CN100481153C - Method for automatically inspecting highway traffic event based on offset minimum binary theory - Google Patents

Abstract

A method for automatically detecting traffic event on expressway based on bias least squares principle includes collecting real time traffic flow information of detection region by detection system according to set step length through ground-sensing coil, judging whether traffic event is occurred on detection region or not by bias least squares regression algorithm based on collected information, issuing alarm and informing traffic management department to treat said event if it is or otherwise collecting data continuously and carrying out next judgment.

Description

Freeway traffic event automatic detection method based on offset minimum binary theory

Technical field

The present invention relates to the method that a kind of freeway traffic event detects automatically, be a kind of freeway traffic event automatic detection method specifically based on offset minimum binary theory, can be used for whether freeway traffic event existed and carry out real time automatic detection, relate to traffic intelligent management and control technology.

Background technology

Traffic events refers to recurrent non-repeatability incident, as traffic hazard, goods be scattered, vehicle casts anchor, pavement maintenance, signal or checkout equipment fault and some other interrupt normal traffic, cause crowded and incur loss through delay, reduce the special event of road passage capability.Because it is crowded that traffic events causes is unexpected, thereby very easily causes second accident, thereby reduces the road safety level.The whole world is annual because cap loss, property loss, personal injury that traffic events causes reach multi-billion dollar.In recent years, along with the raising of the growing with each passing day of socioeconomic fast development, vehicle population and path link car kilometer, running velocity, traffic events frequently takes place, and has become a serious social concern in China.

The early detection of traffic events can reduce delay significantly with shunting early, improves road safety and service level.Traffic events detects (Automatic Incident Detection automatically, AID) be the most basic most important component in the advanced traffic control and management system, its objective is by detecting and clear up traffic events as early as possible, make highway recover normal operation as early as possible, reducing and incur loss through delay, is one of effective way that solves highway burst traffic events.From the seventies in 20th century, the automatic detection of traffic events just is subjected to various countries traffic engineering expert's attention, becomes a research focus in Communication and Transportation Engineering field.

At present change the method that detects traffic events, because of its economic convenient main direction that becomes developing country and most of AID of developed country research based on traffic flow parameter.Invention " is adopted the freeway incident automatic checkout system of artificial neural networks and genetic algorithm ", and (CN 1311880A) discloses the design of the neural network of automatic transport detection incident on a kind of highway, the design of this invention utilization genetic algorithm optimization neural network.Many experimental results show, are better than other algorithm based on the detection algorithm performance of neural network.But neural network algorithm has following shortcoming, and neural network is the black box structure, and knowledge lies in the connection weight, is difficult to extract and understands, and restrain slowlyer, and restrains the setting that depends on learning parameter, has precocious phenomenon, easily sinks into local minimum etc.In addition, genetic algorithm itself except that consuming time, the problem that also has self parameter to select.

In recent years (Support Vector Machine SVM) discerns traffic behavior, thereby realizes the automatic detection of traffic events to have the scholar to propose to use support vector machine again.The kernel function of SVM and parameter thereof have a significant impact the classification performance of SVM, but, how selecting suitable kernel function and parameter thereof is a challenging job, also there is not at present the simple method of determining, generally be to grope by the experiment consuming time of a large amount of efforts, therefore, place one's entire reliance upon user's experience of the effect of SVM, this has just influenced it and has detected the ability of traffic events.

Summary of the invention

Technical matters to be solved by this invention is, overcomes technological deficiency and deficiency that prior art exists, and a kind of freeway traffic event automatic detection method based on offset minimum binary theory is provided, and its verification and measurement ratio, the alert rate of mistake and classification rate all are better than prior art.

The present invention is based on the freeway traffic event automatic detection method of offset minimum binary theory, may further comprise the steps:

----detection system according to setting step-length acquisition testing district real-time traffic stream information, comprises the upstream and downstream magnitude of traffic flow, speed, lane occupancy ratio, time headway by ground induction coil or video equipment, and sampling step length can be made as 30 seconds or 1 minute;

----judges according to the real-time traffic stream information of gathering whether detection zone traffic events takes place, if take place, give the alarm, and the notice vehicle supervision department removing incident that takes appropriate measures, otherwise the continuation image data is judged next time;

It is characterized in that: described real-time traffic stream information according to collection judges whether detection zone traffic events takes place, and is to adopt the partial least squares regression detection algorithm to judge whether detection zone traffic events takes place.

Partial least-square regression method and common least square regression are that it has adopted informix and triage techniques in the regression modeling process in the key distinction on the thinking.It at first extracts some new generalized variables (promptly extracting composition) that system is had the best interpretations ability, utilizes the extraction composition to carry out regression modeling then.

If known dependent variable y and k independent variable x ₁, x ₂..., x _k, sample number is n, composition data Table X=[x ₁, x ₂..., x _k] _N*kAnd y=[y] _N*1The modeling procedure of partial least squares regression comprises:

(1) X and y are carried out standardization, obtain the independent variable matrix E after the standardization ₀And F ₀,

${x_{\mathrm{ij}}}^{*}=\frac{x_{\mathrm{ij}}-\stackrel{\‾}{x_j}}{s_j},$ $y_j^{*}=\frac{y_i-\stackrel{\‾}{y}}{s_y},i=\mathrm{1,2},...,k;j=\mathrm{1,2},...,k---\left(1\right)$

$E_0={\left(x_{\mathrm{ij}}^{*}\right)}_{n*k},$ $F_0={\left(y_j^{*}\right)}_{n*1},i=\mathrm{1,2},...,n$

In the formula,

Be X _jAverage, s _jBe X _jStandard deviation; Y is the average of y; s _yIt is the standard deviation of y.

Make h=1.

(2) compute vector weight w _h,

$w_h=E_{h-1}\′F_{h-1}---\left(2\right)$

(3) extract composition t _h

t _h＝E _h-1w _h (3)

(4) the regression coefficient p of calculating X and y _h, q _h

$p_h=E_{h-1}\′t_h/\left(t_h^{\′}{t}_h\right)---\left(4\right)$

$q_h=F_{h-1}\′t_h/\left(t_h^{\′}{t}_h\right)---\left(5\right)$

(5) calculate residual matrix

$E_h=E_{h-1}-t_h{p}_h^{\′}$

$F_h=F_{h-1}-q_h^{\′}{t}_h---\left(6\right)$

(6) check convergence, available intersection validity is determined.If above-mentioned equation satisfies accuracy requirement, change next step.Otherwise, h=h+1, repeating step (2)-(5), the composition that residual matrix is carried out a new round extracts and regretional analysis.

(7) establish and obtain k composition t ₁, t ₂..., t _k, implement F ₀At t ₁, t ₂..., t _kOn recurrence,

F ₀＝q ₁t ₁+q ₂t ₂+...+q _kt _k (7)

Because t ₁, t ₂..., t _kAll are E ₀Linear combination, therefore,

$F_0=q_1{E}_0{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="C200710023009E00121.png"\; state="\{\{state\}\}">w</figure-callout>}}_1+q_2{E}_1{w}_2+...+q_k{E}_{k-1}{w}_k=q_1{E}_0{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="C200710023009E00121.png"\; state="\{\{state\}\}">w</figure-callout>}}_1^{*}+...+q_k{E}_0{w}_k^{*}---\left(8\right)$

In the formula, $w_h^{*}=\underset{j=1}{\overset{h-1}{\mathrm{\&Pi;}}}(I-w_j{p}_j^{\&prime;})w_h,$ I is a unit matrix.

(8) be reduced into the regression equation of y according to standardized inverse process to X,

$y=\stackrel{\&OverBar;}{y}+s_y\left(\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{x}_i^{*}\right)=\stackrel{\&OverBar;}{y}+s_y\left(\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i\frac{x_i-{\stackrel{\&OverBar;}{x}}_i}{s_i}\right)$

${\mathrm{\&alpha;}}_i=\underset{h=1}{\overset{m}{\mathrm{\&Sigma;}}}{r}_h{w}_{\mathrm{hi}}^{*}---\left(9\right)$

Make X represent the traffic characteristic parameter of gathering, as the magnitude of traffic flow, occupation rate, speed and time headway etc., y represents traffic behavior, value 1 or-1, and presentation of events takes place or does not have an incident respectively.According to the transport information of current collection, utilization following formula (9) calculates y, can judge having or not of traffic events.

Therefore, the concrete steps that adopt the partial least squares regression detection algorithm to judge whether detection zone traffic events takes place are:

1, establish X and represent the traffic characteristic parameter of gathering, as the magnitude of traffic flow, occupation rate, speed and time headway etc., y represents traffic behavior, value 1 or-1, and presentation of events takes place or does not have an incident respectively.With known traffic behavior y=[y] _N*1Telecommunication flow information X=[x ₁, x ₂..., x _k] _N*k, set up (demarcation) partial least squares regression mathematical model

$y=\stackrel{\&OverBar;}{y}+s_y\left(\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{x}_i^{*}\right)=\stackrel{\&OverBar;}{y}+s_y\left(\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i\frac{x_i-{\stackrel{\&OverBar;}{x}}_i}{s_i}\right)$

${\mathrm{\&alpha;}}_i=\underset{h=1}{\overset{m}{\mathrm{\&Sigma;}}}{r}_h{w}_{\mathrm{hi}}^{*}$

2, the detection zone real-time traffic stream information substitution institute established model that will gather in real time determines according to the y value current whether event occurs.If y is not more than 0, be the normal traffic running status in the expression detection zone, otherwise expression generation incident.

The advantage of partial least squares regression model is fast convergence rate, need not parameter adjustment, and can show the relative importance of different variablees, and, implement more or less freely owing to reduced calculated amount.The present invention uses the relation between partial least squares regression model representation incident and the traffic flow, model output is compared with predefined thresholding, thereby judged whether that event occurs.This method verification and measurement ratio height, the alert rate of mistake reduces greatly, detects performance and obviously improves.Therefore, use partial least squares regression model (PLSR) to detect traffic events and have application promise in clinical practice.

Description of drawings

Fig. 1 is that highway regional upstream and downstream to be detected is arranged the checkout equipment synoptic diagram;

Fig. 2 is based on the freeway traffic event automatic detection method flow chart of offset minimum binary theory;

The modeling program block diagram of Fig. 3 partial least squares regression;

Embodiment

Below in conjunction with drawings and Examples, the present invention is described in further detail.

Embodiment: as shown in Figure 1, arrange checkout equipment in highway regional upstream and downstream to be detected.Detecting device can be ground induction coil or video capture device, thinks certain sample unit (as 30 seconds) collection magnitude of traffic flow, occupation rate, speed.Make that X is above-mentioned traffic characteristic parameter, y is a traffic behavior, and the state-event value is 1, and non-state-event value is-1, then,

$X=\left[\begin{array}{cccc}{x}_1& x_2& \&CenterDot;\&CenterDot;\&CenterDot;& x_6\end{array}\right]$

$=\left[\begin{array}{cccccc}{\mathrm{speed}}_{\mathrm{up}1}& {\mathrm{occupancy}}_{\mathrm{up}1}& {\mathrm{volume}}_{\mathrm{up}1}& {\mathrm{<figure-callout\; id="1"\; label="speed\; dn"\; filenames="C200710023009E00121.png"\; state="\{\{state\}\}">speed</figure-callout>}}_{\mathrm{dn}}& {\mathrm{<figure-callout\; id="1"\; label="occupancy\; dn"\; filenames="C200710023009E00121.png"\; state="\{\{state\}\}">occupancy</figure-callout>}}_{\mathrm{dn}}& {\mathrm{<figure-callout\; id="1"\; label="volume\; dn"\; filenames="C200710023009E00121.png"\; state="\{\{state\}\}">volume</figure-callout>}}_{\mathrm{dn}}\\ {\mathrm{speed}}_{\mathrm{up}2}& {\mathrm{occupancy}}_{\mathrm{up}2}& {\mathrm{volume}}_{\mathrm{up}2}& {\mathrm{<figure-callout\; id="1"\; label="speed\; dn"\; filenames="C200710023009E00121.png"\; state="\{\{state\}\}">speed</figure-callout>}}_{\mathrm{dn}}& {\mathrm{occupancy}}_{\mathrm{dn}2}& {\mathrm{volume}}_{\mathrm{dn}2}\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ {\mathrm{speed}}_{\mathrm{upn}}& {\mathrm{occupancy}}_{\mathrm{upn}}& {\mathrm{volume}}_{\mathrm{upn}}& {\mathrm{speed}}_{\mathrm{dnn}}& {\mathrm{occupancy}}_{\mathrm{dnn}}& {\mathrm{volume}}_{\mathrm{dnn}}\end{array}\right]$

$Y=\left[\begin{array}{c}{y}_1\\ y_2\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ y_n\end{array}\right]=\left[\begin{array}{c}1/-1\\ 1/-1\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ 1/-1\end{array}\right]$

Actual detected is resolved process and is divided into model calibration and two processes of model use.

Model calibration: collect or the traffic data in acquisition testing district a period of time, contain incident sample and non-incident sample.In order to guarantee the accuracy of model, can portray the relation between detected parameters and the incident, sample should be enough big.If collect n sample arranged,, thereby obtain the partial least squares regression model according to the parameter of aforementioned modeling procedure computing formula.In order to reduce rate of false alarm, can use T-2, T-1 and the T magnitude of traffic flow, occupation rate, speed is constantly set up the partial least squares regression model, and this moment, X was the two-dimensional matrix of n*18, and y is still the n*1 matrix.

Model uses: the magnitude of traffic flow, occupation rate, the speed of acquisition testing district upstream and downstream in real time, with its substitution institute established model, definite current whether event occurs according to y value.If y is not more than 0, be the normal traffic running status in the expression detection zone, otherwise expression generation incident.For the consideration that reduces rate of false alarm, can adopt the continuation test, have only double model output all greater than 0, just trigger event alarm, the supvr of notice traffic department carries out events corresponding and handles.

Sample calculation analysis

This experiment is tested the performance that partial least square method detects traffic events with gathering the true traffic data (I-880 data) that comes on the I-880 highway in SF Bay area, California.This database has been used in a lot of other similar event detection researchs.Data comprise the magnitude of traffic flow, speed and the lane occupancy ratio of upstream and downstream, gather once every 30 seconds.Mean value with all passages of a measuring station is set up partial least square model.Existing 45 incident examples wherein select 22 examples (2100 examples) as training set at random, and 23 remaining examples (2036 examples) are as detected set.The normal event data (43418 examples) that on February 16th, 1993 gathered are as training set, and the normal event data of gathering on February 17th, 1993 are as detected set.

Because PLSR is responsive to the ratio of positive class in the training set and anti-class, in order to make PLSR more performance is arranged, our picked at random comprises that 12200 examples of all event instance produce a new training set that contains 20.6% event instance.Set up PLSR with this new data set, according to the result of cross validation, PLSR is made up of four major components.For ease of relatively, set up the svm classifier device simultaneously and detect, use radial basis function to make up SVM, parameters C is configured to 1.0, and the parameter gamma value of radial basis function is set to 1.

With the identical test data evaluation partial least square method and the performance of support vector machine method, comparing result is as shown in table 1.

Table 1. partial least square method and support vector machine test performance contrast (containing 20.6% event instance)

	Verification and measurement ratio (%)	The alert rate (%) of mistake	The average detected time (branch)	Classification rate (%)	Area under the ROC (%)	Working time (second)
							PLSR	95.65	0.06	4.66	96.65	94	0.30
SVM	91.30	0.24	2.55	94.65	94	13.69

The verification and measurement ratio of PLSR, the alert rate of mistake and classification rate are more superior than SVM, and the verification and measurement ratio of PLSR is 95.65%, and SVM is 91.30%, and the alert rate of the mistake of PLSR is 0.06%, and SVM is 0.24%; The classification rate of PLSR is 96.65%, and SVM is 94.65%.Another important difference is that PLSR has very fast convergence, contrasts 13.69 seconds of SVM, and as long as it is 0.30 second.Although the average detected time of this PLSR is undesirable, be inferior to 2.55 seconds of SVM in its 4.66 seconds,, we further test and show, suitably improve the ratio of event instance in the training set, can shorten detection time, make this index approach the average detected time of SVM.

From the experimental result of I-880 True Data, PLSR can fine portrayal traffic flow character parameter and incident between internal relation.Comparison shows that between PLSR and the SVM, approaching even surpassed the SVM method based on the detection algorithm performance of PLSR model.In addition, the PLSR convergence is fast, also do not need to regulate any parameter, and the performance of neural network depends on the parameter of the selection and the training of structure, and the performance of SVM depends on the selection of kernel function and some parameters; In addition, especially the convergence of neural network is very slow for SVM.Therefore, use PLSR to detect traffic events and have remarkable advantages.

Claims (3)

1, a kind of freeway traffic event automatic detection method based on offset minimum binary theory may further comprise the steps:

----detection system by ground induction coil or video equipment according to setting step-length acquisition testing district real-time traffic stream information;

----judges according to the real-time traffic stream information of gathering whether detection zone traffic events takes place, if take place, give the alarm, and the notice vehicle supervision department removing incident that takes appropriate measures, otherwise the continuation image data is judged next time;

It is characterized in that: described real-time traffic stream information according to collection judges whether detection zone traffic events takes place, and is to adopt the partial least squares regression detection algorithm to judge whether detection zone traffic events takes place.

2, according to the described freeway traffic event automatic detection method based on offset minimum binary theory of claim 1, it is characterized in that: described employing partial least squares regression detection algorithm judges that the concrete steps whether detection zone traffic events takes place are:

A, with known traffic behavior y=[y] _N*lTelecommunication flow information X=[x ₁, x ₂..., x _k] _N*k, as modeling sample, set up the partial least squares regression mathematical model, traffic behavior value 1 or-1, promptly $Y=\left[\begin{array}{c}{y}_1\\ y_2\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ y_n\end{array}\right]=\left[\begin{array}{c}1/-1\\ 1/-1\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ 1/-1\end{array}\right],$ 1 or-1 respectively presentation of events take place or do not have incident to take place; Wherein: n is a sample number, i.e. the line number of traffic behavior y; K is the number of telecommunication flow information, i.e. the columns of X;

B, the detection zone real-time traffic stream information substitution institute established model that will gather in real time determine according to model output current whether event occurs; If model output is not more than 0, be the normal traffic running status in the expression detection zone, otherwise expression generation incident.

Images