阅读说明：本技术 一种用于分布式目标跟踪的自适应一致性信息滤波方法 (Self-adaptive consistency information filtering method for distributed target tracking ) 是由 王奕迪 李钊 郑伟 于 2019-08-23 设计创作，主要内容包括：本发明提供一种用于分布式目标跟踪的自适应一致性信息滤波方法,首先利用无迹变换对非线性模型进行线性化并进行时间更新,然后根据当前时刻的测量信息构建检验统计量,以此判断每一个传感器的测量值是否出现异常,依据相邻传感器中出现异常的总数判断是否出现野值或机动,基于上述判断结果,若传感器测量出现野值,则将野值剔除,若目标出现机动,则利用衰减因子抑制动力学模型误差的影响,最后对网络中多个传感器所得到的信息进行一致性迭代,得到更新后的目标状态估计值及方差。本发明方法针对分布式传感器网络中的目标跟踪问题,在存在有色噪声和目标机动的情况下,实现稳定且高精度的目标跟踪。(The invention provides a self-adaptive consistency information filtering method for distributed target tracking, which comprises the steps of firstly utilizing unscented transformation to linearize a nonlinear model and carrying out time updating, then constructing test statistics according to the measurement information at the current moment to judge whether the measurement value of each sensor is abnormal or not, judging whether a wild value or maneuvering occurs or not according to the total number of the abnormal values occurring in the adjacent sensors, based on the judgment result, if the wild value occurs in the measurement of the sensor, rejecting the wild value, if the maneuvering occurs in the target, utilizing an attenuation factor to inhibit the influence of errors of a dynamic model, and finally carrying out consistency iteration on the information obtained by a plurality of sensors in a network to obtain an updated target state estimation value and variance. Aiming at the target tracking problem in the distributed sensor network, the method realizes stable and high-precision target tracking under the conditions of colored noise and target maneuvering.)

1. a kind of adaptive consensus information filtering method for distributed object tracking, which is characterized in that in sensor network N number of sensor node is set in network, and communication topological structure is indicated by non-directed graph G=(C, E), wherein C={ 1,2 ..., N } table Show the set of node, E={ (i, j) | i, j ∈ C } indicates that the set on side, the node set adjacent with i-th of node are expressed as N_i =j ∈ C | (i, j) ∈ E } and have N in the set_iA node；

Target dynamics model and sensor measurement model are expressed as form:

x_k=f (x_k-1)+w_k-1

In formula, x_kIt is state vector of the target k-th of moment, f is the nonlinear function determined by target dynamics characteristic, w_k-1 It is kinetic model noise,It is measured value of i-th of sensor at the kth moment, hⁱIt is the measurement functions of i-th of sensor,It is measurement noise, in addition, w_k-1WithIt is zero mean Gaussian white noise and the two is uncorrelated, w_k-1Variance be Q_k-1, Variance be

Described method includes following steps:

Step 1) linearizes nonlinear model using Unscented transform, and carries out time update；

Step 2) constructs test statistics according to the metrical information at current time, then judges that the measured value of each sensor is No appearance is abnormal, last to judge whether outlier or motor-driven occur according to the sum for occurring exception in adjacent sensors；

Step 3) is based on above-mentioned judging result, if there is outlier in sensor measurement, by unruly-value rejecting, and if target appearance is motor-driven, Then utilize the influence of decay factor inhibition dynamics model error；

Step 4) carries out consistency iteration to the obtained information of sensors multiple in network；

Step 5) updates Target state estimator value and variance.

2. the adaptive consensus information filtering method according to claim 1 for distributed object tracking, feature It is, detailed process is as follows for the step 1):

1., Sigma point sampling

Sigma point set is combined into { χ_i,k-1| τ=0 ..., 2n, k >=1 }, as follows:

In formula, λ=α²(n+K)-n, 0≤α≤1 and the distribution for controlling Sigma point, K=3-n,Cholsky because The row of son, the mean value of stochastic regime variable and the weights omega of covariance value are as follows:

In formula, β is parameter related with the prior distribution of state, and the subscript of ω represents the type of weight, when being above designated as m Shi Weijun The weight of value is the weight of covariance value when being above designated as c；

2. the time updates

Set k moment dbjective state estimated value and variance asWithThe then predicted value of information stateWith information square The predicted value of battle arrayIt is calculate by the following formula to obtain:

And

3. the adaptive consensus information filtering method according to claim 2 for distributed object tracking, feature It is, during Sigma point sampling, β is set as 2 in the case of a gaussian distribution.

