CN103646533B - Traffic accident modeling and control method based on sparse multi-output regression - Google Patents

Abstract

The present invention discloses a kind of traffic accident modeling and control method based on sparse multi-output regression, can be used to the situation of certain selected area generation traffic accident to set up model. Multi-output regression hazard model is set up in certain area, can automatically detect the characteristic parameter irrelevant with accident and distinguish the impact of parameters, bring the characteristic parameter relevant to traffic into model, can also predict the following situation that traffic accident occurs in this area. The present invention utilizes the history and current situation data of traffic accident, the order of severity of traffic accident is carried out to modeling, the characteristic parameter that energy automatic rejection and accident are irrelevant, distinguishes the key factor that affects traffic safety, takes measure more targetedly while contributing to roading. This method is being predicted traffic accident, is analyzing cause of accident, is being instructed road construction improvement to have actual engineering using value.

Description

Traffic accident modeling and control method based on sparse multi-output regression

Technical field

The invention belongs to traffic intelligent management and control technology field, use the sparse multi-output regression model proposing, set up a kind of roadRoad traffic accident method, and can automatically detect the key factor that affects traffic safety, in prediction traffic accident, analysis thingTherefore reason, instruct road construction to improve to there is actual engineering using value.

Background technology

Along with the develop rapidly of China's road traffic cause, traffic accident is surged and is become the serious problem of society today receiving much concern nowOne of. Just because of the serious consequence that traffic accident causes, a large amount of human and material resources and wealth have poured into accident prevention and countermeasure in ChinaPower, formulated comparatively perfect control of traffic and road law, regulation and relevant policies, but the road traffic accident of China causesFactual Damage but still serious, and the trend in rising always, annual toll on traffic occupy first place, the world.Therefore, in order to ensure economic development and social stability, ensure people's safety and cargo transportation security, to Chinese traffic peaceHolotype gesture is made the prediction of science, becomes very necessary and extremely urgent.

At present, there is several different methods to be applied to accident modeling both at home and abroad. Existing research has been set up accident and has been occurred and traffic stream characteristics,Road environment condition, the size of population, the relation between the data such as vehicle guaranteeding organic quantity, but great majority are all to enter from macroscopical angleRow is analyzed, and the analysis result drawing is the prediction of the traffic accident number of times in a region, and this rises for prevention and minimizing traffic accidentLess than too large effect. Although also there is part Study impact on traffic accident from the factor of the angle analysis road of microcosmic,,Ignored that to affect the factor of traffic accident a lot, and some factors wherein exist as NF, if do not reject thisA little characteristic parameters and carry out traffic accident modeling, this may hide real feature and the inherent law of traffic accident.

Traffic accident is the complicated system being produced by many factors acting in conjunction, and relation between these factors is difficult to separatingThe method of analysing is described, and passes through the study to historical data based on sparse multi-output regression model, can build with degree of precisionThe model of vertical traffic accident.

For the accuracy of modeling, do not omit important information, we can consider as far as possible many factors, also must reject and thing simultaneouslyTherefore irrelevant characteristic parameter, therefore, the present invention has introduced sparse analytic approach, initial data is carried out to sparse analysis, thereby distinguishAffect the key factor of traffic safety, carry out quantitative analysis thereby effectively utilize a large amount of statistics, disclose between variableInternal relation, and the improvement that can be road construction provides targetedly and instructs.

Summary of the invention

The problem to be solved in the present invention is the above-mentioned deficiency that overcomes prior art, provides a kind of simple to operate, and accuracy is high, is suitable forThe method of traffic accident modeling in road safety checks.

Technical scheme of the present invention is:

A) obtain the historical traffic accident data sample N group { X in somewhere₁,X₂,...,X_N, each X_I（I=1,2,3,...,N）Comprise following population, vehicle, road and environmental information: total number of people x₁, schooling x₂, income level x₃, car steeringMember's quantity x₄, truck conspicuity marking quantity x₅, cyclecar quantity x₆, bicycle quantity x₇, main section total length x₈，Main road section traffic volume lamp quantity x₉, average weather conditions x₁₀。

B) obtain corresponding traffic accident result sample { Y₁,Y₂,...,Y_N, each Y_I(I=1,2,3 ..., N) comprise accident frequencyy₁, death toll y₂, number of injured people y₃And direct economic loss y₄。

C) traffic accident data pretreatment. By above-mentioned N group traffic accident data sample { X₁,X₂,...,X_NNormalization, and leave square inIn battle array X. Above-mentioned N is organized to corresponding traffic accident result sample { Y₁,Y₂,...,Y_NLeave in matrix Y.

D) supposition X_IWith Y_I(I=1,2,3 ..., N) meet and be related to Y_I=X_IW+ ε, W is coefficient, ε is Gaussian noise. If λ isBalance parameters, can be taken as 0.01≤λ≤100, and W can obtain by solving following sparse multi-output regression model

$\min J\left(W\right)=\underset{I=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|X_{I}W-Y_I\left|\right|}^2+\mathrm{\λ}{\left|\right|W\left|\right|}_{21}={\left|\right|\mathrm{XW}-Y\left|\right|}^2+\mathrm{\λ}{\left|\right|W\left|\right|}_{21}$

E) choose coefficient matrix W, make J (W) reach minimum of a value, establishing D is unit matrix.

F) by J (W) to W differentiation, and make result equal 0, obtain

X^T(XW-Y)+λDW=0

G) solve above formula and obtain W=(X^TX+λD)^-1X^TY。

If D' is diagonal matrix, diagonal element is the inverse of the L2 norm of W corresponding row.

H) whether evaluation algorithm restrains, || and whether D-D'|| is less than given several ε (as ε=0.001), if so, exports W,Otherwise D' is copied in D, forward step F to).

I) identify the factor irrelevant to traffic accident modeling according to the row that in W is entirely zero, according to the positive and negative of the capable value of non-zero in W andThe impact of the different factors of size discrimination on traffic accident. As driver's quantity x₄Coefficient for negative, road that should be afterwardsIn the traffic programme of road, increase driver's quantity to reduce following traffic accident.

J) according to the traffic accident characteristic parameter X'=(x in the above-mentioned area gathering later₁,x₂,...,x₁₀), predict that the accident in above-mentioned area occursNumber of times, death toll, number of injured people and direct economic loss are WX'.

Technique effect of the present invention is as follows:

Based on sparse multi-output regression model. More existing research methods directly utilize initial data to instruct as input variablePractice, but these initial data are deposited problem both ways, the one, data volume is large, and dimension is many, and the one, between data, there is correlation,Information has redundancy, and this will cause the reduction of the not high and training speed of the precision of result. Initial data is carried out to sparse analysis,The characteristic parameter that automatic rejection and accident are irrelevant, carries out quantitative analysis thereby effectively utilize a large amount of statistics, disclose variable itBetween internal relation.

Choose influence factor from the angle of microcosmic and carry out modeling. Traditional Predictive Methods of Road Accidents is all to utilize some macroscopical shadowsThe factor of sound is analyzed as input variable, and what obtain is the data of an overall traffic accident. The present invention is micro-by roadSight factor, as input data, is analyzed the relation of these factors and traffic accident, thus obtain these factors to traffic accident respectivelyFrom impact, this conclusion can be applied in Highway traffic safety assessment, for design and the trimming of road and surrounding enviroment provideFoundation.

Brief description of the drawings

Fig. 1 is the traffic accident model FB(flow block) based on sparse multi-output regression of the present invention.

Detailed description of the invention

At society, traffic safety gets more and more people's extensive concerning, and research that road safety is evaluated is for traffic accidentReduce and have very large effect. The present invention is from the angle of microcosmic, and the modeling method of application based on sparse multi-output regression is to a certain areaRelation etc. multiple variable and traffic accident is analyzed and researched.

Introduce in detail workflow of the present invention below:

One, choosing of data

Road traffic is the dynamical system being made up of key elements such as people, car, road, environment. Traffic accident is in road traffic system,The event accidentally happening suddenly due to the cooperation imbalance of people, car, road, all key elements of environment. Therefore, choosing the impact of traffic accidentBecause of the period of the day from 11 p.m. to 1 a.m, analyze from above aspect.

1. human factor: according to the statistics of traffic accidents result of 2002, the death toll causing due to people's reason accounts for worked as year traffic88.98% of death by accident sum, what wherein automobile driver was main cause accounts for 78.56%, and bicycle driver account for4.20%, pedestrian and rider account for 6.22%. Visible, in general, the key of traffic accident is automobile driver.Therefore chosen a certain regional periphery total number of people and average schooling and income, four indexs of driver's quantity.

2. the factor of car: causing the second largest key element of traffic accident is vehicle. The vehicle travelling on road, existing motor vehicle,Also have bicycle, wherein motor vehicle (especially truck conspicuity marking) is the one vehicles fast, energy maximum, protectionProperty is also best. But this protective is only protected driver and occupant, therefore, with respect to bicycle and other bicycles,Motor vehicle is traffic powerhouse. But bicycle quantity increase also can to traffic accident number exert an influence, therefore chosen largeType vehicles number, cyclecar quantity, three indexs of bicycle quantity.

3. road and environmental factor: the environmental factor that affects traffic safety can be divided into natural environment and artificial environment, natural environmentMainly comprise geographical position, meteorological condition, the time etc., artificial environment comprises that situation of land use, trackside disturb, road barricadeThing etc. The present invention chosen the main section total length in above-mentioned area, main road section traffic volume lamp quantity and average weather conditions this threeIndividual index is as factor of influence.

Two, set up road accident risk model:

A) obtain the historical traffic accident data sample N group { X in somewhere₁,X₂,...,X_N, each X_I(I=1,2,3 ..., N) comprisePopulation, vehicle, road and environmental information: total number of people x₁, schooling x₂, income level x₃, driver's quantity x₄，Truck conspicuity marking quantity x₅, cyclecar quantity x₆, bicycle quantity x₇, main section total length x₈, hand in main sectionLogical lamp quantity x₉, average weather conditions x₁₀。

B) obtain corresponding traffic accident result sample { Y₁,Y₂,...,Y_N, each Y_I(I=1,2,3 ..., N) comprise accident frequencyy₁, death toll y₂, number of injured people y₃And direct economic loss y₄。

C) traffic accident data pretreatment. By above-mentioned N group traffic accident data sample { X₁,X₂,...,X_NNormalization, and leave square inIn battle array X. Above-mentioned N is organized to corresponding traffic accident result sample { Y₁,Y₂,...,Y_NLeave in matrix Y.

D) supposition X_IWith Y_I(I=1,2,3 ..., N) meet and be related to Y_I=X_IW+ ε, W is coefficient, ε is Gaussian noise. If λ isBalance parameters, can be taken as 0.01≤λ≤100, and W can obtain by solving following sparse multi-output regression model

$\min J\left(W\right)=\underset{I=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|X_{I}W-Y_I\left|\right|}^2+\mathrm{\λ}{\left|\right|W\left|\right|}_{21}={\left|\right|\mathrm{XW}-Y\left|\right|}^2+\mathrm{\λ}{\left|\right|W\left|\right|}_{21}$

E) choose coefficient matrix W, make J (W) reach minimum of a value, establishing D is unit matrix.

F) by J (W) to W differentiation, and make result equal 0, obtain

X^T(XW-Y)+λDW=0

G) solve above formula and obtain W=(X^TX+λD)^-1X^TY。

If D' is diagonal matrix, diagonal element is the inverse of the L2 norm of W corresponding row.

H) whether evaluation algorithm restrains, || and whether D-D'|| is less than given several ε (as ε=0.001), if so, exports W,Otherwise D' is copied in D, forward step F to).

I) identify the factor irrelevant to traffic accident modeling according to the row that in W is entirely zero, according to the positive and negative of the capable value of non-zero in W andThe impact of the different factors of size discrimination on traffic accident. As driver's quantity x₄Coefficient for negative, road that should be afterwardsIn the traffic programme of road, increase driver's quantity to reduce following traffic accident.

J) according to the traffic accident characteristic parameter X'=(x in the above-mentioned area gathering later₁,x₂,...,x₁₀), predict that the accident in above-mentioned area occursNumber of times, death toll, number of injured people and direct economic loss are WX'.

Practice process is divided into model learning and model uses two processes.

Model learning: as Fig. 1, collect traffic accident data and corresponding traffic accident result data in a certain regional certain hour.The traffic accident function of setting up sparse multi-output regression according to aforementioned modeling procedure, calculates Y_I=X_IThe sparse matrix W of W+ ε.

Model uses: identify the factor irrelevant to traffic accident modeling according to the row that in W is entirely zero, and capable according to non-zero in WThe impact of positive and negative different with the size discrimination factors of value on traffic accident. As driver's quantity x₄Coefficient for negative, shouldIn traffic planninng afterwards, increase driver's quantity to reduce following traffic accident. And, according to same groundThe traffic accident feature X'=(x gathering in certain time of district₁,x₂,...,x₁₀), utilize the accident frequency of W to this area, deadThe number of dying, number of injured people and direct economic loss are predicted X'W.

Claims (3)

1. the traffic accident modeling method based on sparse multi-output regression, is characterized in that, concrete steps are:

A) obtain the historical traffic accident data sample N group { X in somewhere₁,X₂,...,X_N, each X_IComprise following population, vehicle,Road and environmental information: total number of people x₁, schooling x₂, income level x₃, driver's quantity x₄, large-scale motor-drivenCar quantity x₅, cyclecar quantity x₆, bicycle quantity x₇, main section total length x₈, main road section traffic volume lamp numberAmount x₉, average weather conditions x₁₀; Wherein, I=1,2,3 ..., N;

B) obtain corresponding traffic accident result sample { Y₁,Y₂,...,Y_N, each Y_IComprise accident frequency y₁, death tolly₂, number of injured people y₃And direct economic loss y₄；

C) traffic accident data pretreatment; By above-mentioned N group traffic accident data sample { X₁,X₂,...,X_NNormalization, and leave square inIn battle array X; Above-mentioned N is organized to corresponding traffic accident result sample { Y₁,Y₂,...,Y_NLeave in matrix Y;

D) supposition X_IWith Y_IMeet and be related to Y_I＝X_IW+ ε, W is coefficient matrix, ε is Gaussian noise; If λ is balance parameters,Be taken as 0.01≤λ≤100, W obtains by solving following sparse multi-output regression model

$\min J\left(W\right)=\underset{I=1}{\overset{N}{\Σ}}\left|\right|X_{I}W-Y_I|{|}^2+\mathrm{\λ}|\left|W\right|{|}_{21}=\left|\right|XW-Y|{|}^2+\mathrm{\λ}|\left|W\right|{|}_{21};$

E) choose coefficient matrix W, make J (W) reach minimum of a value, establishing D is unit matrix;

F) by J (W) to W differentiation, and make result equal 0, obtain

X^T(XW-Y)+λDW＝0；

G) solve above formula and obtain W=(X^TX+λD)^-1X^TY；

If D' is diagonal matrix, diagonal element is the inverse of the L2 norm of W corresponding row;

H) whether evaluation algorithm restrains, || and whether D-D'|| is less than given several ε, if so, exports W, otherwise D' is copiedIn D, forward step F to);

I) identify the factor irrelevant to traffic accident modeling according to the row that in W is entirely zero, according to the positive and negative of the capable value of non-zero in W andThe impact of the different factors of size discrimination on traffic accident;

J) according to the traffic accident characteristic parameter X'=(x in the above-mentioned area gathering later₁,x₂,…,x₁₀), predict that the accident in above-mentioned area is sent outRaw number of times, death toll, number of injured people and direct economic loss are WX'.

2. the traffic accident modeling method based on sparse multi-output regression according to claim 1, is characterized in that described stepH) in, given several ε: 0.001≤ε≤0.1.

3. the traffic accident modeling method based on sparse multi-output regression according to claim 1, is characterized in that described stepRapid I) in, the method for the impact according to the different factors of positive and negative and size discrimination of the capable value of non-zero in W on traffic accident is: ifThe value that in W, non-zero is capable is for just, and this correspondence factor has negative effect to accident forecast, should in the time of roading, consider to reduce thisFactor; If the value that in W, non-zero is capable is for negative, this correspondence factor has positive influences to accident forecast, should in the time of roading, examineConsider this factor that increases; As driver's quantity x₄Coefficient for negative, should in traffic planninng afterwards, increase vapourCar driver's quantity is to reduce following traffic accident.