CN116311905A - Game theory-based signal-free right-hand fork pedestrian track prediction method - Google Patents

Game theory-based signal-free right-hand fork pedestrian track prediction method Download PDF

Info

Publication number
CN116311905A
CN116311905A CN202310048086.4A CN202310048086A CN116311905A CN 116311905 A CN116311905 A CN 116311905A CN 202310048086 A CN202310048086 A CN 202310048086A CN 116311905 A CN116311905 A CN 116311905A
Authority
CN
China
Prior art keywords
pedestrian
vehicle
game
track
function
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
Application number
CN202310048086.4A
Other languages
Chinese (zh)
Inventor
李文礼
唐远航
张祎楠
龚小豪
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University of Technology
Original Assignee
Chongqing University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Chongqing University of Technology filed Critical Chongqing University of Technology
Priority to CN202310048086.4A priority Critical patent/CN116311905A/en
Publication of CN116311905A publication Critical patent/CN116311905A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/042—Backward inferencing
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
    • G08G1/0125—Traffic data processing
    • G08G1/0129—Traffic data processing for creating historical data or processing based on historical data
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
    • G08G1/0137—Measuring and analyzing of parameters relative to traffic conditions for specific applications
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/09—Arrangements for giving variable traffic instructions
    • G08G1/0962—Arrangements for giving variable traffic instructions having an indicator mounted inside the vehicle, e.g. giving voice messages
    • G08G1/0967—Systems involving transmission of highway information, e.g. weather, speed limits
    • G08G1/096708—Systems involving transmission of highway information, e.g. weather, speed limits where the received information might be used to generate an automatic action on the vehicle control
    • G08G1/096725—Systems involving transmission of highway information, e.g. weather, speed limits where the received information might be used to generate an automatic action on the vehicle control where the received information generates an automatic action on the vehicle control
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/16—Anti-collision systems
    • G08G1/166—Anti-collision systems for active traffic, e.g. moving vehicles, pedestrians, bikes
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Chemical & Material Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Analytical Chemistry (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Atmospheric Sciences (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention belongs to the technical field of pedestrian track prediction, and particularly relates to a signal-free right-hand-off fork pedestrian track prediction method based on game theory, which comprises the following steps: s1, acquiring historical data of pedestrians and vehicles at a right-hand fork without signals; s2, analyzing the human-vehicle game factors of the signal-free right-turn intersection, and constructing a corresponding human-vehicle game model; s3, inserting the man-vehicle game model into a preset S-GAN model to obtain an SDG-GAN model for predicting the track of the pedestrian; s4, training the SDG-GAN model by using the historical data acquired in the S1; s5, predicting the pedestrian track of the right-hand fork without the signal in real time by using a trained SDG-GAN model. The method can ensure the prediction accuracy of the pedestrian track of the signal-free right-turn intersection and the effectiveness of the auxiliary driving decision of the signal-free right-turn intersection, thereby considering the efficiency and the safety of the vehicle passing through the signal-free right-turn intersection.

Description

Game theory-based signal-free right-hand fork pedestrian track prediction method
Technical Field
The invention belongs to the technical field of pedestrian track prediction, and particularly relates to a signal-free right-hand-off fork pedestrian track prediction method based on game theory.
Background
In the field of assisted driving, the passing efficiency and safety of vehicles passing through an intersection are very important considerations. At the crossing where the traffic light exists, the traffic light is indicated, and only the traffic light is indicated to be followed for automatic driving or auxiliary driving, so that the occurrence probability of collision between the human body and the vehicle is low.
However, at special intersections where traffic lights are not present, such as no-signal right-hand intersections, the correlation between vehicles and people becomes particularly complex, and both behaviors are full of uncertainty and contact, and both sides are often affected by a plurality of uncertainty factors, such as both-side character, environmental factors, and the like. When the vehicle passes through the right-hand crossing without the signal, the right steering is not controlled by traffic signals, and the random running of the right-hand motor vehicle and the interaction of traffic flows in different directions generate interference to pedestrians, so that the traffic danger of the pedestrians is greatly improved. For pedestrians, as the pedestrians pass through the right-hand fork without signals, the pedestrians have strong randomness and mobility, and the passing intentions among different pedestrians are obviously different. The subjective judgment is inaccurate only by the driver, and the human-vehicle interaction risk cannot be reduced.
Therefore, in order to achieve both the efficiency and the safety of the non-signal right-hand intersection during the auxiliary driving/automatic driving, accurate prediction of the trajectory of the pedestrian at the non-signal right-hand intersection is required, and only then can the effectiveness of the auxiliary driving decision at the non-signal right-hand intersection be ensured. At present, the track prediction of pedestrians is mainly divided into model-driven prediction and history-data-driven deep learning prediction. Model drivers include social force models, markov models, kalman filter models, and the like. Because the deep learning can better solve some defects of the model driven prediction method, the deep learning prediction based on historical data driving gradually becomes the main stream research direction. For example, the S-GAN model (i.e., social-GAN) for assisting driving adds the idea of interactive game of vehicles and people on the basis of deep learning prediction.
Gambling theory is an important branch of modern mathematics, and is a process for studying two parties or multiple objects as players to make decisions in a state of interaction. The advent of game theory provides a precise perspective for analyzing collaboration and competition status and mutual decision-making and losing. However, in the existing research on the collision between vehicles and persons at the traffic intersection, the use of the concept of the game of vehicles and persons is focused on the research on the collision safety, the traffic style, the interaction behavior and the like of vehicles and persons. The method for predicting the track of the pedestrian is almost not available for analyzing various risk factors in the collision of the human and the vehicle by adding the game idea. In an actual interaction scene, human-vehicle interaction is a dynamic change process, and motion factors and macroscopic game decisions are required to be combined. This results in a general need for improvement in the accuracy of existing studies, which is difficult to use for the actual driving assistance decision for the signalless right-hand fork.
In summary, how to ensure the accuracy of predicting the pedestrian track of the signal-free right-hand-over fork, thereby ensuring the effectiveness of the auxiliary driving decision of the signal-free right-hand-over fork, realizing the efficiency and the safety of the vehicle passing through the signal-free right-hand-over fork, and becoming the current problem to be solved urgently.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a method for predicting the pedestrian track of the right-hand-off intersection without signals based on the game theory, which can ensure the accuracy of predicting the pedestrian track of the right-hand-off intersection without signals and the effectiveness of auxiliary driving decision of the intersection without signals, thereby considering the efficiency and the safety of vehicles passing through the intersection without signals.
In order to solve the technical problems, the invention adopts the following technical scheme:
a method for predicting the track of a pedestrian with a right-hand fork without a signal based on game theory comprises the following steps:
s1, acquiring historical data of pedestrians and vehicles at a right-hand fork without signals;
s2, analyzing the human-vehicle game factors of the signal-free right-turn intersection, and constructing a corresponding human-vehicle game model;
s3, inserting the man-vehicle game model into a preset S-GAN model to obtain an SDG-GAN model for predicting the track of the pedestrian;
s4, training the SDG-GAN model by using the historical data acquired in the S1;
s5, predicting the pedestrian track of the right-hand fork without the signal in real time by using a trained SDG-GAN model.
Preferably, in S2, the process of building the man-car game model includes:
s21, dividing a game stage, designing an observation area and a conflict area, and recognizing that a game starts once pedestrians and vehicles enter the observation area; the dangerous degree of collision of the human and the vehicle is represented by the rear intrusion time, wherein the rear intrusion time is the time difference of the pedestrian and the vehicle entering the collision area, and the dangerous degree is higher when the rear intrusion time is shorter;
s22, constructing a man-vehicle game payment matrix based on decision strategies of pedestrians and vehicles in the game; the decision strategy comprises the steps of simultaneous passing of pedestrian vehicles, waiting for passing of pedestrians and waiting for passing of vehicles and simultaneous passing of pedestrians;
s23, obtaining an expected function and a corresponding loss function of the human-vehicle game model based on the characteristics of the human-vehicle game payment matrix;
s24, constructing a man-vehicle game model based on the expected function and the loss function obtained in the S23.
Preferably, in S22, when the pedestrian vehicles simultaneously pass,
the payment function of the pedestrian is:
Figure BDA0004056452490000021
the payment function of the vehicle is:
Figure BDA0004056452490000022
wherein v is v Indicating pedestrian passing speed, a v Representing pedestrian acceleration v p Indicating the passing speed of the vehicle, a p Indicating vehicle acceleration, alpha 1 Representing the common influencing factor of the speed and acceleration of the vehicle, alpha 2 Representing the co-influencing factor, sigma, of the speed and acceleration of a pedestrian v Representing a crash severity factor, sigma, of a vehicle p Represents a collision severity factor for a pedestrian, and:
Figure BDA0004056452490000031
Figure BDA0004056452490000032
when a pedestrian waits for the vehicle to pass, the payment function of the pedestrian is:
Figure BDA0004056452490000033
wherein alpha is 4 To wait for the suppression coefficient, t, to pass p Waiting time for pedestrians;
the payment function of the vehicle is:
Figure BDA0004056452490000034
wherein alpha is 3 Is the velocity excitation coefficient when passing;
when the pedestrian passes through the vehicle waiting, the payment function of the pedestrian is as follows:
Figure BDA0004056452490000035
wherein alpha is 3 V is the velocity excitation coefficient when passing p The pedestrian passing speed;
the payment function of the vehicle is:
Figure BDA0004056452490000036
wherein alpha is 4 To wait for the suppression coefficient, t, to pass v For the vehicle to be waiting for a time period,0.75s is the driver reaction time;
when the pedestrian and the vehicle wait simultaneously, the payment function of the pedestrian is that
Figure BDA0004056452490000037
The payment function of the vehicle is
Figure BDA0004056452490000038
Wherein alpha is 5 Wait suppression coefficient for the vehicle under common loss; alpha 6 Wait for the suppression coefficient of pedestrian under the common loss; k is a remorse factor of both parties, the remorse factor is related to starting acceleration in the waiting process, and represents remorse of the waiting strategy, and the larger the starting acceleration is, the larger the remorse degree is;
the man-vehicle game payment matrix is as follows:
Figure BDA0004056452490000039
preferably, in S23, the expected function is a mixed expected benefit of both the vehicle and the pedestrian when the vehicle passes through a nash balance point of a mixed dominant strategy of waiting for the vehicle and the pedestrian;
wherein the expected benefit when the vehicle selects to pass
Figure BDA00040564524900000310
The method comprises the following steps:
Figure BDA0004056452490000041
wherein (1)>
Figure BDA0004056452490000042
Representing the probability of passing a pedestrian, & lt & gt>
Figure BDA0004056452490000043
Probability of waiting for a pedestrian;
expected benefits when a vehicle selects waiting
Figure BDA0004056452490000044
The method comprises the following steps:
Figure BDA0004056452490000045
analyzing pure strategy profits of pedestrians, and expected profits of pedestrians passing through
Figure BDA0004056452490000046
The method comprises the following steps:
Figure BDA00040564524900000423
wherein (1)>
Figure BDA0004056452490000047
Representing the probability of a vehicle passing, < > or >>
Figure BDA0004056452490000048
Representing a probability of waiting for the vehicle;
expected benefits when pedestrians select waiting
Figure BDA0004056452490000049
The method comprises the following steps:
Figure BDA00040564524900000410
when the vehicle passes through the expected benefits
Figure BDA00040564524900000411
And wait for the expected benefit->
Figure BDA00040564524900000412
The same Nash equalization occurs and the probability combination of pedestrian passage and waiting is as follows:
Figure BDA00040564524900000413
Figure BDA00040564524900000414
when pedestrian passing expected income
Figure BDA00040564524900000415
And wait for the expected benefit->
Figure BDA00040564524900000416
The same Nash equalization occurs and the probability combination of vehicle passing and waiting is as follows:
Figure BDA00040564524900000417
Figure BDA00040564524900000418
preferably, when the SDG-GAN model predicts the track of the pedestrian, the track of the pedestrian is defined as two-dimensional coordinate position change under the time sequence, and the coordinate of the pedestrian u at the time t is defined as
Figure BDA00040564524900000419
The coordinates of vehicle j at time t are +.>
Figure BDA00040564524900000420
Historical track set X of each step length of pedestrians u from 1 to u The method comprises the following steps:
Figure BDA00040564524900000421
wherein, 1 to is the observation frame of the history track of the pedestrian, and to is the length of the observation frame;
pedestrians u from to+1 to t p Predicted track set for each step in
Figure BDA00040564524900000422
The method comprises the following steps:
Figure BDA0004056452490000051
in which, to+1 to to+t p Predicted frame for pedestrian history trajectory, t p To predict frame length.
Pedestrians u from to+1 to t p Real history track set Y of each step length u The method comprises the following steps:
Figure BDA0004056452490000052
pedestrian u is from 1 to t p Within the true history trace
Figure BDA0004056452490000053
And predictive generation track->
Figure BDA0004056452490000054
Respectively, [ X ] u ,Y u ]And
Figure BDA0004056452490000055
preferably, in S3, the SDG-GAN model includes a human-vehicle game model, a track generator, and a track discriminator; the track generator is used for encoding and decoding the output result of the human-vehicle game model and the historical track of the pedestrian and outputting the predicted track of the pedestrian; the trajectory discriminator is used for discriminating the probability that the predicted trajectory of the pedestrian is the true trajectory.
Preferably, the track generator comprises a track encoder, a gaming mechanism module, a pooling module, and a track decoder;
the track encoder is used for embedding the pedestrian coordinate position and the vehicle coordinate position at each time step into an embedded function containing a Relu nonlinear activation functionPhi, a fixed length vector is obtained
Figure BDA0004056452490000056
And->
Figure BDA0004056452490000057
Obtaining a pedestrian history track feature vector by LSTM unit coding>
Figure BDA0004056452490000058
And vehicle history trajectory feature vector->
Figure BDA0004056452490000059
Figure BDA00040564524900000510
Figure BDA00040564524900000511
In which the embedded function phi is a fully connected neural network layer,
Figure BDA00040564524900000512
weight parameters for the embedding function, +.>
Figure BDA00040564524900000513
The LSTM unit weight parameter is adopted;
the game mechanism module is used for extracting game related data of the two people and vehicles based on game payment functions of the two people and vehicles; the game related data comprises speed, acceleration, relative distance and waiting time; the method is also used for obtaining the post-intrusion time at the moment by utilizing the speeds of the two parties and the interval distance between the two parties and the conflict area at each time step; judging the interaction risk degree of the two parties under the time step by utilizing the post invasion time value; judging whether the two parts are in the observation area or the conflict area according to the position coordinates of the two parts at the moment;
the game mechanism module is also used for controlling the game mechanism module according to the trueThe interactive decisions of both sides of the person and the vehicle in the real world are calibrated, and specific expectations of both sides under the decision are obtained
Figure BDA00040564524900000514
And->
Figure BDA00040564524900000515
The specific expectation loss influences the historical track feature vectors of the two sides of the human and the vehicle to obtain the game feature vector of the two sides +.>
Figure BDA00040564524900000516
And fusing the characteristic vectors into the human-vehicle hybrid game characteristic vector +.>
Figure BDA00040564524900000517
Figure BDA00040564524900000518
In the method, in the process of the invention,
Figure BDA00040564524900000519
for the respective crash severity factor of pedestrian u and vehicle j +.>
Figure BDA00040564524900000520
For each real-time coordinate area of pedestrian u and vehicle j, f pet F is a collision degree determination function R For real-time zone determination function, t pet Is the post-invasion time omega of human-vehicle interaction pet Omega as a collision degree weighting parameter R The real-time regional weight parameter;
Figure BDA0004056452490000061
Figure BDA0004056452490000062
Figure BDA0004056452490000063
wherein f pay The function is calculated for the game and,
Figure BDA0004056452490000064
for specific payment functions under different strategies of both parties, f inf F is a game influencing function mix Is a game mixing function;
the pooling module is used for embedding the relative distance between the vehicles into the embedded function phi to obtain the relative position feature vector of the vehicles
Figure BDA0004056452490000065
Feature vector for hybrid game with human-vehicle>
Figure BDA0004056452490000066
Connected and finally output through a multi-layer perceptron to obtain game pooling vectors
Figure BDA0004056452490000067
Figure BDA0004056452490000068
Figure BDA0004056452490000069
Wherein Mlp GP For the pooling module multi-layer perceptron network layer, cat is used to connect feature vectors,
Figure BDA00040564524900000610
the weight parameters are corresponding to the respective network layers;
track decoder for pooling game into vector
Figure BDA00040564524900000611
Historical track characteristics of pedestriansVector->
Figure BDA00040564524900000612
Connecting and passing through decoder module multi-layer perceptron Mlp GD Connecting with the network random Gaussian noise Z to finally obtain the decoder feature vector +.>
Figure BDA00040564524900000613
Figure BDA00040564524900000614
Pedestrian coordinate vector encoded by embedded function>
Figure BDA00040564524900000615
Initializing the cell all zero vector->
Figure BDA00040564524900000616
Inputting the vectors into the LSTM neural network layer to obtain vectors>
Figure BDA00040564524900000617
Finally will->
Figure BDA00040564524900000618
Finally obtaining pedestrian track prediction coordinates through a decoder multi-layer perceptron>
Figure BDA00040564524900000619
Figure BDA00040564524900000620
Figure BDA00040564524900000621
Figure BDA00040564524900000622
Figure BDA00040564524900000623
In the method, in the process of the invention,
Figure BDA00040564524900000624
ω λ is a weight parameter of the corresponding network layer.
Preferably, in S3, the operation of the track identifier includes: predicted trajectory of pedestrian output by trajectory generator
Figure BDA00040564524900000625
And a true pedestrian history trajectory Y u Inputting the high-dimensional feature vectors into a discriminator neural network, and outputting the high-dimensional feature vectors by embedding the high-dimensional feature vectors into a fully-connected neural network layer>
Figure BDA0004056452490000071
Sequentially passing through an LSTM neural network layer and a discriminator multilayer perceptron, and finally outputting to obtain the probability of being a real track;
Figure BDA0004056452490000072
Figure BDA0004056452490000073
Figure BDA0004056452490000074
in the method, in the process of the invention,
Figure BDA0004056452490000075
output vector for LSTM network layer, +.>
Figure BDA0004056452490000076
Corresponding to the weight parameters of the respective network layer, P real To be distinguished as a summary of a real trackThe rate.
Preferably, in S3, the loss function L of the SDG-GAN model SDG - GAN Cross entropy loss function L comprising a generator and a discriminator GAN (G, D) minimum difference loss function L from pedestrian real track and pedestrian predicted track L2 (G)；