4. the adaptive consensus information filtering method according to claim 2 for distributed object tracking, feature It is, detailed process is as follows for the step 2):

1. constructing test statistics

It is as follows in the test statistics at k moment for i-th of sensor:

Wherein,

In formula,For test statistics,Newly to cease,To estimate measured value,To measure autocorrelation matrix；

2. judging whether measured value exception occurs

Due toChi square distribution is obeyed when without exception, and measurement value sensor can be detected according to the following formula:

Wherein

In formula,Represent whether measured value of i-th of sensor at the k moment exception occurs, TH is preset detection threshold, and m is The dimension of measurement amount；

3. outlier and motor-driven differentiation

Adjacent sensors are transmitted mutuallyThe adjacent exception that i-th of sensor can then be acquired is total:

According toValue to determine whether there is outlier or motor-driven, expression formula is as follows:

In formula, Ω_THFor judgement threshold, value is obtained by comparing outlier probability and the relative size of motor-driven probability.

5. the adaptive consensus information filtering method according to claim 4 for distributed object tracking, feature It is, detailed process is as follows for the step 3):

1. unruly-value rejecting

When determining outlier occur in the step 2), then pass through following formula for unruly-value rejecting:

2. motor-driven processing

When determining that target occurs motor-driven in the step 2), then by the way that decay factor amplification target status predication variance square is added Battle array, and then weaken the influence of Dynamic model error bring, decay factorCalculating process it is as follows:

Wherein

In formula, σ is forgetting factor, for controlling previous momentValue for currentThe size that value influences, tr are indicated To Matrix Calculating mark, the i.e. summation of matrix leading diagonal each element, T representing matrix transposition；

3. calculating the common recognition factorWith

Wherein

When without motor-driven,Take 1.

6. the adaptive consensus information filtering method according to claim 5 for distributed object tracking, feature It is, in motor-driven treatment process, formulaThe value of middle σ is 0.95.

7. the adaptive consensus information filtering method according to claim 5 for distributed object tracking, feature It is, detailed process is as follows for the step 4):

For l=1 to L

I. it sendsWithTo neighbor node j ∈ Ν_i

II. it receivesWithFrom neighbor node j ∈ Ν_i

III. the common recognition factor updates:

end for

Wherein, L is consistency the number of iterations, and θ is that consistency adopts rate and 0 < θ <, 1/ Δ_max, Δ_maxIt is sensor network figure Maximum degree.

8. the adaptive consensus information filtering method according to claim 7 for distributed object tracking, feature It is, in the step 5), target can be calculate by the following formula in the estimated value and variance of k moment state:

Technical field

The invention belongs to wireless sensor network target tracking technique fields, and particularly, target machine can be reduced by being related to one kind The adaptive consensus information filtering method for distributed object tracking that dynamic and sensor measurement outlier influences.

Background technique

Target following, which refers to, obtains metrical information relevant to dbjective state using single or multiple sensors, then passes through Certain estimation method determines the status information of target, such as location information, velocity information etc. that target is current.With sensor The development of technology, wireless sensor network are widely applied in target tracking domain, by using wireless sensor network, Such as radar netting, optical camera network, the tracking accuracy and tracking stability of target are available to be significantly improved.State estimation Method is the key that realize target following, is roughly divided into two classes, i.e. centralized approach and distributed method.Estimate in collected state In meter method, the measurement data of all the sensors, which requires to be sent to fusion center, to be further processed, and in turn results in fusion The calculating at center and communications burden are very heavy；And distributed state estimation method does not need any fusion center then, sensor Measurement data information exchange is achieved that between adjacent node.Thus, distributed state estimation method have it is good can Scalability, calculation amount is small, has many advantages, such as robustness to the failure of sensor node.

In distributed state estimation method, the method based on consistency has global convergence and is easily achieved, common The Kalman filtering (CKF, consensus-based Kalman filter) having based on consistency and the letter based on consistency Breath filtering (CIF, consensus-based information filter)；Wherein, CIF has higher computational efficiency, more Field is tracked suitable for distributed object.And in order to realize the distributions estimation to nonlinear system, common CIF includes Extended information filter (CEIF, consensus-based extend information filter) and base based on consistency In consistency without mark information filter (CUIF, consensus-based unscented information filte)；Its In, higher tracking accuracy may be implemented in CUIF, because it is equal using the posteriority that Unscented transform carrys out approximate random state variable Value and covariance, and do not need to derive Jacobian matrix, it is easier to it is applied among not homologous ray.

Tracking for maneuvering target is a very challenging problem.When target occurs motor-driven, due to filter Kinetic model used in wave device cannot really reflect the motion state of target, and traditional method for estimating state may go out Now dissipate.In addition, sensor measurement outlier is also one of the unfavorable factor for causing performance of target tracking to reduce.In order to make distribution Method for estimating state is more applicable in, it would be desirable to be improved regarding to the issue above to conventional method.

Existing improved technology scheme is as follows:

Chinese patent CN201810020071.6 disclose a kind of maneuvering target based on distributed sensor consistency with Track method, this method use interactive multi-model (IMM, interacting multiple model) on the basis of CUIF Thought, multiple motion models are combined into system model collection, the probability of each model is constantly adjusted during tracking, so that being System model is more nearly actual conditions during target maneuver, to realize the tracking to maneuvering target.

Chinese patent CN201811136618.5 discloses the strong tracking fading factor meter in a kind of distributed fusion structure Calculation method, core are (also known as to be declined when target state is mutated due to motor-driven generation by calculating strong tracking fading factor Subtracting coefficient) the predicting covariance battle array of dbjective state is adjusted, the motor-driven influence for target tracking accuracy can be reduced, And then effective tracking of the realization to maneuvering target.

Chinese patent CN201510341115.1 discloses a kind of method of abnormal value removing and correction and device, utilizes i-th of sensor Actual measured value and estimate measured value building test value, and judge whether the test value at each moment is less than pre-determined threshold, when When test value is greater than pre-determined threshold, then the measured value at the moment is determined for outlier and is rejected, by sensor measurement open country The detection and rejecting of value, can effectively improve tracking accuracy.

But there are still following defects for above-mentioned improvement project: (1) sensor measurement outlier and target maneuver can cause mesh The mutation of estimated state is marked, but prior art can not accurately distinguish both of these case and take corresponding treatment measures；(2) IMM method can not be modeled to using the target of Impulse maneuver, and the calculating in strong tracking technology for fading factor is all base It is realized in Kalman filtering, is unsuitable for applying in distributed method.

Summary of the invention

It is suitable for distributed object tracking the purpose of the present invention is to provide one kind and reduces by sensor outlier and target The adaptive consensus information filtering method of motor-driven brought adverse effect, to solve the problems, such as to propose in background technique.

To achieve the above object, the present invention provides a kind of adaptive consensus information filters for distributed object tracking N number of sensor node is arranged in wave method in sensor network, and communication topological structure is indicated by non-directed graph G=(C, E), In, C={ 1,2 ..., N } indicates the set of node, and E={ (i, j) | i, j ∈ C } indicates the set on side, adjacent with i-th of node Node set be expressed as N_i=j ∈ C | (i, j) ∈ E } and have N in the set_iA node；

Target dynamics model and sensor measurement model are expressed as form:

x_k=f (x_k-1)+w_k-1

In formula, x_kIt is state vector of the target k-th of moment, f is the non-linear letter determined by target dynamics characteristic Number, w_k-1It is kinetic model noise,It is measured value of i-th of sensor at the kth moment, hⁱIt is the measurement of i-th of sensor Function,It is measurement noise, in addition, w_k-1WithIt is zero mean Gaussian white noise and the two is uncorrelated, w_k-1Variance be Q_k-1,Variance be

Described method includes following steps:

Step 1) linearizes nonlinear model using Unscented transform, and carries out time update；

Step 2) constructs test statistics according to the metrical information at current time, then judges the measurement of each sensor Whether value there is exception, last to judge whether outlier or motor-driven occur according to the sum for occurring exception in adjacent sensors；

Step 3) is based on above-mentioned judging result, if outlier occurs in sensor measurement, by unruly-value rejecting, if machine occurs in target It is dynamic, then utilize the influence of decay factor inhibition dynamics model error；

Step 4) carries out consistency iteration to the obtained information of sensors multiple in network；

Step 5) updates Target state estimator value and variance.

Preferably, detailed process is as follows for the step 1):

1., Sigma point sampling

Sigma point set is combined into { χ_i,k-1| τ=0 ..., 2n, k >=1 }, as follows:

In formula, λ=α²(n+K)-n, 0≤α≤1 and the distribution for controlling Sigma point, K=3-n,It is The row of the cholsky factor, the mean value of stochastic regime variable and the weights omega of covariance value are as follows:

In formula, β is parameter related with the prior distribution of state, and the subscript of ω represents the type of weight, when being above designated as m It is the weight of covariance value when being above designated as c for the weight of mean value；

2. the time updates

Set k moment dbjective state estimated value and variance asWithThe then predicted value of information stateAnd letter Cease the predicted value of matrixIt is calculate by the following formula to obtain:

And

Preferably, during Sigma point sampling, β is set as 2 in the case of a gaussian distribution.

Preferably, detailed process is as follows for the step 2):

1. constructing test statistics

It is as follows in the test statistics at k moment for i-th of sensor:

Wherein,

In formula,For test statistics,Newly to cease, (filtering essential term, for actual measurement amount and estimates measurement and measures it Between difference),To estimate measured value,To measure autocorrelation matrix；

2. judging whether measured value exception occurs

Due toChi square distribution is obeyed when without exception, and measurement value sensor can be detected according to the following formula:

Wherein

In formula,Representing whether measured value of i-th of sensor at the k moment exception occurs, TH is preset detection threshold, M is the dimension of measurement amount；

3. outlier and motor-driven differentiation

Adjacent sensors are transmitted mutuallyThe adjacent exception that i-th of sensor can then be acquired is total:

According toValue to determine whether there is outlier or motor-driven, expression formula is as follows:

In formula, Ω_THFor judgement threshold, value is obtained by comparing outlier probability and the relative size of motor-driven probability.

Preferably, detailed process is as follows for the step 3):

1. unruly-value rejecting

When determining outlier occur in the step 2), then pass through following formula for unruly-value rejecting:

2. motor-driven processing

When determining that target occurs motor-driven in the step 2), then by the way that decay factor amplification target status predication side is added Poor matrix, and then weaken the influence of Dynamic model error bring, decay factorCalculating process it is as follows:

Wherein

In formula, σ is forgetting factor, for controlling previous momentValue for currentThe size that value influences, Tr is indicated to Matrix Calculating mark, the i.e. summation of matrix leading diagonal each element, T representing matrix transposition；

3. calculating the common recognition factorWith

Wherein

When without motor-driven,Take 1.

Preferably, in motor-driven treatment process, formulaThe value of middle σ is 0.95。

Preferably, detailed process is as follows for the step 4):

For l=1to L

I. it sendsWithTo neighbor node j ∈ Ν_i

II. it receivesWithFrom neighbor node j ∈ Ν_i

III. the common recognition factor updates:

end for

Wherein, L is consistency the number of iterations, and θ is that consistency adopts rate and 0 < θ <, 1/ Δ_max, Δ_maxIt is sensor network The maximum degree of figure.

Preferably, in the step 5), target can be calculate by the following formula in the estimated value and variance of k moment state:

Technical solution provided by the invention at least has the following beneficial effects:

The method of the present invention is for the Target Tracking Problem in distributed sensor networks, there are coloured noises and target machine In the case where dynamic, outlier can be reduced and motor-driven to target with effective district dividing line value and motor-driven and take corresponding treatment measures The adverse effect that estimated state generates reduces target state estimator state and the case where being mutated or dissipating occurs, and then realizes to target Stable and high precision tracking.

The method of the present invention, without mark information filter, discriminates whether occur by measured value abnormality detection based on consistency Outlier is motor-driven, inhibits to improve target to Dynamic model error caused by target maneuver further through decay factor is introduced Tracking accuracy also improves the applicability of distributed state estimation method.

Detailed description of the invention

To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment Attached drawing is briefly described, it should be apparent that, drawings discussed below is only some embodiments of the present invention, for this For the those of ordinary skill of field, without creative efforts, it can also be obtained according to these attached drawings others Attached drawing, in which:

Fig. 1 is the flow chart in the present invention for the adaptive consensus information filtering method of distributed object tracking；

Fig. 2 is the Communication topology of the radar netting in the embodiment of the present invention 1；

Fig. 3 is simulation result diagram when outlier occurs in sensor in the embodiment of the present invention 1；

Fig. 4 is simulation result diagram when target occurs motor-driven in the embodiment of the present invention 1.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that the described embodiments are merely a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts all other Embodiment shall fall within the protection scope of the present invention.

Referring to Fig. 1, a kind of adaptive consensus information filtering method for distributed object tracking is specifically included as follows Step:

N number of sensor node is arranged in step 1) in sensor network, communicates topological structure by non-directed graph G=(C, E) It indicates, wherein C={ 1,2 ..., N } indicates the set of node, and E={ (i, j) | i, j ∈ C } indicates the set on side, with i-th of section The adjacent node set of point is expressed as N_i=j ∈ C | (i, j) ∈ E } and have N in the set_iA node；

Target dynamics model and sensor measurement model are expressed as form:

x_k=f (x_k-1)+w_k-1

In formula, x_kIt is state vector of the target k-th of moment, f is the non-linear letter determined by target dynamics characteristic Number, w_k-1It is kinetic model noise,It is measured value of i-th of sensor at the kth moment, hⁱIt is the measurement of i-th of sensor Function,It is measurement noise, in addition, w_k-1WithIt is zero mean Gaussian white noise and the two is uncorrelated, w_k-1Variance be Q_k-1,Variance be

Time update is linearized and carried out to nonlinear model using Unscented transform, detailed process is as follows:

1., Sigma point sampling

Sigma point set is combined into { χ_i,k-1| τ=0 ..., 2n, k >=1 }, as follows:

In formula, λ=α²(n+K)-n, 0≤α≤1 and the distribution for controlling Sigma point, K=3-n,It is The row of the cholsky factor, the mean value of stochastic regime variable and the weights omega of covariance value are as follows:

In formula, β is parameter related with the prior distribution of state, and β is set as the upper of 2, ω in the case of a gaussian distribution Mark represents the type of weight, is the weight of mean value when being above designated as m, is the weight of covariance value when being above designated as c；

2. the time updates

Set k moment dbjective state estimated value and variance asWithThe then predicted value of information stateAnd letter Cease the predicted value of matrixIt is calculate by the following formula to obtain:

And

Step 2) constructs test statistics according to the metrical information at current time, then judges the measurement of each sensor Whether value there is exception, last to judge whether outlier or motor-driven occur according to the sum for occurring exception in adjacent sensors, specifically Process is as follows:

1. constructing test statistics

It is as follows in the test statistics at k moment for i-th of sensor:

Wherein,

In formula,For test statistics,Newly to cease,To estimate measured value,To measure autocorrelation matrix；

2. judging whether measured value exception occurs

Due toChi square distribution is obeyed when without exception, and measurement value sensor can be detected according to the following formula:

Wherein

In formula,Representing whether measured value of i-th of sensor at the k moment exception occurs, TH is preset detection threshold, M is the dimension of measurement amount；

3. outlier and motor-driven differentiation

Adjacent sensors are transmitted mutuallyThe adjacent exception that i-th of sensor can then be acquired is total:

According toValue to determine whether there is outlier or motor-driven, expression formula is as follows:

In formula, Ω_THFor judgement threshold, value is obtained by comparing outlier probability and the relative size of motor-driven probability.

Step 3) is based on above-mentioned judging result, if outlier occurs in sensor measurement, by unruly-value rejecting, if machine occurs in target It is dynamic, then the influence of decay factor inhibition dynamics model error is utilized, detailed process is as follows:

1. unruly-value rejecting

When determining outlier occur in the step 2), then pass through following formula for unruly-value rejecting:

2. motor-driven processing

When determining that target occurs motor-driven in the step 2), then by the way that decay factor amplification target status predication side is added Poor matrix, and then weaken the influence of Dynamic model error bring, decay factorCalculating process it is as follows:

Wherein

In formula, σ is forgetting factor, for controlling previous momentValue for currentThe size that value influences, It is rule of thumb 0.95 by σ value, is conducive to the stability for enhancing algorithm, tr is indicated to Matrix Calculating mark, i.e. matrix leading diagonal The summation of each element, T representing matrix transposition；

3. calculating the common recognition factorWith

Wherein

When without motor-driven,Take 1.

Step 4) carries out consistency iteration to the obtained information of sensors multiple in network, and detailed process is as follows:

For l=1to L

I. it sendsWithTo neighbor node j ∈ Ν_i

II. it receivesWithFrom neighbor node j ∈ Ν_i

III. the common recognition factor updates:

end for

Wherein, L is consistency the number of iterations, and θ is that consistency adopts rate and 0 < θ <, 1/ Δ_max, Δ_maxIt is sensor network The maximum degree of figure.

Step 5) updates Target state estimator value and variance, and detailed process is as follows:

Target can be calculate by the following formula in the estimated value and variance of k moment state: