CN115096305A - Intelligent driving automobile path planning system and method based on generation of countermeasure network and simulation learning - Google Patents
Intelligent driving automobile path planning system and method based on generation of countermeasure network and simulation learning Download PDFInfo
- Publication number
- CN115096305A CN115096305A CN202111119055.0A CN202111119055A CN115096305A CN 115096305 A CN115096305 A CN 115096305A CN 202111119055 A CN202111119055 A CN 202111119055A CN 115096305 A CN115096305 A CN 115096305A
- Authority
- CN
- China
- Prior art keywords
- track
- trajectory
- longitudinal
- transverse
- generator
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
- G01C21/20—Instruments for performing navigational calculations
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
- G01C21/26—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for navigation in a road network
- G01C21/34—Route searching; Route guidance
- G01C21/3407—Route searching; Route guidance specially adapted for specific applications
- G01C21/343—Calculating itineraries, i.e. routes leading from a starting point to a series of categorical destinations using a global route restraint, round trips, touristic trips
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention discloses an intelligent driving automobile path planning system and method based on generation of an confrontation network and simulation learning. The track point generation countermeasure network takes scene characteristics and random noise as input, takes the driving track of an experienced driver as a corresponding sample, and simulates and generates the transverse and longitudinal track point state of the end state of the experienced track; the track generation part uses the generated track horizontal and vertical end state and the current horizontal and vertical end state of the vehicle to fit the fifth-order polynomial of the horizontal and vertical tracks and carry out the combination of the horizontal and vertical tracks. The invention solves the problem that a single sample possibly corresponds to a plurality of data labels and simulation learning is difficult to train due to variable driving styles of drivers; in addition, horizontal and longitudinal tracks are separately fitted through a fifth-order polynomial, the difficulty of learning the whole driving track by simulating learning is reduced, and the smoothness of the generated track is ensured. Meanwhile, the potential risk of generating the track is avoided by adopting track evaluation.
Description
Technical Field
The invention belongs to the field of automatic driving of intelligent automobiles, and relates to an intelligent driving automobile path planning system and method based on generation of a countermeasure network and simulation learning.
Background
The automatic driving is used as a main research content in the field of intelligent traffic control, and various key technologies such as perception, prediction, planning and control are integrated. The technology of automatic driving has been rapidly developed in recent years. It has not only great potential economically but also great advantages in improving traffic efficiency and driving safety. The path planning is an indispensable technical module in an automatic driving system and has important significance for the research of the whole automatic driving vehicle. How to accurately avoid surrounding obstacles according to upper layer sensing and prediction results and carry out safe and efficient driving is a basic requirement on an automatic driving vehicle. A reliable automatic driving path planning algorithm can safely avoid surrounding obstacles in real time, has higher safety and comfort, and greatly improves the travel efficiency. Most of the existing planning algorithms are sampling and searching methods based on artificial rules. The general sampling method similar to the grid method is difficult to perform complete sampling, and only a better driving track can be sampled. The complete search method is difficult to consider the dynamic constraint of the automatic driving vehicle, and has higher requirements on the calculation power of a vehicle-mounted computer. Therefore, most autonomous driving researchers in academia and industry are focusing on more intelligent, safe and reliable path planning methods.
The simulation learning is realized in a machine learning mode, and a behavior model is trained by using expert experience and is mainly used for solving the control planning problem of a complex scene. Currently, more and more scholars apply this to the field of autopilot. A model for directly mapping the traffic environment and the experience track is trained by using experience data of an experience driver, so that a feasible intelligent track planning mode is formed. However, the behaviors of experienced drivers often have complex diversity, and the same environmental characteristics often have corresponding relations with a plurality of experience tracks, which also brings great challenges to common regression models.
Disclosure of Invention
In order to solve the problems, the invention designs an intelligent driving automobile path planning system and method based on a generation countermeasure network and simulation learning, and trains a trajectory planning model simulating various empirical drivers by using the generation countermeasure network, so that the driving safety and the efficiency of an automatic driving vehicle are improved, the diversity of planned trajectories is kept, and the driving trajectory of the automatic driving vehicle is more intelligent.
The invention provides an intelligent driving automobile path planning system based on generation of a countermeasure network and simulation learning.
The generation confrontation network module comprises a track point generator and a track point discriminator. The input of the track point generator is random noise z and scene characteristics y extracted by a scene dependency graph, and the output is the transverse and longitudinal end states of the track after the time t
And
the input of the track point discriminator is the scene characteristic y extracted by the scene dependency graph and the transverse and longitudinal end states generated by the randomly selected generator
Or the final state label of the original driving track of the experienced driver
And the output is the discrimination result True/False.
Further, the random noise z is Gaussian noise, generalAfter training, the driving styles of various experienced drivers can be mapped to a gaussian distribution. Multiple sampling in random noise z, the generator can generate different horizontal and vertical end states
Further, the scene characteristics y are derived from the environmental information O t While extracting the environmental information O t And end state label of driving track of experienced driver
Acquiring from the collected data set;
further, the specific training process for generating the confrontation network module comprises two parts. Training a discriminator by using real samples and generated data, and improving the capacity of a generator when the discriminator can correctly distinguish the samples and the generated data; and secondly, training the generator, namely, putting the track points of the pseudo-end state output by the generator into the discriminator, performing back propagation by using the error of a discrimination result to update the training generator, and further improving the capability of the discriminator when the generator cannot correctly distinguish samples and generate data. In the training process, the track point discriminator improves the capability of discriminating empirical driver data labels and pseudo data generated by the generator as much as possible; the trace point generator generates a vivid transverse and longitudinal end state as much as possible, so that the trace point generator is attempted to be deceived.
The loss function for generating the countermeasure network is:
min G max D loss＝E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))]
wherein G is a track point generator, D is a track point discriminator, and x is an end state track point
y is a scene characteristic, and z is random Gaussian noise; e x～p(x) (f (x)) is f (x) with a probability p (x), where f (x) logD (x | y); e z～p(z) (G (z)) is the expectation for G (z) at probability p (z), where G (z) is log (1-D (G (z | y))).
The corresponding track generation module comprises a track beam generation module and a track evaluation module;
further, the generation module of the track bundle comprises a transverse track generation module and a longitudinal track generation module;
further, the lateral trajectory generation module: using the current lateral state quantity of the vehicle
And the generated end state transverse state quantity
As a boundary condition, a fifth order polynomial l of l with respect to time t is established trajectory ：
Further, the coefficient of the lateral trajectory quintic polynomial solved by the boundary condition is:
further, the longitudinal trajectory generation module: using current longitudinal state quantities of the vehicle
And the generated final state longitudinal state quantity
As a boundary condition, a fifth order polynomial s of s with respect to time t is established trajectory ：
Further, the coefficient of the longitudinal trajectory quintic polynomial solved by the boundary condition is:
furthermore, a fifth-order polynomial l corresponding to the plurality of transverse tracks trajectory Fifth order polynomial s corresponding to longitudinal trajectory trajectory One-to-one correspondence is stored into track bundle<l trajectory ，s trajectory >；
Traversing all the cross and longitudinal trajectory quintic polynomial pairs in the trajectory bundle<l trajectory ，s trajectory >Traversing the transverse trajectory quintic polynomial l at time intervals Δ t trajectory All track points in
And longitudinal trajectory quintic polynomial s trajectory All track points in
Combining the transverse and longitudinal track points into track points
Finally, all track points are combined according to the time interval delta t to generate the final track project i And store in the track bundle<trajectory i >；
Further, the trajectory evaluation module is configured to perform online evaluation on all trajectories in the trajectory bundle: traversing track bundle<trajectory i >All tracks in the track bundle are subjected to value sorting by using a set value function R, the tracks which are possibly collided and the tracks which contain track points with transverse and longitudinal speeds not conforming to the physical characteristics of the vehicle are deleted, and finally the track reject with the optimal value is selected max To a downstream control module. When the tracks in the track bundle do not pass the evaluation, namely, no legal track exists, the process is carried out at the moment for ensuring the safetyThe driver takes over.