L SDG-GAN ＝L GAN (G,D)+L L2 (G)；
Figure BDA0004056452490000077
Figure BDA0004056452490000078
Wherein E is the expected value, r is the inputted real track data of the pedestrian, S is the sampling number based on the model generating result, Z is the inputted Gaussian noise distribution, D is the discriminator, and G is the generator.
Preferably, in S1, the history data includes: the number of frames of video, coordinates of pedestrians, accelerations of pedestrians, waiting time of pedestrians, coordinates of vehicles, accelerations of vehicles, waiting time of vehicles, distance of vehicles and post-intrusion time; the man-vehicle distance is the Euclidean distance between the pedestrian and the vehicle.
Compared with the prior art, the invention has the following beneficial effects:
1. according to the invention, on the basis of S-GAN, microscopic data and game ideas are utilized to analyze, so as to obtain a macroscopic interaction strategy of the human and vehicle without the signal right-hand fork at different moments; meanwhile, the human-vehicle interaction is divided into stages, including an observation area and a conflict area. While in the observation area, the two-party game pays the benefit and obtains its expected benefit through probability. When the two parties are in the conflict area, the existing decision is judged, the strategy benefits of the two parties are directly obtained, and more detailed analysis on the game process of the people and the vehicles is realized. And on the basis, an SDG-GAN model is constructed. And then, the SDG-GAN model is driven by the microscopic man-vehicle interaction data and macroscopic man-vehicle game strategy benefits to predict the track of the pedestrian. In such a way, the invention combines micro motion factors and macro game decisions in the actual interaction scene, adds the game idea into the human-vehicle interaction state, and can accurately predict the pedestrian track of the right-hand fork without signals.
In conclusion, the method and the device can ensure the prediction accuracy of the pedestrian track of the signal-free right-hand intersection and the effectiveness of the auxiliary driving decision of the signal-free right-hand intersection, so that the efficiency and the safety of vehicles passing through the signal-free right-hand intersection are considered.
2. The SDG-GAN algorithm constructed by the invention integrates macroscopic game and microscopic motion parameters, so that the relationship between the two people and the vehicle is more comprehensively considered, the true reflection of the descending people in the game state can be expressed, and the accuracy and the interpretability of predicting the pedestrian track are improved.
3. The invention analyzes the specific human-vehicle game factors in the special interaction scene of the signal-free right-hand-off fork, divides the game stage, distinguishes the dangerous degree of human-vehicle conflict by utilizing the time index after infringement, characterizes the remorse degree of both game sides by the remorse factor related to the starting acceleration, comprehensively considers various factors of both game sides as much as possible, and establishes the human-vehicle game payment matrix on the basis. And the Nash equilibrium idea of the complete information game is combined to be embedded into a deep learning network model, so that the prediction of the pedestrian track is realized. By the aid of the processing, the prediction accuracy of the pedestrian track of the right-hand fork without the signal can be guaranteed.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings, in which:
FIG. 1 is a flow chart in an embodiment;
FIG. 2 is a schematic diagram of a gaming scenario in an embodiment;
FIG. 3 is a schematic diagram of the structure of an SDG-GAN model in an embodiment;
fig. 4 is a schematic diagram of a data acquisition scenario in an embodiment.
Detailed Description
The following is a further detailed description of the embodiments:
examples:
as shown in fig. 1, the embodiment discloses a method for predicting a pedestrian track of a right-hand handover fork without signals based on game theory, which comprises the following steps:
s1, acquiring historical data of pedestrians and vehicles at a right turn intersection without signals.
In particular implementations, the history data includes: the number of frames of video, coordinates of pedestrians, accelerations of pedestrians, waiting time of pedestrians, coordinates of vehicles, accelerations of vehicles, waiting time of vehicles, distance of vehicles and post-intrusion time; the man-vehicle distance is the Euclidean distance between the pedestrian and the vehicle.
And establishing a man-vehicle data set according to the collected historical data:
(frame,id,x P ,y P ,v P ,a P ,w P ,x V ,yV,v V ,a V ,w V r, PET), wherein v, p in the right subscript refer to the number of frames of the video for the vehicle and pedestrian, frame, x/y represents coordinates, v represents speed, a represents acceleration, w represents waiting time, R represents man-car distance, PET represents post-intrusion time (i.e., the time difference between the pedestrian and vehicle reaching the conflict area), respectively.
S2, analyzing the human-vehicle game factors of the signal-free right-turn intersection, and constructing a corresponding human-vehicle game model.
Through actual scene observation and data acquisition, the vehicle is required to be decelerated and slowly driven in a specific interaction scene according to local traffic regulations, and the vehicle speed of the right-hand vehicle is slower than that of the straight-going vehicle, so that the danger feeling is not strong when the pedestrians pass through, the passing intention of most pedestrians is strong, and the time is not wasted. Through statistical analysis of pedestrian traffic behavior, pedestrians commonly select traffic due to no clear road right distribution in the right turn no-signal intersection, drivers commonly select speed reduction and yield, and delay time is obviously higher than pedestrian delay time due to failure of vehicle game.
The invention provides a man-vehicle game model, which is used for establishing a game formula, giving a macroscopic decision probability and specifically analyzing influence factors when pedestrians pass through.
And (5) defining a model. (1) Pedestrian and driver have rational thinking, and can select strategies according to self conditions. (2) And once the pedestrians and vehicles enter the judging area, the game is determined to start. (3) The pedestrian and the driver decide whether to statically game by obeying the complete information, and both sides decide at the same time. (4) When the pedestrian and the vehicle enter the conflict area at the same time, the collision of the pedestrian and the vehicle is determined.
And (5) establishing a model. (1) players: in the game process, two players are contained in total, i=1, 2, i=1 represents a pedestrian, and i=2 represents a vehicle. (2) strategy: in the process of the pedestrian and vehicle game, the pedestrian and the vehicle have two same strategies { passgae, wait }. (3) benefit: during the game, each player will obtain the effect of other players on itself and the result of its decision. Including delay loss when the player waits, loss in the event of a collision between two parties, revenue when one party waits for one to pass, and the like. The player benefits in different situations will be analyzed specifically.
The pedestrians and vehicles can generate four decision results in the game together, and the set is set as s n ，n∈[1,2,3,4]Corresponding to the simultaneous passing of the pedestrian vehicles, the waiting of the pedestrian passing vehicles and the simultaneous waiting of the pedestrian vehicles respectively. When a pedestrian and a vehicle enter the observation area simultaneously, the driver and the pedestrian start to analyze and select the strategy, and execute the strategy in the next step. Both parties in the game will eventually enter the conflict area, and the relative time difference between the two parties entering the conflict area is the post-intrusion time, which symbolizes the risk of the game, as shown in fig. 2.
Based on the analysis, in the implementation, in S2, the process of constructing the man-vehicle game model includes:
s21, dividing a game stage, designing an observation area and a conflict area, and recognizing that a game starts once pedestrians and vehicles enter the observation area; the dangerous degree of collision of the human and the vehicle is represented by the rear intrusion time, wherein the rear intrusion time is the time difference of the pedestrian and the vehicle entering the collision area, and the dangerous degree is higher when the rear intrusion time is shorter;
s22, constructing a man-vehicle game payment matrix based on decision strategies of pedestrians and vehicles in the game; the decision strategy comprises the steps of simultaneous passing of pedestrian vehicles, waiting for passing of pedestrians and waiting for passing of vehicles and simultaneous passing of pedestrians;
when the pedestrian vehicles are simultaneously passing by,
the payment function of the pedestrian is:
Figure BDA0004056452490000091
the payment function of the vehicle is:
Figure BDA0004056452490000092
wherein v is v Indicating pedestrian passing speed, a v Representing pedestrian acceleration v p Indicating the passing speed of the vehicle, a p Indicating vehicle acceleration, alpha 1 Representing the common influencing factor of the speed and acceleration of the vehicle, alpha 2 Representing the co-influencing factor, sigma, of the speed and acceleration of a pedestrian v Representing a crash severity factor, sigma, of a vehicle p Represents a collision severity factor for a pedestrian, and:
Figure BDA0004056452490000101
Figure BDA0004056452490000102
when a pedestrian waits for the vehicle to pass, the payment function of the pedestrian is:
Figure BDA0004056452490000103
wherein alpha is 4 To wait for the suppression coefficient, t, to pass p Waiting time for pedestrians;
the payment function of the vehicle is:
Figure BDA0004056452490000104
wherein alpha is 3 Is the velocity excitation coefficient when passing;
when the pedestrian passes through the vehicle waiting, the payment function of the pedestrian is as follows:
Figure BDA0004056452490000105
wherein alpha is 3 V is the velocity excitation coefficient when passing p The pedestrian passing speed;
the payment function of the vehicle is:
Figure BDA0004056452490000106
wherein alpha is 4 To wait for the suppression coefficient, t, to pass v For vehicle waiting time, 0.75s is driver reaction time;
when the pedestrian and the vehicle wait simultaneously, the payment function of the pedestrian is that
Figure BDA0004056452490000107
The payment function of the vehicle is
Figure BDA0004056452490000108
Wherein alpha is 5 Wait suppression coefficient for the vehicle under common loss; alpha 6 Wait for the suppression coefficient of pedestrian under the common loss; k is a remorse factor of both parties, the remorse factor is related to starting acceleration in the waiting process, and represents remorse of the waiting strategy, and the larger the starting acceleration is, the larger the remorse degree is;
the man-vehicle game payment matrix is as follows:
Figure BDA0004056452490000109
the two strategies of vehicle selection waiting by pedestrian selection and vehicle selection waiting by pedestrian selection are more reasonable as can be known from the human-vehicle game payment matrix. The two strategies are more dominant and are unstable in the evolution process, the Nash equilibrium principle is adopted to know that Nash equilibrium points are necessarily present if a mixed dominant strategy combination exists in a game, and the mixed expected benefits of both vehicles and pedestrians are known through calculation.
S23, obtaining an expected function and a corresponding loss function of the human-vehicle game model based on the characteristics of the human-vehicle game payment matrix;
in the specific implementation, the expected function is the mixed expected benefit of the vehicle and the pedestrian when the vehicle passes through a Nash equilibrium point of a mixed advantage strategy of waiting by the vehicle and the pedestrian passes through the vehicle;
wherein the expected benefit when the vehicle selects to pass
Figure BDA0004056452490000111
The method comprises the following steps:
Figure BDA0004056452490000112
wherein (1)>
Figure BDA0004056452490000113
Representing the probability of passing a pedestrian, & lt & gt>
Figure BDA0004056452490000114
Probability of waiting for a pedestrian;
expected benefits when a vehicle selects waiting
Figure BDA0004056452490000115
The method comprises the following steps:
Figure BDA0004056452490000116
analyzing pure strategy profits of pedestrians, and expected profits of pedestrians passing through
Figure BDA0004056452490000117
The method comprises the following steps:
Figure BDA0004056452490000118
wherein (1)>
Figure BDA0004056452490000119
Representing the probability of a vehicle passing, < > or >>
Figure BDA00040564524900001110
Representing a probability of waiting for the vehicle;
expected benefits when pedestrians select waiting
Figure BDA00040564524900001111
The method comprises the following steps:
Figure BDA00040564524900001112
when the vehicle passes through the expected benefits
Figure BDA00040564524900001113
And wait for the expected benefit->
Figure BDA00040564524900001114
The same Nash equalization occurs and the probability combination of pedestrian passage and waiting is as follows:
Figure BDA00040564524900001115
Figure BDA00040564524900001116
when pedestrian passing expected income
Figure BDA00040564524900001117
And wait for the expected benefit->
Figure BDA00040564524900001118
The same Nash equalization occurs and the probability combination of vehicle passing and waiting is as follows:
Figure BDA00040564524900001119
Figure BDA00040564524900001120
s24, constructing a man-vehicle game model based on the expected function and the loss function obtained in the S23.
S3, inserting the man-vehicle game model into a preset S-GAN model to obtain an SDG-GAN model for predicting the track of the pedestrian. The structure of the SDG-GAN model is shown in FIG. 3.
When the SDG-GAN model predicts the track of the pedestrian, the track of the pedestrian is defined as two-dimensional coordinate position change under the time sequence, and the coordinate of the pedestrian u at the t moment is defined as
Figure BDA0004056452490000121
The track coordinate of the vehicle j is the same as the pedestrian coordinate, and the coordinate of the vehicle j at the t moment is +.>
Figure BDA0004056452490000122
Historical track set X of each step length of pedestrians u from 1 to u The method comprises the following steps:
Figure BDA0004056452490000123
wherein, 1 to is the observation frame of the history track of the pedestrian, and to is the length of the observation frame;
pedestrians u from to+1 to t p Predicted track set for each step in
Figure BDA0004056452490000124
The method comprises the following steps:
Figure BDA0004056452490000125
in which, to+1 to to+t p Predicted frame for pedestrian history trajectory, t p To predict frame length.
Pedestrians u from to+1 to t p Real history track set Y of each step length u The method comprises the following steps:
Figure BDA0004056452490000126
pedestrian u is from 1 to t p Within the true history trace
Figure BDA0004056452490000127
And predictive generation track->
Figure BDA0004056452490000128
Respectively, [ X ] u ,Y u ]And
Figure BDA0004056452490000129
when the method is implemented, the SDG-GAN model comprises a man-vehicle game model, a track generator and a track discriminator; the track generator is used for encoding and decoding the output result of the human-vehicle game model and the historical track of the pedestrian and outputting the predicted track of the pedestrian; the trajectory discriminator is used for discriminating the probability that the predicted trajectory of the pedestrian is the true trajectory.
The track generator comprises a track encoder, a game mechanism module, a pooling module and a track decoder;
the track encoder is used for embedding the pedestrian coordinate position and the vehicle coordinate position at each time step into an embedding function phi containing a Relu nonlinear activation function to obtain a fixed length vector
Figure BDA00040564524900001210
And->
Figure BDA00040564524900001211
Obtaining a pedestrian history track feature vector by LSTM unit coding>
Figure BDA00040564524900001212
And vehicle history trajectory feature vector->
Figure BDA00040564524900001213
Figure BDA00040564524900001214
Figure BDA00040564524900001215
In which the embedded function phi is a fully connected neural network layer,
Figure BDA00040564524900001216
weight parameters for the embedding function, +.>
Figure BDA00040564524900001217
Is the LSTM unit weight parameter.
The game mechanism module is used for extracting game related data of the two people and vehicles based on game payment functions of the two people and vehicles; the game related data comprises speed, acceleration, relative distance and waiting time; the method is also used for obtaining the post-intrusion time at the moment by utilizing the speeds of the two parties and the interval distance between the two parties and the conflict area at each time step; judging the interaction risk degree of the two parties under the time step by utilizing the post invasion time value; judging whether the two parts are in the observation area or the conflict area according to the position coordinates of the two parts at the moment;
the game mechanism module is also used for calibrating according to the interactive decisions of both sides of the person and the vehicle in the real world, and obtaining the specific expected loss of both sides under the decisions
Figure BDA0004056452490000131
And->
Figure BDA0004056452490000132
The specific expectation loss influences the historical track feature vectors of the two sides of the human and the vehicle to obtain the game feature vector of the two sides +.>
Figure BDA0004056452490000133
And fusing the characteristic vectors into the human-vehicle hybrid game characteristic vector +.>
Figure BDA0004056452490000134
Figure BDA0004056452490000135
In the method, in the process of the invention,
Figure BDA0004056452490000136
for the respective crash severity factor of pedestrian u and vehicle j +.>
Figure BDA0004056452490000137
For each real-time coordinate area of pedestrian u and vehicle j, f pet F is a collision degree determination function R For real-time zone determination function, t pet Is the post-invasion time omega of human-vehicle interaction pet Omega as a collision degree weighting parameter R The real-time regional weight parameter;
Figure BDA0004056452490000138
Figure BDA0004056452490000139
Figure BDA00040564524900001310
wherein f pay The function is calculated for the game and,
Figure BDA00040564524900001311
for specific payment functions under different strategies of both parties, f inf F is a game influencing function mix Is a game mixing function;
the pooling module is used for embedding the relative distance between the vehicles into the embedded function phi to obtain the relative position feature vector of the vehicles
Figure BDA00040564524900001312
Feature vector for hybrid game with human-vehicle>
Figure BDA00040564524900001313
Connected and finally output through a multi-layer perceptron to obtain game pooling vectors
Figure BDA00040564524900001314
Figure BDA00040564524900001315
Figure BDA00040564524900001316
Wherein Mlp GP For the pooling module multi-layer perceptron network layer, cat is used to connect feature vectors,
Figure BDA00040564524900001317
the weight parameters are corresponding to the respective network layers; />
Track decoder for pooling game into vector
Figure BDA00040564524900001318
Characteristic vector of history track of pedestrian->
Figure BDA00040564524900001319
Connecting and passing through decoder module multi-layer perceptron Mlp GD Connecting with network random Gaussian noise Z to obtain decoderFeature vector->
Figure BDA00040564524900001320
Figure BDA00040564524900001321
Pedestrian coordinate vector encoded by embedded function>
Figure BDA00040564524900001322
Initializing the cell all zero vector->
Figure BDA00040564524900001323
Inputting the vectors into the LSTM neural network layer to obtain vectors>
Figure BDA00040564524900001324
Finally will->
Figure BDA00040564524900001325
Finally obtaining pedestrian track prediction coordinates through a decoder multi-layer perceptron>
Figure BDA00040564524900001326
Figure BDA00040564524900001327
Figure BDA0004056452490000141
Figure BDA0004056452490000142
Figure BDA0004056452490000143
In the method, in the process of the invention,
Figure BDA0004056452490000144
ω λ is a weight parameter of the corresponding network layer.
The track discriminator comprises the following working procedures: predicted trajectory of pedestrian output by trajectory generator
Figure BDA0004056452490000145
And a true pedestrian history trajectory Y u Inputting the high-dimensional feature vectors into a discriminator neural network, and outputting the high-dimensional feature vectors by embedding the high-dimensional feature vectors into a fully-connected neural network layer>
Figure BDA0004056452490000146
Sequentially passing through an LSTM neural network layer and a discriminator multilayer perceptron, and finally outputting to obtain the probability of being a real track;
Figure BDA0004056452490000147
Figure BDA0004056452490000148
Figure BDA0004056452490000149
in the method, in the process of the invention,
Figure BDA00040564524900001410
output vector for LSTM network layer, +.>
Figure BDA00040564524900001411
Corresponding to the weight parameters of the respective network layer, P rea l is the probability of discriminating as a true track.
Loss function L of SDG-GAN model SDG - GAN Cross entropy loss function L comprising a generator and a discriminator GAN (G, D) minimum difference loss function L from pedestrian real track and pedestrian predicted track L2 (G)；
L SDG-GAN ＝L GAN (G,D)+L L2 (G)；
Figure BDA00040564524900001412
Figure BDA00040564524900001413
Wherein E is the expected value, r is the inputted real track data of the pedestrian, S is the sampling number based on the model generating result, Z is the inputted Gaussian noise distribution, D is the discriminator, and G is the generator.
And S4, training the SDG-GAN model by using the historical data acquired in the S1.
S5, predicting the pedestrian track of the right-hand fork without the signal in real time by using a trained SDG-GAN model.
When the technical scheme is actually implemented, the inventor selects a right-hand intersection without a signal at one of the X major roads in the BN zone of C city as a data acquisition place, and mainly acquires basic indexes of pedestrians and vehicles. Because the people flow is dense in the place, the people and the vehicles are frequently interacted, so that the occurrence frequency of the people and the vehicles conflict is high, the first right-turning vehicle in the scene is provided with a parking space, and the right-turning vehicle can only drive into the second vehicle, so that the people and the vehicles interaction area is more accurate. This location is therefore more suitable for capturing the man-car gaming process, as shown in fig. 4.
And inputting the collected man-vehicle data set into an SDG-GAN model for training to obtain a result. In the observation area, the vehicles and pedestrians are not decided, the SDG-GAN algorithm obtains expected benefits of pedestrian games through the actual passing probability, and the prediction results of the pedestrian tracks are influenced by combining the relative positions of the vehicles and the pedestrians, so that the prediction results are good due to the fact that the prediction mechanism is more perfect. And when the pedestrian game is successful and the vehicle waits, the pedestrian dynamic game payment is positive, and the vehicle dynamic game payment is negative. The gaming mechanism would indicate a reduced risk to the vehicle and would encourage the pedestrian to walk forward. And when the pedestrian game fails and waits, the pedestrian dynamic game payment is negative, and the vehicle dynamic game payment is positive. The gaming mechanism inhibits pedestrians from proceeding and increases the risk of vehicles passing. When people and vehicles collide, the dynamic game payment of pedestrians and vehicles is negative at the same time. The game mechanism shows that the collision probability of the two parties is rapidly increased, and the influence on the pedestrian track is further refined according to the microscopic data of the two parties. When people and vehicles wait simultaneously, traffic jam and traffic efficiency are reduced, both dynamic game payments are negative, and a game mechanism influences pedestrian track prediction through remorse factors.
According to the invention, on the basis of S-GAN, microscopic data and game ideas are utilized to analyze, so as to obtain a macroscopic interaction strategy of the human and vehicle without the signal right-hand fork at different moments; meanwhile, the human-vehicle interaction is divided into stages, including an observation area and a conflict area. While in the observation area, the two-party game pays the benefit and obtains its expected benefit through probability. When the two parties are in the conflict area, the existing decision is judged, the strategy benefits of the two parties are directly obtained, and more detailed analysis on the game process of the people and the vehicles is realized. In addition, the invention analyzes the specific human-vehicle game factors under the special interaction scene of the signal-free right-hand transfer fork, divides the game stage, distinguishes the dangerous degree of human-vehicle conflict by utilizing the time index after infringement, characterizes the remorse degree of both game sides by the remorse factor related to the starting acceleration, comprehensively considers various factors of both game sides as much as possible, and establishes the human-vehicle game payment matrix on the basis. And the Nash equilibrium idea of the complete information game is combined to be embedded into a deep learning network model, so that the prediction of the pedestrian track is realized. And on the basis, an SDG-GAN model is constructed. The SDG-GAN algorithm fuses macroscopic game and microscopic motion parameters, so that the relationship between the two persons and the vehicles is considered more comprehensively, the true reflection of the downstream persons in the game state can be expressed, and the accuracy and the interpretability of predicting the pedestrian track are improved. And then, the SDG-GAN model is driven by the microscopic man-vehicle interaction data and macroscopic man-vehicle game strategy benefits to predict the track of the pedestrian. In such a way, the invention combines micro motion factors and macro game decisions in the actual interaction scene, adds the game idea into the human-vehicle interaction state, and can accurately predict the pedestrian track of the right-hand fork without signals. The method can ensure the prediction accuracy of the pedestrian track of the signal-free right-turn intersection and the effectiveness of the auxiliary driving decision of the signal-free right-turn intersection, thereby considering the efficiency and the safety of the vehicle passing through the signal-free right-turn intersection.
Finally, it should be noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the technical solution, and those skilled in the art should understand that modifications and equivalents may be made to the technical solution of the present invention without departing from the spirit and scope of the present invention, and all such modifications and equivalents are included in the scope of the claims.

Claims (10)

1. A method for predicting the track of a pedestrian with a right-hand fork without a signal based on a game theory is characterized by comprising the following steps:
s1, acquiring historical data of pedestrians and vehicles at a right-hand fork without signals;
s2, analyzing the human-vehicle game factors of the signal-free right-turn intersection, and constructing a corresponding human-vehicle game model;
s3, inserting the man-vehicle game model into a preset S-GAN model to obtain an SDG-GAN model for predicting the track of the pedestrian;
s4, training the SDG-GAN model by using the historical data acquired in the S1;
s5, predicting the pedestrian track of the right-hand fork without the signal in real time by using a trained SDG-GAN model.
2. The method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 1, wherein the method comprises the following steps of: s2, the construction process of the man-vehicle game model comprises the following steps:
s21, dividing a game stage, designing an observation area and a conflict area, and recognizing that a game starts once pedestrians and vehicles enter the observation area; the dangerous degree of collision of the human and the vehicle is represented by the rear intrusion time, wherein the rear intrusion time is the time difference of the pedestrian and the vehicle entering the collision area, and the dangerous degree is higher when the rear intrusion time is shorter;
s22, constructing a man-vehicle game payment matrix based on decision strategies of pedestrians and vehicles in the game; the decision strategy comprises the steps of simultaneous passing of pedestrian vehicles, waiting for passing of pedestrians and waiting for passing of vehicles and simultaneous passing of pedestrians;
s23, obtaining an expected function and a corresponding loss function of the human-vehicle game model based on the characteristics of the human-vehicle game payment matrix;
s24, constructing a man-vehicle game model based on the expected function and the loss function obtained in the S23.
3. The method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 2, wherein the method comprises the following steps of: in S22, when the pedestrian vehicles simultaneously pass,
the payment function of the pedestrian is:
Figure FDA0004056452480000011
the payment function of the vehicle is:
Figure FDA0004056452480000012
wherein v is v Indicating pedestrian passing speed, a v Representing pedestrian acceleration v p Indicating the passing speed of the vehicle, a p Indicating vehicle acceleration, alpha 1 Representing the common influencing factor of the speed and acceleration of the vehicle, alpha 2 Representing the co-influencing factor, sigma, of the speed and acceleration of a pedestrian v Representing a crash severity factor, sigma, of a vehicle p Represents a collision severity factor for a pedestrian, and:
Figure FDA0004056452480000013
Figure FDA0004056452480000014
when a pedestrian waits for the vehicle to pass, the payment function of the pedestrian is:
Figure FDA0004056452480000015
wherein alpha is 4 To wait for the suppression coefficient, t, to pass p Waiting time for pedestrians;
the payment function of the vehicle is:
Figure FDA0004056452480000021
wherein alpha is 3 Is the velocity excitation coefficient when passing;
when the pedestrian passes through the vehicle waiting, the payment function of the pedestrian is as follows:
Figure FDA0004056452480000022
wherein alpha is 3 V is the velocity excitation coefficient when passing p The pedestrian passing speed;
the payment function of the vehicle is:
Figure FDA0004056452480000023
wherein alpha is 4 To wait for the suppression coefficient, t, to pass v For vehicle waiting time, 0.75s is driver reaction time;
when the pedestrian and the vehicle wait simultaneously, the payment function of the pedestrian is that
Figure FDA0004056452480000024
The payment function of the vehicle is
Figure FDA0004056452480000025
Wherein alpha is 5 Wait suppression coefficient for the vehicle under common loss; alpha 6 Wait for the suppression coefficient of pedestrian under the common loss; k is the regret factor of both parties, regret factorThe son is related to starting acceleration in the waiting process, and represents the remorse degree of the waiting strategy, wherein the remorse degree is larger as the starting acceleration is larger;
the man-vehicle game payment matrix is as follows:
Figure FDA0004056452480000026
4. the signaling-free right-hand-of-hand-off crossing pedestrian trajectory prediction method based on game theory as claimed in claim 3, wherein: in S23, when the expectation function is a nash equilibrium point of a hybrid advantage policy that the vehicle waits for the pedestrian and the pedestrian waits for the pedestrian, the vehicle and the pedestrian are both mixed to expect benefits;
wherein the expected benefit when the vehicle selects to pass
Figure FDA0004056452480000027
The method comprises the following steps:
Figure FDA0004056452480000028
wherein (1)>
Figure FDA0004056452480000029
Representing the probability of passing a pedestrian, & lt & gt>
Figure FDA00040564524800000210
Probability of waiting for a pedestrian;
expected benefits when a vehicle selects waiting
Figure FDA00040564524800000211
The method comprises the following steps:
Figure FDA00040564524800000212
analyzing pure strategy profits of pedestrians, and expected profits of pedestrians passing through
Figure FDA00040564524800000213
The method comprises the following steps:
Figure FDA0004056452480000031
wherein (1)>
Figure FDA0004056452480000032
Representing the probability of a vehicle passing, < > or >>
Figure FDA0004056452480000033
Representing a probability of waiting for the vehicle;
expected benefits when pedestrians select waiting
Figure FDA0004056452480000034
The method comprises the following steps:
Figure FDA0004056452480000035
when the vehicle passes through the expected benefits
Figure FDA0004056452480000036
And wait for the expected benefit->
Figure FDA0004056452480000037
The same Nash equalization occurs and the probability combination of pedestrian passage and waiting is as follows:
Figure FDA0004056452480000038
Figure FDA0004056452480000039
when pedestrian passing expected income
Figure FDA00040564524800000310
And wait for the expected benefit->
Figure FDA00040564524800000311
The same Nash equalization occurs and the probability combination of vehicle passing and waiting is as follows:
Figure FDA00040564524800000312
Figure FDA00040564524800000313
5. the method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 4, wherein the method comprises the following steps: when the SDG-GAN model predicts the track of the pedestrian, the track of the pedestrian is defined as two-dimensional coordinate position change under the time sequence, and the coordinate of the pedestrian u at the t moment is defined as
Figure FDA00040564524800000314
The coordinates of vehicle j at time t are +.>
Figure FDA00040564524800000315
Historical track set X of each step length of pedestrians u from 1 to u The method comprises the following steps:
Figure FDA00040564524800000316
wherein, 1 to is the observation frame of the history track of the pedestrian, and to is the length of the observation frame;
pedestrianu is from to+1 to t p Predicted track set for each step in
Figure FDA00040564524800000317
The method comprises the following steps:
Figure FDA00040564524800000318
in which, to+1 to to+t p Predicted frame for pedestrian history trajectory, t p To predict frame length;
pedestrians u from to+1 to t p Real history track set Y of each step length u The method comprises the following steps:
Figure FDA00040564524800000319
pedestrian u is from 1 to t p Within the true history trace
Figure FDA00040564524800000320
And predictive generation track->
Figure FDA00040564524800000321
Respectively, [ X ] u ,Y u ]And->
Figure FDA00040564524800000322
6. The method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 5, wherein the method comprises the following steps: s3, the SDG-GAN model comprises a man-vehicle game model, a track generator and a track discriminator; the track generator is used for encoding and decoding the output result of the human-vehicle game model and the historical track of the pedestrian and outputting the predicted track of the pedestrian; the trajectory discriminator is used for discriminating the probability that the predicted trajectory of the pedestrian is the true trajectory.
7. The method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 6, wherein: the track generator comprises a track encoder, a game mechanism module, a pooling module and a track decoder;
the track encoder is used for embedding the pedestrian coordinate position and the vehicle coordinate position at each time step into an embedding function phi containing a Relu nonlinear activation function to obtain a fixed length vector
Figure FDA0004056452480000041
And->
Figure FDA0004056452480000042
Obtaining a pedestrian history track feature vector by LSTM unit coding>
Figure FDA0004056452480000043
And vehicle history trajectory feature vector->
Figure FDA0004056452480000044
Figure FDA0004056452480000045
Figure FDA0004056452480000046
In which the embedded function phi is a fully connected neural network layer,
Figure FDA0004056452480000047
weight parameters for the embedding function, +.>
Figure FDA0004056452480000048
The LSTM unit weight parameter is adopted;
the game mechanism module is used for extracting game related data of the two people and vehicles based on game payment functions of the two people and vehicles; the game related data comprises speed, acceleration, relative distance and waiting time; the method is also used for obtaining the post-intrusion time at the moment by utilizing the speeds of the two parties and the interval distance between the two parties and the conflict area at each time step; judging the interaction risk degree of the two parties under the time step by utilizing the post invasion time value; judging whether the two parts are in the observation area or the conflict area according to the position coordinates of the two parts at the moment;
the game mechanism module is also used for calibrating according to the interactive decisions of both sides of the person and the vehicle in the real world, and obtaining the specific expected loss of both sides under the decisions
Figure FDA0004056452480000049
And->
Figure FDA00040564524800000410
The specific expectation loss influences the historical track feature vectors of the two sides of the human and the vehicle to obtain the game feature vector of the two sides +.>
Figure FDA00040564524800000411
And fusing the characteristic vectors into the human-vehicle hybrid game characteristic vector +.>
Figure FDA00040564524800000412
Figure FDA00040564524800000413
In the method, in the process of the invention,
Figure FDA00040564524800000414
for the respective crash severity factor of pedestrian u and vehicle j +.>
Figure FDA00040564524800000415
For each real-time coordinate area of pedestrian u and vehicle j, f pet F is a collision degree determination function R Is real-timeRegion judgment function, t pet Is the post-invasion time omega of human-vehicle interaction pet Omega as a collision degree weighting parameter R The real-time regional weight parameter;
Figure FDA00040564524800000416
Figure FDA00040564524800000417
Figure FDA0004056452480000051
wherein f pay The function is calculated for the game and,
Figure FDA0004056452480000052
for specific payment functions under different strategies of both parties, f inf F is a game influencing function mix Is a game mixing function;
the pooling module is used for embedding the relative distance between the vehicles into the embedded function phi to obtain the relative position feature vector of the vehicles
Figure FDA0004056452480000053
Feature vector for hybrid game with human-vehicle>
Figure FDA0004056452480000054
Connecting and finally outputting the game pooling vector by a multi-layer perceptron>
Figure FDA0004056452480000055
Figure FDA0004056452480000056
Figure FDA0004056452480000057
Wherein Mlp GP For the pooling module multi-layer perceptron network layer, cat is used to connect feature vectors,
Figure FDA0004056452480000058
the weight parameters are corresponding to the respective network layers;
track decoder for pooling game into vector
Figure FDA0004056452480000059
Characteristic vector of history track of pedestrian->
Figure FDA00040564524800000510
Connecting and passing through decoder module multi-layer perceptron Mlp GD Connecting with the network random Gaussian noise Z to obtain the decoder feature vector
Figure FDA00040564524800000511
Figure FDA00040564524800000512
Pedestrian coordinate vector encoded by embedded function>
Figure FDA00040564524800000513
Initializing the cell all zero vector->
Figure FDA00040564524800000514
Inputting the vectors into the LSTM neural network layer to obtain vectors>
Figure FDA00040564524800000515
Finally will->
Figure FDA00040564524800000516
Finally obtaining pedestrian track prediction coordinates through a decoder multi-layer perceptron>
Figure FDA00040564524800000517
Figure FDA00040564524800000518
Figure FDA00040564524800000519
Figure FDA00040564524800000520
Figure FDA00040564524800000521
In the method, in the process of the invention,
Figure FDA00040564524800000522
ω λ is a weight parameter of the corresponding network layer.
8. The method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 7, wherein: in S3, the track discriminator includes: predicted trajectory of pedestrian output by trajectory generator
Figure FDA00040564524800000523
And a true pedestrian history trajectory Y u Inputting the high-dimensional feature vectors into a discriminator neural network, and outputting the high-dimensional feature vectors by embedding the high-dimensional feature vectors into a fully-connected neural network layer>
Figure FDA00040564524800000524
Sequentially passing through an LSTM neural network layer and a discriminator multilayer perceptron, and finally outputting to obtain the probability of being a real track;
Figure FDA00040564524800000525
Figure FDA00040564524800000526
Figure FDA0004056452480000061
in the method, in the process of the invention,
Figure FDA0004056452480000062
output vector for LSTM network layer, +.>
Figure FDA0004056452480000063
Corresponding to the weight parameters of the respective network layer, P real Is the probability of discriminating as a true trajectory.
9. The method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 8, wherein: s3, a loss function L of the SDG-GAN model SDG-GAN Cross entropy loss function L comprising a generator and a discriminator GAN (G, D) minimum difference loss function L from pedestrian real track and pedestrian predicted track L2 (G)；
L SDG-GAN ＝L GAN (G,D)+L L2 (G)；
Figure FDA0004056452480000064
Figure FDA0004056452480000065
Wherein E is the expected value, r is the inputted real track data of the pedestrian, S is the sampling number based on the model generating result, Z is the inputted Gaussian noise distribution, D is the discriminator, and G is the generator.
10. The method for predicting the right-hand cross pedestrian trajectory without signals based on game theory as claimed in claim 9, wherein: in S1, the history data includes: the number of frames of video, coordinates of pedestrians, accelerations of pedestrians, waiting time of pedestrians, coordinates of vehicles, accelerations of vehicles, waiting time of vehicles, distance of vehicles and post-intrusion time; the man-vehicle distance is the Euclidean distance between the pedestrian and the vehicle.
CN202310048086.4A 2023-01-31 2023-01-31 Game theory-based signal-free right-hand fork pedestrian track prediction method Pending CN116311905A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310048086.4A CN116311905A (en) 2023-01-31 2023-01-31 Game theory-based signal-free right-hand fork pedestrian track prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310048086.4A CN116311905A (en) 2023-01-31 2023-01-31 Game theory-based signal-free right-hand fork pedestrian track prediction method

Publications (1)

Publication Number Publication Date
CN116311905A true CN116311905A (en) 2023-06-23

Family

ID=86833164

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310048086.4A Pending CN116311905A (en) 2023-01-31 2023-01-31 Game theory-based signal-free right-hand fork pedestrian track prediction method

Country Status (1)

Country Link
CN (1) CN116311905A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117227763A (en) * 2023-11-10 2023-12-15 新石器慧通（北京）科技有限公司 Automatic driving behavior decision method and device based on game theory and reinforcement learning
CN117373235A (en) * 2023-09-28 2024-01-09 合肥工业大学 Interactive game balance strategy exploration method for pedestrians and automatic driving vehicles
CN117475090A (en) * 2023-12-27 2024-01-30 粤港澳大湾区数字经济研究院(福田) Track generation model, track generation method, track generation device, terminal and medium
CN117765737A (en) * 2024-02-21 2024-03-26 天津大学 Traffic abnormality detection method, device, apparatus, medium, and program product

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117373235A (en) * 2023-09-28 2024-01-09 合肥工业大学 Interactive game balance strategy exploration method for pedestrians and automatic driving vehicles
CN117227763A (en) * 2023-11-10 2023-12-15 新石器慧通（北京）科技有限公司 Automatic driving behavior decision method and device based on game theory and reinforcement learning
CN117227763B (en) * 2023-11-10 2024-02-20 新石器慧通（北京）科技有限公司 Automatic driving behavior decision method and device based on game theory and reinforcement learning
CN117475090A (en) * 2023-12-27 2024-01-30 粤港澳大湾区数字经济研究院(福田) Track generation model, track generation method, track generation device, terminal and medium
CN117475090B (en) * 2023-12-27 2024-06-11 粤港澳大湾区数字经济研究院(福田) Track generation model, track generation method, track generation device, terminal and medium
CN117765737A (en) * 2024-02-21 2024-03-26 天津大学 Traffic abnormality detection method, device, apparatus, medium, and program product
CN117765737B (en) * 2024-02-21 2024-05-14 天津大学 Traffic abnormality detection method, device, apparatus, medium, and program product

Similar Documents

Publication Publication Date Title
CN116311905A (en) Game theory-based signal-free right-hand fork pedestrian track prediction method
CN110298131B (en) Method for establishing automatic driving lane change decision model in hybrid driving environment
CN110362910B (en) Game theory-based automatic driving vehicle lane change conflict coordination model establishment method
Toromanoff et al. End-to-end model-free reinforcement learning for urban driving using implicit affordances
CN112907967B (en) Intelligent vehicle lane change decision-making method based on incomplete information game
CN112258841B (en) Intelligent vehicle risk assessment method based on vehicle track prediction
CN110843789B (en) Vehicle lane change intention prediction method based on time sequence convolution network
Hou et al. Driver intention recognition method using continuous hidden Markov model
Casas et al. The importance of prior knowledge in precise multimodal prediction
US11345368B2 (en) Guidance system, guidance method, and storage medium
CN114495046A (en) Driver behavior risk assessment and pedestrian awareness
CN113516846B (en) Vehicle lane change behavior prediction model construction, prediction and early warning method and system
CN114118349A (en) Method, system and apparatus for user understandable interpretable learning models
CN112949597A (en) Vehicle track prediction and driving manipulation identification method based on time mode attention mechanism
CN114511999A (en) Pedestrian behavior prediction method and device
CN115743179A (en) Vehicle probability multi-mode expected trajectory prediction method
CN116501820A (en) Vehicle track prediction method, device, equipment and storage medium
CN116842340A (en) Pedestrian track prediction method and equipment based on ATT-GRU model
CN115973179A (en) Model training method, vehicle control method, device, electronic equipment and vehicle
Sharma et al. Wad: A deep reinforcement learning agent for urban autonomous driving
WO2023004698A1 (en) Method for intelligent driving decision-making, vehicle movement control method, apparatus, and vehicle
Li et al. Pedestrian interaction algorithm for autonomous vehicles considering crossing intention and driving style
CN117709602B (en) Urban intelligent vehicle personification decision-making method based on social value orientation
CN117208019B (en) Longitudinal decision method and system under perceived occlusion based on value distribution reinforcement learning
Hsieh et al. Trajectory Prediction at Unsignalized Intersections using Social Conditional Generative Adversarial Network

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