CN109726804B - Intelligent vehicle driving behavior personification decision-making method based on driving prediction field and BP neural network - Google Patents
Intelligent vehicle driving behavior personification decision-making method based on driving prediction field and BP neural network Download PDFInfo
- Publication number
- CN109726804B CN109726804B CN201910071774.6A CN201910071774A CN109726804B CN 109726804 B CN109726804 B CN 109726804B CN 201910071774 A CN201910071774 A CN 201910071774A CN 109726804 B CN109726804 B CN 109726804B
- Authority
- CN
- China
- Prior art keywords
- driving
- vehicle
- intelligent vehicle
- behavior
- prediction field
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 33
- 230000006399 behavior Effects 0.000 claims abstract description 107
- 230000001133 acceleration Effects 0.000 claims abstract description 17
- 238000004891 communication Methods 0.000 claims abstract description 16
- 238000010606 normalization Methods 0.000 claims abstract description 9
- 230000002093 peripheral effect Effects 0.000 claims abstract description 9
- 238000012545 processing Methods 0.000 claims abstract description 5
- 210000002569 neuron Anatomy 0.000 claims description 32
- 238000012549 training Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000005516 engineering process Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 238000004422 calculation algorithm Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 2
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 239000010410 layer Substances 0.000 description 42
- 230000008859 change Effects 0.000 description 9
- 230000008569 process Effects 0.000 description 7
- 230000003542 behavioural effect Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000003062 neural network model Methods 0.000 description 3
- 239000003795 chemical substances by application Substances 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 241000282412 Homo Species 0.000 description 1
- 108010001267 Protein Subunits Proteins 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 210000005069 ears Anatomy 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 230000004438 eyesight Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000000056 organ Anatomy 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 230000003238 somatosensory effect Effects 0.000 description 1
- 238000000859 sublimation Methods 0.000 description 1
- 230000008022 sublimation Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 230000016776 visual perception Effects 0.000 description 1
Images
Landscapes
- Traffic Control Systems (AREA)
Abstract
The invention discloses an intelligent vehicle driving behavior personification decision method based on a driving prediction field and a BP neural network, which comprises the following steps of1: dividing the peripheral vehicle behaviors into 9 typical behaviors bi according to the combination of the transverse aspect and the longitudinal aspect, and fitting each vehicle behavior b i Setting a corresponding similarity track, and setting a region through which the target vehicle runs according to the similarity track; step2: taking the front and rear vehicles of the current lane and the front and rear vehicles of the adjacent lanes of the intelligent vehicle as surrounding vehicles, and acquiring the position, the speed and the acceleration of each vehicle at each moment in real time by using V2V communication; step3: establishing a safety prediction field, an efficiency prediction field and a driving comfort prediction field; step4: the intelligent vehicle acquires each behavior b according to real time i And (3) inputting the sub-prediction field intensity sum under the driving prediction field into a BP neural network driving behavior decision model after normalization processing, and outputting a y vector for decoding to obtain the most reasonable driving behavior decision result.
Description
Technical Field
The invention belongs to the field of intelligent driving decision, and particularly relates to an intelligent vehicle driving behavior personification decision method based on a driving prediction field and a BP neural network.
Background
The intelligent driving vehicle is a vehicle with autonomous driving capability, and has the capabilities of environment perception, behavior decision, motion planning, vehicle control, automatic obstacle avoidance and other similar behaviors aiming at traffic scenes besides the capability of completing conventional automobile driving actions. As an important embodiment of the intelligent level of the unmanned vehicle, driving behavior decision has become the key point and difficulty of research of expert students of all parties. For the complex behavior of driving a vehicle, a driver often collects road environment information by means of external sensing organs such as eyes, ears and the like, and selects driving behavior according to driving experience and expected pursuit so as to formulate a sport plan. The intelligent driving vehicle should also have a "personification" decision mechanism, and according to the current driving state, driving task and road environment perception information, the intelligent driving vehicle decides the driving behavior of the vehicle at the next moment.
At present, the research of the driving behavior decision model usually classifies driving traffic scenes first and then makes behavior decisions aiming at the driving scenes, and although good decision results can be obtained for certain specific driving scenes, the traffic elements are complex, random and uncertain factors are more in the real traffic environment, and a decision system cannot completely cover all possible driving scenes, so that the method generally cannot achieve good robustness and adaptability. In addition, the existing intelligent vehicle driving behavior decision method is often focused on the capability of the vehicle to avoid driving danger, and the behavior decision of a normal rational driver during driving can be abstracted into a process of continuously pursuing benefit maximization, so that the method ignores expected benefits of the driver or passengers on other performances of the vehicle, such as high efficiency, driving comfort and the like. In fact, intelligent vehicle "driving brains" can only achieve true "intelligent" driving just like humans, reaching sublimation of increasingly perceived traffic environments. In recent years, with the development of communication technology, communication equipment such as the internet of vehicles, vehicle-to-vehicle (V2V), smart phones and the like can also help vehicles to accurately acquire surrounding additional information, which also brings convenience for establishing a driving prediction field which can adapt to multiple driving scenes.
Disclosure of Invention
Aiming at the complex and changeable driving scenes and the improvement of the requirements of people on driving safety, high efficiency and comfort, the invention provides a driving prediction field based on driving safety, high efficiency and comfort, and provides an intelligent vehicle behavior decision method based on the driving prediction field and a BP neural network, which can reasonably make behavior personification decisions according to the surrounding traffic environment and provide reference basis for the motion trail planning of the intelligent vehicle. The object of the invention can be achieved by the following technical scheme. An intelligent vehicle driving behavior personification decision-making method based on a driving prediction field specifically comprises the following steps:
step1: the peripheral vehicle behavior is divided into 9 typical behaviors bi according to the combination of the transverse aspect and the longitudinal aspect. Fitting each vehicle behaviour b i Corresponding similarity track, setting the area where the target vehicle runs according to the similarity track as
Step2: the method comprises the steps that a front vehicle and a rear vehicle of a current lane of an intelligent vehicle and a front vehicle and a rear vehicle of an adjacent lane are taken as surrounding vehicles, and each traffic environment participating vehicle acquires the position (x, y) and the speed (V) of a vehicle at each moment in real time by using a vehicle-mounted GPS and IMU combined positioning system x ,V y ) Acceleration (a) x ,a y ). A real-time V2V communication network of surrounding vehicle groups is constructed, and a host vehicle (intelligent vehicle) acquires state information of surrounding vehicles in a traffic environment in real time by using a D2D (Device-To-Device) proximity communication service (ProSe) of an LTE module in a V2V communication technology.
Step3: the area travelled by the similarity track is
And the state information of vehicles around the traffic environment, and a driving prediction field is established, including a safety prediction field E S Efficiency prediction field E E Driving comfort prediction field E C 。
(1) Unit safety potential value of any point position in intelligent vehicle running area affected by surrounding jth vehicle
Wherein, (X, Y) is any point position in the intelligent vehicle driving area; g S A field undetermined constant is predicted for driving safety; delta j Vehicle type coefficients of a j-th vehicle surrounding the intelligent vehicle; m is M j The equivalent mass ratio of the jth vehicle around is the reciprocal of the product of the length, width and height of the jth vehicle; (x) [j] ,y [j] ) The position vector is the position vector of the current moment of the surrounding jth vehicle;
the current moment speed vector of the surrounding jth vehicle;
The acceleration vector is the acceleration vector of the current moment of the surrounding jth vehicle; delta T is the intelligent vehicle behavior decision execution time; i 2 Is a euclidean norm symbol.
Obtaining a unit safety potential value of any point position in the intelligent vehicle driving area
Then a certain behavior b of the intelligent vehicle i Is the sum of the safety prediction field strengths of (2)
(2) Unit efficiency potential value of any point in intelligent vehicle driving area
Y is the longitudinal position of any point in the intelligent vehicle driving area; g E Predicting a field undetermined constant for driving efficiency; m is M 0 The equivalent mass ratio of the intelligent vehicle is the reciprocal of the product of the length, the width and the height of the intelligent vehicle; y is [0] The longitudinal position of the intelligent vehicle at the current moment;
the efficiency of a certain behavior bi of the intelligent vehicle predicts the field intensity sum as
(3) Unit driving comfort potential value of any point position in intelligent vehicle driving area
(X, Y) is any point position in the intelligent vehicle driving area; g C Predicting a field undetermined constant for driving comfort; (x) [0] ,y [0] ) The position vector is the position vector of the intelligent vehicle at the current moment;
the speed vector is the speed vector of the intelligent vehicle at the current moment; delta T is peripheral intelligent vehicle behavior prediction cycle time; i 2 Is the euclidean norm symbol of the vector. />
Then a certain behavior b of the intelligent vehicle i Is the sum of the driving comfort predictive field strengths of (2)
Step4: aiming at the high nonlinearity of the rationality relation between each expected pursuit factor of driving behavior and the driving behavior caused by complex and changeable driving environment, the three-layer BP neural network is used for self-adaptively obtaining the driving behavior decision result.
The neural network structure is 27-16-9. With each vehicle behaviour b i The sum of the sub-predicted field strengths of (2) forming a 9 x 3 matrix T for the elements, each column representing a different vehicle behavior b i Each row represents a respective row sub-prediction field. Converting matrix T into vector form
Adopts the maximum value method for->
Normalization process->
The input vector is found to be x= (x 1 ,x 2 ,……,x 27 ). Hidden layer jth neuron outputs as
Wherein the connection weight, a, between the input layer neuron i and the hidden layer neuron j is represented j Representing the bias value of hidden layer neuron j, f () is the activation function of hidden layer neuron, and adopts Sigmoid function in the form of
The vehicle behavior decision types 1 to 9 are encoded by using one hot, the number of neurons of an output layer is 9, and the output vector y= (y 1, y2, … …, y 9). Output layer kth neuron outputs as
Wherein mu jk Representing the connection weights between hidden layer neuron j and output layer neuron k, b k Representing the bias value of the output layer neuron j.
The driving simulator is used for collecting behavior decision selection of human drivers with rich driving experience in various traffic scenes, and the driving prediction field established by the invention is combined for calculation and data preprocessing to generate a sample training set L. Training the BP neural network:
(1) Initializing a BP network, namely randomly reducing the connection weight between layers and the threshold value (-1, 1) of hidden layer and output layer nodes;
(2) Calculating h for each sample j And y k Obtaining a mean square error
Wherein o is k Deciding a result vector element value for actual driving behavior;
(3) The connection weights and offsets are updated by error back propagation so that the error function value decreases. Connection weight mu jk 、ω ij And bias b k 、a j The update formula is:
μ jk ＝μ jk +ηh j e k
b k ＝b k +ηe k
wherein the eta learning rate is 0.35.
(4) Setting training precision epsilon=0.01, repeatedly inputting samples in a sample set L, executing an algorithm to calculate an error E (L), and stopping connecting the weight and the bias updating party when the difference between two adjacent iteration errors E (L) is lower than epsilon or the iteration times reaches 100 times to obtain a trained BP neural network driving behavior decision model.
Step4: the intelligent vehicle acquires each behavior b in real time according to Step1-3 i And (3) taking the sum of sub-prediction field strengths under the driving prediction field as an x vector to input a trained BP neural network driving behavior decision model after normalization processing, and outputting a y vector to decode to obtain the most reasonable driving behavior decision result.
The invention has the beneficial effects that:
(1) The three elements of driving expectations are comprehensively considered, a driving prediction field is established in a driving area of the intelligent vehicle, and the requirements of people on increasing driving safety, high efficiency and comfort in the driving process can be met.
(2) The driving behavior decision of the intelligent vehicle is carried out by using the driving prediction field, so that the intelligent vehicle can adapt to complex and changeable traffic environments, the connection of a human-vehicle-road closed-loop system is realized, the accuracy of the behavior decision is high, the environmental adaptability is strong, and the robustness is good.
(3) The real-time V2V communication network of the surrounding vehicle group is constructed, the host vehicle directly acquires the real-time state information of the self-test of other vehicles, the data processing process of the vision system is avoided, the data accuracy is high, and the real-time performance is good.
(4) The specific BP neural network suitable for the invention is established, and the limit that the field intensity and the weight coefficient of each driving sub prediction field are difficult to select is broken through.
(5) The BP neural network is used for learning a driving behavior decision mechanism of an excellent human driver, so that an intelligent vehicle can realize 'personification' decision, and the driving behavior decision is high in rationality.
Drawings
FIG. 1 is a method for intelligent vehicle driving behavior personification decision-making based on a driving prediction field and a BP neural network;
FIG. 2 is a discretized division of peripheral vehicle behavior;
FIG. 3 peripheral target vehicle behavior b i A driven-through area;
FIG. 4 is a traffic environment H;
FIG. 5 safely predicts field strength distribution in an H-traffic environment;
FIG. 6 predicts field strength distribution of efficiency under an H-traffic environment;
FIG. 7 illustrates a driving comfort predictive field strength distribution in an H-traffic environment;
FIG. 8BP neural network model;
fig. 9BP neural network model training process.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
As shown in fig. 1, the implementation steps of the present invention include the following:
step1: the possible behavior decision of the intelligent vehicle is divided into two directions of transverse behavior and longitudinal behavior to be combined and divided. By changing left Lane (Lane Change to Left), lane keeping (Lane Ke)ep), right lane change (Lane Change to Right), acceleration (Speed Increase), maintenance Speed (Speed Keep), deceleration (Speed Decrease) in longitudinal behavior to discretize the surrounding vehicle behavior into N typical behaviors b i N=9, left lane change deceleration (LCL-SD), left lane change uniform velocity (LCL-SK), left lane change acceleration (LCL-SI), lane keeping deceleration (LK-SD), lane keeping uniform velocity (LK-SK), lane keeping acceleration (LK-SI), right lane change deceleration (LCR-SD), right lane change deceleration (LCR-SK), right lane change deceleration (LCR-SI), respectively. Each trace corresponds to each behavior b as shown in FIG. 2 i Wherein i is more than or equal to 1 and less than or equal to 9. Dividing a safety area in a drivable area into 9 behavior hotspots according to behavior types, and simulating a decision of an intelligent vehicle to make a certain vehicle behavior b i And (3) fuzzy estimation of the balance influence of the field intensity sum of the prediction fields of all driving sub-units, and taking the central point of each behavior hot zone as the position of the target vehicle at the end moment of each behavior type according to the similarity principle. Fitting each vehicle behaviour b i The corresponding similarity track is taken as an integral area
As shown in fig. 3.
Step2: each traffic environment participating vehicle has an independent ID, and the vehicle-mounted GPS and IMU combined positioning system is used for acquiring the position (x, y) and the speed (V) of the own vehicle at each moment in real time x ,V y ) Acceleration (a) x ,a y ). In consideration of the real-time performance and robustness of data transmission in the subsequent steps, the data acquisition frequency is 50Hz, namely 0.02s is the time length of taking the two data intervals before and after. The method comprises the steps of constructing a surrounding vehicle group real-time V2V communication network, enabling a host vehicle To access the V2V communication network through a PC5 interface, and acquiring state time sequence information of surrounding vehicles in a traffic environment in real time by using a D2D (Device-To-Device) proximity communication service (ProSe) in an LTE module. The V2V communication technology is called Vehicle-To-Vehicle Technology, and the D2D module can establish mutual communication between the adjacent Vehicle terminals without a base station. In a real traffic environment, a human driver can acquire surrounding traffic information through a forward visible area, a rearview mirror and a rearview camera,the intelligent driving system can achieve the same purpose through the communication module or the visual perception module, so that the influence of traffic information on the target vehicle is transmitted adjacently and alternately front and back. When the surrounding traffic environment of the intelligent vehicle is built, the front and rear vehicles of the current lane of the intelligent vehicle and the front and rear vehicles of the adjacent lane are taken as the influencers of the behavior occurrence of the front and rear vehicles, and the number of surrounding influencers is set as h. a is that the target vehicle is in a middle lane, and h is 6; b is that the target vehicle is in a left lane, and h is 4 at the moment; c is that the target vehicle is in the right lane, and h is 4.
Step3: whether a traffic vehicle driven by a person or a traffic vehicle with automatic driving capability is an agent which can respond to the excitation of the cooperative traffic environment of the surrounding vehicle road in a corresponding way of "trending away the interests", and the agent is influenced by the balance of various expected benefits of the self-vehicle. According to the driving behavior, the expected income three elements (safety, efficiency and driving comfort) are decided, the driving prediction field is divided into three sub prediction fields, namely a safety prediction field E S Efficiency prediction field E E Prediction field E for driving comfort C . Taking a certain traffic environment H as shown in FIG. 4 as an example, a safety prediction field E is established S Efficiency prediction field E E Prediction field E for driving comfort C The following are provided:
(1) Establishing a security prediction field
As shown in fig. 5, the security prediction field characterizes the desired pursuit of security when the intelligent vehicle makes a behavioural decision. The front-back direction h surrounding traffic vehicles of the intelligent vehicle are used as 'electric charges' for generating safety field potential, and the positions, the speeds and the accelerations of the front-back direction h surrounding traffic vehicles are used as main variables for influencing the safety potential value.
Writing out unit safety potential value of any point position affected by the jth vehicle around in the intelligent vehicle driving area
Wherein, (X, Y) is any point position in the intelligent vehicle driving area; g S Predicting field pending constant for driving safety；δ j Vehicle type coefficients of surrounding jth vehicles; m is M j The equivalent mass ratio of the jth vehicle around is the reciprocal of the product of the length, width and height of the jth vehicle; (x) [j] ,y [j] ) The position vector is the position vector of the jth vehicle around the intelligent vehicle at the current moment;
the current moment speed vector of the surrounding jth vehicle;
The acceleration vector is the acceleration vector of the current moment of the surrounding jth vehicle; delta T is the intelligent vehicle behavior decision execution time; i 2 Is a euclidean norm symbol.
Unit safety potential value of any point position in intelligent vehicle driving area
The sum of the safety prediction field strengths of a certain behavior bi of the intelligent vehicle is
(2) Establishing an efficiency prediction field
As shown in fig. 6, the efficiency prediction field characterizes the desire for efficiency when the intelligent vehicle makes a behavioral decision. The intelligent vehicle is taken as the 'charge' for generating the efficiency field potential, and the longitudinal position of the intelligent vehicle is taken as the main variable for influencing the efficiency potential value.
Writing out unit efficiency potential value of any point position in intelligent vehicle driving area
Y is the longitudinal position of any point in the intelligent vehicle driving area; g E Prediction of driving efficiencyA field pending constant; m is M 0 The equivalent mass ratio of the intelligent vehicle is the reciprocal of the product of the length, the width and the height of the intelligent vehicle; y is [0] The longitudinal position of the intelligent vehicle at the current moment;
then a certain behavior b of the intelligent vehicle i The sum of the efficiency predictive field strengths of (2) is
(3) Establishing a driving comfort prediction field
As shown in fig. 7, the driving comfort prediction field characterizes the desired pursuit of driving comfort when the intelligent vehicle makes a behavioral decision. The intelligent vehicle is used as the 'charge' for generating the driving comfort field potential, and the transverse and longitudinal acceleration of the intelligent vehicle to a certain position of a driving area is used as a main variable for influencing the driving comfort potential value.
Writing out a unit driving comfort potential value of any point position in an intelligent vehicle driving area
(X, Y) is any point position in the intelligent vehicle driving area; g C Predicting a field undetermined constant for driving comfort; (x) [0] ,y [0] ) The position vector is the position vector of the intelligent vehicle at the current moment;
the speed vector is the speed vector of the intelligent vehicle at the current moment; delta T is the intelligent vehicle behavior decision execution time; i 2 Is a euclidean norm symbol.
The sum of the field strengths of the predicted driving comfort of a certain behavior bi of the intelligent vehicle is
Step4: aiming at the high nonlinearity of the rationality relation between each expected pursuit factor of driving behavior and the driving behavior caused by complex and changeable driving environment, an artificial neural network is used for obtaining a driving behavior decision result.
As shown in fig. 8, the neural network used is a three-layer BP neural network with only one hidden layer. According to each vehicle behaviour b i The sum of the sub-predicted field strengths of (2) can be used to obtain a driving prediction field matrix T of driving behavior decision, the matrix is 9 multiplied by 3, and each column represents different vehicle behaviors b i Each row represents a respective row sub-prediction field.
Converting matrix T into vector form
The number of input layer neurons n= 9*3 =27. Adopts the maximum value method for->
Normalization process->
The input vector is found to be x= (x 1 ,x 2 ,……,x 27 ) Wherein->
Representing the value before normalization, x i Represents the value before normalization,>
represents maximum value>
Representing a minimum value. For vehicle behavior decision types 1-9, one hot is adopted for coding, the number m of neurons of an output layer is 9, and an output vector y= (y) 1 ,y 2 ,……,y 9 ) For example when the driving behavior of the vehicle is decided as b 4 When y is 4 =1, other y k (k.epsilon.1, 2,4, … …, 9) is 0. According to the empirical function of neural network of the hidden layer of the monolayer +.>
And
the number of hidden layer nodes s is determined, where f is an integer between 1 and 10. Empirically, the hidden layer node number s is set to 16. The BP neural network structure n-s-m is 27-16-9.
Hidden layer jth neuron outputs as
Wherein omega ij Representing the connection weights, a, between input layer neuron i and hidden layer neuron j j The threshold value of the hidden layer neuron j is represented, g () is the activation function of the hidden layer neuron, and the form is adopted as a Sigmoid function
Output layer kth neuron outputs as
Wherein mu jk Representing the connection weights between hidden layer neuron j and output layer neuron k, b k Representing the bias value of the output layer neuron j.
25 excellent human drivers with the ages of 24-48 years and rich driving experience, five years and more, no serious diseases and no major accidents in the driving ages are selected as experimental participants, and the 6-degree-of-freedom SCANER II simulated driver platform is used for collecting the behavior decision selection of the human drivers in different traffic scenes. The 6-degree-of-freedom SCANER II simulation driver platform can enable experimental participants to feel visual, auditory and somatosensory automobile driving experience close to real effects in a virtual driving environment, and the experimental participants approach the real driving environment infinitely. Eliminating obvious accident data from the collected data, and selecting each vehicle behavior b i Is effective data of the (c). The driving prediction field established by combining the invention is calculated to obtain eachAnd predicting field intensity and matrix of the sub-driving and converting the field intensity and matrix into vector form, and generating a sample training set L after normalization processing.
As shown in fig. 9, the gradient descent method is used to train the driving behavior personification decision BP neural network model, and the specific process is as follows:
(1) Initializing BP neural network, taking the connection weight between layers and random small quantity between hidden layer and output layer node bias (-1, 1). The learning rate eta is between 0.3 and 0.5, and eta=0.35 is selected.
(2) Each sample in the training set
Sequentially inputting training, and obtaining h according to the above formula j And y k . Calculating the mean square error E, the formula is +.>
Wherein o is k Decision result vector element value for actual driving behavior, e k ＝(o k -y k )。
(3) The BP training process is a process of error back propagation, with the goal of making the error function minimum, i.e., min E, by updating the connection weights and offsets.
a. Implicit layer to output layer weight update gradient
Then weight mu jk The updated formula of (1) is mu jk ＝μ jk +ηh j e k
b. Input layer to hidden layer weight update gradient
Then weight omega ij The updated formula of (2) is
c. Implicit layer to output layer bias update gradient
Bias b k The updated formula of (b) is b k ＝b k +ηe k
d. Bias update gradient from input layer to hidden layer
Then bias a j The updated formula of (2) is
(4) Setting training precision epsilon=0.01, repeatedly inputting samples in a sample set L, executing an algorithm to calculate an error E (L), gradually updating a weight value of each neuron connecting line and a bias value of each layer in the BP neural network through error reverse feedback, and stopping updating by an updater when the difference between two adjacent iteration errors E (L) is lower than epsilon or the number of iterations reaches 100, so as to obtain a trained BP neural network driving behavior decision model.
Step4: the intelligent vehicle acquires each behavior b in real time according to Step1-3 i Sub-prediction field intensity sum under driving prediction field and normalizationAnd (3) inputting the processed result as an x vector into a trained BP neural network driving behavior decision model, outputting a y vector, and decoding to obtain the most reasonable driving behavior decision result.
The above list of detailed descriptions is only specific to practical embodiments of the present invention, and they are not intended to limit the scope of the present invention, and all equivalent embodiments or modifications that do not depart from the spirit of the present invention should be included in the scope of the present invention.
Claims (5)
1. An intelligent vehicle driving behavior personification decision-making method based on a driving prediction field and a BP neural network is characterized by comprising the following steps:
step1: dividing the peripheral vehicle behaviors according to the combination of the transverse aspect and the longitudinal aspect, discretizing the peripheral vehicle behaviors into 9 typical behaviors bi, and fitting each vehicle behavior b i Corresponding similarity track, setting the area where the target vehicle runs according to the similarity track as
Step2: the method comprises the steps that a front vehicle and a rear vehicle of a current lane of an intelligent vehicle and a front vehicle and a rear vehicle of an adjacent lane are taken as surrounding vehicles, and each traffic environment participating vehicle acquires the position (x, y) and the speed (V) of a vehicle at each moment in real time by using a vehicle-mounted GPS and IMU combined positioning system x ,V y ) Acceleration (a) x ,a y )；
Step3: the area travelled by the similarity track is
The state information of surrounding vehicles is used for establishing a driving prediction field, wherein the driving prediction field comprises a safety prediction field ES, an efficiency prediction field EE and a driving comfort prediction field EC;
in the step3, the method for establishing the safe prediction field ES is as follows:
taking the front-back h peripheral traffic vehicles of the intelligent vehicle as 'charge' for generating a safety field potential, and taking the positions, the speeds and the accelerations of the front-back h peripheral traffic vehicles as main variables for influencing the safety potential value;
writing out unit safety potential value of any point position affected by the jth vehicle around in the intelligent vehicle driving area
Wherein, (X, Y) is any point position in the intelligent vehicle driving area; g S A field undetermined constant is predicted for driving safety; delta j Vehicle type coefficients of surrounding jth vehicles; m is M j The equivalent mass ratio of the jth vehicle around is the reciprocal of the product of the length, width and height of the jth vehicle; (x) [j] ,y [j] ) The position vector is the position vector of the jth vehicle around the intelligent vehicle at the current moment;
the current moment speed vector of the surrounding jth vehicle;
The acceleration vector is the acceleration vector of the current moment of the surrounding jth vehicle; delta T is the intelligent vehicle behavior decision execution time; i 2 Is a euclidean norm symbol;
unit safety potential value of any point position in intelligent vehicle driving area
The sum of the safety prediction field strengths of a certain behavior bi of the intelligent vehicle is
In the step3, the method for establishing the efficiency prediction field EE is as follows:
taking the intelligent vehicle as a 'charge' for generating an efficiency field potential, and taking the longitudinal position of the intelligent vehicle as a main variable for influencing the efficiency potential value;
writing out unit efficiency potential value of any point position in intelligent vehicle driving area
Y is the longitudinal position of any point in the intelligent vehicle driving area; g E Predicting a field undetermined constant for driving efficiency; m is M 0 The equivalent mass ratio of the intelligent vehicle is the reciprocal of the product of the length, the width and the height of the intelligent vehicle; y is [0] The longitudinal position of the intelligent vehicle at the current moment;
then a certain behavior b of the intelligent vehicle i The sum of the efficiency predictive field strengths of (2) is
In the step3, the method for establishing the driving comfort prediction field EC includes:
the intelligent vehicle is used as 'charge' for generating a driving comfort field potential, and the transverse and longitudinal acceleration of the intelligent vehicle to a certain position of a driving area is used as a main variable for influencing the driving comfort potential value;
writing out a unit driving comfort potential value of any point position in an intelligent vehicle driving area
(X, Y) is any point position in the intelligent vehicle driving area; g C Predicting a field undetermined constant for driving comfort; (x) [0] ,y [0] ) The position vector is the position vector of the intelligent vehicle at the current moment;
the speed vector is the speed vector of the intelligent vehicle at the current moment; delta T is the intelligent vehicle behavior decision execution time; i 2 Is a euclidean norm symbol;
the sum of the field strengths of the predicted driving comfort of a certain behavior bi of the intelligent vehicle is
Step4: the intelligent vehicle acquires each behavior b according to real time i And (3) taking the sum of sub-prediction field strengths under the driving prediction field as an x vector to input a trained BP neural network driving behavior decision model after normalization processing, and outputting a y vector to decode to obtain the most reasonable driving behavior decision result.
2. The intelligent vehicle driving behavior personification decision method based on the driving prediction field and the BP neural network according to claim 1, wherein in the step2, a V2V communication network is constructed between the intelligent vehicle and surrounding vehicles, and the intelligent vehicle acquires the state information of the surrounding vehicles in the traffic environment in real time by using the D2D proximity communication service of the LTE module in the V2V communication technology.
3. The intelligent vehicle driving behavior personification decision method based on the driving prediction field and the BP neural network according to claim 1, wherein in the step4, the BP neural network is of a three-layer 27-16-9 structure.
4. The intelligent vehicle driving behavior personification decision method based on a driving prediction field and a BP neural network according to claim 3, wherein the BP neural network: with each vehicle behaviour b i The sum of the sub-predicted field strengths of (2) forming a 9 x 3 matrix T for the elements, each column representing a different vehicle behavior b i Each row represents a respective driving sub-prediction field; converting matrix T into vector form
Adopts the maximum value method for->
Normalization process->
The input vector is found to be x= (x 1 ,x 2 ,……,x 27 ) The method comprises the steps of carrying out a first treatment on the surface of the Hidden layer jth neuron outputs as
Wherein the connection weight, a, between the input layer neuron i and the hidden layer neuron j is represented j Representing the bias value of hidden layer neuron j, f () is the activation function of hidden layer neuron, and adopts Sigmoid function in the form of
For the vehicle behavior decision types 1-9, the 'one hot' is adopted for encoding, the number of neurons of an output layer is 9, the output vector y= (y 1, y2, … …, y 9), and the output of the kth neuron of the output layer is
Wherein mu jk Representing the connection weights between hidden layer neuron j and output layer neuron k, b k Representing the bias value of the output layer neuron j.
5. The intelligent vehicle driving behavior personification decision method based on the driving prediction field and the BP neural network according to claim 4, wherein the BP neural network training method adopts a gradient descent method, and specifically comprises the following steps:
initializing a BP network, namely selecting a learning rate eta=0.35 for random small quantity between a connection weight value between layers and a threshold value (-1, 1) of hidden layer and output layer nodes;
step (2) calculating h for each sample j And y k Obtaining a mean square error
Wherein o is k Deciding a result vector element value for actual driving behavior;
step (3), updating the connection weight and the bias value through error back propagation, so that the error function value is reduced; connection weight mu jk 、ω ij And bias b k 、a j Updating; the method comprises the following steps:
a. implicit layer to output layer weight update gradient
Then weight mu jk The updated formula of (1) is mu jk ＝μ jk +ηh j e k ；
b. Input layer to hidden layer weight update gradient
Then weight omega ij The updated formula of (2) is
c. Implicit layer to output layer bias update gradient
Bias b k The updated formula of (2) isb k ＝b k +ηe k ；
d. Bias update gradient from input layer to hidden layer
Then bias a j The updated formula of (2) is
And (4) setting training precision epsilon=0.01, repeatedly inputting samples in the sample set L, executing algorithm calculation error E (L), and when the difference between two adjacent iteration errors E (L) is lower than epsilon or the iteration times reach 100 times, stopping updating the connection weight and the bias to obtain a trained BP neural network driving behavior decision model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910071774.6A CN109726804B (en) | 2019-01-25 | 2019-01-25 | Intelligent vehicle driving behavior personification decision-making method based on driving prediction field and BP neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910071774.6A CN109726804B (en) | 2019-01-25 | 2019-01-25 | Intelligent vehicle driving behavior personification decision-making method based on driving prediction field and BP neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109726804A CN109726804A (en) | 2019-05-07 |
| CN109726804B true CN109726804B (en) | 2023-06-13 |
Family
ID=66300285
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910071774.6A Active CN109726804B (en) | 2019-01-25 | 2019-01-25 | Intelligent vehicle driving behavior personification decision-making method based on driving prediction field and BP neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109726804B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110347155B (en) * | 2019-06-26 | 2020-11-06 | 北京理工大学 | Intelligent vehicle automatic driving control method and system |
| CN110377982B (en) * | 2019-07-01 | 2023-07-11 | 深圳中集智能科技有限公司 | Automatic driving performance testing method and system, electronic equipment and storage medium |
| CN110427536B (en) * | 2019-08-12 | 2022-03-04 | 深圳忆海原识科技有限公司 | Brain-like decision and motion control system |
| CN110834644B (en) * | 2019-10-30 | 2021-01-19 | 中国第一汽车股份有限公司 | Vehicle control method and device, vehicle to be controlled and storage medium |
| CN113741412B (en) * | 2020-05-29 | 2023-09-01 | 杭州海康威视数字技术股份有限公司 | Control method and device for automatic driving equipment and storage medium |
| CN111753371B (en) * | 2020-06-04 | 2024-03-15 | 纵目科技（上海）股份有限公司 | Training method, system, terminal and storage medium for vehicle body control network model |
| CN114148344B (en) * | 2020-09-08 | 2023-06-02 | 华为技术有限公司 | Vehicle behavior prediction method and device and vehicle |
| CN112406694A (en) * | 2020-11-16 | 2021-02-26 | 东风汽车集团有限公司 | Automatic control system and method for automobile steering lamp |
| CN112967493A (en) * | 2021-01-16 | 2021-06-15 | 北京工业大学 | Neural network-based prediction method for vehicle passing intersection travel time |
| CN113033902B (en) * | 2021-03-31 | 2024-03-19 | 中汽院智能网联科技有限公司 | Automatic driving lane change track planning method based on improved deep learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9434389B2 (en) * | 2013-11-18 | 2016-09-06 | Mitsubishi Electric Research Laboratories, Inc. | Actions prediction for hypothetical driving conditions |
| CN107609502A (en) * | 2017-09-05 | 2018-01-19 | 百度在线网络技术（北京）有限公司 | Method and apparatus for controlling automatic driving vehicle |
-
2019
- 2019-01-25 CN CN201910071774.6A patent/CN109726804B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN109726804A (en) | 2019-05-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109726804B (en) | Intelligent vehicle driving behavior personification decision-making method based on driving prediction field and BP neural network | |
| CN110796856B (en) | Vehicle lane change intention prediction method and training method of lane change intention prediction network | |
| Duan et al. | Hierarchical reinforcement learning for self‐driving decision‐making without reliance on labelled driving data | |
| CN108595823B (en) | Autonomous main vehicle lane changing strategy calculation method combining driving style and game theory | |
| Chen et al. | Stabilization approaches for reinforcement learning-based end-to-end autonomous driving | |
| Hang et al. | Driving conflict resolution of autonomous vehicles at unsignalized intersections: A differential game approach | |
| CN111222630A (en) | Autonomous driving rule learning method based on deep reinforcement learning | |
| JP2020059497A (en) | Method and device of controlling automatic driving introducing "detection acquisition" model, and program constituted thereby | |
| CN112183288B (en) | Multi-agent reinforcement learning method based on model | |
| CN112700642B (en) | Method for improving traffic passing efficiency by using intelligent internet vehicle | |
| CN109727490A (en) | A kind of nearby vehicle behavior adaptive corrective prediction technique based on driving prediction field | |
| Ye et al. | Meta reinforcement learning-based lane change strategy for autonomous vehicles | |
| Liu et al. | A three-level game-theoretic decision-making framework for autonomous vehicles | |
| Tang | Towards learning multi-agent negotiations via self-play | |
| Wang et al. | High-level decision making for automated highway driving via behavior cloning | |
| CN113901718A (en) | Deep reinforcement learning-based driving collision avoidance optimization method in following state | |
| CN114399743A (en) | Method for generating future track of obstacle | |
| Vasquez et al. | Multi-objective autonomous braking system using naturalistic dataset | |
| Yuan et al. | Multi-reward architecture based reinforcement learning for highway driving policies | |
| Sun et al. | Human-like highway trajectory modeling based on inverse reinforcement learning | |
| Liu et al. | Impact of sharing driving attitude information: A quantitative study on lane changing | |
| Yun et al. | Parallelized and randomized adversarial imitation learning for safety-critical self-driving vehicles | |
| Jia et al. | Interactive decision-making with switchable game modes for automated vehicles at intersections | |
| Hou et al. | Hybrid residual multiexpert reinforcement learning for spatial scheduling of high-density parking lots | |
| Gutiérrez-Moreno et al. | Hybrid decision making for autonomous driving in complex urban scenarios |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |