CN106485090A - A kind of Driver Vision perception forecast model generation method based on recurrence learning - Google Patents

Abstract

The present invention discloses a kind of perception forecast model generation method of the Driver Vision based on recurrence learning, including：Obtain sample data, with point of fixation, speed and the relation in dynamic vision and speed and the visual field, regression modeling is carried out to speed, described regression modeling comprises：Logarithm regression modeling, index return modeling, polynomial regression modeling, linear regression modeling；Regression diagnostics is carried out to described different regression modeling, draws best forecast model.Using technical scheme, by to the knowledge acquisition of driving behavior and expression, making intelligent vehicle, there is more preferable autonomy and intelligent, the theory of cognitive science is applied to driving behavior research field, the inherent mechanism of analysis driver behavior, can preferably explanation and understanding intelligent driving behavior.

Description

A kind of Driver Vision perception forecast model generation method based on recurrence learning

Technical field

The invention belongs to visually-perceptible field, more particularly, to a kind of Driver Vision perception prediction mould based on recurrence learning Type generation method.

Background technology

Driver when driving, relies primarily on sensory organ eye, ear, nose, tongue, skin and perceives in time and obtain each Plant transport information.Acquisition information is not enough or misses necessary information, is likely to lead to vehicle accident to occur.Therefore obtain information be Most important psychological process in driving behavior.Human perceptual model mainly solves three problems：First in actual driving-activity In, the visual attention location focus of driver is where；The visual range of kid has much.Driver is total in straight-line travelling It is on the road surface focus attention on front, and speed is higher, note more concentrating, be more difficult to shift.Carrying with speed Height, can make driver's seat narrow, the point of fixation of driver also moves forward simultaneously.Dynamic in 3rd driver's startup procedure regards Power is much.Driver passes through the information being continually changing in visually-perceptible highway environment in driving procedure, and carries out corresponding Speed controlling is to ensure the purpose driving safely.In driving, the referred to as dynamic vision of vision of human eye.Driver and outside ambient ring Border is in relative motion, and vision occurs decline phenomenon.With the increase of speed, the amplitude of dynamic visual deterioration can increase, car In the case of running at high speed, the reduction of dynamic vision can cause certain impression to human eye resolution distance, and resultant interference is driven The perception of member.

For point of fixation, when driving, the kinetic energy that becomes of point of fixation shows the selection of vision attention well, continues And transfer.Selective attention is a key character of mankind's natural vision：One side natural vision is not to complete in scene That portion's object is carried out is careful, identify exactly, often simply purposeful, selectively the subregion in scene is carried out perception with Understand, pay close attention to interesting target；Another aspect selective attention time to time change, that is, nonuniform sampling, natural vision Sample frequency to scene is not unalterable, but changes in time.According to psychophysicss law weber-Fei Xi Receive law S=k₁Log R=k₂Ln R (wherein S is sensation intensity, and R is stimulus intensity, and K is constant).The all feeling of people, Including vision, audition, dermal sensation (containing pain, itch, touch, temperature), the sense of taste, olfactory sensation, electric shock feel etc., all defer to feel be not with corresponding The intensity of physical quantity is directly proportional, but is directly proportional with the common logarithm of the intensity of corresponding physical quantity.

Research shows, the acquisition of human eye retina's information to external world is heterogeneous, that is, foveal region of retina have one high " macular area " (fovea) of resolution, and the resolution of macular area periphery gradually reduces with the increase of centre distance, and regard The mapping of nethike embrane to strip epidermis (Striate Cortex) can be with log-polar coordinate mapping Lai approximate.Log-polar is one Individual conformal projection, it is heterogeneous to live sampling, so can reduce the information of redundancy for characteristic target.And reflect Penetrate with ratio and rotational invariance, this characteristic based on log-polar for the present invention is to speed and point of fixation and speed Carried out the nonlinear regression of machine learning-logarithm regression modeling with dynamic vision, speed and the visual field, and with linear regression, multinomial Formula returns and index return modeling is compared, and draws best forecast model.

Content of the invention

It is an object of the invention to provide one kind is with regard between speed and the visual field, speed and dynamic vision and speed and point of fixation The Driver Vision perception forecast model generation method of relation, to solve intelligent driving behavior problem.Intelligent driving behavior analysiss The always key of intelligent vehicle research and difficulties, and the sense of the stability of intelligent driving ride-on vehicles and comfortable type and people Feel that characteristic is relevant.

For achieving the above object, the present invention adopts the following technical scheme that：

A kind of Driver Vision perception forecast model method based on recurrence learning, comprises the steps：

Step one, acquisition sample data, the relation to speed and point of fixation, speed and dynamic vision and speed and the visual field Carry out regression modeling, described regression modeling comprises：Logarithm regression modeling, index return modeling, polynomial regression model, linearly return Return modeling；

Step 2, regression diagnostics is carried out to described different regression modeling, draw best forecast model.

Preferably, it is as follows with regard to the relation between speed and the visual field, speed and dynamic vision and speed and point of fixation：

Relation between speed and the visual field is：

v_s=311.331-58.557lnv_f,

Wherein, v_sRepresent speed, v_fRepresent the visual field.

Relation between speed and point of fixation is：

v_s=-235.312+52.856lnf_p,

Wherein, v_sRepresent speed, f_pRepresent point of fixation.

Relation between speed and dynamic vision is：

v_s=46.982-120.071lnv_d,

Wherein, v_sRepresent speed, v_dRepresent dynamic vision.

Preferably, step one specifically includes：

Step 1, calculate the correlation coefficient r of dependency relation between variable in regression analyses_XY,

Wherein, (X_i, Y_i) represent i-th sample,By speed and the visual field, speed and watch attentively X in observation sample table in point and speed and dynamic vision_i、Y_i、Respectively substitute into above formula come to judge independent variable with because become If appropriate for doing linear regression between amount, if r_XY>Following relation is met between 0.5 X and Y：Y=alpha+beta X, if the value of α and β is Know, then provide corresponding X value, the predictive value y of corresponding Y can be obtained according to Y ≈ alpha+beta X_i.

Step 2, using square error and weigh predictive value y_iWith actual value y_iGap, if actual value is y, predictive value isThen square error is exactlyFind suitable parameter so that square error andMinimum：

RSS is the function with regard to α and β, respectively α and β is sought local derviation and makes local derviation be equal to 0 it is possible to draw the value of α and β：

It is estimated as using sample value：

Hence for each x_i, can pass throughPredict corresponding y value, be more than to utilize linear regression method The linear regression model (LRM) obtaining；

Step 3, by speed and the visual field, speed and point of fixation and speed and dynamic vision relation Using different functions such as logarithmic function, index letter Number, multinomial are modeled trying to achieve different regression functions respectively.

Preferably, regression diagnostics comprises in step 2：Judgment sample whether meet normal distribution, judge whether from Group's value, judge that regression model is whether reasonable and whether error in judgement meets normal distribution hypothesis.

The present invention is by the knowledge acquisition of driving behavior and expression, making intelligent vehicle have more preferable autonomy and intelligence Energy property, the theory of cognitive science is applied to driving behavior research field, analyzes the inherent mechanism of driver behavior, can be inherently Preferably explanation and understanding intelligent driving behavior.

Brief description

Fig. 1 is that the Driver Vision based on recurrence learning for the present invention perceives forecast model method flow diagram；

Fig. 2 is speed and visual field exception Distribution value；

Fig. 3 is speed and point of fixation exception Distribution value；

Fig. 4 is speed and dynamic eyesight abnormality Distribution value.

Specific embodiment

As shown in figure 1, the embodiment of the present invention discloses a kind of perception of the Driver Vision based on recurrence learning forecast model side Method, comprises the steps：

Step one, acquisition sample data, the relation to speed and point of fixation, speed and dynamic vision and speed and the visual field Carry out regression modeling, described regression modeling comprises：Logarithm regression modeling, index return modeling, polynomial regression model, linearly return Return modeling；

Step 2, regression diagnostics is carried out to described different regression modeling, draw best forecast model.

Described step 2 specifically includes following steps：

S1, for the recurrence side between the above-mentioned speed drawing and the visual field, speed and dynamic vision and speed and point of fixation Cheng Jinhang regression diagnostics, first has to whether judgment sample meets normal distribution, substantially conforms to normal distribution through judgment sample. (judgement carrying out sample normal distribution is mainly used for the F inspection followed by regression coefficient test and regression equation inspection institute Test and T inspection).

Step S2, outlier to be judged whether lead to model to produce larger error.It is based on for doing regression analyses These sample datas all by measurement, sampling get since measurement get, data will produce error, some Error is rational, and model will not be had a great impact, and the error of some data is larger, and the error of some data is less, number Can cancel out each other according to more and more, positive errors and negative error and deviate very big error it is also possible to exist, generation deviates normal It is worth far data, be called outlier.

As shown in Fig. 2 there being two point deflection path.However, this two points are without departing from a lot.It can be said that, do not have Have and there is outlier data model and have a huge impact.

As shown in figure 3, there being a point deflection path, equally it is also without departing from a lot.Therefore, it can draw such knot By：There is not exceptional value model is had a great impact.

As shown in figure 4, there is not an original path of deviation.Therefore, it can conclude that：Do not deposit exceptional value pair Model produces impact.

Rationally whether, it is not necessarily linear that nature has a lot of relations for step S3, regression model to be judged.Possibly secondary Multinomial is it may be possible to exponential function is also likely to be other more complicated relation, or the expression formula that can not write out parsing.How to sentence Whether disconnected regression model is reasonable？Whether model is rationally also performed to further statistical test.The inspection being related to has F to check And T inspection.Here with dynamic vision relation, logarithm letter has all been carried out with the visual field, speed and point of fixation and speed respectively to speed Number, exponential function, multinomial, one-variable linear regression regression modeling, and by the hypothesis testing of regression coefficient and recurrence side The significance test of journey compares judgement.Finally draw and use logarithmic function between speed and point of fixation, speed and dynamic vision The significance being returned is significantly many compared to other exponential functions, multinomial etc., and uses unitary between speed and the visual field The significance that linear function is returned is significantly many compared to logarithmic function, exponential function etc..In order to be better understood from following table, Relevant physical quantity is explained as follows：

Significance：It is significance labelling.The quantity of " * " is more, and reliability is higher.* * represents highly significant；* represents height Degree is notable；* represent notable.* quantity is also related to Pr value.(statistics is constructed according to theory of probability and mathematics statistical knowledge Amount t.Then the value of calculation Pr value corresponding with T value).If labelling 1* Pr value is 0.05, equally, if Pr value is about 0.01 We are labeled as 2*；If Pr value is about 0.001, it is labeled as 3*.For the convenience of expression, if the quantity of * is 0, use “0*”.If the number of * is 2, use " 2* " labelling.If the number of * is 3, using 3* labelling.

P-value：According to probability mathematical statistics, construct a statistic " T ".Then calculate T value corresponding Pr region.The area in this region is exactly P value.This p value is less, and result is better.

Correlation coefficient square：The value of correlation coefficient square is higher, represents corresponding model also more accurate.

For significance, it will be clear that linear regression function and logarithmic function are better than multinomial from upper table Function.Then eliminate polynomial function.As for correlation coefficient square, correlation coefficient square value is bigger, and representative is corresponding model Probability higher.Can see that linear regression function is better than logarithmic function.However, according to the knowledge of statistics it is impossible to all Depend on the result of calculation of used software.The theory of forefathers be also us it is also contemplated that emphasis, according to Weber Fechner Law, all feeling of people, including vision, audition etc., all deferring to sensation is not to be directly proportional with the intensity of corresponding physical quantity, and Be with corresponding physical quantity the common logarithm of intensity be directly proportional.Importantly, for logarithmic function, correlation coefficient is put down Side is 0.9566.For to a certain extent, this is that comparison is high.It follows that using logarithmic function v_s=311.331- 58.557lnv_f(v_sRepresent the speed of vehicle, v_fRepresent field range) can preferably represent the relation of speed and the visual field.

For significance, can see from upper, be better than linear regression function and index letter using logarithmic function modeling Number.And for correlation coefficient square, also see that and be higher than with linear regression and exponential function with the value of logarithmic function modeling The value of modeling.Therefore it may be concluded that using logarithmic function v_{s=-235.3+1252.8}f_p(v_sThe car speed representing, f_pRepresent note View distance) can preferably represent the relation of speed and point of fixation.

For significance, it will be clear that modeled excellent with linear regression modeling and logarithmic function from upper table In polynomial function modeling.Then eliminate Polynomial modeling.For correlation coefficient square, the value of correlation coefficient square is got over Greatly, the probability being modeled with corresponding function is higher.Can see that logarithmic function is better than linear regression.Therefore, it can Go out conclusion, it is preferably using v_s=46.982-120.071lnv_dTo express the relation between car speed and dynamic vision.

Step S4, want error in judgement whether meet independence, etc. variance (in theory with that dependent variable it does not matter whether , error will not change with the size of dependent variable), the assumed condition such as normal distribution.

The present invention is by the knowledge acquisition of driving behavior and expression, making intelligent vehicle have more preferable autonomy and intelligence Property, the theory of cognitive science is applied to driving behavior research field, analyzes the inherent mechanism of driver behavior, can inherently more Good explanation and understanding intelligent driving behavior.

Claims (4)

1. a kind of Driver Vision perception forecast model method based on recurrence learning is it is characterised in that comprise the steps：

Step one, acquisition sample data, are carried out with point of fixation, speed and the relation in dynamic vision and speed and the visual field to speed Regression modeling, described regression modeling comprises：Logarithm regression modeling, index return modeling, polynomial regression modeling, linear regression are built Mould；

Step 2, regression diagnostics is carried out to described different regression modeling, draw best forecast model.

2. Driver Vision according to claim 1 perception forecast model generation method it is characterised in that with regard to speed with Relation between the visual field, speed and dynamic vision and speed and point of fixation is as follows：

Relation between speed and the visual field is：

v_s=311.331-58.557lnv_f,

Wherein, v_sRepresent speed, v_fRepresent field range.

Relation between speed and point of fixation is：

v_s=-235.312+52.856lnf_p,

Wherein, v_sRepresent speed, f_pRepresent point of fixation.

Relation between speed and dynamic vision is：

v_s=46.982-120.071lnv_d,

Wherein, v_sRepresent speed, v_dRepresent dynamic vision.

3. Driver Vision perception forecast model generation method according to claim 1 is it is characterised in that step one is concrete Including：

Step 1, calculate the correlation coefficient r of dependency relation between variable in regression analyses_XY,

$r_{XY}=\frac{\underset{i=1}{\overset{n}{\Σ}}(X_i-\stackrel{\‾}{X})(Y_i-\stackrel{\‾}{Y})}{\sqrt{\underset{i=1}{\overset{n}{\Σ}}{(X_i-\stackrel{\‾}{X})}^2\underset{i=1}{\overset{n}{\Σ}}{(Y_i-\stackrel{\‾}{Y})}^2}}$

Wherein, (X_i, Y_i) represent i-th sample,By speed and the visual field, speed and point of fixation and X in observation sample table in speed and dynamic vision_i、Y_i、Substitute into above formula respectively to judge between independent variable and dependent variable If appropriate for doing linear regression, if r_XY>Following relation is met between 0.5 X and Y：Y=alpha+beta X, if the value of α and β it is known that, give Go out corresponding X value, the predictive value y of corresponding Y can be obtained according to Y ≈ alpha+beta X_i.

${\hat{y}}_i=\mathrm{\α}+{\mathrm{\βX}}_i;$

Step 2, using square error and weigh predictive value y_iWith actual value y_iGap, if actual value is y, predictive value isThen Square error is exactlyFind suitable parameter so that square error andMinimum：

RSS is the function with regard to α and β, respectively α and β is sought local derviation and makes local derviation be equal to 0 it is possible to draw the value of α and β：

$\mathrm{\β}=\frac{{\Σ}_{i=1}^n(X_i-\stackrel{\‾}{X})(Y_i-\stackrel{\‾}{Y})}{{\Σ}_{i=1}^n{(X_i-\stackrel{\‾}{X})}^2}$

$\mathrm{\α}=\stackrel{\‾}{Y}-\mathrm{\β}\stackrel{\‾}{X})$

It is estimated as using sample value：

$b=\stackrel{\‾}{\mathrm{\β}}=\frac{{\Σ}_{i=1}^n(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})}{{\Σ}_{i=1}^n{(x_i-\stackrel{\‾}{x})}^2}$

$a=\widehat{\mathrm{\α}}=\stackrel{\‾}{y}-b\stackrel{\‾}{x}$

Hence for each x_i, can pass throughPredict corresponding y value, be more than to be obtained using linear regression method Linear regression model (LRM)；

Step 3, by the X in speed and the visual field, speed and point of fixation and speed and dynamic vision relation_i(i=1,2,3,4,5,6),Y_i(i=1,2,3,4,5,6),It is modeled asking respectively using different functions such as logarithmic function, exponential function, multinomial Obtain different regression functions.

4. Driver Vision perception forecast model generation method according to claim 1 is it is characterised in that return in step 2 Diagnosis is returned to comprise：Judgment sample whether meet normal distribution, judge whether outlier, judge regression model whether rationally with And whether error in judgement meets normal distribution and assumes.