CN105372659A - Road traffic monitoring multi-target detection tracking method and tracking system - Google Patents

Abstract

The invention discloses a road traffic monitoring multi-target detection tracking method. The method includes the following steps that: a radar detects a plurality of moving targets; Kalman filtering and prediction are performed on feedback data of the detection of the plurality of moving targets by the radar, so that the motion states of the plurality of moving targets at the current time point can be obtained, and the motion states of the plurality of moving targets in future time points are predicted; and an interactive multi-model (IMM) algorithm with Markov switching coefficient is adopted to track the plurality of moving targets detected by the radar. According to the road traffic monitoring multi-target detection tracking method of the invention, based on Kalman filtering and the establishment of a road traffic monitoring model, the motion situations of a plurality of moving targets in a monitoring range can be monitored in real time without losing the targets.

Description

Traffic monitoring multiple target detection tracking and tracker

Technical field

The present invention relates to a kind of traffic monitoring technology, be specifically related to a kind of traffic monitoring multiple target detection tracking and tracker.

Background technology

Along with China's economy and the develop rapidly of road traffic and the increasing of vehicle, traffic monitoring is more and more subject to people's attention.In multiple target tracking, first will set up potential threat target and identify target, for reducing the false alarm rate of system, the reliability of raising system provides theories integration; By the comparison to board techniques for distance measurement, select millimeter wave frequency modulation on pulse radar Doppler; Analyse in depth the complex environment that road target is followed the tracks of, the data correlation of radar data process and the initial of target following have been conducted a preliminary study with termination.Data correlation is one of gordian technique of multiple target tracking, is also one of gordian technique of multi-sensor information fusion.Its object is to set up a kind of relation measured between flight path (target), to determine whether each metric data comes from same target, the key problem of data correlation measures to be similar to the estimated value of much degree and flight path (target), the result of data correlation directly has influence on the estimation to dbjective state, therefore the concern of extremely people.What current application was more has the methods (JPDA) such as arest neighbors method, many assumption methods, probabilistic data association algorithm (PDA), JPDA method.

Summary of the invention

The invention provides a kind of traffic monitoring multiple target detection tracking and tracker, by the foundation of Kalman filtering and traffic monitoring model, can monitor in real time the motion conditions of the multiple targets in monitoring range and not lose objects.

For achieving the above object, the invention discloses a kind of traffic monitoring multiple target detection tracking, be characterized in, the method comprises:

Radar detects some moving targets;

Kalman filtering and prediction are carried out to the feedback data of the some moving targets of radar detection, obtains the motion state of the some moving targets of current time, and predict the motion state of the some moving targets of future time instance;

The some moving targets of interactive multi-model process to radar detection with Markov Switching coefficient are adopted to follow the tracks of.

Above-mentioned moving target has two degree of freedom in road traffic, is respectively: around the rotation of ground normal and the translation along road direction.

In above-mentioned Kalman filtering, the state equation of moving target is such as formula (27):

X(k+1)＝FX(k)+GW(k)(27)

In formula (27):

$F=\left[\begin{array}{cccccc}1& T& T^2/2& 0& 0& 0\\ 0& 1& T& 0& 0& 0\\ 0& 0& 1& 0& 0& 0\\ 0& 0& 0& 1& T& T^2/2\\ 0& 0& 0& 0& 1& T\\ 0& 0& 0& 0& 0& 1\end{array}\right],X\left(k\right)=\left[\begin{array}{c}{x}_1\left(k\right)\\ x_2\left(k\right)\\ x_3\left(k\right)\\ y_1\left(k\right)\\ y_2\left(k\right)\\ y_3\left(k\right)\end{array}\right],$

$G=\left[\begin{array}{cc}{T}^2/4& 0\\ T/2& 0\\ 1& 0\\ 0& T^2/4\\ 0& T/2\\ 0& 0\end{array}\right];$ W(k)＝[w ₁(k)w ₂(k)]，

Wherein, F is state-transition matrix; T is the sampling interval; Variable X (k) is dbjective state vector, comprises radial distance r, relative velocity relative acceleration angle beta, angular velocity and angular acceleration w (k) is system noise vector; G is the acting matrix of system noise; w ₁(k), w ₂k () is average is zero, and is mutual incoherent white noise sequence, and its variance is the covariance matrix of W (k) is such as formula (28):

$Q\left(k\right)=E\[W\left(k\right)W^T\left(k\right)\]=\left[\begin{array}{cc}{\mathrm{\σ}}_{\mathrm{\α}}^2& 0\\ 0& {\mathrm{\σ}}_{\mathrm{\β}}^2\end{array}\right]---\left(28\right).$

The fundamental equation of the one-step prediction of above-mentioned Kalman filtering is such as formula (7), (8), (9):

$\hat{X}(k+1/k)=\mathrm{\Φ}(k+1/k)\hat{X}(k/k-1)+K_p\left(k\right)\[Y\left(k\right)\·H\left(k\right)\hat{X}(k/k-1)\]---\left(7\right)$

K _p(k)＝Φ(k+1/k)P(k/k-1)HT[H(k)P(k/k-1)HT(k)+R(k)] ^-1(8)

P(k+1/k)＝[Φ(k+1/k)-K _p(k)H(k)]P(k/k-1)+G(k)Q(k)GT(k)(9)

K in formula (7), (8), (9) _pk () is one-step prediction gain battle array; Φ (k+1, k) (Φ (k+1, k) ∈ R ^{n × n}) be state-transition matrix; Y (k) (Y (k) ∈ R ^{m × 1}) for measuring vector; H (k) (H (k) ∈ R ^{m × n}) be observing matrix; Formula (9) is one-step prediction error in equation.

The above-mentioned measurement model with the interactive multi-model process of Markov Switching coefficient is:

Z(k)＝H(k)X(k)+V(k)(29)

Wherein:

$Z\left(k\right)=\left[\begin{array}{c}{z}_1\left(k\right)\\ z_2\left(k\right)\\ z_3\left(k\right)\end{array}\right];H\left(k\right)=\left[\begin{array}{cccccc}1& 0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0& 0\\ 0& 0& 0& 1& 0& 0\end{array}\right];V\left(k\right)=\left[\begin{array}{c}{v}_1\left(k\right)\\ v_2\left(k\right)\\ v_3\left(k\right)\end{array}\right];$

In formula: z ₁(k), z ₂(k), z ₃k () represents the distance of the target that radar records, relative velocity and position angle respectively; H (k) is observing matrix; V (k) is measurement noise; Wherein v ₁(k), v ₂(k), v ₃k () is three mutual incoherent random white noise sequences, its equation is respectively with $E{\[v_3\left(k\right)\]}^2={\mathrm{\σ}}_3^2;$

The covariance matrix of V (k) is formula (30):

$R\left(k\right)=E{\[V\left(k\right)V^T\left(k\right)\]}^2=\left[\begin{array}{ccc}{\mathrm{\σ}}_1^2& 0& 0\\ 0& {\mathrm{\σ}}_2^2& 0\\ 0& 0& {\mathrm{\σ}}_3^2\end{array}\right]---\left(30\right)$

Here, for the variance of distance measure error, for the variance of the error of relative velocity (Doppler) measured value, for the variance of measurement of azimuth value error.

Be applicable to a tracker for above-mentioned traffic monitoring multiple target detection tracking, be characterized in, this tracker comprises:

Radar, it detects some moving targets and uploads feedback data;

Data handling system, the feedback data that its receiving radar is uploaded, carries out Kalman filtering and prediction to feedback data, obtains the motion state of the some moving targets of current time, and predicts the motion state of the some moving targets of future time instance; The some moving targets of interactive multi-model process control radar with Markov Switching coefficient are adopted to follow the tracks of.

Above-mentioned radar is millimeter wave frequency modulation on pulse radar Doppler.

Traffic monitoring multiple target detection tracking of the present invention and tracker are compared with the tracking of the road traffic moving target of prior art, its advantage is, the present invention, by the foundation of Kalman filtering and traffic monitoring model, can monitor the motion conditions of the multiple moving targets in monitoring range and not lose objects in real time;

The present invention adopts radar as monitoring and target following sensor, be particularly suitable for the traffic monitoring under rainy day or bad weather circumstances, accurately can carry out the monitoring of motion state to multiple goal, so that traffic police carries out traffic administration, be with a wide range of applications in traffic monitoring field.

Accompanying drawing explanation

Fig. 1 is the operational flowchart of interactive multi-model process;

Fig. 2 is the method flow diagram of traffic monitoring multiple target detection tracking of the present invention.

Embodiment

Below in conjunction with accompanying drawing, further illustrate specific embodiments of the invention.

The data that radar detection obtains contain measuring error and noise, therefore by data processing process, i.e. tracking filter, by error and noise reduction, thus will improve radar to the accuracy of environment sensing.Kalman filter only use the previous estimated value of signal and a nearest observed reading just can under linear unbiased minimum variance estimation criterion the currency of estimated signal, and necessarily whole observed reading in the past.

Kalman filtering estimates desired signal by calculating from the measurement relevant with being extracted signal.Be wherein the random response caused by white-noise excitation by estimated signal, system equation is known as the transferring structure between excitation and response, and measurement equation is also known as measurement amount and by the funtcional relationship between estimator.In estimation procedure, utilize the statistical property of known system equation, measurement equation, white-noise excitation, the statistical property of error in measurement.Kalman filtering designs in time domain, and information used is also all the amount in time domain.And Kalman filter is applicable to multidimensional.Thus the scope of application of Kalman filtering widely.It has following features:

1) Kalman filtering process to as if random signal;

2) processed signal dividing without useful and interference, the object of filtering to estimate all processed signals;

3) white-noise excitation of system and measurement noise are not the objects wanting filtering, and their statistical property needs the information utilized just in estimation procedure.

Kalman filtering generally can be divided into linear in nonlinear according to the system equation of physical system and the character of measurement equation, can be divided into again continuous print and discrete in each.Non-linear Kalman filtering is also called EKF.At this moment, system equation is nonlinear, or system equation and measurement equation are all nonlinear.The system equation that this patent is set up and measurement equation are all nonlinear, and are discrete.

Kalman filtering in motor vehicle target following and prediction:

The object of filter and predication estimates motion state that is current and future time instance target, comprises position, speed and angle etc.The criterion of Kalman filtering and prediction is that root-mean-square error is minimum.In addition, it also has other advantages many in maneuvering target tracking:

1) Kalman filtering of the motor-driven and measurement noise model of based target and prediction gain sequence can automatically be selected.This means by changing some key parameters, identical wave filter goes for different maneuvering targets and measures environment;

2) Kalman filtering and measurement gain sequence automatically can adapt to the change of testing process, comprise the change in sampling period and undetected situation;

3) Kalman filtering can be measured estimated accuracy by covariance matrix easily with prediction.In multiple maneuvering target tracking, this measurement facility also can be used for the formation of tracking gate and the determination of thresholding size;

4) by the change of Kalman filtering with residual vector d (k) in prediction, can judge whether the object module of former supposition and the kinetic characteristic of realistic objective meet.Thus, d (k) can be used as motor-driven detection and the motor-driven a kind of means debating knowledge.Meanwhile, also can be used for consistency analysis etc.

(5), in the multi-machine scheduling under heavy clutter environment, by the use of Kalman filtering and Forecasting Methodology, the impact of uncertain correlated error can partly be compensated.

Kalman filtering fundamental equation:

The state equation of target and measurement equation can be represented by formula (1):

$\left\{\begin{array}{c}X(k+1)=\mathrm{\Φ}(k+1,k)X\left(k\right)+G\left(k\right)W\left(k\right)\\ Y\left(k\right)=H\left(k\right)X\left(k\right)+V\left(k\right)\end{array}\right.---\left(1\right)$

In formula (1):

X (k) (X (k) ∈ R ^{n × 1}) be dbjective state vector; Y (k) (Y (k) ∈ R ^{m × 1}) for measuring vector; Φ (k+1, k) (Φ (k+1, k) ∈ R ^{n × n}) be state-transition matrix; G (k) (G (k) ∈ R ^{n × n}) be the acting matrix of system noise; H (k) (H (k) ∈ R ^{m × n}) be observing matrix; W (k) (W (k) ∈ R ^{n × 1}) average is E [W (k)]=0, covariance is E [W (k) W ^t(j)]=Q (k) δ _kj, be called system noise; V (k) (V (k) ∈ R ^{m × 1}), average is E [V (k)]=0, and covariance is E [V (k) V ^t(j)]=R (k) δ _kj, it is mutual incoherent white noise, i.e. E [W (k) V with W (k) ^t(j)]=0, be measurement noise.

When original state, X ₀independent with W (k), V (k), that is: E [X ₀w ^t(k)]=0, E [X ₀v ^t(k)]=0.Corresponding Kalman filtering fundamental equation is as follows:

State estimation, such as formula (2):

$\hat{X}(k/k)=\hat{X}(k/k-1)+K\left(k\right)\[Y\left(k\right)-H\left(k\right)\hat{X}(k/k-1)\]---\left(2\right)$

This state is predicted further, such as formula (3):

$\hat{X}(k/k-1)=\mathrm{\Φ}(k+1,k)\hat{X}(k-1/k-1)---\left(3\right)$

Its filter gain, such as formula (4):

K(k)＝P(k/k-1)HT(k)[H(k)P(k/k-1)HT(k)+R(k)] ^-1(4)

One-step prediction error in equation, such as formula (5)

P(k/k-1)＝Φ(k+l,k)P(k-1/k-1)Φ ^T(k+l,k)+G(k-1)Q(k-1)GT(k-1)(5)

Optimal estimation square error, such as formula (6)

P(k/k)＝[I-K(k)H(k)]P(k/k-l)(6)

Wherein, d (k)=Y (k)-H (k) X (k/k-1) is defined as residual vector, and its covariance matrix is: S (k)=H (k) p (k/k-1) HT (k)+R (k).

At maneuvering target tracking, particularly in multiple maneuvering target tracking, filter and predication is very important.The fundamental equation of Kalman filtering one-step prediction, such as formula (7), (8), (9).

$\hat{X}(k+1/k)=\mathrm{\Φ}(k+1/k)\hat{X}(k/k-1)+K_p\left(k\right)\[Y\left(k\right)\·H\left(k\right)\hat{X}(k/k-1)\]---\left(7\right)$

K _p(k)＝Φ(k+1/k)P(k/k-1)HT[H(k)P(k/k-1)HT(k)+R(k)] ^-1(8)

P(k+1/k)＝[Φ(k+1/k)-K _p(k)H(k)]P(k/k-1)+G(k)Q(k)GT(k)(9)

K in formula (7), (8), (9) _pk () is one-step prediction gain battle array.

Kalman filtering has two counter circuits: gain counter circuit and filtering counter circuit.Wherein gain counter circuit is independently, and filtering counter circuit depends on gain counter circuit, there are two renewal processes in one cycle: time renewal process and measurement renewal process.When predicted state estimated value and the measuring value in k moment in known k-1 moment, just can obtain the optimal estimation value of k moment state vector with the one-step prediction mean square deviation in k-1 moment, and state estimation and the measuring value in k-1 moment can be predicted.

Moving-target, such as automobile, when road travels, often will carry out accelerating, to slow down and turning causes its motion state constantly to change.For traffic monitoring radar, accurately follow the tracks of multiple target, the state of such as vehicle, estimate that the motion state of each target is the vital task of traffic monitoring radar in time.At present, most of multi-object tracking method designs for following the tracks of the less aerial target of clutter, and the environment residing for road vehicle, as the maneuvering condition of target, ground unrest and environment clutter etc., differ widely with the environment of aerial target, they all can have an impact to the output of police radar.Therefore, the conventional aerial multiobject tracking of tracking is also not suitable for the tracking of road target, and traffic monitoring multiple target detection system needs the actual state considering road target, take a kind of newly or the multi-object tracking method that improves.

On the basis of generalized Bayes method, take Kalman filtering as starting point, a kind of interactive multi-model process (IMM method) with Markov Switching coefficient proposed, wherein multiple model concurrent working, Target state estimator is the interactive result of multiple wave filter.The method does not need motor-driven detection, reaches comprehensive adaptive ability simultaneously.Compared with single model sef-adapting filter, the method has the following advantages:

1) carrying out refinement modeling by expanding model rightly, describing because it have employed multi-model to parameter space;

2) in filtering, adaptive structure changes is realized by the variation of model probability.Additionally by increase and decrease in real time or change model, adaptive structural capacity can be strengthened;

3) under the condition meeting a priori assumption, be the optimal estimation under square errors sense.Thus the rationality of Research Hypothesis of can concentrating one's energy, find and more reasonably suppose;

4) the method has obvious parallel organization, is convenient to Parallel Implementation effectively.

As shown in Figure 1, the implementation procedure of interactive multi-model process is as follows:

R model is carved with, then target state mode, such as formula (10) when supposing k:

X(k+1)＝Φ _jX(k)+G _jW _j(k),j＝1,…,r.(10)

Wherein, W _jk () is average is zero, covariance matrix is Q _jwhite noise sequence.Control the conversion between these models by a Markov chain, the transition probability matrix of Markov chain is: measurement model is formula (11):

Z(k)＝H _jX _j(k)+V _j(k)(11)

This interactive multi-model process step can be summarized as follows:

Input is mutual, such as formula (12), (13):

${\hat{X}}^{oj}(k-1/k-1)=\underset{i=1}{\overset{r}{\Σ}}{\hat{X}}^i(k/k-1){\mathrm{\μ}}_{ij}(k-1/k-1),j=1,...,r---\left(12\right)$

$\begin{array}{c}{P}^{oj}\left(k-1/k-1\right)=\underset{i=1}{\overset{r}{\Σ}}{\mathrm{\μ}}_{ij}\left(k-1/k-1\right)\\ \left\{P^i\left(k-1/k-1\right)+\[{\hat{X}}^i\left(k/k-1\right)-{\hat{X}}^{oj}\left(k-1/k-1\right)\]{\[{\hat{X}}^i\left(k/k-1\right)-{\hat{X}}^{oj}\left(k-1/k-1\right)\]}^T\right\}\end{array}---\left(13\right)$

Wherein, ${\mathrm{\μ}}_{ij}(k-1/k-1)=P\{M_j(k-1)/M_j\left(k\right),Z^{k-1}\}=p_{ij}{\mathrm{\μ}}_i(k-1)/{\stackrel{\‾}{c}}_j,$ P _ijmodel i forwards the transition probability of model j to, for normaliztion constant,

For model M _j(k), with p ^oj(k-1/k-1) and Z (k) carry out Kalman filtering as input.

Predictive equation, such as formula (14):

${\hat{X}}^j(k-1/k-1)={\mathrm{\Φ}}_j{\hat{X}}^{oj}(k-1/k-1)---\left(14\right)$

Predicting covariance, such as formula (15):

$P^j(k/k-1)={\mathrm{\Φ}}_j{P}^{oj}(k-1/k-1){\mathrm{\Φ}}_j^T+G_j{Q}_j{G}_j^T---\left(15\right)$

Kalman gain, such as formula (16):

K _j(k)＝P ^j(k/k-1)H ^T[HP ^j(k/k-1)H ^T+R] ^-1(16)

Filtering, such as formula (17):

${\hat{X}}^j(k/k)={\hat{X}}^j(k/k-1)+K_j\left(k\right)\[Z\left(k\right)-H{\hat{X}}^j(k/k-1)\]---\left(17\right)$

Filtering covariance, such as formula (18):

P ^j(k/k)＝[I-K _j(k)]P ^j(k/k-1)(18)

Model probability upgrades, such as formula (19):

$\begin{array}{c}{\mathrm{\μ}}_j\left(k\right)=P\left\{Z\left(k\right)/M_j\left(k\right)/M_j\left(k\right),Z^{k-1}\right\}P\left\{M_j\left(k\right)/Z^{k-1}\right\}\\ =\frac{1}{c}{\mathrm{\Λ}}_j\left(k\right)\underset{i=1}{\overset{r}{\Σ}}{p}_{ij}{\mathrm{\μ}}_i\left(k-1\right)={\mathrm{\Λ}}_j\left(k\right){\stackrel{\‾}{c}}_j/c\end{array}---\left(19\right)$

Wherein, c is normaliztion constant, and and Λ _jk likelihood function that () is observing matrix Z (k), ${\mathrm{\Λ}}_j\left(k\right)=P\left\{Z\left(k\right)/M_i\left(k\right),Z^{k-1}\right\}=\frac{1}{{\left(2\mathrm{\π}\right)}^{n/2}{\left|S_j\left(k\right)\right|}^{1/2}}\exp \left\{-\frac{1}{2}{v_j}^T{s_j}^{-1}\left(k\right)v_j\right\},$ Wherein, $v_j\left(k\right)=Z\left(k\right)-H{\hat{X}}^j(k/k-1),$ S _j(k)＝HP ^j(k/k-1)H ^T+R。

Export mutual, such as formula (20), (21):

$\hat{X}(k/k)=\underset{j=1}{\overset{r}{\Σ}}{\hat{X}}^j(k/k){\mathrm{\μ}}_j\left(k\right)---\left(20\right)$

$P(k/k)=\underset{j=1}{\overset{r}{\Σ}}{\mathrm{\μ}}_j\left(k\right)\{P^j(k/k)+\[{\hat{X}}^j(k/k)-\hat{X}(k/k)\]{\[{\hat{X}}^j(k/k)-\hat{X}(k/k)\]}^T\}---\left(21\right)$

Based on said method, below illustrate a kind of embodiment of traffic monitoring multiple target detection tracking.

First road target motion model in the present embodiment is determined, i.e. the state equation of tracking target:

The form of the dynamic model differential equation describes the kinematic parameter relevant with estimating forecasting process.Object can be defined as the synthesis result of rotation and translation in three-dimensional motion.The road head fall monitored due to road monitoring is less, can be approximated to be level in short distance, so be level at hypothesis road plane herein, namely road does not have the fluctuating in vertical direction.This hypothesis prerequisite under, according to the dynamics of vehicle movement, consider two degree of freedom of target travel in the present embodiment, comprising: 1. target around ground normal rotate, make target travel direction and road direction not parallel; 2. along the translation of road direction.For the tracking of moving target, it is required to determine that it: the distance r between target and radar, and the rate of change of distance, i.e. speed of related movement distance r between target and radar is speed of related movement integration.Target is relative to the transverse movement angular velocity of radar determine target in the position of subsequent time on road with angle beta, this is very important to the prediction of system.In the present embodiment, radar adopts Millimeter PD Radar, measures the distance of target, relative velocity and angle radar, needs the signal estimated to be distance r, relative velocity in subsequent time target and angle beta.

As shown in Figure 2, this traffic monitoring multiple target detection tracking specifically comprises following steps:

S1, radar carry out detecting periodically, monitor the some moving targets in road traffic, and the feedback data upload process that will receive.

S2, by Kalman filtering, filter and predication is carried out to the feedback data that radar receives, estimate current and motion state that is future time instance target, comprise the distance of target and radar, speed and angle.

In the present embodiment, if moving target distance, speed and the angle in k moment when scanning (radar certain) be respectively r (k), β (k), be respectively at the distance of k+1 moment (when radar scans), speed and position angle next time r (k+1), with β (k+1).If the time T of the adjacent twice sweep of radar is enough little, then what can be similar to thinks:

$r(k+1)=r\left(k\right)+T\stackrel{\·}{r}\left(k\right)+\frac{1}{2}{T}^2\stackrel{\·\·}{r}\left(k\right)---\left(22\right)$

$\stackrel{\·}{r}(k+1)=\stackrel{\·}{r}\left(k\right)+T\stackrel{\·\·}{r}\left(k\right)---\left(23\right)$

$\mathrm{\β}(k+1)=\mathrm{\β}\left(k\right)+\stackrel{\·}{\mathrm{\β}}\left(k\right)---\left(24\right)$

In formula (22), (23), (24): with be respectively target in the radial acceleration in k moment and azimuthal variation speed.

Suppose the impact of the enchancement factor such as irregular change due to the change of environment, condition of road surface, coefficient of road adhesion, make the radial acceleration of moving target and azimuthal variation acceleration that random change occur:

$\stackrel{\·\·}{r}(k+1)=\stackrel{\·\·}{r}\left(k\right)+u_1\left(k\right)---\left(25\right)$

$\stackrel{\·\·}{\mathrm{\β}}(k+1)=\stackrel{\·\·}{\mathrm{\β}}\left(k\right)+u_2\left(k\right)---\left(26\right)$

In formula (25), (26), u ₁(k) for average be zero, variance is random white noise sequence.U ₂(k) for radial acceleration and azimuthal variation speed are from the k moment to the increment in k+1 moment, it and u ₁k () is incoherent random white noise sequence.

The state equation of now moving target motion can describe such as formula (27):

X(k+1)＝FX(k)+GW(k)(27)

In formula (27):

$F=\left[\begin{array}{cccccc}1& T& T^2/2& 0& 0& 0\\ 0& 1& T& 0& 0& 0\\ 0& 0& 1& 0& 0& 0\\ 0& 0& 0& 1& T& T^2/2\\ 0& 0& 0& 0& 1& T\\ 0& 0& 0& 0& 0& 1\end{array}\right],X\left(k\right)=\left[\begin{array}{c}{x}_1\left(k\right)\\ x_2\left(k\right)\\ x_3\left(k\right)\\ y_1\left(k\right)\\ y_2\left(k\right)\\ y_3\left(k\right)\end{array}\right],$

$G=\left[\begin{array}{cc}{T}^2/4& 0\\ T/2& 0\\ 1& 0\\ 0& T^2/4\\ 0& T/2\\ 0& 0\end{array}\right];$ W(k)＝[w ₁(k)w ₂(k)]，

Wherein, F is state-transition matrix, and T is the sampling interval, variable X (k) be dbjective state vector, comprise radial distance, relative velocity, relative acceleration, angle, angular velocity and angular acceleration, be masked as respectively r, β, with w (k) is system noise vector, w ₁(k), w ₂k () is average is zero, and is mutual incoherent white noise sequence, and its variance is the covariance matrix of W (k) is such as formula (28):

$Q\left(k\right)=E\[W\left(k\right)W^T\left(k\right)\]=\left[\begin{array}{cc}{\mathrm{\σ}}_{\mathrm{\α}}^2& 0\\ 0& {\mathrm{\σ}}_{\mathrm{\β}}^2\end{array}\right]---\left(28\right).$

S3, be starting point with Kalman filtering, a kind of interactive multi-model process with Markov Switching coefficient proposed is as road traffic multi-object tracking method, have the interactive multi-model process of Markov Switching coefficient by this, operational radar is followed the tracks of the multiple goal in road traffic.

In the present embodiment, radar adopts Millimeter PD Radar (ARS100 millimetre-wave radar), and its distance to moving target, relative velocity and azimuthal measurement equation can be write as formula (29):

Z(k)＝H(k)X(k)+V(k)(29)

Wherein:

$Z\left(k\right)=\left[\begin{array}{c}{z}_1\left(k\right)\\ z_2\left(k\right)\\ z_3\left(k\right)\end{array}\right];H\left(k\right)=\left[\begin{array}{cccccc}1& 0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0& 0\\ 0& 0& 0& 1& 0& 0\end{array}\right];V\left(k\right)=\left[\begin{array}{c}{v}_1\left(k\right)\\ v_2\left(k\right)\\ v_3\left(k\right)\end{array}\right];$

In formula: z ₁(k), z ₂(k), z ₃k () represents the distance of the target that radar records, relative velocity and position angle respectively; H (k) is observing matrix; V (k) is measurement noise; Wherein v ₁(k), v ₂(k), v ₃k () is three mutual incoherent random white noise sequences, its equation is respectively with $E{\[v_3\left(k\right)\]}^2={\mathrm{\σ}}_3^2.$

The covariance matrix of V (k) is formula (30):

$R\left(k\right)=E{\[V\left(k\right)V^T\left(k\right)\]}^2=\left[\begin{array}{ccc}{\mathrm{\σ}}_1^2& 0& 0\\ 0& {\mathrm{\σ}}_2^2& 0\\ 0& 0& {\mathrm{\σ}}_3^2\end{array}\right]---\left(30\right)$

Here: for the variance of distance measure error, for the variance of the error of relative velocity (Doppler) measured value, for the variance of measurement of azimuth value error.

The invention also discloses a kind of tracker being applicable to above-mentioned traffic monitoring multiple target detection tracking, this tracker comprises:

Radar, it adopts millimeter wave frequency modulation on pulse radar Doppler, detects and upload feedback data to some moving targets;

Data handling system, the feedback data that its receiving radar is uploaded, carries out Kalman filtering and prediction to feedback data, obtains the motion state of the some moving targets of current time, and predicts the motion state of the some moving targets of future time instance; The some moving targets of interactive multi-model process control radar with Markov Switching coefficient are adopted to follow the tracks of.

Current traffic monitoring mainly relies on the means such as transmission of video to judge the motion conditions of target vehicle, so that traffic police carries out traffic administration, but can not reach this effect under rainy day or bad weather circumstances.The present invention adopts radar as monitoring and target following sensor, is particularly suitable for the traffic monitoring under bad weather circumstances, accurately can carries out the monitoring of motion state to multiple goal.

Although content of the present invention has done detailed introduction by above preferred embodiment, will be appreciated that above-mentioned description should not be considered to limitation of the present invention.After those skilled in the art have read foregoing, for multiple amendment of the present invention and substitute will be all apparent.Therefore, protection scope of the present invention should be limited to the appended claims.

Claims (7)

1. a traffic monitoring multiple target detection tracking, is characterized in that, the method comprises:

Radar detects some moving targets;

Kalman filtering and prediction are carried out to the feedback data of the some moving targets of radar detection, obtains the motion state of the some moving targets of current time, and predict the motion state of the some moving targets of future time instance;

The some moving targets of interactive multi-model process to radar detection with Markov Switching coefficient are adopted to follow the tracks of.

2. traffic monitoring multiple target detection tracking as claimed in claim 1, it is characterized in that, described moving target has two degree of freedom in road traffic, is respectively: around the rotation of ground normal and the translation along road direction.

3. traffic monitoring multiple target detection tracking as claimed in claim 1 or 2, it is characterized in that, in described Kalman filtering, the state equation of moving target is such as formula (27):

X(k+1)＝FX(k)+GW(k)(27)

In formula (27):

$F=\left[\begin{array}{cccccc}1& T& T^2/2& 0& 0& 0\\ 0& 1& T& 0& 0& 0\\ 0& 0& 1& 0& 0& 0\\ 0& 0& 0& 1& T& T^2/2\\ 0& 0& 0& 0& 1& T\\ 0& 0& 0& 0& 0& 1\end{array}\right],$ $X\left(k\right)=\left[\begin{array}{c}{x}_1\left(k\right)\\ x_2\left(k\right)\\ x_3\left(k\right)\\ y_1\left(k\right)\\ y_2\left(k\right)\\ y_3\left(k\right)\end{array}\right],$

$G=\left[\begin{array}{cc}{T}^2/4& 0\\ T/2& 0\\ 1& 0\\ 0& T^2/4\\ 0& T/2\\ 0& 0\end{array}\right];$ W(k)＝[w ₁(k)w ₂(k)]，

Wherein, F is state-transition matrix; T is the sampling interval; Variable X (k) is dbjective state vector, comprises radial distance r, relative velocity relative acceleration angle beta, angular velocity and angular acceleration w (k) is system noise vector; G is the acting matrix of system noise; w ₁(k), w ₂k () is average is zero, and is mutual incoherent white noise sequence, and its variance is the covariance matrix of W (k) is such as formula (28):

$Q\left(k\right)=E\[W\left(k\right)W^T\left(k\right)\]=\left[\begin{array}{cc}{\mathrm{\σ}}_{\mathrm{\α}}^2& 0\\ 0& {\mathrm{\σ}}_{\mathrm{\β}}^2\end{array}\right]---\left(28\right).$

4. traffic monitoring multiple target detection tracking as claimed in claim 3, it is characterized in that, the fundamental equation of the one-step prediction of described Kalman filtering is such as formula (7), (8), (9):

K _p(k)＝Φ(k+1/k)P(k/k-1)HT[H(k)P(k/k-1)HT(k)+R(k)] ^-1(8)

P(k+1/k)＝[Φ(k+1/k)-K _p(k)H(k)]P(k/k-1)+G(k)Q(k)GT(k)(9)

K in formula (7), (8), (9) _pk () is one-step prediction gain battle array; Φ (k+1, k) (Φ (k+1, k) ∈ R ^{n × n}) be state-transition matrix; Y (k) (Y (k) ∈ R ^{m × 1}) for measuring vector; H (k) (H (k) ∈ R ^{m × n}) be observing matrix; Formula (9) is one-step prediction error in equation.

5. traffic monitoring multiple target detection tracking as claimed in claim 1 or 2, is characterized in that, described in there is the interactive multi-model process of Markov Switching coefficient measurement model be:

Z(k)＝H(k)X(k)+V(k)(29)

Wherein:

$Z\left(k\right)=\left[\begin{array}{c}{z}_1\left(k\right)\\ z_2\left(k\right)\\ z_3\left(k\right)\end{array}\right];$ $H\left(k\right)=\left[\begin{array}{cccccc}1& 0& 0& 0& 0& 0\\ 0& 1& 0& 0& 0& 0\\ 0& 0& 0& 1& 0& 0\end{array}\right];$ $V\left(k\right)=\left[\begin{array}{c}{v}_1\left(k\right)\\ v_2\left(k\right)\\ v_3\left(k\right)\end{array}\right];$

In formula: z ₁(k), z ₂(k), z ₃k () represents the distance of the target that radar records, relative velocity and position angle respectively; H (k) is observing matrix; V (k) is measurement noise; Wherein v ₁(k), v ₂(k), v ₃k () is three mutual incoherent random white noise sequences, its equation is respectively $E{\[v_2\left(k\right)\]}^2={\mathrm{\σ}}_2^2$ With $E{\[v_3\left(k\right)\]}^2={\mathrm{\σ}}_3^2;$

The covariance matrix of V (k) is formula (30):

$R\left(k\right)=E{\[V\left(k\right)V^T\left(k\right)\]}^2=\left[\begin{array}{ccc}{\mathrm{\σ}}_1^2& 0& 0\\ 0& {\mathrm{\σ}}_2^2& 0\\ 0& 0& {\mathrm{\σ}}_3^2\end{array}\right]---\left(30\right)$

Here, for the variance of distance measure error, for the variance of the error of relative velocity (Doppler) measured value, for the variance of measurement of azimuth value error.

6. be applicable to a tracker for traffic monitoring multiple target detection tracking described in any one claim in claim 1 to 5, it is characterized in that, this tracker comprises:

Radar, it detects some moving targets and uploads feedback data;

Data handling system, the feedback data that its receiving radar is uploaded, carries out Kalman filtering and prediction to feedback data, obtains the motion state of the some moving targets of current time, and predicts the motion state of the some moving targets of future time instance; The some moving targets of interactive multi-model process control radar with Markov Switching coefficient are adopted to follow the tracks of.

7. tracker as claimed in claim 6, it is characterized in that, described radar is millimeter wave frequency modulation on pulse radar Doppler.