CN114299607A - Human-vehicle collision risk degree analysis method based on automatic driving of vehicle - Google Patents
Human-vehicle collision risk degree analysis method based on automatic driving of vehicle Download PDFInfo
- Publication number
- CN114299607A CN114299607A CN202111517555.XA CN202111517555A CN114299607A CN 114299607 A CN114299607 A CN 114299607A CN 202111517555 A CN202111517555 A CN 202111517555A CN 114299607 A CN114299607 A CN 114299607A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- vehicle
- collision
- track
- determining
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention discloses a pedestrian and vehicle collision risk degree analysis method based on an automatic driving automobile, which comprises the steps of firstly dividing street crossing behaviors into different habit types through Gaussian clustering according to street crossing characteristics in a pedestrian street crossing historical data set, utilizing a joint probability distribution function to help an automatic driving vehicle to obtain a possible future state set according to the current motion state of a pedestrian, considering the dynamic space-time relation of the pedestrian and the vehicle through a GCN graph convolution neural network to obtain a possible future trajectory set of the pedestrian, finally considering the trajectory collision probability and the minimum meeting distance, utilizing an object element extension theory to perform characteristic dimension reduction, establishing a risk degree function, and realizing real-time judgment of the pedestrian and vehicle collision risk degree. The method focuses on the characteristic that the movement state of the pedestrian has sudden change, more particularly considers the influence of the space-time relation of the pedestrian and the vehicle on the track, improves the accuracy and the reliability of the collision judgment result of the pedestrian and the vehicle through the fusion of multiple indexes, further improves the intelligence of automatic driving, and improves the riding comfort and the safety.
Description
Technical Field
The invention belongs to the field of automatic driving decision-making algorithms, and particularly relates to a method for analyzing the collision risk degree of a pedestrian crossing a street based on an automatic driving vehicle.
Background
The rapid development of science and technology pushes the whole society to a more intelligent direction. In recent years, the technology of auto-driving automobiles has been rapidly developed, and companies such as google, tesla, and hundredth have made continuous attempts in this field. Automatic driving is closer to people's life, but meanwhile, people still have a great question about driving safety. The human-vehicle conflict judgment is a core module in the automatic driving technology. The accuracy of human-vehicle conflict assessment is improved, the calculation complexity is reduced, the safety of the automatic driving vehicle in driving is effectively ensured, and the accident prevention and handling capacity is enhanced.
At present, most of researches on human-vehicle collision early warning continue to adopt a deterministic model, namely that human and vehicles are considered to continuously move according to the current motion state, and collision risks are quantified through indexes such as collision time and the like. However, pedestrians are characterized by abrupt changes in speed and direction unlike automobiles, which is particularly apparent in games with vehicles. The deterministic model does not reflect the true state of the human-vehicle game most truly. And the judgment of the pedestrian-vehicle conflict is more reasonable based on the possible future track of the pedestrian. On the other hand, since the speed of the pedestrian can be greatly changed suddenly, the possible travel track, speed and direction of the pedestrian are all possible. However, in the existing human-vehicle conflict discrimination method considering the pedestrian track, a plurality of track sets possibly adopted by the pedestrian are not considered in the track prediction stage, and the discrimination of the human-vehicle conflict is further influenced by the inaccurate track prediction methods.
Disclosure of Invention
The invention aims to provide a pedestrian and vehicle collision risk degree judging method based on an automatic driving automobile, which can avoid the problems of low collision conflict prediction precision and incapability of predicting sudden change of street behavior which can be generated due to the fact that the traditional pedestrian track single-mode prediction is carried out, considers the future track set of pedestrians, constructs an interaction graph network of pedestrians and vehicles, improves the collision risk prediction precision, enriches the prediction scenes, provides a pedestrian and vehicle collision risk degree judging result with high reliability and accuracy, provides an accurate data base for an automatic driving vehicle decision module and a path planning module, further improves the intelligence of automatic driving, improves the riding comfort and the safety, and promotes the progress of an automatic driving technology.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a human-vehicle collision risk analysis method based on an automatic driving vehicle comprises the following steps:
step 1, acquiring street crossing characteristics in a pedestrian street crossing historical data set, wherein the street crossing characteristics comprise speed, acceleration and a street crossing course angle, and constructing a Gaussian mixture distribution function by using a Gaussian mixture clustering GMM method to express three different pedestrian street crossing habits;
step 2, acquiring the motion parameters of the pedestrian in front of the automatic driving vehicle, and matching the corresponding pedestrian crossing habits determined in the step 1 based on the motion characteristics of the pedestrian;
step 3, adding the current real moving speed of the pedestrianTaking the speed and the heading angle as centers, taking positive and negative sigma as a value interval, acquiring all possible speed, acceleration and heading angle sets and joint probabilities in the joint Gaussian distribution, and acquiring n through uniform resampling3A set of group state vectors;
step 4, preprocessing the pedestrian crossing motion state data acquired by the vehicle-mounted sensor, the motion state data of the automatic driving vehicle and the spatial position relation of the pedestrian and the vehicle pair, importing the preprocessed data into a GCN model, and training the structure weight and the offset parameter of the GCN model;
step 5, predicting a track set of the pedestrian under different data combinations through the trained GCN model, and determining a future track of the vehicle when the vehicle keeps the motion state unchanged;
step 6, determining the collision condition and the minimum meeting distance of the future track set of the pedestrian and the future track of the vehicle;
and 7, comprehensively considering the collision condition and the minimum meeting distance by using an object element extension theory to obtain the collision risk of the people and the vehicles.
Compared with the prior art, the invention has the remarkable advantages that:
(1) according to the technical scheme, the characteristic that the speed direction of the pedestrian can be suddenly changed is considered in the track prediction stage, a Gaussian mixture distribution function is constructed, three different pedestrian crossing habits are represented, and therefore a more real pedestrian future track set can be obtained by the method;
(2) the technical scheme of the invention adopts the GCN graph convolution neural network to predict the pedestrian track, which is different from the prior art, additionally and specifically considers the spatial relationship between the vehicles and the pedestrians, and establishes the graph network formed by the pedestrians and the vehicles, so that the method can extract more accurate space-time characteristics to further obtain more accurate future track; in addition, the method can also predict the trajectories of multiple pedestrians and multiple vehicles at the same time;
(3) according to the technical scheme, the collision risk and the shortest distance risk of the collision risk and the shortest distance risk are comprehensively considered in the collision risk estimation method, and the one-dimensional risk variable is obtained through the matter element extension algorithm, so that the collision risk estimation method is more specific and accurate compared with the prior art.
The invention is further described with reference to the following figures and detailed description.
Drawings
FIG. 1 is a flow chart of steps of a human-vehicle collision risk analysis method based on an automatic driving vehicle.
FIG. 2 is a schematic diagram of a pedestrian and vehicle relationship in an embodiment of the invention.
Fig. 3 is a schematic diagram of a two-dimensional extension set according to an embodiment of the present invention.
FIG. 4 is a one-dimensional risk function diagram according to an embodiment of the present invention.
Detailed Description
A human-vehicle collision risk analysis method based on an automatic driving vehicle comprises the following steps:
step 1, acquiring street crossing characteristics in a pedestrian street crossing historical data set, wherein the street crossing characteristics comprise speed, acceleration and a street crossing course angle, and constructing a Gaussian mixture distribution function by using a Gaussian mixture clustering GMM method to express three different pedestrian street crossing habits, specifically:
acquiring pedestrian crossing motion parameters with pedestrian and vehicle games through an existing data set, and extracting street crossing characteristics including speed, acceleration and street crossing course angle in pedestrian crossing historical data sets;
the pedestrian track course angle is as follows: the pedestrian track and the curb of the road form an included angle, and because China adopts a right-side driving system, the course angle is an acute angle when the track is consistent with the specified advancing direction, and the course angle is an obtuse angle when the track is opposite to the specified advancing direction.
The vehicle motion state information includes: the course angle θ of the vehicle, the traveling speed γ of the vehicle, and the traveling acceleration ω of the vehicle in the past seconds.
The vehicle course angle indicates: the vehicle track and the road edge of the road form an included angle, and because China adopts a right-side driving system, the course angle is an acute angle when the track is consistent with the specified advancing direction, and the course angle is an obtuse angle when the track is opposite to the specified advancing direction.
The human-vehicle spatial relationship information comprises: the absolute position of the autonomous vehicle and the absolute position of the pedestrian.
Step 1-1, constructing a probability density function:
wherein μ represents a 3-dimensional mean vector, ∑ represents a 3 × 3 covariance matrix determined by the heading angle, velocity, and acceleration of the pedestrian, and p (x) represents a probability density function of a random vector χ in a 3-dimensional sample space χ subject to gaussian distribution;
step 1-2, determining a Gaussian mixture distribution function according to 3 pedestrian street crossing behavior habits:
wherein, muiAnd ΣiIs a parameter of the ith Gaussian mixture component, k is 3, αi> 0 is the corresponding mixing coefficient,
step 1-3, obtaining an optimal Gaussian distribution parameter after iteration by using an EM algorithm, wherein the optimization solution of the EM algorithm firstly calculates the posterior probability (step E) of each sample belonging to each Gaussian component according to the current parameter in each iteration step, and updates the model parameter (step M) by using maximum likelihood estimation and a Larlang multiplier method, specifically:
step 1-3-1, initializing Gaussian parameter mui、∑iAnd alphaiFor each sample ZjDetermining the posterior probability that it belongs to each gaussian distribution:
step 1-3-2, determining each sample xjCluster mark of (2)j：
Step 1-3-3, marking lambda by clusterjPartitioning a sample set into 3 clusters C ═ C1,C2,C3}；
Step 1-3-4, constructing a Lagrange function through a maximum likelihood method to update Gaussian distribution parameters of each cluster: mu.si、∑iAnd alphaiFor m samples and k clusters, the lagrange function formula is:
the mean vector mu of each Gaussian distributioniThe update formula of (2) is:
the covariance matrix sigma of each Gaussian distributioniThe update formula of (2) is:
the mixing coefficient alpha of each Gaussian distributioniThe update formula of (2) is:
then, the current iteration is ended, and the newly acquired mu 'is obtained'i、∑′iAnd alpha'iAs an initial parameter for the next iteration until a set iteration stop requirement is met.
Step 2, obtaining the motion parameters of the pedestrian in front of the automatic driving vehicle, and matching the corresponding pedestrian crossing habits determined in the step 1 based on the motion characteristics of the pedestrian, wherein the method specifically comprises the following steps:
step 2-1, collecting the course angle and the historical track of the pedestrian and the current speed and acceleration parameters in real time by utilizing a laser radar, a camera and/or a millimeter wave radar sensor carried by an automatic driving automobile through information fusion of multiple sensors to form a state vector eta;
the camera carried by the automatic driving vehicle is used for collecting the posture and the head action of the pedestrian, and accordingly the course angle of the pedestrian is obtained.
The laser radar equipment carried by the automatic driving vehicle is used for acquiring the position and kinematic parameters of a pedestrian and obtaining the historical track, the current speed and the acceleration parameters of the pedestrian according to the position and kinematic parameters;
step 2-2, substituting the state vector eta of the pedestrian into Gaussian mixture distribution, and respectively calculating the posterior probability, p, of the vector eta belonging to three types of Gaussian distributionM(C＝1|η)、pM(C＝2|η)、pMAnd (C is 3| eta), the Gaussian distribution with the maximum posterior probability is the Gaussian distribution to which the current state vector of the pedestrian belongs, and the street crossing habit of the pedestrian is further determined.
Step 3, taking the actual motion speed, acceleration and course angle of the current pedestrian as the center, taking positive and negative sigma as a value interval, acquiring all possible speed, acceleration and course angle sets and joint probabilities in the joint Gaussian distribution, and acquiring n3 groups of state vector sets through uniform resampling, wherein the method specifically comprises the following steps:
taking a three-dimensional state vector eta formed by the real motion speed, the acceleration and the course angle of the current pedestrian as a center, and taking positive and negative sigma as a value interval, and acquiring all possible speed, acceleration and course angle sets and joint probabilities in the joint Gaussian distribution;
respectively averagely dividing the acquired possible speed, acceleration and course angle set interval and joint probability interval into n parts, and further determining n through cross matching3Sets of three-dimensional vectors that group the possible states of the pedestrian, and the probability of each set.
Step 4, preprocessing the pedestrian movement state data acquired by the vehicle-mounted sensor, the movement state data of the automatic driving vehicle and the spatial position relation of the pedestrian and the vehicle pair, importing the preprocessed data into a GCN (graph convolution neural network) model, and iteratively training the structural weight and the offset parameter of the GCN model for multiple times according to a gradient descent method;
the GCN model comprises a graph structure, a convolution layer, a pooling layer and a full-connection layer.
The graph structure refers to a directed graph formed by a space domain or a vertex domain and is used for extracting the space characteristics of the topological graph.
The convolution layer is a network layer formed by a plurality of convolution units through convolution kernel transformation, and the parameters of each convolution unit are obtained through optimization of a back propagation algorithm.
The pooling layer is a form of down-sampling which divides an input image into a plurality of rectangular areas and outputs a maximum value to each sub-area.
The full connection layer means that each node is connected with all nodes of the previous layer.
Step 5, predicting a track set of the pedestrian under different data combinations through the trained GCN model, and determining a future track of the vehicle when the vehicle keeps the motion state unchanged, wherein the method specifically comprises the following steps:
step 5-1, acquiring the speed, the acceleration and the position of the pedestrian crossing the street, the speed and the acceleration of the vehicle and the position of the vehicle of the current vehicle and the pedestrian crossing the street, obtaining a real-time track sequence of the pedestrian and the vehicle through data normalization operation, and combining the n acquired in the step 33One group in the group state vector set is respectively input into the trained GCN model, and n is output3Grouping the prediction data, and performing inverse normalization on the output prediction data to obtain n3The movement track of the pedestrian crossing the street in the future 2 s;
and 5-2, determining the track of the vehicle in the future 2s according to the vehicle dynamics.
Step 6, determining the collision condition and the minimum meeting distance of the pedestrian future track set and the vehicle future track, specifically:
step 6-1, determining a collision condition, considering a real outline of the vehicle, simulating pedestrians into a circle with the diameter of 0.5 meter, determining the collision condition of each future track of the pedestrians and the future track of the vehicle, and determining a comprehensive collision probability rho based on the probability of each track;
and 6-2, determining the minimum meeting distance, namely considering the real outline of the vehicle and simulating the pedestrian into a circle with the diameter of 0.5 meter, determining the minimum meeting distance y of each pedestrian future track which can not collide and the vehicle future track, and determining the weighted average minimum meeting distance y based on the probability of each track.
Step 7, comprehensively considering the collision situation and the minimum meeting distance by using an object element extension theory to obtain the collision risk of the people and the vehicles, which specifically comprises the following steps:
step 7-1, constructing a two-dimensional space consisting of rho axis and y axis, and respectively setting a rectangular classical domain phi1And setting the extension field phi2Determining a point in the two-dimensional space according to the weighted average minimum meeting distance y and the comprehensive collision probability rho determined in the step 6;
7-2, based on the points determined in the step 7-1, calculating an extension distance by using an extension theory to determine a correlation function value, wherein when the correlation degree is greater than 1, no conflict exists; when the correlation degree is between 0 and 1, a small conflict exists, and deceleration avoidance is adopted; when the degree of association is less than 0, there is a large conflict, and a parking avoidance measure should be taken.
A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the following steps when executing the computer program:
step 1, acquiring street crossing characteristics in a pedestrian street crossing historical data set, wherein the street crossing characteristics comprise speed, acceleration and a street crossing course angle, and constructing a Gaussian mixture distribution function by using a Gaussian mixture clustering GMM method to express three different pedestrian street crossing habits;
step 2, acquiring the motion parameters of the pedestrian in front of the automatic driving vehicle, and matching the corresponding pedestrian crossing habits determined in the step 1 based on the motion characteristics of the pedestrian;
step 3, taking the actual movement speed, the acceleration and the course angle of the current pedestrian as the center, taking positive and negative sigma as a value interval, and obtaining all the values in the combined Gaussian distributionPossible speed, acceleration and course angle set and joint probability, and obtaining n through uniform resampling3A set of group state vectors;
step 4, preprocessing the pedestrian crossing motion state data acquired by the vehicle-mounted sensor, the motion state data of the automatic driving vehicle and the spatial position relation of the pedestrian and the vehicle pair, importing the preprocessed data into a GCN model, and training the structure weight and the offset parameter of the GCN model;
step 5, predicting a track set of the pedestrian under different data combinations through the trained GCN model, and determining a future track of the vehicle when the vehicle keeps the motion state unchanged;
step 6, determining the collision condition and the minimum meeting distance of the future track set of the pedestrian and the future track of the vehicle;
and 7, comprehensively considering the collision condition and the minimum meeting distance by using an object element extension theory to obtain the collision risk of the people and the vehicles.
A computer-storable medium having stored thereon a computer program which, when executed by a processor, performs the steps of:
step 1, acquiring street crossing characteristics in a pedestrian street crossing historical data set, wherein the street crossing characteristics comprise speed, acceleration and a street crossing course angle, and constructing a Gaussian mixture distribution function by using a Gaussian mixture clustering GMM method to express three different pedestrian street crossing habits;
step 2, acquiring the motion parameters of the pedestrian in front of the automatic driving vehicle, and matching the corresponding pedestrian crossing habits determined in the step 1 based on the motion characteristics of the pedestrian;
step 3, taking the actual motion speed, acceleration and course angle of the current pedestrian as the center, taking positive and negative sigma as a value interval, acquiring all possible speed, acceleration and course angle sets and joint probabilities in the joint Gaussian distribution, and acquiring n through uniform resampling3A set of group state vectors;
step 4, preprocessing the pedestrian crossing motion state data acquired by the vehicle-mounted sensor, the motion state data of the automatic driving vehicle and the spatial position relation of the pedestrian and the vehicle pair, importing the preprocessed data into a GCN model, and training the structure weight and the offset parameter of the GCN model;
step 5, predicting a track set of the pedestrian under different data combinations through the trained GCN model, and determining a future track of the vehicle when the vehicle keeps the motion state unchanged;
step 6, determining the collision condition and the minimum meeting distance of the future track set of the pedestrian and the future track of the vehicle;
and 7, comprehensively considering the collision condition and the minimum meeting distance by using an object element extension theory to obtain the collision risk of the people and the vehicles.
The invention is further illustrated by the following examples and figures.
Examples
With reference to fig. 1, a method for analyzing the risk of collision between a person and a vehicle based on an autonomous vehicle includes the following steps:
step 1, acquiring street crossing characteristics in a pedestrian street crossing historical data set, wherein the street crossing characteristics comprise speed, acceleration and a street crossing course angle, and constructing a Gaussian mixture distribution function by using a Gaussian mixture clustering GMM method to express three different pedestrian street crossing habits, specifically:
acquiring pedestrian crossing motion parameters with pedestrian and vehicle games through an existing data set, and extracting street crossing characteristics including speed, acceleration and street crossing course angle in pedestrian crossing historical data sets;
the pedestrian track course angle is as follows: the pedestrian track and the curb of the road form an included angle, and because China adopts a right-side driving system, the course angle is an acute angle when the track is consistent with the specified advancing direction, and the course angle is an obtuse angle when the track is opposite to the specified advancing direction.
The vehicle motion state information includes: the course angle θ of the vehicle, the traveling speed γ of the vehicle, and the traveling acceleration ω of the vehicle in the past seconds.
The vehicle course angle indicates: the vehicle track and the road edge of the road form an included angle, and because China adopts a right-side driving system, the course angle is an acute angle when the track is consistent with the specified advancing direction, and the course angle is an obtuse angle when the track is opposite to the specified advancing direction.
The human-vehicle spatial relationship information comprises: the absolute position of the autonomous vehicle and the absolute position of the pedestrian.
Step 1-1, constructing a probability density function:
wherein μ represents a 3-dimensional mean vector, ∑ represents a 3 × 3 covariance matrix determined by the heading angle, velocity, and acceleration of the pedestrian, and p (x) represents a probability density function of a random vector χ in a 3-dimensional sample space χ subject to gaussian distribution;
step 1-2, determining a Gaussian mixture distribution function according to 3 pedestrian street crossing behavior habits:
wherein, muiAnd ΣiIs a parameter of the ith Gaussian mixture component, k is 3, αi> 0 is the corresponding mixing coefficient,
step 1-3, obtaining an optimal Gaussian distribution parameter after iteration by using an EM algorithm, wherein the optimization solution of the EM algorithm firstly calculates the posterior probability (step E) of each sample belonging to each Gaussian component according to the current parameter in each iteration step, and updates the model parameter (step M) by using maximum likelihood estimation and a Larlang multiplier method, specifically:
step 1-3-1, initializing Gaussian parameter mui、∑iAnd alphaiFor each sample ZjDetermining the posterior probability that it belongs to each gaussian distribution:
step 1-3-2, determinationEach sample xjCluster mark of (2)j：
Step 1-3-3, marking lambda by clusterjPartitioning a sample set into 3 clusters C ═ C1,C2,C3}；
Step 1-3-4, constructing a Lagrange function through a maximum likelihood method to update Gaussian distribution parameters of each cluster: mu.si、∑iAnd alphaiFor m samples and k clusters, the lagrange function formula is:
the mean vector mu of each Gaussian distributioniThe update formula of (2) is:
the covariance matrix sigma of each Gaussian distributioniThe update formula of (2) is:
the mixing coefficient alpha of each Gaussian distributioniThe update formula of (2) is:
then, the current iteration is ended, and the newly acquired mu 'is obtained'i、∑′iAnd alphai' as an initial parameter for the next iteration until a set iteration stop requirement is met.
With reference to fig. 2, a signal-free and zebra crossing section under free flow is selected as an implementation area, and for the zebra crossing under the control of the signal-free lights, the real track of pedestrians crossing the street and the track of vehicles playing games with the pedestrian can be obtained through video shooting. The method comprises the steps that an existing data set is preprocessed, so that pedestrian motion state information, vehicle motion state information and pedestrian-vehicle spatial relationship information in front of a vehicle are obtained; and extracting the festival features from the preprocessed data through Gaussian mixture clustering to obtain three different street-crossing habits and feature distributions thereof.
Step 2, obtaining the motion parameters of the pedestrian in front of the automatic driving vehicle, and matching the corresponding pedestrian crossing habits determined in the step 1 based on the motion characteristics of the pedestrian, wherein the method specifically comprises the following steps:
step 2-1, collecting the course angle and the historical track of the pedestrian and the current speed and acceleration parameters in real time by utilizing a laser radar, a camera and/or a millimeter wave radar sensor carried by an automatic driving automobile through information fusion of multiple sensors to form a state vector eta;
the camera carried by the automatic driving vehicle is used for collecting the posture and the head action of the pedestrian, and accordingly the course angle of the pedestrian is obtained.
The laser radar equipment carried by the automatic driving vehicle is used for acquiring the position and kinematic parameters of a pedestrian and obtaining the historical track, the current speed and the acceleration parameters of the pedestrian according to the position and kinematic parameters;
step 2-2, substituting the state vector eta of the pedestrian into Gaussian mixture distribution, and respectively calculating the posterior probability, p, of the vector eta belonging to three types of Gaussian distributionM(C＝1|η)、pM(C＝2|η)、pMAnd (C is 3| eta), the Gaussian distribution with the maximum posterior probability is the Gaussian distribution to which the current state vector of the pedestrian belongs, and the street crossing habit of the pedestrian is further determined.
Step 3, taking the actual motion speed, acceleration and course angle of the current pedestrian as the center, taking positive and negative sigma as a value interval, acquiring all possible speed, acceleration and course angle sets and joint probabilities in the joint Gaussian distribution, and acquiring n through uniform resampling3The group state vector set specifically includes:
taking a three-dimensional state vector eta formed by the real motion speed, the acceleration and the course angle of the current pedestrian as a center, and taking positive and negative sigma as a value interval, and acquiring all possible speed, acceleration and course angle sets and joint probabilities in the joint Gaussian distribution;
respectively averagely dividing the acquired possible speed, acceleration and course angle set interval and joint probability interval into n parts, and further determining n through cross matching3Sets of three-dimensional vectors that group the possible states of the pedestrian, and the probability of each set.
Step 4, preprocessing the pedestrian movement state data acquired by the vehicle-mounted sensor, the movement state data of the automatic driving vehicle and the spatial position relation of the pedestrian and the vehicle pair, importing the preprocessed data into a GCN (graph convolution neural network) model, and iteratively training the structural weight and the offset parameter of the GCN model for multiple times according to a gradient descent method;
the GCN model comprises a graph structure, a convolution layer, a pooling layer and a full-connection layer.
The graph structure refers to a directed graph formed by a space domain or a vertex domain and is used for extracting the space characteristics of the topological graph.
The convolution layer is a network layer formed by a plurality of convolution units through convolution kernel transformation, and the parameters of each convolution unit are obtained through optimization of a back propagation algorithm.
The pooling layer is a form of down-sampling which divides an input image into a plurality of rectangular areas and outputs a maximum value to each sub-area.
The full connection layer means that each node is connected with all nodes of the previous layer.
Step 5, predicting a track set of the pedestrian under different data combinations through the trained GCN model, and determining a future track of the vehicle when the vehicle keeps the motion state unchanged, wherein the method specifically comprises the following steps:
step 5-1, acquiring the speed, the acceleration and the position of the pedestrian crossing the street, the speed and the acceleration of the vehicle and the position of the vehicle of the current vehicle and the pedestrian crossing the street, obtaining a real-time track sequence of the pedestrian and the vehicle through data normalization operation, and combining the stepsN obtained in 33One group in the group state vector set is respectively input into the trained GCN model, and n is output3Grouping the prediction data, and performing inverse normalization on the output prediction data to obtain n3The movement track of the pedestrian crossing the street in the future 2 s;
and 5-2, determining the track of the vehicle in the future 2s according to the vehicle dynamics.
Step 6, determining the collision condition and the minimum meeting distance of the pedestrian future track set and the vehicle future track, specifically:
step 6-1, determining a collision condition, considering a real outline of the vehicle, simulating pedestrians into a circle with the diameter of 0.5 meter, determining the collision condition of each future track of the pedestrians and the future track of the vehicle, and determining a comprehensive collision probability rho based on the probability of each track;
and 6-2, determining the minimum meeting distance, namely considering the real outline of the vehicle and simulating the pedestrian into a circle with the diameter of 0.5 meter, determining the minimum meeting distance y of each pedestrian future track which can not collide and the vehicle future track, and determining the weighted average minimum meeting distance y based on the probability of each track.
Step 7, comprehensively considering the collision situation and the minimum meeting distance by using an object element extension theory to obtain the collision risk of the people and the vehicles, which specifically comprises the following steps:
step 7-1, constructing a two-dimensional space consisting of rho axis and y axis, and respectively setting a rectangular classical domain phi1And setting the extension field phi2Determines a point P3 in the two-dimensional space according to the weighted average minimum encounter distance y and the sum collision probability ρ determined in step 6, as shown in fig. 3;
7-2, calculating an extension distance by using an extension theory based on the points determined in the step 7-1 to determine a correlation function value, and achieving the purpose of reducing from two dimensions to one dimension as shown in fig. 4;
then the extension distance from point P3 to the classical domain and the extension domain is denoted as P (P)3，(P4，P1) P and P (P)3，(P5，P2))：
The degree of association K (S) is expressed as:
when the association degree is greater than 1, no conflict exists; when the correlation degree is between 0 and 1, a small conflict exists, and deceleration avoidance is adopted; when the degree of association is less than 0, there is a large conflict, and a parking avoidance measure should be taken.
In summary, the pedestrian and vehicle collision risk analysis method based on pedestrian track prediction and based on the automatic driving vehicle, provided by the invention, uses the high-precision detection equipment carried by the automatic driving vehicle to acquire the pedestrian motion data in front of the vehicle, predicts the pedestrian future track set, and combines the vehicle track and the pedestrian track set to comprehensively consider the safety and the comfort, so that more accurate judgment of the pedestrian and vehicle collision risk can be realized.
The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.
Claims (10)
1. A human-vehicle collision risk analysis method based on an automatic driving automobile is characterized by comprising the following steps:
step 1, acquiring street crossing characteristics in a pedestrian street crossing historical data set, wherein the street crossing characteristics comprise speed, acceleration and a street crossing course angle, and constructing a Gaussian mixture distribution function by using a Gaussian mixture clustering GMM method to express three different pedestrian street crossing habits;
step 2, acquiring the motion parameters of the pedestrian in front of the automatic driving vehicle, and matching the corresponding pedestrian crossing habits determined in the step 1 based on the motion characteristics of the pedestrian;
step 3, taking the actual motion speed, acceleration and course angle of the current pedestrian as the center, taking positive and negative sigma as a value interval, acquiring all possible speed, acceleration and course angle sets and joint probabilities in the joint Gaussian distribution, and acquiring n through uniform resampling3A set of group state vectors;
step 4, preprocessing the pedestrian crossing motion state data acquired by the vehicle-mounted sensor, the motion state data of the automatic driving vehicle and the spatial position relation of the pedestrian and the vehicle pair, importing the preprocessed data into a GCN model, and training the structure weight and the offset parameter of the GCN model;
step 5, predicting a track set of the pedestrian under different data combinations through the trained GCN model, and determining a future track of the vehicle when the vehicle keeps the motion state unchanged;
step 6, determining the collision condition and the minimum meeting distance of the future track set of the pedestrian and the future track of the vehicle;
and 7, comprehensively considering the collision condition and the minimum meeting distance by using an object element extension theory to obtain the collision risk of the people and the vehicles.
2. The method for analyzing the risk of collision between a pedestrian and a vehicle based on an autonomous vehicle as claimed in claim 1, wherein the step 1 of constructing a gaussian mixture distribution function represents three different pedestrian crossing habits, specifically:
step 1-1, constructing a probability density function:
wherein μ represents a 3-dimensional mean vector, ∑ represents a 3 × 3 covariance matrix determined by the heading angle, velocity, and acceleration of the pedestrian, and p (x) represents a probability density function of a random vector χ in a 3-dimensional sample space χ subject to gaussian distribution;
step 1-2, determining a Gaussian mixture distribution function according to 3 pedestrian street crossing behavior habits:
wherein, muiAnd ΣiIs a parameter of the ith Gaussian mixture component, k is 3, αi> 0 is the corresponding mixing coefficient,
and 1-3, acquiring the iterated optimal Gaussian distribution parameters by using an EM (effective electromagnetic radiation) algorithm.
3. The method for analyzing the risk of collision between a human and a vehicle based on an autonomous vehicle as claimed in claim 2, wherein the optimal gaussian distribution parameters after iteration, specifically the optimal gaussian distribution parameters after iteration, are obtained by using the EM algorithm in the steps 1 to 3
Step 1-3-1, initializing Gaussian parameter mui、∑iAnd alphaiFor each sample ZjDetermining the posterior probability that it belongs to each gaussian distribution:
step 1-3-2, determining each sample xjCluster mark of (2)j：
Step 1-3-3, marking lambda by clusterjPartitioning a sample set into 3 clusters C ═ C1,C2,C3}；
Step 1-3-4, constructing the Lagrange function by the maximum likelihood methodGaussian distribution parameters for the new clusters: mu.si、∑iAnd alphaiFor m samples and k clusters, the lagrange function formula is:
the mean vector mu of each Gaussian distributioniThe update formula of (2) is:
the covariance matrix sigma of each Gaussian distributioniThe update formula of (2) is:
the mixing coefficient alpha of each Gaussian distributioniThe update formula of (2) is:
then, the current iteration is ended, and the newly acquired mu 'is obtained'i、∑′iAnd alpha'iAs an initial parameter for the next iteration until a set iteration stop requirement is met.
4. The method for analyzing the risk of collision between a pedestrian and a vehicle based on an autonomous vehicle as claimed in claim 1, wherein the step 2 of matching the pedestrian crossing habits acquired in the step 1 by using the motion characteristics of the pedestrian specifically comprises:
step 2-1, collecting the course angle and the historical track of the pedestrian in real time, and the current speed and acceleration parameters to form a state vector eta;
step 2-2, setting the state of the pedestrianThe vector eta is substituted into Gaussian mixture distribution, and the posterior probability, p, of the vector eta belonging to three kinds of Gaussian distribution is respectively calculatedM(C＝1|η)、pM(C＝2|η)、pMAnd (C is 3| eta), the Gaussian distribution with the maximum posterior probability is the Gaussian distribution to which the current state vector of the pedestrian belongs, and the street crossing habit of the pedestrian is further determined.
5. The method for analyzing the risk of collision between a human and a vehicle based on an autonomous vehicle as claimed in claim 1, wherein the step 3 of obtaining the set of discrete state vectors specifically comprises:
taking a three-dimensional state vector eta formed by the real motion speed, the acceleration and the course angle of the current pedestrian as a center, and taking positive and negative sigma as a value interval, and acquiring all possible speed, acceleration and course angle sets and joint probabilities in the joint Gaussian distribution;
respectively averagely dividing the acquired possible speed, acceleration and course angle set interval and joint probability interval into n parts, and further determining n through cross matching3Sets of three-dimensional vectors that group the possible states of the pedestrian, and the probability of each set.
6. The method for analyzing the risk of collision between a pedestrian and a vehicle based on an autonomous vehicle as claimed in claim 1, wherein the determining the set of trajectories of the pedestrian and the trajectory of the vehicle in step 5 comprises:
step 5-1, acquiring the speed, the acceleration and the position of the pedestrian crossing the street, the speed and the acceleration of the vehicle and the position of the vehicle of the current vehicle and the pedestrian crossing the street, obtaining a real-time track sequence of the pedestrian and the vehicle through data normalization operation, and combining the n acquired in the step 33One group in the group state vector set is respectively input into the trained GCN model, and n is output3Grouping the prediction data, and performing inverse normalization on the output prediction data to obtain n3The movement track of the pedestrian crossing the street in the future 2 s;
and 5-2, determining the track of the vehicle in the future 2s according to the vehicle dynamics.
7. The method for analyzing the risk of collision between a human and a vehicle based on an autonomous vehicle as claimed in claim 1, wherein the step 6 of determining the collision situation and the minimum meeting distance comprises:
step 6-1, determining a collision condition, considering a real outline of the vehicle, simulating pedestrians into a circle with the diameter of 0.5 meter, determining the collision condition of each future track of the pedestrians and the future track of the vehicle, and determining a comprehensive collision probability rho based on the probability of each track;
and 6-2, determining the minimum meeting distance, namely considering the real outline of the vehicle and simulating the pedestrian into a circle with the diameter of 0.5 meter, determining the minimum meeting distance y of each pedestrian future track which can not collide and the vehicle future track, and determining the weighted average minimum meeting distance y based on the probability of each track.
8. The method for analyzing the human-vehicle collision risk based on the autonomous driving vehicle as claimed in claim 1, wherein the step 7 of obtaining the human-vehicle collision risk specifically comprises:
step 7-1, constructing a two-dimensional space consisting of rho axis and y axis, and respectively setting a rectangular classical domain phi1And setting the extension field phi2Determining a point in the two-dimensional space according to the weighted average minimum meeting distance y and the comprehensive collision probability rho determined in the step 6;
7-2, based on the points determined in the step 7-1, calculating an extension distance by using an extension theory to determine a correlation function value, wherein when the correlation degree is greater than 1, no conflict exists; when the correlation degree is between 0 and 1, a small conflict exists, and deceleration avoidance is adopted; when the degree of association is less than 0, there is a large conflict, and a parking avoidance measure should be taken.
9. A computer arrangement comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the steps of the method according to any of claims 1-8 are implemented by the processor when executing the computer program.
10. A computer-storable medium having a computer program stored thereon, wherein the computer program is adapted to carry out the steps of the method according to any one of claims 1-8 when executed by a processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111517555.XA CN114299607A (en) | 2021-12-13 | 2021-12-13 | Human-vehicle collision risk degree analysis method based on automatic driving of vehicle |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111517555.XA CN114299607A (en) | 2021-12-13 | 2021-12-13 | Human-vehicle collision risk degree analysis method based on automatic driving of vehicle |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114299607A true CN114299607A (en) | 2022-04-08 |
Family
ID=80968075
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111517555.XA Pending CN114299607A (en) | 2021-12-13 | 2021-12-13 | Human-vehicle collision risk degree analysis method based on automatic driving of vehicle |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114299607A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114644018A (en) * | 2022-05-06 | 2022-06-21 | 重庆大学 | Game theory-based man-vehicle interaction decision planning method for automatic driving vehicle |
| CN114802307A (en) * | 2022-05-23 | 2022-07-29 | 哈尔滨工业大学 | Intelligent vehicle transverse control method under automatic and manual hybrid driving scene |
| CN114863685A (en) * | 2022-07-06 | 2022-08-05 | 北京理工大学 | Traffic participant trajectory prediction method and system based on risk acceptance degree |
-
2021
- 2021-12-13 CN CN202111517555.XA patent/CN114299607A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114644018A (en) * | 2022-05-06 | 2022-06-21 | 重庆大学 | Game theory-based man-vehicle interaction decision planning method for automatic driving vehicle |
| CN114802307A (en) * | 2022-05-23 | 2022-07-29 | 哈尔滨工业大学 | Intelligent vehicle transverse control method under automatic and manual hybrid driving scene |
| CN114802307B (en) * | 2022-05-23 | 2023-05-05 | 哈尔滨工业大学 | Intelligent vehicle transverse control method under automatic and manual mixed driving scene |
| CN114863685A (en) * | 2022-07-06 | 2022-08-05 | 北京理工大学 | Traffic participant trajectory prediction method and system based on risk acceptance degree |
| CN114863685B (en) * | 2022-07-06 | 2022-09-27 | 北京理工大学 | Traffic participant trajectory prediction method and system based on risk acceptance degree |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Xin et al. | Intention-aware long horizon trajectory prediction of surrounding vehicles using dual LSTM networks | |
| US11537134B1 (en) | Generating environmental input encoding for training neural networks | |
| CN114299607A (en) | Human-vehicle collision risk degree analysis method based on automatic driving of vehicle | |
| CN112133089B (en) | Vehicle track prediction method, system and device based on surrounding environment and behavior intention | |
| Chen et al. | Driving maneuvers prediction based autonomous driving control by deep Monte Carlo tree search | |
| Jia et al. | Long short‐term memory and convolutional neural network for abnormal driving behaviour recognition | |
| CN106599827A (en) | Small target rapid detection method based on deep convolution neural network | |
| Wang et al. | A learning-based personalized driver model using bounded generalized Gaussian mixture models | |
| Kooij et al. | Analysis of pedestrian dynamics from a vehicle perspective | |
| CN112487954B (en) | Pedestrian crossing behavior prediction method for plane intersection | |
| US20210276547A1 (en) | Multi-agent trajectory prediction | |
| CN114446049B (en) | Traffic flow prediction method, system, terminal and medium based on social value orientation | |
| CN111027627A (en) | Vibration information terrain classification and identification method based on multilayer perceptron | |
| Zhang et al. | Multi-vehicle interaction scenarios generation with interpretable traffic primitives and gaussian process regression | |
| Kawasaki et al. | Multimodal trajectory predictions for autonomous driving without a detailed prior map | |
| Qing et al. | A novel particle filter implementation for a multiple-vehicle detection and tracking system using tail light segmentation | |
| Prasetyo et al. | Spatial Based Deep Learning Autonomous Wheel Robot Using CNN | |
| Zhou et al. | Spatiotemporal attention-based pedestrian trajectory prediction considering traffic-actor interaction | |
| CN116258058A (en) | Online generation method of key edge test scene for automatic driving acceleration test | |
| Neloy et al. | Alpha-N-V2: Shortest path finder automated delivery robot with obstacle detection and avoiding system | |
| Raj et al. | Lane Prediction by Autonomous Vehicle in Highway Traffic using Artificial Neural Networks | |
| CN114187577A (en) | System and method for predicting pedestrian crossing track | |
| Geng et al. | Dynamic-learning spatial-temporal Transformer network for vehicular trajectory prediction at urban intersections | |
| Tiwari et al. | Deep learning based lateral control system | |
| Li et al. | A probabilistic framework for estimating the risk of pedestrian-vehicle conflicts at intersections |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |