within the prediction time window are predicted; the output sequence Y of the LSTM-based decoder is:

[0022] Where F is the prediction time window length of the traffic participant uncertainty kinematic information prediction module;

[0023] 1.4 The Hybrid Density Network (MDN) describes the uncertainty of traffic participant predicted location, i.e., the motion uncertainty of traffic participants, through a mixture of six Gaussian distributions. Taking traffic participant j as an example, the modeling method of MDN for the uncertainty of traffic participant predicted location is explained: assuming that the discretized position of traffic participant j follows a Markov process, then at decoding time step t... f Given the previous decoding time step t f -1 location information In the case of traffic participants at decoding time step t f Located in The probability is expressed as:

[0024]

[0025] in:

[0026]

[0027]

[0028] in, and are the true longitudinal and lateral positions of traffic participant j, respectively; μ and σ are the mean vector and covariance matrix of the two-dimensional Gaussian distribution, respectively; π is the weight coefficient of a Gaussian distribution in the Gaussian mixture distribution; ρ is the correlation coefficient of the longitudinal and lateral positions;

[0029] The basic parameters μ, σ, π, and ρ of the Gaussian distribution are generated by the MDN (the parameters with superscript “~” in equation (6)):

[0030]

[0031] The output sequence Ω after MDN processing is:

[0032]

[0033] 1.5, by predicting the longitudinal position sequence of traffic participant j Taking the gradient thereof to obtain the longitudinal velocity prediction sequence of traffic participant j by predicting the lateral position sequence of traffic participant j Taking the gradient thereof to obtain the lateral velocity prediction sequence of traffic participant j by predicting the longitudinal velocity sequence of traffic participant j Taking the gradient thereof to obtain the longitudinal acceleration prediction sequence of traffic participant j by predicting the lateral velocity sequence of traffic participant j Taking the gradient thereof to obtain the lateral acceleration prediction sequence of traffic participant j Combining Ω and and to obtain the final output sequence Ψ of the traffic participant uncertainty kinematic information prediction module:

[0034]

[0035] Ψ contains the position, velocity, acceleration of the predicted traffic participant within the prediction time window, and the probability of being in the position, velocity, and acceleration;

[0036] The traffic participant uncertainty kinematic information prediction module utilizes equations (1)-(2) and (7)-(8) to predict the kinematic information of dynamic traffic participants in each traffic participant, and the kinematic information of static traffic participants does not need to be predicted.

[0037] The traffic participant uncertainty kinematic information prediction module needs to be trained based on natural driving data, and the training loss function Loss of the traffic participant uncertainty kinematic information prediction module is designed as:

[0038]

[0039] wherein, For the predicted traffic participant j in t f The actual location at any given moment For the predicted traffic participant j in t f Predicted location at time J k The number of dynamic traffic participants around the vehicle is ω1 and ω2, which are weighting coefficients designed based on development experience and actual needs. The first term in Equation (9) is used to measure the difference between the predicted value and the true value output by the traffic participant uncertainty kinematic information prediction module. The second term is a parameter regularization term in the form of L2 norm. The third term is a logarithmic MDN. Through the above method, those skilled in the art can complete the construction and training of the traffic participant uncertainty kinematic information prediction module.

[0040] Furthermore, the traffic participant anisotropic risk field module, based on whether the traffic participants are moving, divides the traffic participant anisotropic risk field into a dynamic anisotropic risk field and a static anisotropic risk field:

[0041] 2.1 Dynamic Anisotropic Risk Field

[0042] The anisotropy of the dynamic anisotropic risk field is determined by the motion state of the dynamic traffic participants. If j is a static traffic participant, then its relevant variables are labeled using the superscript k. In the world coordinate system, at time t, the dynamic traffic participants... With bicycle Distance vector d k,e (t) is:

[0043]

[0044] The anisotropic risk field module for traffic participants establishes a coordinate system X fixed on the dynamic traffic participant k. k O k Y k O k For the center of k, OX k OY represents the direction of k's movement. k The normal vector represents the direction in which k moves forward.

[0045] Furthermore, the traffic participant anisotropic risk field module will use the distance vector d in the world coordinate system. k,e Convert to X k O k Y k Equivalent distance vector in coordinate system

[0046]

[0047] in, and OX in world coordinate system k and OY k The unit vector of direction, α k and β k OX k and OY k Distance scaling factor in direction, α k and β k The lengths of the major and minor axes of the circumscribed ellipse of k are respectively equal to the lengths of the minor and major axes; then the field strength of the dynamic anisotropic risk field generated by the dynamic traffic participant k on its own vehicle EV at time t is... for:

[0048]

[0049] in, The direction is from the center of the dynamic traffic participant k to the center of the vehicle EV; R is the road state factor. The lower the road visibility and adhesion coefficient, and the greater the slope, the larger the road state factor, which represents the higher the driving risk from the road state. and OX k and OY k Acceleration in the direction of λ K1 and λ K2 Here, μ is the acceleration coefficient, and μ0 is the peak field strength occurring at the center of traffic participants. They are respectively With OX k OY k The angle between directions, M k To consider the equivalent mass of dynamic traffic participants' actual mass and speed:

[0050]

[0051] Where, m k For the true quality of dynamic traffic participant k, v k Let k be the velocity in the forward direction, and b be the velocity. M and c M It is a constant;

[0052] 2.2 Static Anisotropic Risk Field

[0053] The anisotropy of the static anisotropic risk field is determined by the geometric shape of the static traffic participants. If j is a static traffic participant, then the superscript s is used to label its relevant variables. In the world coordinate system, at time t, the static traffic participant s(x s ,y s ) and bicycle Distance vector d s,e (t) is:

[0054]

[0055] The traffic participant anisotropic risk field module establishes a coordinate system X s O s Y s , wherein O s is the center of s, OX s and OY s respectively represent the long axis and short axis directions of the circumscribed ellipse of s.

[0056] Further, the traffic participant anisotropic risk field module converts the distance vector d s,e in the world coordinate system into the equivalent vector s in the X s O s Y

[0057]

[0058] wherein, and are unit vectors in the OX s and OY s directions of the world coordinate system, and α s and β s are distance scaling factors in the OX s and OY s directions, and α s and β s are respectively equal to the lengths of the long axis and short axis of the circumscribed ellipse of the static obstacle s. By rewriting equation (12), the static anisotropic risk field strength generated by the static obstacle s on the ego vehicle EV at time t is obtained as:

[0059]

[0060] wherein, M s is the equivalent mass of s, and m s is the actual mass of s;

[0061] 2.3. Anisotropic risk field expression considering motion uncertainty of traffic participants

[0062] The output of the traffic participant anisotropic risk field module is the risk field strength generated by each traffic participant on the EV at each time in the past and future:

[0063] ​When the risk field strength generated by each traffic participant to the EV in the history is calculated, the kinematic information of each traffic participant and the EV history needs to be known, which is the known quantity that can be directly obtained by the EV sensor, so the risk field strength generated by each traffic participant to the EV in the history time window can be directly calculated according to formula (12) or (16);

[0064] When the risk field strength generated by each traffic participant to the EV in the future is calculated, the kinematic information of the EV and each traffic participant in the future needs to be known, wherein the kinematic information of the EV in the future is directly obtained according to the motion planning system of the EV; for the static traffic participant, the future position is unchanged, and the speed and acceleration are always 0, so the field strength generated by the future static traffic participant to the EV is directly calculated according to formula (12) or (16); for the dynamic traffic participant, the kinematic information and uncertainty in the future, i.e. the prediction time window, are output by the traffic participant uncertainty kinematic information prediction module, so the probability weighted risk field strength generated by the dynamic traffic participant k to the EV considering the motion uncertainty is calculated according to formula (12) and (3) at the future time step t f

[0065]

[0066] Further, when the driving risk space distribution pseudo picture is constructed, the driving risk space distribution pseudo picture construction module first selects eight traffic participants around the EV according to the position of the EV:

[0067] 3.1, for the dynamic traffic participant in the scene, the 6-dimensional vector for describing the relative risk state of the dynamic traffic participant to the EV at time t is:

[0068]

[0069] wherein: ① and are respectively the risk field strength generated by the dynamic traffic participant k to the EV along the forward direction of the EV and the component module along the normal direction, is the risk field strength generated by the dynamic traffic participant k to the EV and the angle between the forward direction of the EV; ② and are respectively the risk repulsion acting on the forward direction of the EV and the risk repulsion acting on the normal direction of the EV, v e is the speed of the forward direction of the EV, and m e ​​For the ego vehicle real mass, D1 and D2 are the risk sensitivity coefficients of the ego vehicle in the forward direction and the normal direction respectively, and γ is the speed coefficient of the ego vehicle, by adopting the Sigmoid function structure, the values of D1 and D2 are limited in the range of [0, 1]; ③ and are respectively the acceleration mutation markers of the traffic participant k in the forward direction and the normal direction, and the mutation markers are 0 or 1; and are respectively the acceleration mutation markers of the traffic participant k in the forward direction and the normal direction, and the mutation markers are 0 or 1;

[0070] 3.2, for the static traffic participant in the figure, the 6-dimensional vector for describing the relative risk state of the ego vehicle at time t is:

[0071]

[0072] wherein, ① and are respectively the risk field intensity generated by the static traffic participant s to the ego vehicle along the component value of the ego vehicle in the forward direction and the component value along the normal direction, is the risk field intensity generated by the static traffic participant s to the ego vehicle and the angle with the forward direction of the ego vehicle; ② and are respectively the risk repulsion of the static traffic participant s acting on the ego vehicle in the forward direction and the risk repulsion acting on the ego vehicle in the normal direction; ③ since the static traffic participant s is always stationary, the acceleration mutation markers in the forward direction and the normal direction are always 0;

[0073] 3.3, at time t, for the ego vehicle EV, the 6-dimensional vector for describing the state of the ego vehicle is:

[0074]

[0075] wherein, ① is the speed of the ego vehicle in the forward direction, and since the normal speed of the ego vehicle is 0, the second element of is always 0; ② and are respectively the acceleration of the ego vehicle in the forward direction and the normal direction; ③ and are respectively the acceleration mutation markers of the ego vehicle in the forward direction and the normal direction, and the mutation markers are 0 or 1;

[0076] 3.4, at time t, the driving risk space distribution pseudo picture is described as:

[0077]

[0078] ​​The driving risk space distribution pseudo-picture construction module combines the driving risk space distribution pseudo-pictures of each moment in t0-H to t0+F into a history and future driving risk space distribution pseudo-picture sequence according to a time sequence relationship.

[0079] Further, the space-time coupling driving risk assessment module is a deep learning model comprising a CNN and an LSTM, receives the history and future (t0-H to t0+F) driving risk space distribution pseudo-picture sequence, extracts the driving risk space distribution information contained in each moment pseudo-picture in the history and future by a convolutional neural network (CNN), processes the time sequence dependence relationship and dynamic evolution trend between each moment risk by a long-short term memory network (LSTM), and finally outputs a time and space coupling intelligent automobile driving risk level:

[0080] 4.1, the space-time coupling driving risk assessment module processes each moment driving risk space distribution pseudo-picture by a CNN, the CNN comprises a convolution layer and a pooling layer, the convolution layer extracts the driving risk space distribution characteristics in the pseudo-picture by a convolution kernel, and the pooling layer performs secondary sampling on the output of the convolution layer to ensure that the model can extract sufficient driving risk space dependence; each moment driving risk space distribution pseudo-picture is processed into a semantic vector containing the driving risk space distribution characteristics of the moment by the CNN; the semantic vector of each moment is input into the LSTM part of the space-time coupling driving risk assessment module according to the forward time sequence of t0-H to t0+F;

[0081] 4.2, the space-time coupling driving risk assessment module processes the t0-H to t0+F driving risk space distribution characteristic semantic vectors output by the CNN by an LSTM to extract the time sequence dependence relationship and dynamic evolution trend between each moment risk; the final output of the LSTM is converted into a space-time coupling self-vehicle driving risk level after being processed by a fully connected layer and a Softmax layer.

[0082] Advantages of the present application:

[0083] 1. The present application describes the risk generated by other traffic participants as a risk source in the traffic environment by constructing an anisotropic driving risk field, and describes the risk state of the traffic participant relative to the ego vehicle using 6-dimensional risk features composed of the risk field strength and its derivative variables, effectively solving the problem of inaccurate description of objective risk in the traffic environment in the prior art.

[0084] 2. The present application introduces the uncertainty prediction of the future kinematic information of other traffic participants, so as to consider the uncertain interaction coupling relationship between the ego vehicle and other traffic participants and the dynamic evolution mechanism of the traffic situation in the driving risk assessment, further improve the risk assessment accuracy, and solve the problem that the prior art does not fully consider the motion uncertainty of other traffic participants in the traffic environment.

[0085] 3. The present application constructs a pseudo picture covering the risk space distribution information of the ego vehicle and eight traffic participants around the ego vehicle based on the six-dimensional risk feature, and further composes a history and future driving risk space distribution pseudo picture sequence based on the uncertainty prediction results of the future kinematic information of the traffic participants. Finally, taking the pseudo picture sequence as the input, the driving risk space distribution information contained in each time pseudo picture of the history and future is extracted through the convolutional neural network (CNN), the time sequence dependence relationship and evolution trend of the history and future driving risk are processed through the long-short term memory network (LSTM), and finally the spatio-temporal coupled driving risk level is output, which effectively solves the technical problem that the spatio-temporal coupled accurate driving risk assessment is difficult to realize in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0086] Figure 1 It is a schematic diagram of the overall architecture of the present application.

[0087] Figure 2 It is a schematic diagram of the surrounding traffic participants that have an impact on the driving risk of the ego vehicle.

[0088] Figure 3 It is a schematic diagram of the structure of the traffic participant uncertainty kinematic information prediction module of the present application.

[0089] Figure 4 It is a schematic diagram of the pseudo picture construction principle of the driving risk space distribution of the present application.

[0090] Figure 5 It is a schematic diagram of the network structure used by the spatio-temporal coupled driving risk assessment module of the present application. DETAILED DESCRIPTION

[0091] Referring to Figure 1 The present application provides a spatio-temporal coupled intelligent automobile risk assessment system considering motion uncertainty, which comprises a traffic participant uncertainty kinematic information prediction module, a traffic participant anisotropic risk field module, a driving risk space distribution pseudo picture construction module and a spatio-temporal coupled driving risk assessment module.

[0092] 1. The traffic participant uncertainty kinematic information prediction module takes the historical kinematic information of each traffic participant and the ego vehicle as input, and outputs the future kinematic information of each traffic participant considering its motion uncertainty within a prediction time window in the form of a probability distribution.

[0093] Referring to Figure 2 As shown in the figure, the traffic participants considered by the present application to have an impact on the driving risk of the ego vehicle (EV) are a total of 8 traffic participants located in the front (Fr), rear (Re), left (L), left front (LF), left rear (LR), right (Ri), right front (RF), and right rear (RR) of the EV. The longitudinal range of the selected traffic participants is 150 m in front and behind the EV, and the lateral range of the selected traffic participants is the current lane of the EV and the adjacent lanes on the left and right. According to whether the traffic participants are moving or not, the traffic participants are divided into two categories: dynamic traffic participants (the main dynamic traffic participants considered by the present application are traffic vehicles) and static traffic participants (such as cones, roadblocks, etc.).

[0094] Referring to Figure 3 As shown in the figure, the traffic participant uncertainty kinematic information prediction module is composed of an LSTM-based encoder, an attention mechanism, an LSTM-based decoder, and a mixture density network (MDN) 4 parts;

[0095] 1.1. The input of the LSTM-based encoder is the time series of the historical kinematic information of the ego vehicle and the traffic participants. At the current time t0, the historical kinematic information time series X is:

[0096]

[0097] where H is the length of the historical time window, and the present application takes H = 5 s; x e and y e are the longitudinal position and lateral position of the ego vehicle EV, respectively; v ex and v ey are the longitudinal speed and lateral speed of the EV, respectively; x j and y j are the longitudinal position and lateral position of the traffic participant j, respectively; v jx and v jy are the longitudinal speed and lateral speed of the traffic participant j, respectively;

[0098] At each encoding time step t hThe LSTM unit of the LSTM-based encoder receives the previous encoding time step t. h -1 hidden state and the current encoding time step t h Input Output the current encoding time step t h Hidden state Until the entire input sequence X is encoded into a hidden state sequence

[0099] 1.2, at each decoding time step t f The attention mechanism calculates the decoder's performance at the previous decoding time step t. f -1 hidden state and The correlation between them is used to calculate the current decoding time step t. f Below, encoder hidden state sequence The weight coefficients of each hidden state are then used to obtain the current decoding time step t through weighted summation. f The semantic vector; the introduction of the Attention mechanism enables the traffic participant uncertainty kinematic information prediction module to fully and efficiently extract the temporal dependency features in the input sequence X, thereby improving the prediction accuracy.

[0100] 1.3, at each decoding time step t f The LSTM unit of the LSTM-based decoder receives the hidden state of the previous decoding time step tf-1. and the current decoding time step t f The semantic vector is used to output the current decoding time step t. f Hidden state Until decoding is completed for all decoding time steps, the positions of relevant traffic participants around the EV within the prediction time window are predicted; the output sequence Y of the LSTM-based decoder is:

[0101]

[0102] Wherein, F is the prediction time window length of the traffic participant uncertainty kinematic information prediction module; in this invention, F = 3s.

[0103] 1.4 The Hybrid Density Network (MDN) describes the uncertainty of traffic participant predicted location (i.e., the uncertainty of traffic participant motion) through a mixture of six Gaussian distributions. Taking traffic participant j as an example, the modeling method of MDN for the uncertainty of traffic participant predicted location is explained: Assuming that the discretized position of traffic participant j follows a Markov process, then at decoding time step t... f Given the previous decoding time step t f -1 location information In this case, the traffic participant decodes the time step t f is located at position The probability is expressed as:

[0104]

[0105] where:

[0106]

[0107]

[0108] where, and are the true longitudinal and lateral positions of the traffic participant j; μ and σ are the mean vector and covariance matrix of the two-dimensional Gaussian distribution, respectively; π is the weight coefficient of a Gaussian distribution in the mixed Gaussian distribution; ρ is the correlation coefficient of the longitudinal and lateral positions;

[0109] The basic parameters μ, σ, π, and ρ of the Gaussian distribution are generated by the MDN (the parameters with superscript “~” in equation (6)):

[0110]

[0111] Therefore, the output sequence Ω after the MDN processing is:

[0112]

[0113] 1.5, the longitudinal position prediction sequence of the traffic participant j is obtained by Taking the gradient, the longitudinal velocity prediction sequence of the traffic participant j is obtained The lateral position prediction sequence of the traffic participant j is obtained by Taking the gradient, the lateral velocity prediction sequence of the traffic participant j is obtained The longitudinal velocity sequence of the traffic participant j is obtained by Taking the gradient, the longitudinal acceleration prediction sequence of the traffic participant j is obtained The lateral velocity sequence of the traffic participant j is obtained by Taking the gradient, the lateral acceleration prediction sequence of the traffic participant j is obtained Combining Ω and and obtains the final output sequence Ψ of the traffic participant uncertainty kinematics information prediction module:

[0114]

[0115] Ψ contains the position, velocity, and acceleration of the traffic participant within the prediction time window, as well as the probability of being at that position, velocity, and acceleration; therefore, the traffic participant uncertainty kinematic information prediction module realizes the prediction of the future kinematic information of traffic participants and the quantitative description of their uncertainty.

[0116] It should be noted that although all eight traffic participants around the EV are listed in equations (1)-(2) and (7)-(8), this is only for the convenience of mathematical expression. In fact, for the traffic participant uncertainty kinematic information prediction module, the kinematic information of static traffic participants does not need to be predicted, only the kinematic information of dynamic traffic participants needs to be predicted.

[0117] The traffic participant uncertainty kinematics information prediction module requires model parameter training based on natural driving data. This invention designs the training loss function Loss for the traffic participant uncertainty kinematics information prediction module as follows:

[0118]

[0119] in, For the predicted traffic participant j in t f The actual location at any given moment For the predicted traffic participant j in t f Predicted location at time J k The number of dynamic traffic participants around the vehicle is ω1 and ω2, which are weighting coefficients designed based on development experience and actual needs. The first term in Equation (9) is used to measure the difference between the predicted value and the true value output by the traffic participant uncertainty kinematic information prediction module. The second term is a parameter regularization term in the form of L2 norm. The third term is a logarithmic MDN. Through the above method, those skilled in the art can complete the construction and training of the traffic participant uncertainty kinematic information prediction module.

[0120] Furthermore, this invention does not elaborate on the process of calculating the hidden state by the LSTM unit in the LSTM-based encoder and the LSTM-based decoder, because this is a basic method that is well known to those skilled in the art, and this invention will not elaborate on it.

[0121] 2. The traffic participant anisotropic risk field module takes the historical kinematic information of each traffic participant and the future kinematic information of each traffic participant considering motion uncertainty output by the traffic participant uncertainty kinematic information prediction module as input. It calculates the risk generated by each traffic participant at each moment in the past and future by constructing anisotropic risk field, and outputs the risk field strength of each traffic participant to the vehicle in the past and future.

[0122] The traffic participant anisotropic risk field module divides the traffic participant anisotropic risk field into a dynamic anisotropic risk field and a static anisotropic risk field, depending on whether the traffic participants are moving.

[0123] 2.1 Dynamic Anisotropic Risk Field

[0124] The anisotropy of the dynamic anisotropic risk field is determined by the motion state of dynamic traffic participants. This invention comprehensively considers various working conditions such as going straight, changing lanes, and turning, and constructs a dynamic anisotropic risk field that can comprehensively and accurately reflect the defined attributes and motion states of dynamic traffic participants. If j is a static traffic participant, its relevant variables are labeled using the superscript k. In the world coordinate system, at time t, the dynamic traffic participant... With bicycle Distance vector d k,e (t) is:

[0125]

[0126] The anisotropic risk field module for traffic participants establishes a coordinate system X fixed on the dynamic traffic participant k. k O k Y k O k For the center of k, OX k OY represents the direction of k's movement. k The normal vector represents the direction in which k moves forward.

[0127] Furthermore, the traffic participant anisotropic risk field module will use the distance vector d in the world coordinate system. k,e Convert to X k O k Y k Equivalent distance vector in coordinate system

[0128]

[0129] in, and OX in world coordinate system k and OY k The unit vector of direction, α k and β k OX k and OY k Distance scaling factor in direction, α k and β k The lengths of the major and minor axes of the circumscribed ellipse of k are respectively equal to the lengths of the minor and major axes; then the field strength of the dynamic anisotropic risk field generated by the dynamic traffic participant k on its own vehicle EV at time t is... for:

[0130]

[0131] in, The direction is from the center of the dynamic traffic participant k to the center of the vehicle EV; R is the road state factor. The lower the road visibility and adhesion coefficient, and the greater the slope, the larger the road state factor, which represents the higher the driving risk from the road state. and OX k and OY k Acceleration in the direction of λ K1 and λ K2 Here, μ is the acceleration coefficient, and μ0 is the peak field strength occurring at the center of traffic participants. They are respectively With OX k OY k The angle between directions, M k To consider the equivalent mass of dynamic traffic participants' actual mass and speed:

[0132]

[0133] Where, m k For the true quality of dynamic traffic participant k, v k Let k be the velocity in the forward direction, and b be the velocity. M and c M It is a constant;

[0134] 2.2 Static Anisotropic Risk Field

[0135] The anisotropy of the static anisotropic risk field is determined by the geometric shape of the static traffic participants. If j is a static traffic participant, then the superscript s is used to label its relevant variables. In the world coordinate system, at time t, the static traffic participant s(x s ,y s ) and bicycle Distance vector d s,e (t) is:

[0136]

[0137] The traffic participant anisotropic risk field module establishes a coordinate system X fixed on s. s O s Y s O s OX is the center of s s and OY s These represent the directions of the major and minor axes of the circumscribed ellipse s, respectively.

[0138] Further, the traffic participant anisotropic risk field module converts the distance vector d s,e into X s O s Y s equivalent vector in the world coordinate system

[0139]

[0140] wherein, and are the unit vectors in the OX s and OY s directions in the world coordinate system, and α s and β s are the distance scaling factors in the OX s and OY s directions, respectively, and α s and β s are equal to the lengths of the long axis and the short axis of the circumscribed ellipse of the static obstacle s, respectively. Rewriting equation (12), the static anisotropic risk field strength generated by the static obstacle s on the ego vehicle EV at time t is:

[0141]

[0142] wherein M s is the equivalent mass of s, and m s is the real mass of s;

[0143] 2.3. Anisotropic risk field expression considering motion uncertainty of traffic participants

[0144] The output of the traffic participant anisotropic risk field module is the risk field strength generated by each traffic participant on the EV at each time in the past and in the future:

[0145] ① When calculating the risk field strength generated by each traffic participant on the EV in the past, the kinematic information of each traffic participant and the EV in the past needs to be known, and these information are known quantities that can be directly obtained by the EV sensors, so the risk field strength generated by each traffic participant on the EV in the historical time window can be directly calculated according to equation (12) or (16);

[0146] ​②When calculating the risk field strength generated by each traffic participant to the EV in the future, the kinematic information of the EV in the future needs to be known, wherein the kinematic information of the EV in the future is directly obtained according to the motion planning system of the EV; for the static traffic participant, the future position of the static traffic participant is constant, and the speed and acceleration thereof are always 0, so the field strength generated by the static traffic participant to the EV in the future is directly calculated according to formula (12) or formula (16); for the dynamic traffic participant, the kinematic information and uncertainty of the dynamic traffic participant in the future, i.e., the prediction time window, are output by the traffic participant uncertainty kinematic information prediction module, and then the probability weighted risk field strength generated by the dynamic traffic participant k to the EV in the future time step t f , considering the motion uncertainty, is calculated based on formula (12) and formula (3).

[0147]

[0148] 3、The driving risk space distribution pseudo picture construction module takes, as input, the risk field strength generated by each traffic participant to the EV at each moment in the history and the future (t-H to t+F), the kinematic information of the EV at each moment in the history and the future (t-H to t+F), and the kinematic information of each traffic participant at each moment in the history and the future (t-H to t+F), and uses the risk field strength, the risk field strength derivative variable, and the kinematic information to form a 6-dimensional risk feature.

[0149] Further, at each moment in the history and the future, the driving risk space distribution pseudo picture construction module divides a certain space region around the EV into a 3x3 grid region according to the position of the EV at the moment, and uses the 6-dimensional risk feature at the moment to describe the risk state of each traffic participant in each grid relative to the EV and the risk state of the EV, thereby constructing a 3x3x6 pseudo picture to describe the driving risk space distribution at the moment, and finally outputting a pseudo picture sequence composed of the driving risk space distribution pseudo pictures with a dimension of 3x3x6 at each moment in the history and the future (t-H to t+F).

[0150] Referring to the accompanying Figure 4 , when constructing the driving risk space distribution pseudo picture, the driving risk space distribution pseudo picture construction module first selects eight traffic participants around the EV according to the position of the EV. Figure 2 The traffic scene shown in FIG. 3 is taken as an example to illustrate the construction principle of the driving risk space distribution pseudo picture at each moment:

[0151] 3.1, for the dynamic traffic participants Fr, Re, RF, RR, and L in the scene, the 6-dimensional vector for describing the risk state of each dynamic traffic participant relative to the EV at the moment t is:

[0152] ​

[0153] wherein: ① and are the risk field strength generated by dynamic traffic participant k to the ego vehicle along the ego vehicle's advancing direction and normal direction, respectively are the component modulus of the risk field generated by dynamic traffic participant k to the ego vehicle along the ego vehicle's advancing direction and normal direction, respectively is the risk field strength generated by dynamic traffic participant k to the ego vehicle is the angle between the ego vehicle's advancing direction and the risk field generated by dynamic traffic participant k; ② and are the risk repulsion force generated by the risk field to the ego vehicle along the ego vehicle's advancing direction and normal direction, respectively e is the speed of the ego vehicle along the ego vehicle's advancing direction, m e is the real mass of the ego vehicle, D1 and D2 are the risk sensitivity coefficients of the ego vehicle along the advancing direction and normal direction, respectively, and γ is the speed coefficient of the ego vehicle, and by adopting a Sigmoid function structure, the values of the and are limited in the range of [0, 1]; ③ and are the acceleration mutation markers of traffic participant k along the advancing direction and normal direction, respectively; when the absolute value of the advancing direction acceleration of traffic participant k is greater than 3 m / s 2 , otherwise when the absolute value of the normal direction acceleration of traffic participant k is greater than 2 m / s 2 , otherwise

[0154] 3.2, for the static traffic participants LF, LR and Ri in the figure, the 6-dimensional vector used to describe the relative risk state of the ego vehicle at time t is:

[0155]

[0156] wherein: ① and are the risk field strength generated by static traffic participant s to the ego vehicle along the ego vehicle's advancing direction and normal direction, respectively are the component values of the risk field generated by static traffic participant s to the ego vehicle along the ego vehicle's advancing direction and normal direction, respectively is the risk field strength generated by static traffic participant s to the ego vehicle is the angle between the ego vehicle's advancing direction and the risk field generated by static traffic participant s; ② and are the risk repulsion force generated by the risk field to the ego vehicle along the ego vehicle's advancing direction and normal direction, respectively; ③ since static traffic participant s is always stationary, the acceleration mutation markers of static traffic participant s along the advancing direction and normal direction are always 0.​

[0157] 3.3、At time t, for the ego vehicle EV, the 6-dimensional vector used to describe its state is

[0158]

[0159] where ① is the speed of the ego vehicle in the forward direction, and since the normal speed of the ego vehicle is 0, thus the second element of is always 0; ② and are the accelerations of the ego vehicle in the forward direction and the normal direction, respectively; ③ and are the acceleration jerk of the ego vehicle in the forward direction and the normal direction, respectively; when the ego vehicle has an absolute value greater than 3 m / s 2 forward direction acceleration, otherwise when the ego vehicle has an absolute value greater than 2 m / s2normal direction acceleration, otherwise

[0160] 3.4、At time t, the pseudo picture of the driving risk space distribution can be described as:

[0161]

[0162] The driving risk space distribution pseudo picture construction module combines the driving risk space distribution pseudo picture of each time from t0-H to t0+F in chronological order into a sequence of historical and future driving risk space distribution pseudo pictures.

[0163] 4、The space-time coupled driving risk assessment module is a deep learning model containing CNN and LSTM, which receives the sequence of historical and future (t0-H to t0+F) driving risk space distribution pseudo pictures, extracts the driving risk space distribution information contained in each pseudo picture at each time through a convolutional neural network (CNN), processes the time sequence dependence relationship and dynamic evolution trend between the risk at each time through a long-short term memory network (LSTM), and finally outputs the intelligent vehicle driving risk level coupled with time and space.

[0164] ​4.1, the space-time coupling driving risk assessment module processes the driving risk space distribution pseudo picture of each moment by CNN, the CNN is composed of convolution layer and pooling layer, the convolution layer extracts the driving risk space distribution characteristics in the pseudo picture through the convolution kernel, the pooling layer carries out secondary sampling to the output of the convolution layer, to ensure that the model can extract sufficient driving risk space dependence; the driving risk space distribution pseudo picture of each moment is processed by CNN into a semantic vector containing the driving risk space distribution characteristics at this moment; the semantic vector of each moment is input into the LSTM part of the space-time coupling driving risk assessment module according to the positive time sequence of t0-H to t0+F;

[0165] 4.2, the space-time coupling driving risk assessment module processes the t0-H to t0+F driving risk space distribution characteristic semantic vector output by CNN to extract the time sequence dependence and dynamic evolution trend between the driving risk of each moment; the final output of LSTM is converted into the space-time coupling driving risk level of the ego vehicle after being processed by a fully connected layer and a Softmax layer.

[0166] It should be noted that the CNN and LSTM used in the space-time coupling driving risk assessment module are general technologies mastered by technicians in the relevant field, therefore, the basic principles of CNN and LSTM will not be described herein.

[0167] The space-time coupling intelligent automobile risk assessment system considering motion uncertainty described in the application is trained and verified on the natural driving data set "100-cars", the driving risk level determination accuracy of the system described in the application on the training set reaches 97.19%, and the driving risk level determination accuracy on the test set reaches 94.91%.

Claims

1. A spatiotemporally coupled intelligent vehicle risk assessment system considering motion uncertainties, characterized in that: It includes a traffic participant uncertainty kinematic information prediction module, a traffic participant anisotropic risk field module, a driving risk spatial distribution pseudo-image construction module, and a spatiotemporally coupled driving risk assessment module. (1) The traffic participant uncertainty kinematic information prediction module takes the historical kinematic information of each traffic participant and the vehicle as input, and outputs the future kinematic information of each traffic participant in the form of a probability distribution, taking into account the uncertainty of their motion within the prediction time window. (2) The traffic participant anisotropic risk field module takes the historical kinematic information of each traffic participant and the future kinematic information of each traffic participant considering motion uncertainty output by the traffic participant uncertainty kinematic information prediction module as input. It calculates the risk generated by each traffic participant at each moment in the past and future by constructing anisotropic risk fields, and outputs the risk field strength of each traffic participant to the vehicle in the past and future. The traffic participant anisotropic risk field module divides the traffic participant anisotropic risk field into dynamic anisotropic risk field and static anisotropic risk field according to whether the traffic participant is moving. (3) The pseudo-image construction module for the spatial distribution of driving risk takes the risk field strength of each traffic participant to the vehicle at each historical and future moment, the kinematic information of the vehicle at each historical and future moment, and the kinematic information of each traffic participant at each historical and future moment as input, and uses the risk field strength, the risk field strength derived variables and kinematic information to form a 6-dimensional risk feature. At each moment in the past and future, the pseudo-image construction module for the spatial distribution of driving risk divides a certain spatial area around the vehicle into a 3×3 gridded area based on the vehicle's position at that moment. It then uses the 6-dimensional risk features at that moment to describe the risk status of traffic participants relative to the vehicle and the risk status of the vehicle itself within each grid, thereby constructing a 3×3×6 pseudo-image to describe the spatial distribution of driving risk at that moment. Finally, it outputs a pseudo-image sequence composed of pseudo-images of the spatial distribution of driving risk at each moment in the past and future. (4) The spatiotemporal coupled driving risk assessment module is a deep learning model consisting of two parts: CNN and LSTM. It receives a sequence of pseudo-images showing the spatial distribution of driving risks in the past and future. It extracts the spatial distribution information of driving risks contained in the pseudo-images at each moment in the past and future through a convolutional neural network (CNN). It processes the temporal dependency relationship and dynamic evolution trend between risks at each moment through a long short-term memory network (LSTM). Finally, it outputs the driving risk level of the intelligent vehicle that is coupled in time and space.

2. The spatiotemporally coupled intelligent vehicle risk assessment system considering motion uncertainties according to claim 1, characterized in that: The traffic participants include eight traffic participants located in front of, behind, to the left of, to the left front of, to the left rear of, to the right of, to the right front of, and to the right rear of the EV. The longitudinal range of the selected traffic participants is 150m in front of and behind the EV, and the lateral range of the selected traffic participants is the current lane of the EV and the adjacent lanes to the left and right. According to whether they are moving, the traffic participants are divided into dynamic traffic participants and static traffic participants.

3. The spatiotemporally coupled intelligent vehicle risk assessment system considering motion uncertainties according to claim 1, characterized in that: The traffic participant uncertainty kinematics information prediction module includes an LSTM-based encoder, an attention mechanism, an LSTM-based decoder, and a hybrid density network (MDN). 1.1 The input to the LSTM-based encoder is a time series of historical kinematic information of the vehicle and traffic participants. At the current time t0, the historical kinematic information time series X is: Where H is the length of the historical time window; x e and y e These refer to the longitudinal and lateral positions of the EV, respectively; v ex and v ey These represent the longitudinal and lateral velocities of the EV, respectively; x j and y j These represent the longitudinal and lateral positions of traffic participant j, respectively; v jx and v jy The longitudinal and lateral velocities of traffic participant j are respectively; Fr, Re, RF, Ri, RR, LF, L, and LR are the eight traffic participants on the front, rear, right front, right, right rear, left front, left, and left rear sides of the vehicle EV, respectively. At each encoding time step t h The LSTM unit of the LSTM-based encoder receives the previous encoding time step t. h -1 hidden state and the current encoding time step t h Input Output the current encoding time step t h Hidden state Until the entire input sequence X is encoded into a hidden state sequence 1.2, at each decoding time step t f The attention mechanism calculates the decoder's performance at the previous decoding time step t. f -1 hidden state and The correlation between them is used to calculate the current decoding time step t. f Below, encoder hidden state sequence The weight coefficients of each hidden state are summed using weighted averages to obtain the current decoding time step t. f semantic vector; 1.3, at each decoding time step t f The LSTM unit of the LSTM-based decoder receives the previous decoding time step t. f -1 hidden state and the current decoding time step t f The semantic vector is used to output the current decoding time step t. f Hidden state Until decoding is completed for all time steps, the positions of relevant traffic participants around the EV within the prediction time window are predicted; the output sequence Y of the LSTM-based decoder is: Where F is the prediction time window length of the traffic participant uncertainty kinematic information prediction module; 1.4 The Hybrid Density Network (MDN) describes the uncertainty of traffic participant predicted location, i.e., the uncertainty of traffic participant motion, through a mixture of six Gaussian distributions. The modeling method is as follows: assuming the discretized position of traffic participant j follows a Markov process, then at decoding time step t... f Given the previous decoding time step t f -1 location information In the case of traffic participants at decoding time step t f Located in The probability is expressed as: in: in, and denoted as the actual longitudinal and lateral positions of traffic participant j; μ and σ are the two-dimensional mean vector and covariance matrix, respectively; π is the weighting coefficient of a Gaussian distribution in the mixture Gaussian distribution; and ρ is the correlation coefficient between the longitudinal and lateral positions. The basic parameters μ, σ, π, and ρ of the Gaussian distribution mentioned above are generated using MDN: The output sequence Ω after MDN processing is: 1.

5. Predicting sequences based on the longitudinal position of traffic participant j The gradient is used to obtain the longitudinal velocity prediction sequence. Predicting the lateral position sequence of traffic participant j The gradient is used to obtain the lateral velocity prediction sequence. By analyzing the longitudinal velocity sequence of traffic participant j The gradient is used to obtain the longitudinal acceleration prediction sequence. By analyzing the lateral velocity sequence of traffic participant j The gradient is used to obtain the lateral acceleration prediction sequence. Ω and and By combining these sequences, the final output sequence Ψ of the traffic participant uncertainty kinematic information prediction module is obtained: Ψ contains the location, speed, and acceleration of the traffic participant being predicted within the prediction time window, as well as the probability of being at that location, speed, and acceleration. The traffic participant uncertainty kinematic information prediction module uses equations (1)-(2) and (7)-(8) to predict the kinematic information of dynamic traffic participants among all traffic participants, while the kinematic information of static traffic participants does not need to be predicted.

4. A spatiotemporally coupled intelligent vehicle risk assessment system considering motion uncertainties according to claim 1 or 3, characterized in that: The traffic participant uncertainty kinematic information prediction module trains model parameters based on natural driving data. The training loss function Loss of the traffic participant uncertainty kinematic information prediction module is designed as follows: in, For the predicted traffic participant j in t f The actual location at any given moment For the predicted traffic participant j in t f Predicted location at time J k The number of dynamic traffic participants around the vehicle is ω1 and ω2, which are weighting coefficients designed based on development experience and actual needs. The first term in Equation (9) is used to measure the difference between the predicted value and the true value output by the traffic participant uncertainty kinematic information prediction module. The second term is a parameter regularization term in the form of L2 norm. The third term is a logarithmic MDN. The construction and training of the traffic participant uncertainty kinematic information prediction module are completed through the above method.

5. A spatiotemporally coupled intelligent vehicle risk assessment system considering motion uncertainties according to claim 1 or 3, characterized in that: 2.1 The anisotropy of the dynamic anisotropic risk field is determined by the motion state of the dynamic traffic participants. If j is a static traffic participant, then the superscript k is used to label its relevant variables. In the world coordinate system, at time t, the dynamic traffic participants... With bicycle Distance vector d k,e (t) is: The anisotropic risk field module for traffic participants establishes a coordinate system X fixed on the dynamic traffic participant k. k O k Y k O k For the center of k, OX k OY represents the direction of k's movement. k The normal vector representing the direction of k's movement; The anisotropic risk field module for traffic participants will use the distance vector d in the world coordinate system. k,e Convert to X k O k Y k Equivalent distance vector in coordinate system in, and OX in world coordinate system k and OY k The unit vector of direction, α k and β k OX k and OY k Distance scaling factor in direction, α k and β k The lengths of the major and minor axes of the circumscribed ellipse of k are respectively equal to the lengths of the minor and major axes; then the field strength of the dynamic anisotropic risk field generated by the dynamic traffic participant k on its own vehicle EV at time t is... for: in, The direction is from the center of the dynamic traffic participant k to the center of the vehicle EV; R is the road state factor. The lower the road visibility and adhesion coefficient, and the greater the slope, the larger the road state factor, which represents the higher the driving risk from the road state. and OX k and OY k Acceleration in the direction of λ K1 and λ K2 Here, μ is the acceleration coefficient, and μ0 is the peak field strength occurring at the center of traffic participants. They are respectively With OX k OY k The angle between directions, M k To consider the equivalent mass of dynamic traffic participants' actual mass and speed: Where, m k For the true quality of dynamic traffic participant k, v k Let k be the velocity in the forward direction, and b be the velocity. M and c M It is a constant; 2.2 The anisotropy of the static anisotropic risk field is determined by the geometric shape of the static traffic participant. If j is a static traffic participant, then the superscript s is used to label its relevant variables; in the world coordinate system, at time t, the static traffic participant s(x s ,y s ) and bicycle Distance vector d s,e (t) is: The traffic participant anisotropic risk field module establishes a coordinate system X fixed on s. s O s Y s O s OX is the center of s s and OY s These represent the directions of the major and minor axes of the circumscribed ellipse s, respectively. The anisotropic risk field module for traffic participants will use the distance vector d in the world coordinate system. s,e Convert to X s O s Y s Equivalent vector in coordinate system in, and OX in world coordinate system s and OY s The unit vector of direction, α s and β s OX s and OY s Distance scaling factor in direction, α s and β s These are equal to the lengths of the major and minor axes of the circumscribed ellipse of the static obstacle s, respectively. Rewriting equation (12), we obtain the field strength of the static anisotropic risk field generated by the static obstacle s on the vehicle EV at time t. for: Among them, M s For the equivalent mass of s, m s The true mass of s; 2.

3. Anisotropic risk field representation considering the uncertainty of traffic participant motion: The output of the traffic participant anisotropic risk field module is the risk field strength generated by each traffic participant for the EV at each historical and future moment: ① When calculating the risk field strength of each traffic participant to the EV in history, the kinematic information of each traffic participant and the EV in history is obtained through the EV sensor. Therefore, the risk field strength of each traffic participant to the EV in the historical time window is directly calculated according to formula (12) or (16). ② When calculating the risk field strength generated by each traffic participant to the EV in the future, the future kinematic information of the EV is directly obtained from the EV's motion planning system; for static traffic participants, the field strength generated by the static traffic participants to the vehicle in the future is directly calculated according to equation (12) or equation (16); for dynamic traffic participants, their kinematic information and uncertainty in the future, i.e., the prediction time window, are output by the traffic participant uncertainty kinematic information prediction module, and then based on equation (12) and equation (3), in the future time step t f Considering the probability-weighted risk field strength of dynamic traffic participant k for vehicles with motion uncertainty.

6. The spatiotemporally coupled intelligent vehicle risk assessment system considering motion uncertainties according to claim 1, characterized in that: When constructing a pseudo image of the spatial distribution of driving risks, the driving risk spatial distribution pseudo image construction module first selects eight traffic participants around the vehicle based on the vehicle's location. 3.1 For dynamic traffic participants in the scenario, at time t, a 6-dimensional vector is used to describe their risk state relative to the vehicle. for: Among them: ① and The field strength of the risk field generated by dynamic traffic participant k for the vehicle is respectively. The component magnitude along the vehicle's forward direction and the component magnitude along the normal direction, For dynamic traffic participant k, the risk field strength posed by the vehicle The angle between the vehicle's direction of travel and its direction of travel; ② and These are the risk repulsion forces acting in the vehicle's forward direction and the risk repulsion forces acting in the vehicle's normal direction, generated by the risk field, respectively. e Let m be the velocity of the vehicle in the direction of its forward movement. e The actual mass of the vehicle is given, D1 and D2 are the risk sensitivity coefficients in the forward and normal directions of the vehicle, respectively, and γ is the vehicle's speed coefficient. By employing a Sigmoid function structure, the... and The range of values ​​for is restricted to [0, 1]; ③ and These are the acceleration abrupt markers for traffic participant k in the forward direction and normal direction, respectively, with the abrupt marker being 0 or 1; 3.2 For the static traffic participants in the figure, at time t, there is a 6-dimensional vector describing their risk state relative to the vehicle. for: Among them, ① and The risk field strength of static traffic participant s to its own vehicle is respectively... The component values ​​along the direction of vehicle travel and the component values ​​along the normal direction, The field strength of the risk field posed by static traffic participant s to its own vehicle. The angle between the vehicle's direction of travel and its direction of travel; ② and These are the risk repulsion forces acting on the vehicle's forward direction and the risk repulsion forces acting on the vehicle's normal direction, generated by the risk field, respectively; ③ Since the static traffic participant s remains stationary, the acceleration abrupt changes in its forward direction and normal direction are always 0. 3.

3. At time t, for the vehicle EV, a 6-dimensional vector is used to describe its state. for: Among them, ① Let be the velocity in the direction the vehicle is moving forward, and since the vehicle's normal velocity is 0, therefore The second element is always 0; ② and These are the accelerations in the direction of the vehicle's forward movement and the normal acceleration, respectively; ③ and These are acceleration abrupt markers for the vehicle's forward direction and normal direction, respectively, with the abrupt marker being 0 or 1; 3.

4. At time t, the pseudo-image of the spatial distribution of driving risks is described as follows: Fr, Re, RF, Ri, RR, LF, L, and LR represent the eight traffic participants on the front, rear, right front, right, right rear, left front, left, and left rear sides of the vehicle EV, respectively. The pseudo-image construction module for the spatial distribution of driving risks combines the pseudo-images of the spatial distribution of driving risks at each time from t0-H to t0+F into a sequence of historical and future pseudo-images of the spatial distribution of driving risks according to their temporal relationship.

7. The spatiotemporally coupled intelligent vehicle risk assessment system considering motion uncertainties according to claim 1, characterized in that: 4.1 The spatiotemporal coupled driving risk assessment module processes the pseudo image of driving risk spatial distribution at each moment through CNN. The CNN consists of convolutional layers and pooling layers. The convolutional layers extract the driving risk spatial distribution characteristics in the pseudo image through convolutional kernels. The pooling layers perform secondary sampling on the output of the convolutional layers to ensure that the model can extract sufficient driving risk spatial dependencies. The pseudo-image of the spatial distribution of driving risk at each moment is processed by CNN into a semantic vector containing the characteristics of the spatial distribution of driving risk at that moment; the semantic vector at each moment is input into the LSTM part of the spatiotemporal coupled driving risk assessment module in a forward temporal sequence from t0-H to t0+F. 4.2 The spatiotemporal coupled driving risk assessment module uses LSTM to process the semantic vector of the spatial distribution characteristics of driving risk from t0-H to t0+F output by CNN to extract the temporal dependency and dynamic evolution trend between driving risks at each moment; the final output of LSTM is processed by a fully connected layer and a Softmax layer and then transformed into a spatiotemporally coupled vehicle driving risk level.

Images