CN104050641A - Centralized multi-sensor column target particle filtering algorithm based on shape and direction descriptors - Google Patents

Abstract

In order to meet the engineering demand that targets in a column are accurately tracked through multiple sensors under the complicated background of cloud and rain clutter, banding interference and the like and overcome the defect that a traditional multi-sensor multi-target tracking algorithm and an existing column target tracking algorithm both have difficulty in achieving an ideal tracking effect, the invention provides a centralized multi-sensor column target particle filtering algorithm based on shape and direction descriptors according to the characteristic that true echo space structures of the targets in the same non-motorized column at adjacent moments are fixed relatively. According to the centralized multi-sensor column target particle filtering algorithm based on the shape and direction descriptors, state update of multi-dimensional trace points of the targets in the column is achieved based on particle filtering according to redundancy trace points in graph similarity removal state prediction, and the targets in the centralized multi-sensor column are tracked accurately.

Description

Based on the centralized multisensor formation target particle filter algorithm of shape orientation descriptor

One, technical field

The invention belongs to multiple-sensor and multiple-object information integration technology field, the centralized multisensor formation target track algorithm under a kind of complex background is provided.

Two, background technology

Traditional multiple-sensor and multiple-object track algorithm is very limited to the tracking effect of formation target.This type of algorithm is directly built boat to formation internal object based on measuring conventionally, but because formation internal object spacing is less, each target following ripple door can be seriously overlapping, and data interconnection difficulty increases; And similar because of formation internal object behavior pattern, the track initiation of mistake and peace preservation association are along with time integral causes the serious confusion of overall situation.

Recent domestic scholar has proposed a series of formation target track algorithms, and basic ideas are mostly: the equivalence that utilizes various technology to set up formation measures, and measures and realizes the entirety tracking of forming into columns based on equivalence.This type of algorithm has reduced to a certain extent follows the tracks of occurrence probability chaotic and calculated amount blast, improve the stability of whole tracker, save a large amount of radar resources, but along with the raising of sensor resolution, progressively show following deficiency: the first, the derivation environment of algorithm is mostly fairly simple, conventionally in hypothesis formation, individual goal can be distinguished completely, but under actual conditions, because of blocking mutually and the factor such as environmental interference of target, formation target normally part can be distinguished; The second, in some practical engineering application, in following the tracks of whole formation, need to follow the tracks of separately the interior individual goal of forming into columns; The 3rd, for effectively improving the accurate tracking effect of formation internal object, in engineering, need to utilize multi-section sensor, observe formation targets from different direction findings, but existing algorithm is only considered single-sensor situation, and complicated multisensor situation is not studied.

Three, summary of the invention

1. the technical matters that will solve

Compared with traditional multiple target tracking, the accurate tracking of formation internal object is more complicated, utilizes traditional multiple target tracking algorithm to maintain the situations such as formation internal object flight path there will be leakage to follow, mistake is followed, followed more, and tracking effect is needed improvement badly.Existing formation target track algorithm is followed the tracks of based on formation entirety mostly, does not consider the accurate tracking problem of formation internal object; The algorithm application environment that fraction consideration formation internal object flight path maintains is relatively single again, is difficult to be applicable to the complex background such as sexual intercourse clutter, banded interference; In addition,, while utilizing networking sensor to survey formation target in engineering, must need to carry out data interconnection and fusion treatment, and for centralized multisensor formation target tracking technique, still there is no at present document research.

For the problems referred to above, be necessary to analyse in depth the measurement characteristic of internal object of forming into columns in centralized multisensor syste under the complex environment such as sexual intercourse clutter, banded interference, how research eliminates to greatest extent sexual intercourse clutter and the banded adverse effect of disturbing under the prerequisite that does not affect formation internal object tracking effect, the multidimensional point of realizing formation internal object navigates interconnected and measures merging, completes the accurate tracking of centralized multisensor formation internal object under complex environment.

2. technical scheme

Under complex background, the non-mobile formation metric data of multisensor has following characteristics: each measurement at most may be from a target; Each target has at most a true echo in each moment; Certain sensor may not provide metric data under sky at a time.For above feature and the relatively-stationary characteristic of the true echo space structure of the same formation internal object of adjacent moment, the centralized multisensor formation target particle filter algorithm based on shape orientation descriptor is proposed.This algorithm comprises following techniqueflow: the rejecting of the foundation of formation target shape vector, the foundation of similarity model, redundant image, the state based on particle filter upgrade.

3. beneficial effect

(1) than existing formation target track algorithm, the present invention can be good at overcoming sexual intercourse clutter, banded interference and target occlusion and flight path is maintained to the impact causing, effectively realize the effective interconnected and fusion of data between different system sensors, the positioning precision of raising formation internal object, realize the meticulous tracking of formation target;

(2) than existing formation target track algorithm, the present invention has higher target following rate and shorter consuming time, and algorithm has stronger robustness simultaneously, along with the increase of clutter number in environment, it is higher that the target following rate of algorithm keeps, and increase consuming time is relatively little.

Four, brief description of the drawings

Fig. 1 is the centralized multisensor formation target particle filter algorithm process flow diagram based on shape orientation descriptor;

Fig. 2 is formation target G ^t(k-1) shape schematic diagram;

Fig. 3 is G ^t(k-1) image orientation frame schematic diagram;

Fig. 4 is G ^t(k-1) image constraint frame schematic diagram.

Five, embodiment

In conjunction with algorithm flow chart shown in Fig. 1, the centralized multisensor formation target particle filter algorithm embodiment based on shape orientation descriptor is as follows:

(1) foundation of formation target shape vector

Shape orientation descriptor is a kind of common method of describing space diagram in Digital Image Processing, and 12 components in height, width, area, ratio, least radius, maximum radius, least radius angle, maximum radius angle that it retrains frame by height, width, area, ratio, the image of image orientation frame form.

If the polygon A in Fig. 2 is certain moment formation target G ^t(k-1) flat shape figure, t ₁, t ₂, t ₃, t ₄for the position of the interior each target of forming into columns.Utilize shape orientation descriptor to set up the shape vector of A

Ω _t(k-1)＝{ω ₁，ω ₂，ω ₃，ω ₄，ω ₅，ω ₆，ω ₇，ω ₈，ω ₉，ω ₁₀，ω ₁₁，ω ₁₂} (1)

Wherein ω ₁for the height of image orientation frame; Image orientation frame is the minimum rectangle of surrounding object along image route, as shown in rectangle B in Fig. 3;

ω ₂for the width of B;

ω ₃for the area of B;

ω ₄for the ratio of B, i.e. the ratio of the area of B and the area of A;

ω ₅for the height of image constraint frame; Image constraint frame is the minimum rectangle of surrounding object along image major axes orientation, as shown in the rectangle C in Fig. 4;

ω ₆for the width of C;

ω ₇for the area of C;

ω ₈for the ratio of C, i.e. the ratio of the area of C and the area of A;

ω ₉for least radius, i.e. minor increment between the center of gravity of A and the boundary element of A;

ω ₁₀for maximum radius, i.e. ultimate range between the center of gravity of A and the boundary element of A;

ω ₁₂for least radius angle, least radius vector is with respect to the angle of horizontal axis;

ω ₁₂for maximum radius angle, maximum radius vector is with respect to the angle of horizontal axis.

It should be noted that in the time processing the special graphs such as straight line Ω _t(k-1) some component in possibly cannot obtain, and now removes these components.Seen from the above description, the expression polygon A that shape vector can be unique, if both the shape vector of two figures was identical, judges that these two figures are consistent.Algorithm of the present invention upgrades each dbjective state vector in formation in each moment, upgrades the shape vector of each formation target simultaneously; The shape vector of note k-1 moment formation t flight path is ${\mathrm{\Ω}}_t(k-1)=\{{\mathrm{\ω}}_n^t(k-1),n=1,...,12\}.$

(2) foundation of similarity model

With G ^t(k-1) in, each target is set up associated ripple door centered by the one-step prediction value in k moment, establishes Z (k) and declines and into the measurement collection of associated ripple door be

In formula, for falling into the measurement collection of i target association Bo Mennei of formation t.If classify by sensor difference,

In formula, for come from the number of sensor; for in come from the measurement number of sensor s.

For the gauge point in sensor s, do traversal combination by the associated ripple door under it, generate J ^sindividual measurement collection guarantee each measurement concentrate have and only have individual gauge point, and be derived from respectively different associated ripple doors,

${\mathrm{\τ}}_l^s\left(k\right)=\{z_{1,j_1}^s\left(k\right),z_{2,j_2}^s\left(k\right),...,z_{N_g^t(k-1),j_{N_g^t(k-1)}}^s\left(k\right)\},l=\mathrm{1,2},...,J^s---\left(4\right)$

In formula for in come from arbitrary measurement of sensor s; If use one-step prediction value replace, and definition

Definition event for measuring collection with G ^t(k-1) event corresponding to each targetpath in; For single-sensor,

$\underset{l=1}{\overset{J^s}{\mathrm{\Σ}}}\mathrm{\ρ}\left({\mathrm{\θ}}_l^s\left(k\right)\right)=1---\left(6\right)$

For convenience, note from the angle of figure, in measurement form a figure based on shape orientation its shape vector of descriptor computation be in adjacent moment, the figure that the true echo of non-mobile formation internal object forms should be similar to the figure that the interior each dbjective state of forming into columns forms; Characterize by building similarity at this with the similarity of A, definition similarity for

$q_l^s\left(k\right)=\frac{1}{\underset{n=1}{\overset{12}{\mathrm{\Σ}}}|{\mathrm{\ω}}_{\mathrm{nl}}^s\left(k\right)-{\mathrm{\ω}}_n^t(k-1)|}---\left(12\right)$

larger, figure is described more similar to A, both for really probability is larger.For ease of comparative descriptions, need to be normalized, in this definition for

${\mathrm{\ρ}}_l^s\left(k\right)=\frac{q_l^s\left(k\right)}{\underset{l=1}{\overset{J^s}{\mathrm{\Σ}}}{q}_l^s\left(k\right)}---\left(8\right)$

From formula (7) and formula (8), only consider figure with the inner structure consistance of A, do not consider to measure collection with G ^t(k-1) relevance of each dbjective state renewal value in, therefore separately based on judgement with G ^t(k-1) interconnected possibility imperfection, further defines similarity probability herein for

${\mathrm{\ρ}}_l^{\′s}\left(k\right)=\frac{q_l^{\′s}\left(k\right)}{\underset{l=1}{\overset{J^s}{\mathrm{\Σ}}}{q}_l^{\′s}\left(k\right)}---\left(9\right)$

In formula,

$q_l^{\′s}\left(k\right)=\frac{1}{\underset{i=1}{\overset{N_g^t(k-1)}{\mathrm{\Σ}}}[{\hat{z}}_{\mathrm{il}}^s\left(k\right)-{\hat{z}}_i^t\left(k\right|k-1)]}---\left(10\right)$

In formula, for in be derived from the gauge point of i ripple door, for i target in formation t is in the state one-step prediction value in k moment.

Finally, consider with definition similarity probability for

${\mathrm{\ρ}}_l^{\′\′s}\left(k\right)=\frac{{\mathrm{\ρ}}_l^s\left(k\right)+{\mathrm{\ρ}}_l^{\′s}\left(k\right)}{2}---\left(11\right)$

(3) rejecting of redundant image

In the time utilizing multisensor to survey formation target, can obtain the multiple image of corresponding same formation target, so need to carry out the rejecting of redundant image, the thought at this based on selecting main website addresses this problem.

Based on similarity Making by Probability Sets definition

$s^{*}=\underset{s=1:N_s}{\arg \max }\{\underset{l=1:J^s}{\max },{\mathrm{\ρ}}_l^{\′\′s}\left(k\right)\}---\left(12\right)$

Represent to select sensor s ^*for main website, and utilize sensor s ^*the measurement collection reporting to G ^t(k-1) carry out state renewal.

(4) state based on particle filter upgrades

Based on set with utilize particle filter to carry out formation G ^t(k-1) in, each targetpath carries out state renewal

${\hat{X}}_i^t\left(k\right|k)=\underset{l=1}{\overset{J^{s^{*}}}{\mathrm{\Σ}}}{\mathrm{\ρ}}_l^{\′\′s^{*}}\left(k\right){\hat{X}}_{\mathrm{il}}^t\left(k\right|k)i=\mathrm{1,2},...,N_g^t(k-1)---\left(13\right)$

$P_i^t\left(k\right|k)=\underset{l=1}{\overset{J^{s^{*}}}{\mathrm{\Σ}}}{\mathrm{\ρ}}_l^{\′\′s^{*}}\left(k\right)P_{\mathrm{il}}^t\left(k\right|k)i=\mathrm{1,2},...,N_g^t(k-1)---\left(14\right)$

In formula, be respectively based on state renewal value and the covariance renewal value of utilizing particle filter method to obtain.

Claims (1)

1. the present invention is for the meticulous tracking of centralized multisensor formation target under complex background, and technical characteristics is the construction method of the formation target shape vector similarity model based on shape orientation descriptor:

Based on shape orientation descriptor computation formation target G ^t(k-1) shape vector with formation G ^t(k-1) in, each target is set up associated ripple door centered by the one-step prediction value in k moment, and the gauge point in ripple door is reclassified by the sensor under it; The gauge point that same sensor is provided does traversal combination by the associated ripple door under it, guarantees that each measurement concentrates and have and only have the several gauge points of ripple door, and is derived from respectively different associated ripple doors; The gauge point in sensor s can build J ^sindividual measurement collection j ^sfor the k moment form into columns in the measurement of each ripple door product of counting; If measure collection in gauge point form figure calculating its shape vector is definition similarity for

For ease of comparative descriptions, it is right to need be normalized, in this definition for

Consider to measure collection with G ^t(k-1) relevance of each dbjective state renewal value in, definition

In formula for in be derived from the gauge point of i ripple door, for i target in forming into columns is in the state one-step prediction value in k moment,

Consider with definition similarity probability for

。