The invention provides an intelligent driving automobile path planning method based on generation of a countermeasure network and simulation learning.
The invention has the beneficial effects that:
(1) according to the method, the planning track of the automatic driving vehicle is generated by simulating learning, so that the planning track is more intelligent, and the path planning capability of the automatic driving vehicle in a complex scene is improved;
(2) the invention simulates the experienced drivers to generate the track points by generating the confrontation network, thereby keeping the driving styles of various experienced drivers and increasing the diversity of generated tracks. Meanwhile, the problem that a single sample possibly corresponds to a plurality of data labels and simulation learning is difficult to train due to the fact that the driving styles of drivers are variable is solved;
(3) the method separately fits the transverse and longitudinal tracks through the quintic polynomial, reduces the difficulty of learning the whole driving track by simulating learning, and also ensures the smoothness of the generated track. Meanwhile, the corresponding track evaluation method avoids the potential risk of generating the track.
Drawings
FIG. 1 is a flow chart of a path planning method based on generation of countermeasure networks and mock learning;
FIG. 2 is a corresponding diagram of a generated countermeasure network architecture;
FIG. 3 is a flow chart of the corresponding transverse and longitudinal trajectory generation;
Detailed Description
The implementation of the invention comprises two parts of designing a track point generation countermeasure network and a corresponding track generation method. The track point generation countermeasure network takes scene characteristics and random noise as input, takes the driving track of an experienced driver as a corresponding sample, and simulates and generates the transverse and longitudinal track point state of the end state of the experienced track; and the track generation method part is used for fitting a fifth-order polynomial of the transverse and longitudinal tracks by utilizing the generated transverse and longitudinal end states of the tracks and the current transverse and longitudinal end states of the vehicle and carrying out transverse and longitudinal track combination.
The invention will be further explained with reference to the drawings.
Fig. 1 is a flow chart of an intelligent driving automobile path planning method based on generation of a countermeasure network and simulation learning, and the specific inventive method flow specifically includes the following steps:
1) through various styles of manual driving modes of experienced drivers, the vehicle-mounted camera, the laser radar and the GPS navigator are used for collecting the environmental information O at each moment in the driving process t And simultaneously acquiring the transverse track point state of the vehicle after the corresponding time t
And longitudinal track point state
2) Processing the acquired environmental information to construct a planning dependence graph, wherein the graph comprises various factors influencing path planning, including various lane lines, barrier vehicles, pedestrians and traffic lights;
3) training to generate an antagonistic network, inputting Gaussian noise z and scene characteristics, and generating an object as a transverse track point state after t time in each scene
And longitudinal track point state
4) Utilizing a generated confrontation network after training, utilizing Gaussian noise to uniformly sample the transverse end state after t moment, and generating a transverse sampling point
And the current lateral state of the vehicle
Generation of a fifth order polynomial l of the transverse trajectory trajectory (ii) a The longitudinal end state after the time t is uniformly sampled,and according to the generated longitudinal sampling point
And the current longitudinal state of the vehicle
Generating a longitudinal trajectory quintic polynomial s trajectory (ii) a Finally, the generated transverse tracks and the generated longitudinal tracks are stored into the track bundle one by one<l trajectory ，s trajectory >；
5) Merging transverse tracks and longitudinal tracks: traversing all transverse and longitudinal track pairs in the track bundle<l trajectory ，s trajectory >Traversing the transverse trajectory l at time intervals Δ t trajectory All track points in
And a longitudinal trajectory s trajectory All track points in
Combining the transverse and longitudinal track points into track points
Finally, all track points are combined according to the time interval delta t to generate the final track project i And store in the track bundle<trajectory i >；
6) And (3) performing value sorting on all tracks in the track bundle: traversing track bundle<trajectory i >Sorting all the tracks in the track bundle by using a set value function R;
the cost function R includes the following parts, k 1～5 The scaling factor for each part of cost:
R＝k 1 cost speed +k 2 cost jerk +k 3 cost lateral +k 4 cost comfort +k 5 cost var
whereincost speed For speed penalty, the goal is to keep the vehicle speed at the target vehicle speed, v target To a desired target vehicle speed, t total The number of points, v, corresponding to the track in time units t Vehicle speed at the t-th time point:
wherein cost jerk For the longitudinal comfort penalty, the goal is to keep the longitudinal jerk, j _ longitudinal, small t For longitudinal jerk at each time point:
wherein cost lateral At the expense of lateral deviation, the goal is to keep the lateral deviation from the reference line small,/ t For the lateral deviation of each time point from the reference line:
wherein cost comfort For the lateral comfort penalty, the goal is to keep the lateral jerk, j _ average, small t For lateral jerk at each time point:
wherein cost var The goal is to reduce the rate of change between the last frame trajectory and the current frame, at the cost of trajectory variation,
for the lateral displacement at the instant t of the current trajectory,
the lateral displacement at the moment of the last frame track t +1,
is the longitudinal displacement of the current trajectory at the time t,
for the longitudinal displacement at the time of the last frame trajectory t + 1:
7) sequentially traversing and evaluating the tracks in the track bundle according to the value sequence, judging whether collision is possible and whether the transverse and longitudinal speeds of all track points in the tracks accord with the dynamic characteristics of the vehicle or not, and selecting the optimal legal track project max ；
8) When the optimal legal track is picked max Sending it to a downstream control module; when none of the trajectories in the trajectory bundle passes the collision and dynamics evaluation, the driver takes over.
Fig. 2 is a corresponding diagram for generating a confrontation network structure, and the specific structure includes a track point generator and a track point discriminator.
Wherein the input of the track point generator is random Gaussian noise z and scene characteristic y extracted by a scene dependency graph, and the output is the transverse and longitudinal end state of the track after the time t
And
the input of the track point discriminator is the scene characteristic y extracted by the scene dependency graph and the transverse and longitudinal end states generated by the randomly selected generator
Or original empirical driver label
And the output is the discrimination result True/False.
The driving styles of various experienced drivers are diverse, and the driving routes of the same driver in different emotions for the same scene are different. Such diversity characteristics are hidden in random gaussian noise z. By sampling z, the horizontal and vertical end states adopted under the same scene characteristic y are different. Further, the generated trajectory also has diversity.
Randomly generating the horizontal and vertical end states generated by the generator
And transverse and longitudinal end state labels acquired when experienced drivers drive
And sending the data to a track point discriminator for discrimination, and judging whether the current transverse and longitudinal end state is a transverse and longitudinal end state label acquired when the experienced driver drives.
In the training process, the track point discriminator improves the capability of discriminating empirical driver data labels and pseudo data generated by the generator as much as possible; the trace point generator generates a vivid transverse and longitudinal end state as much as possible, so that the trace point generator is attempted to be deceived. The specific training process comprises two parts, namely training of a discriminator by using real samples and generated data; and secondly, training the generator, namely, putting the track points of the pseudo end state output by the generator into the discriminator, and performing back propagation by using the error of the discrimination result to update the training generator. Thus, the loss function of the entire network consists of two parts, specifically as follows:
min G max D loss＝E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))]
wherein G is a track point generator, D is a track point discriminator, and x is an end state track point
y is a scene characteristic, and z is random Gaussian noise; e x～p(x) (f (x)) is f (x) expectation under probability p (x), f (x) logD (x | y); e z～p(z) (G (z)) is the expectation for G (z) at probability p (z), G (z) log (1-D (G (z | y))).
Fig. 3 is a flowchart of the corresponding transverse and longitudinal trajectory generation, which specifically includes the following steps:
1) extracting the surrounding environment information in the form of pictures, and sending the extracted feature y into a generation countermeasure network;
2) a generator for generating a countermeasure network, and generating end state track points according to the samples uniformly sampled by z in Gaussian noise and the current traffic scene characteristics y
3) Several end state trace points generated by a generator
And generating transverse and longitudinal tracks. For the transverse track, the transverse state quantity of the current vehicle is utilized
And the generated end state transverse state quantity
As a boundary condition, there is a fifth order polynomial l of l with respect to time t trajectory Wherein t is 0 As starting time points:
the boundary conditions are as follows:
the fifth order polynomial and boundary conditions according to the transverse trajectory are:
based on the obtained a 1 a 2 a 3 a 4 a 5 Obtaining a fifth-order polynomial l of the transverse trajectory trajectory . From a plurality of hidden states uniformly sampled in the gaussian distribution z, a plurality of different hidden states can be generated
Further, a plurality of different transverse trajectory quintic polynomials l can be generated trajectory ；
For the longitudinal track, the longitudinal state quantity of the current vehicle is utilized
And the generated final state longitudinal state quantity
As a boundary condition, there is a fifth order polynomial s of s with respect to time t trajectory ：
The boundary conditions are as follows:
the fifth order polynomial and boundary conditions according to the longitudinal trajectory are:
based on the obtained b 1 b 2 b 3 b 4 b 5 Longitudinal track quintic polynomials trajectory . From a plurality of hidden states uniformly sampled in the gaussian distribution z, a plurality of different hidden states can be generated
Further, a plurality of different longitudinal trajectory quintic polynomials s may be generated trajectory ；
Fifthly polynomial l of a plurality of transverse tracks to be generated trajectory And longitudinal trajectory quintic polynomial s trajectory One-to-one correspondence is stored into track bundle<l trajectory ，s trajectory >。
The above-listed detailed description is only a specific description of a possible embodiment of the present invention, and it is not intended to limit the scope of the present invention, and equivalents and modifications not departing from the technical spirit of the present invention should be included in the scope of the present invention.
Claims (10)
1. An intelligent driving automobile path planning system based on generation of a countermeasure network and simulation learning is characterized by comprising a generation countermeasure network module and a corresponding track generation module;
the generation confrontation network module comprises a track point generator and a track point discriminator; the input of the track point generator is random noise z and scene characteristics y extracted by a scene dependency graph, and the output is the transverse and longitudinal end states of the track after the time t
And
the input of the track point discriminator is the scene characteristic y extracted from the scene dependency graph and the transverse and longitudinal end states generated by the randomly selected generator
Or the final state label of the original experienced driver's driving track
Outputting the result as a judgment result True/False;
the corresponding track generation module comprises a track beam generation module and a track evaluation module; the track beam generation module comprises a transverse track generation module and a longitudinal track generation module;
the transverse trajectory generation module: using the current lateral state quantity of the vehicle
And the generated end state transverse state quantity
As a boundary condition, a fifth order polynomial l is established for the time t trajectory ：
The longitudinal trajectory generation module: using current longitudinal state quantities of the vehicle
And the generated final state longitudinal state quantity
As a boundary condition, a fifth order polynomial s of s with respect to time t is established trajectory ：
Fifth order polynomial l of the transverse trajectory trajectory And the fifth order polynomial s of the longitudinal trajectory trajectory One-to-one correspondence is stored into track bundle<l trajectory ,s trajectory >(ii) a Traverse the track bundleAll pairs of transverse and longitudinal tracks in the same<l trajectory ,s trajectory >Fifth order polynomial l traversing the transverse trajectory at time intervals Δ t trajectory All track points in
And the fifth order polynomial s of the longitudinal trajectory trajectory All track points in
Combining the transverse and longitudinal track points into track points
Finally, all track points are combined according to the time interval delta t to generate the final track project i And store in the track bundle<trajectory i >；
The trajectory evaluation module: traversing track bundle<trajectory i >All tracks in the track bundle are subjected to value sorting by using a set value function R, the tracks which are possibly collided and the tracks which contain track points with transverse and longitudinal speeds not conforming to the physical characteristics of the vehicle are deleted, and finally the track reject with the optimal value is selected max And sending the trajectory to a downstream control module, and taking over by the driver when none of the trajectories in the trajectory bundle passes the evaluation, namely, no legal trajectory exists.
2. The intelligent driving automobile path planning system based on generation of the countermeasure network and simulation learning of claim 1, wherein the random noise z is gaussian noise, and the driving styles of various experienced drivers can be mapped to gaussian distribution through training; multiple samples are taken in random noise z and the generator can generate different transverse and longitudinal end states
3. The system as claimed in claim 1, wherein the scene characteristics y are derived from environmental information O t While extracting the environmental information O t And end state label of driving track of experienced driver
Is obtained from the collected data set.
4. The system of claim 1, wherein the training of the generate confrontation network module comprises two parts: the method comprises the following steps that firstly, training of a discriminator is carried out, real samples and generated data are used for training, and when the discriminator can correctly distinguish the samples and the generated data, the capacity of a generator is improved; secondly, training the generator, namely, putting the track points of the pseudo-end state output by the generator into the discriminator, performing back propagation by using the error of a discrimination result to update the training generator, and further improving the capability of the discriminator when the generator cannot correctly distinguish samples and generate data;
in the training process, the track point discriminator improves the capability of discriminating the empirical driver data label and the pseudo data generated by the generator as much as possible; the track point generator generates vivid transverse and longitudinal end states as much as possible, so as to attempt to deceive the track point generator;
the loss function for generating the countermeasure network is:
min G max D loss＝E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))]
wherein G is a track point generator, D is a track point discriminator, and x is an end state track point
y is the scene feature and z is random gaussian noise.
5. The intelligent driving automobile path planning system based on generation of the countermeasure network and the mimic learning of claim 1, wherein the coefficients of solving the lateral trajectory quintic polynomial and the longitudinal trajectory quintic polynomial respectively by using boundary conditions are as follows:
6. an intelligent driving automobile path planning method based on generation of a countermeasure network and simulation learning is characterized by comprising the following steps:
s1 collecting environmental information O at each moment in the driving process by vehicle-mounted camera, laser radar and GPS navigator in various styles of manual driving modes of experienced drivers t And simultaneously acquiring the transverse track point state of the vehicle after the corresponding time t
And longitudinal track point state
S2, processing the collected environment information to construct a planning dependency graph, wherein the graph comprises various factors influencing the path planning, including various lane lines, barrier vehicles, pedestrians and traffic lights;
s3 training and generating a confrontation network, inputting Gaussian noise z and scene characteristics, and generating an object as a transverse track point state after t time in each scene
And longitudinal track point state
S4, using the trained generated countermeasure network, using Gaussian noise to uniformly sample the transverse end state after t time, and according to the generated transverse sampling point
And the current lateral state of the vehicle
Generation of a fifth order polynomial l of the transverse trajectory trajectory (ii) a Uniformly sampling the longitudinal end state after t time, and generating longitudinal sampling points
And the current longitudinal state of the vehicle
Generating a longitudinal trajectory quintic polynomial s trajectory (ii) a Finally, the generated transverse tracks and the generated longitudinal tracks are stored into the track bundle one by one<l trajectory ,s trajectory >；
S5 merging transverse and longitudinal tracks: traversing all transverse and longitudinal track pairs in the track bundle<l trajectory ,s trajectory >Traversing the transverse trajectory quintic polynomial l at time intervals Δ t trajectory All track points in
And longitudinal trajectory quintic polynomial s trajectory All track points in
Combining the transverse and longitudinal track points into track points
Finally, all track points are combined according to the time interval delta t to generate the final track project i And store in the track bundle<trajectory i >；
S6 value ranks all tracks in the track bundle: traversing track bundle<trajectory i >Sorting all the tracks in the track bundle by using a set value function R;
s7 traversing and evaluating the track in the track bundle according to the order of value, judging whether collision is possible and whether the transverse and longitudinal speeds of all track points in the track accord with the dynamic characteristics of the vehicle, until selecting the optimal legal track project max ；
S8 when selecting the optimal legal track reject max Sending it to a downstream control module; when none of the trajectories in the trajectory bundle passes the collision and dynamics evaluation, the driver takes over.
7. The intelligent driving automobile path planning method based on generation of countermeasure network and simulation learning of claim 6, wherein in S3, the generation of countermeasure network comprises a track point generator and a track point discriminator;
wherein the input of the track point generator is random Gaussian noise z and scene characteristic y extracted by a scene dependency graph, and the output is the transverse and longitudinal end state of the track after the time t
And
the input of the track point discriminator is the scene characteristic y extracted by the scene dependency graph and the transverse and longitudinal end states generated by the randomly selected generator
Or original experienced driver label
Outputting the result as a judgment result True/False;
randomly generating transverse and longitudinal end states generated by a generator
And transverse and longitudinal end state labels acquired when experienced drivers drive
And sending the data to a track point discriminator for discrimination, and judging whether the current transverse and longitudinal end state is a transverse and longitudinal end state label acquired when the experienced driver drives.
8. An intelligent driving automobile path planning method based on generation of a countermeasure network and imitation learning according to claim 7, characterized in that in the process of training of the generation of the countermeasure network, a track point discriminator improves the capability of discriminating experienced driver data labels and pseudo data generated by a generator as much as possible; the track point generator generates a vivid transverse and longitudinal end state as much as possible so as to attempt to deceive the track point generator; the specific training process comprises two parts, namely training of a discriminator by using real samples and generated data; secondly, training the generator, namely, putting the track points of the pseudo-end state output by the generator into a discriminator, and performing back propagation by using the error of a discrimination result to update the training generator; the loss function of the whole network comprises two parts, specifically as follows:
min G max D loss＝E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))]
wherein G is a track point generator, D is a track point discriminator, and x is an end state track point
y is the scene feature and z is random gaussian noise.
9. The intelligent driving vehicle path planning method based on generation of countermeasure network and simulation learning of claim 6, wherein the lateral trajectory quintic polynomial l is generated in S4 trajectory And longitudinal trajectory quintic polynomial s trajectory The method comprises the following steps:
s4.1, extracting the surrounding environment information in the form of pictures, and sending the feature y to a generation countermeasure network;
s4.2, a generator for generating the confrontation network generates end state track points according to the samples uniformly sampled by z in Gaussian noise and the current traffic scene characteristics y
S4.3 utilizing the generator to generate a plurality of end state track points
Generating a transverse track and a longitudinal track:
for the transverse track, the transverse state quantity of the current vehicle is utilized
And the generated end state transverse state quantity
As a boundary condition, there is a fifth order polynomial l of l with respect to time t trajectory ：
The boundary conditions are as follows:
the fifth order polynomial and boundary conditions according to the transverse trajectory are:
based on the obtained a 1 a 2 a 3 a 4 a 5 Obtaining a fifth-order polynomial l of the transverse trajectory trajectory . From a plurality of hidden states uniformly sampled in the gaussian distribution z, a plurality of different hidden states can be generated
Further, a plurality of different transverse trajectory quintic polynomials l may be generated trajectory ；
For the longitudinal track, the longitudinal state quantity of the current vehicle is utilized
And the generated final state longitudinal state quantity
As a boundary condition, there is a fifth order polynomial s of s with respect to time t trajectory ：
The boundary conditions are as follows:
the fifth order polynomial and boundary conditions according to the longitudinal trajectory are:
based on the obtained b 1 b 2 b 3 b 4 b 5 Obtaining the fifth-order polynomial s of the longitudinal track trajectory From a plurality of hidden states sampled uniformly in the gaussian distribution z, a plurality of different states can be generated
Further, a plurality of different longitudinal trajectory quintic polynomials s may be generated trajectory ；
A fifth-order polynomial l of a plurality of generated transverse tracks trajectory And longitudinal trajectory quintic polynomial s trajectory One-to-one correspondence is stored into track bundle<l trajectory ,s trajectory >。
10. The intelligent driving vehicle path planning method based on generation of countermeasure network and imitation learning of claim 6, wherein the cost function R in S6 is designed as follows:
R＝k 1 cost speed +k 2 cost jerk +k 3 cost lateral +k 4 cost comfort +k 5 cost var wherein cost speed For speed penalty, the goal is to keep the vehicle speed at the target vehicle speed, v target To a desired target vehicle speed, v t Vehicle speed for each time point:
wherein cost jerk For the longitudinal comfort penalty, the goal is to keep the longitudinal jerk, j _ longitudinal, small t For longitudinal jerk at each time point:
wherein cost lateral At the cost of lateral deviationWith the aim of keeping small lateral deviations from the reference line,/ t For the lateral deviation of each time point from the reference line:
wherein cost comfort For lateral comfort penalty, the goal is to keep the lateral jerk, j _ average, small t For lateral jerk at each time point:
wherein cost var At the cost of trajectory variation, the goal is to reduce the rate of change between the previous frame trajectory and the current frame,
for the lateral displacement at the instant t of the current trajectory,
the lateral displacement at the moment of the last frame track t +1,
is the longitudinal displacement of the current trajectory at the moment t,
the longitudinal displacement of the last frame at the moment t +1 is as follows:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111119055.0A CN115096305A (en) | 2021-09-24 | 2021-09-24 | Intelligent driving automobile path planning system and method based on generation of countermeasure network and simulation learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111119055.0A CN115096305A (en) | 2021-09-24 | 2021-09-24 | Intelligent driving automobile path planning system and method based on generation of countermeasure network and simulation learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115096305A true CN115096305A (en) | 2022-09-23 |
Family
ID=83287646
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111119055.0A Pending CN115096305A (en) | 2021-09-24 | 2021-09-24 | Intelligent driving automobile path planning system and method based on generation of countermeasure network and simulation learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115096305A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117590856A (en) * | 2024-01-18 | 2024-02-23 | 北京航空航天大学 | Automatic driving method based on single scene and multiple scenes |
-
2021
- 2021-09-24 CN CN202111119055.0A patent/CN115096305A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117590856A (en) * | 2024-01-18 | 2024-02-23 | 北京航空航天大学 | Automatic driving method based on single scene and multiple scenes |
| CN117590856B (en) * | 2024-01-18 | 2024-03-26 | 北京航空航天大学 | Automatic driving method based on single scene and multiple scenes |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107169567B (en) | Method and device for generating decision network model for automatic vehicle driving | |
| CN113911129B (en) | Traffic vehicle intention identification method based on driving behavior generation mechanism | |
| CN113044064B (en) | Vehicle self-adaptive automatic driving decision method and system based on meta reinforcement learning | |
| CN107229973A (en) | The generation method and device of a kind of tactful network model for Vehicular automatic driving | |
| CN110956851B (en) | Intelligent networking automobile cooperative scheduling lane changing method | |
| CN108268483A (en) | The method for the grid map that generation controls for unmanned vehicle navigation | |
| CN111267830A (en) | Hybrid power bus energy management method, device and storage medium | |
| CN113239986B (en) | Training method and device for vehicle track evaluation network model and storage medium | |
| Valera et al. | Driving cycle and road grade on-board predictions for the optimal energy management in EV-PHEVs | |
| CN108268518A (en) | The device for the grid map that generation controls for unmanned vehicle navigation | |
| Zöldy et al. | Cognitive mobility–cogmob | |
| CN115257745A (en) | Automatic driving lane change decision control method based on rule fusion reinforcement learning | |
| CN113901718A (en) | Deep reinforcement learning-based driving collision avoidance optimization method in following state | |
| CN113722835A (en) | Modeling method for anthropomorphic random lane change driving behavior | |
| CN114852105A (en) | Method and system for planning track change of automatic driving vehicle | |
| CN115096305A (en) | Intelligent driving automobile path planning system and method based on generation of countermeasure network and simulation learning | |
| Koenig et al. | Bridging the gap between open loop tests and statistical validation for highly automated driving | |
| Li et al. | Driver behavior modelling at the urban intersection via canonical correlation analysis | |
| CN111368465A (en) | Unmanned decision-making method based on ID3 decision tree | |
| Wen et al. | Modeling human driver behaviors when following autonomous vehicles: An inverse reinforcement learning approach | |
| CN115935642A (en) | Intelligent vehicle extreme test scene automatic generation method and system based on accident information | |
| Baek et al. | Personalized speed planning algorithm using a statistical driver model in car-following situations | |
| CN107657106A (en) | traffic flow simulation method based on genetic algorithm | |
| CN114148349A (en) | Vehicle personalized following control method based on generation countermeasure simulation learning | |
| Teng et al. | Car Following Model Based on Driving Risk Field for Vehicle Infrastructure Cooperation |
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 |