CN104200492A - Automatic detecting and tracking method for aerial video target based on trajectory constraint - Google Patents

Abstract

The invention relates to an automatic detecting and tracking method for an aerial video target based on trajectory constraint. The automatic detecting and tracking method comprises the steps of firstly, automatically completing moving object detection in an aerial image by use of a forward image and the motion information of the forward image, secondly, removing parallax-induced mistaken static object detection included in the detection result in a trajectory constraint manner, and finally, analyzing the tracking result of a real moving object and removing the influence of a noise target by use of the statistics of motion trajectory information, and then completing automatic detection and tracking on the aerial video target; the accuracy of the detection and tracking result is more than 80%.

Description

The video object of taking photo by plane based on track constraint detects tracking automatically

Technical field

The present invention relates to the one video object of taking photo by plane online and automatically detect tracking, particularly relate to a kind of video object of taking photo by plane based on track constraint and automatically detect tracking.

Background technology

Utilize moving target in space and coherent kinetic characteristic on the time, automatically, effectively moving target is detected and carried out from the video image of taking photo by plane of complexity the online target following of robust, have very important significance.The existing video object of taking photo by plane automatically detects tracking and mainly contains: the video object of taking photo by plane based on background difference detects tracking and the video object of taking photo by plane based on kinetic characteristic statistics detects tracking.

Document " Moving object localization in thermal imagery by forward-backward MHI.Computer Vision and Pattern Recognition Workshop; 133-140,2006. " discloses a kind of video object of taking photo by plane based on forward-backward algorithm motion history image and has detected and positioning and tracing method.The method is obtained the moving target information of different time sections by movable information frame difference image and the backward movable information frame difference image of structure forward direction, and on the fusion basis of two images, completes the video frequency motion target detection of taking photo by plane automatically.But the method is mainly the poor information of gray scale frame based on image adds up, the result of statistics is that the Pixel-level half-tone information on image changes.Take photo by plane while obtaining over the ground video image when unmanned plane utilizes video sensor to carry out high-altitude, three-dimensional spatial information can be transformed into two-dimentional image information, cause the disappearance of picture depth information.Utilize the video object of taking photo by plane based on forward-backward algorithm motion history image to detect and positioning and tracing method except detecting that actual moving target is to picture, also comprised a large amount of static to picture because depth of field parallax in altitude causes.The data acquisition of unmanned plane earth observation is with ground as base plane, and the parallax static target therefore being caused by depth of field parallax in altitude can produce the two dimensional image movement locus that is different from ground.Therefore, the detection meeting of take photo by plane video object detection and positioning and tracing method based on forward-backward algorithm motion history image, because static target is detected as moving target causes a large amount of false-alarms, is followed the tracks of and has been brought very large challenge for follow-up location.

Summary of the invention

The technical matters solving

For fear of the deficiencies in the prior art part, the present invention proposes a kind of video object of taking photo by plane based on track constraint and automatically detects tracking.

Technical scheme

A kind of video object of taking photo by plane based on track constraint detects tracking automatically, first construct forward direction image and backward image, thereby and merge two movable informations on image and form and automatically complete the moving object detection on image when pre-treatment image, comprise actual moving target and the static target being caused by parallax; Secondly, utilize back projection matrix, calculate the respective pixel of each pixel on forward direction image and obtain the movement locus of each pixel; Finally, utilize the pole-face image-forming principle of camera to calculate track constraint matrix, analyze the tracking results of real motion target the impact of removing the static target causing because of parallax, it is characterized in that step is as follows:

Step 1: the method first suppressing by Corner Detection and non-maximum value is to forward direction image I ^fin t two field picture with t-Δ frame image carries out feature point extraction, Δ>=1, and utilize RANSAC method to remove the exterior point on two two field pictures; Then utilize Optical-flow Feature computing method to obtain t frame forward direction image I ^fmotion affine matrix finally by construct image with between affine variation relation:

$I_t^F=P_{(t,t-\mathrm{\Δ})}^F\×I_{t-\mathrm{\Δ}}^F---\left(1\right)$

Adopt the mode of incremental computations to substitute directly calculating with motion affine matrix

$P_{(t,t-\mathrm{\Δ})}^F=P_{(t,t-1)}^F\×P_{(t-1,t-2)}^F\×\·\·\·\×P_{(t-\mathrm{\Δ}+1,t-\mathrm{\Δ})}^F---\left(2\right)$

Wherein, be with motion affine matrix, be with motion affine matrix, be with motion affine matrix;

Step 2: by two two field pictures with obtain forward direction image I ^fthe poor moving image of frame at t frame:

$D{(x,y,t)}^F=|I_t^F-I_{t-\mathrm{\Δ}}^F|---\left(3\right)$

Step 3: according to and D (x, y, t) ^fconstruct forward direction image H (x, y, t) ^fand utilize Gaussian filter function to carry out smoothing processing and eliminate noise; Adopting uses the same method calculates backward image I ^bthe motion affine matrix of t frame with the poor moving image D of frame (x, y, t) ^bconstruct backward image H (x, y, t) ^b:

$H{(x,y,t)}^F=\left\{\begin{array}{cc}\max {(0,P_{(t,t-\mathrm{\&Delta;})}^{F}H(x,y,t-1)-d)}^F& \mathrm{if\; D}{(x,y,t)}^F<T_F\\ 255& \mathrm{if\; D}{(x,y,t)}^F\&GreaterEqual;T_F\end{array}\right.---\left(4\right)$

$H{(x,y,t)}^B=\left\{\begin{array}{cc}\max {(0,P_{(t,t-\mathrm{\&Delta;})}^{B}H(x,y,t-1)-d)}^B& \mathrm{if\; D}{(x,y,t)}^B<T_B\\ 255& \mathrm{if\; D}{(x,y,t)}^B\&GreaterEqual;T_B\end{array}\right.---\left(5\right)$

Wherein, x, y is the pixel coordinate of representative image respectively, and d is pad value parameter, T _f, T _bfor threshold value;

Step 4: by forward direction image H (x, y, t) ^fwith backward image H (x, y, t) ^bcarrying out two-value merges associated:

H(x,y) _t＝max(H(x,y,t) ^F,H(x,y,t) ^B) (6)

And remember H (x, y) _tthe pixel to be detected that all pixel values on image are 255 is

Step 5: the motion affine matrix that utilizes forward direction image each measuring point p to be checked of backwards calculation t frame _t=(x _t, y _t) ∈ Θ _tthe pixel position p of correspondence in t-1 two field picture _t-1:

$p_{t-1}={\left(P_{(t,t-1)}^F\right)}^T{p}_t---\left(7\right)$

Step 6: with p _t-1centered by, in the β image range point set Q that is radius, utilize regional extent search function F (p _t-1, β) find and p _tthere is the pixel coordinate point p' of maximum similarity _t-1, as p _tin the actual corresponding pixel points of t-1 frame, and set vector parameter ε (p' _t-1, p _t) as p _tat the movement locus of t-1 frame and t frame;

F(p _t-1,β)＝maxf(p _t-1,q) (8)

$f(p_{t-1},q)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}(1-\frac{|{\mathrm{\&rho;}}_p\left(i\right)-{\mathrm{\&rho;}}_q\left(i\right)|}{\mathrm{Max}({\mathrm{\&rho;}}_p\left(i\right),{\mathrm{\&rho;}}_q\left(i\right))})---\left(9\right)$

Wherein, q ∈ Q={q|||l (p _t-1)-l (q) || < β }, l (p _t-1) represent with p _t-1centered by the piece image of 3 × 3 pixel sizes of point, l (q) represents the piece image of 3 × 3 pixel sizes of putting centered by q, ρ _pand ρ (i) _q(i) represent respectively l (p _t-1) and the statistics with histogram of l (q), N fixes value 255;

Step 7: utilize track similar function κ (ε (p' _t-1, p _t)) each measuring point p to be checked of calculating t frame _ttrack ε (p' _t-1, p _t) with the associated similarity of the track of all measuring points to be checked, be 1 or 0 to judge moving target pixel point set according to similar value with static scene point set

$\mathrm{\&kappa;}\left({\mathrm{\&epsiv;}}_p\right)=\left\{\begin{array}{cc}1,& \underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_p-{\mathrm{\&epsiv;}}_i\left|\right|>\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_i-{\mathrm{\&epsiv;}}_j\left|\right|\\ 0,& \underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_p-{\mathrm{\&epsiv;}}_i\left|\right|\&le;\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_i-{\mathrm{\&epsiv;}}_j\left|\right|\end{array}\right.---\left(10\right)$

Wherein, ε _prepresent measuring point p to be checked _ttrack ε (p' _t-1, p _t), ε _i, ε _jrepresent respectively the track of each measuring point to be checked in all measuring points to be checked of t frame, as κ (ε _p)=1 o'clock, as κ (ε _p)=0 o'clock,

Step 8: adopt the mode of facing recently cluster, by moving target pixel point set in the movement locus ε (p' of each pixel _t-1, p _t) carry out Cluster merging synthetic image moving target track ε (o _t-1, o _t);

Step 9: to moving target movement locus ε (o _t-1, o _t) judge; If moving target movement locus be nonlinear, this moving target is spurious motion target; If moving target movement locus ε (o _t-1, o _t) be linear, this moving target is real motion target; According to the continuous path of real motion target, complete the video object of taking photo by plane and automatically detect tracking.

The span of described pad value parameter d is [10,30], threshold value T _fspan be [100,150], threshold value T _fspan be [100,150].

The span of described β is [0.1,3].

Beneficial effect

A kind of video object of taking photo by plane based on track constraint that the present invention proposes detects tracking automatically, first the method constructs forward direction image and backward image, and merge two movable informations on image and automatically complete the moving object detection on image, comprise actual moving target and the static target being caused by parallax; Secondly, utilize back projection matrix, calculate the respective pixel of each pixel on forward direction image and obtain the movement locus of each pixel; Finally, analyze the tracking results of real motion target the impact of removing the static target causing because of parallax.More than detecting tracking results rate of accuracy reached to 80%.

Embodiment

The video object of taking photo by plane based on track constraint detects a tracking automatically, comprises the following steps:

In the image that step 1, unmanned plane sensor obtain, be subject to unmanned plane during flying disturbing influence, ground static scene shows on sensor video image to be motor pattern, therefore according to the mode of motion of camera sensor, calculate the motion affine matrix between two two field pictures, obtain ground static scene fixing between two two field pictures.First the method suppressing by Corner Detection and non-maximum value is carried out feature point extraction to two two field pictures, and utilizes RANSAC method to remove the exterior point on two two field pictures; Then utilize the unique point remaining to calculate the motion affine matrix between two two field pictures, utilize on this basis motion affine matrix that former frame image re-projection is carried out to the poor calculating of picture frame to a rear two field picture; Last according to the poor result of calculation of the picture frame between two two field pictures, obtain the static scene between two two field pictures.

Step 2, according to the static scene of two two field pictures, utilize the poor motion compensated image structure of frame between two two field pictures forward direction image, utilize on this basis Gaussian filter function to carry out smoothing processing and eliminate noise and obtain the forward direction image after optimization.Profit uses the same method and calculates backward image.

Step 3, detect overlaid pixel region according to the maximum of forward direction image and backward image, forward direction image and backward image are carried out to two-value and merge and obtain two-value testing result image as pixel to be detected.The Pixel Information that wherein two-value testing result image comprises is real motion object pixel information, motion artifacts information and the static target background pixel information causing because of parallax.

Step 4, utilize each measuring point to be checked of affine motion model backwards calculation corresponding pixel in previous frame image.Utilize regional extent search function correction affine motion model backwards calculation error, determine each measuring point to be checked actual corresponding pixel in previous frame image

Step 5, utilize track similar function to calculate the similarity of track and all locus of points to be detected of each measuring point to be checked.According to the movement locus agreement principle of static scene target, in conjunction with similarity result, each measuring point to be checked is classified, obtain moving target pixel point set and static scene point set.

The mode of cluster is faced in step 6, utilization recently, and adjacent moving target pixel point set is sorted out to all moving targets that obtain on image, wherein comprises real moving target Pixel Information and motion artifacts information.

The movement locus of step 7, statistics moving target obtains real moving target.If the movement locus of moving target is linear, judge that this moving target is real motion target; If the movement locus of moving target is nonlinear, judge that this moving target is spurious motion target (noise, the impact such as shelter).Finally according to the movement locus of real motion target, the moving target after to each cluster is followed the tracks of.

Now the invention will be further described in conjunction with the embodiments:

1, take photo by plane video frequency motion target detect

(a), according to the mode of motion of camera sensor, by calculating the motion affine matrix between two two field pictures, obtain the relative static scene between two two field pictures.The method suppressing by Corner Detection and non-maximum value is to forward direction image I ^fin t two field picture with t-Δ frame image carries out feature point extraction, Δ>=1, and utilize RANSAC method to remove the exterior point on two two field pictures;

(b) basis with unique point utilize Optical-flow Feature computing method to obtain t frame forward direction image I ^fmotion affine matrix by construct image with between affine variation relation:

$I_t^F=P_{(t,t-\mathrm{\&Delta;})}^F\&times;I_{t-\mathrm{\&Delta;}}^F---\left(1\right)$

In order to reduce the error of affine variation function, adopt the mode of incremental computations to substitute directly calculating with motion affine matrix

$P_{(t,t-\mathrm{\&Delta;})}^F=P_{(t,t-1)}^F\&times;P_{(t-1,t-2)}^F\&times;\&CenterDot;\&CenterDot;\&CenterDot;\&times;P_{(t-\mathrm{\&Delta;}+1,t-\mathrm{\&Delta;})}^F---\left(2\right)$

Wherein, it is basis with unique point utilize Optical-flow Feature computing method to obtain motion affine matrix, it is basis with unique point utilize Optical-flow Feature computing method to obtain motion affine matrix, it is basis with unique point utilize Optical-flow Feature computing method to obtain motion affine matrix.

(c) calculate two two field pictures with obtain forward direction image I ^fthe poor moving image of frame at t frame:

$D{(x,y,t)}^F=|I_t^F-I_{t-\mathrm{\&Delta;}}^F|---\left(3\right)$

(d) basis with construct forward direction image H (x, y, t) ^fand utilize Gaussian filter function to carry out smoothing processing and eliminate noise; Adopting uses the same method calculates backward image I ^bthe motion affine matrix of t frame with the poor moving image D of frame (x, y, t) ^bconstruct backward image H (x, y, β) ^b:

$H{(x,y,t)}^F=\left\{\begin{array}{cc}\max {(0,P_{(t,t-\mathrm{\&Delta;})}^{F}H(x,y,t-1)-d)}^F& \mathrm{if\; D}{(x,y,t)}^F<T_F\\ 255& \mathrm{if\; D}{(x,y,t)}^F\&GreaterEqual;T_F\end{array}\right.---\left(4\right)$

$H{(x,y,t)}^B=\left\{\begin{array}{cc}\max {(0,P_{(t,t-\mathrm{\&Delta;})}^{B}H(x,y,t-1)-d)}^B& \mathrm{if\; D}{(x,y,t)}^B<T_B\\ 255& \mathrm{if\; D}{(x,y,t)}^B\&GreaterEqual;T_B\end{array}\right.---\left(5\right)$

Wherein, x, y is the pixel coordinate of representative image respectively, pad value parameter d=25, threshold value T _f=125, threshold value T _b=125.

(e) by forward direction image H (x, y, t) ^fwith backward image H (x, y, t) ^bcarrying out two-value merges associated:

H(x,y) _t＝max(H(x,y,t) ^F,H(x,y,t) ^B) (6)

And remember H (x, y) _tthe pixel to be detected that all pixel values on image are 255 is

2, take photo by plane video frequency motion target track constraint

(a) utilize the motion affine matrix of forward direction image each measuring point p to be checked of backwards calculation t frame _t=(x _t, y _t) ∈ Θ _tthe pixel position p of correspondence in t-1 two field picture _t-1:

$p_{t-1}={\left(P_{(t,t-1)}^F\right)}^T{p}_t---\left(7\right)$

(b) with p _t-1centered by, in the image range point set Q that is radius of β=1.2, utilize regional extent search function F (p _t-1, β) find and p _tthere is the pixel coordinate point p' of maximum similarity _t-1, as p _tin the actual corresponding pixel points of t-1 frame, and set vector parameter ε (p' _t-1, p _t) as p _tat the movement locus of t-1 frame and t frame;

F(p _t-1,β)＝maxf(p _t-1,q) (8)

$f(p_{t-1},q)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}(1-\frac{|{\mathrm{\&rho;}}_p\left(i\right)-{\mathrm{\&rho;}}_q\left(i\right)|}{\mathrm{Max}({\mathrm{\&rho;}}_p\left(i\right),{\mathrm{\&rho;}}_q\left(i\right))})---\left(9\right)$

Wherein, q ∈ Q={q|||l (p _t-1)-l (q) || < β }, l (p _t-1) represent with p _t-1centered by the piece image of 3 × 3 pixel sizes of point, l (q) represents the piece image of 3 × 3 pixel sizes of putting centered by q, ρ _pand ρ (i) _q(i) represent respectively l (p _t-1) and the statistics with histogram of l (q), N fixes value 255.

(c) utilize track similar function κ (ε (p' _t-1, p _t)) each measuring point p to be checked of calculating t frame _ttrack ε (p' _t-1, p _t) with the associated similarity of the track of all measuring points to be checked, obtain moving target pixel point set with static scene point set ${\mathrm{\&Theta;}}_t^{\mathrm{No}-\mathrm{Object}};$

$\mathrm{\&kappa;}\left({\mathrm{\&epsiv;}}_p\right)=\left\{\begin{array}{cc}1,& \underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_p-{\mathrm{\&epsiv;}}_i\left|\right|>\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_i-{\mathrm{\&epsiv;}}_j\left|\right|\\ 0,& \underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_p-{\mathrm{\&epsiv;}}_i\left|\right|\&le;\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_i-{\mathrm{\&epsiv;}}_j\left|\right|\end{array}\right.---\left(10\right)$

Wherein, ε _prepresent measuring point p to be checked _ttrack ε (p' _t-1, p _t), ε _i, ε _jrepresent respectively the track of each measuring point to be checked in all measuring points to be checked of t frame.As κ (ε _p)=1 o'clock, as κ (ε _p)=0 o'clock,

3, take photo by plane video frequency motion target follow the tracks of

(a) adopt the mode of facing recently cluster, by moving target pixel point set in the movement locus ε (p' of each pixel _t-1, p _t) carry out Cluster merging synthetic image moving target track ε (o _t-1, o _t);

(b) to moving target movement locus ε (o _t-1, o _t) judge; If moving target movement locus ε (o _t-1, o _t) be linear, judge that this moving target is real motion target; If moving target movement locus be nonlinear, judge that this moving target is spurious motion target (noise, the impact such as shelter); According to the continuous path of real motion target, complete the video object of taking photo by plane and automatically detect tracking.

Claims (3)

1. the video object of taking photo by plane based on track constraint detects a tracking automatically, it is characterized in that step is as follows:

Step 1: the method first suppressing by Corner Detection and non-maximum value is to forward direction image I ^fin t two field picture with t-Δ frame image carries out feature point extraction, Δ>=1, and utilize RANSAC method to remove the exterior point on two two field pictures; Then utilize Optical-flow Feature computing method to obtain t frame forward direction image I ^fmotion affine matrix finally by construct image with between affine variation relation:

$I_t^F=P_{(t,t-\mathrm{\&Delta;})}^F\&times;I_{t-\mathrm{\&Delta;}}^F---\left(1\right)$

Adopt the mode of incremental computations to substitute directly calculating with motion affine matrix

$P_{(t,t-\mathrm{\&Delta;})}^F=P_{(t,t-1)}^F\&times;P_{(t-1,t-2)}^F\&times;\&CenterDot;\&CenterDot;\&CenterDot;\&times;P_{(t-\mathrm{\&Delta;}+1,t-\mathrm{\&Delta;})}^F---\left(2\right)$

Wherein, be with motion affine matrix, be with motion affine matrix, be with motion affine matrix;

Step 2: by two two field pictures with obtain forward direction image I ^fthe poor moving image of frame at t frame:

$D{(x,y,t)}^F=|I_t^F-I_{t-\mathrm{\&Delta;}}^F|---\left(3\right)$

Step 3: according to and D (x, y, t) ^fconstruct forward direction image H (x, y, t) ^fand utilize Gaussian filter function to carry out smoothing processing and eliminate noise; Adopting uses the same method calculates backward image I ^bthe motion affine matrix of t frame with the poor moving image D of frame (x, y, t) ^bconstruct backward image H (x, y, t) ^b:

$H{(x,y,t)}^F=\left\{\begin{array}{cc}\max {(0,P_{(t,t-\mathrm{\&Delta;})}^{F}H(x,y,t-1)-d)}^F& \mathrm{if\; D}{(x,y,t)}^F<T_F\\ 255& \mathrm{if\; D}{(x,y,t)}^F\&GreaterEqual;T_F\end{array}\right.---\left(4\right)$

$H{(x,y,t)}^B=\left\{\begin{array}{cc}\max {(0,P_{(t,t-\mathrm{\&Delta;})}^{B}H(x,y,t-1)-d)}^B& \mathrm{if\; D}{(x,y,t)}^B<T_B\\ 255& \mathrm{if\; D}{(x,y,t)}^B\&GreaterEqual;T_B\end{array}\right.---\left(5\right)$

Wherein, x, y is the pixel coordinate of representative image respectively, and d is pad value parameter, T _f, T _bfor threshold value;

Step 4: by forward direction image H (x, y, t) ^fwith backward image H (x, y, t) ^bcarrying out two-value merges associated:

H(x,y) _t＝max(H(x,y,t) ^F,H(x,y,t) ^B) (6)

And remember H (x, y) _tthe pixel to be detected that all pixel values on image are 255 is

Step 5: the motion affine matrix that utilizes forward direction image each measuring point p to be checked of backwards calculation t frame _t=(x _t, y _t) ∈ Θ _tthe pixel position p of correspondence in t-1 two field picture _t-1:

$p_{t-1}={\left(P_{(t,t-1)}^F\right)}^T{p}_t---\left(7\right)$

Step 6: with p _t-1centered by, in the β image range point set Q that is radius, utilize regional extent search function F (p _t-1, β) find and p _tthere is the pixel coordinate point p' of maximum similarity _t-1, as p _tin the actual corresponding pixel points of t-1 frame, and set vector parameter ε (p' _t-1, p _t) as p _tat the movement locus of t-1 frame and t frame;

F(p _t-1,β)＝maxf(p _t-1,q) (8)

$f(p_{t-1},q)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}(1-\frac{|{\mathrm{\&rho;}}_p\left(i\right)-{\mathrm{\&rho;}}_q\left(i\right)|}{\mathrm{Max}({\mathrm{\&rho;}}_p\left(i\right),{\mathrm{\&rho;}}_q\left(i\right))})---\left(9\right)$

Wherein, q ∈ Q={q|||l (p _t-1)-l (q) || < β }, l (p _t-1) represent with p _t-1centered by the piece image of 3 × 3 pixel sizes of point, l (q) represents the piece image of 3 × 3 pixel sizes of putting centered by q, ρ _pand ρ (i) _q(i) represent respectively l (p _t-1) and the statistics with histogram of l (q), N fixes value 255;

Step 7: utilize track similar function κ (ε (p' _t-1, p _t)) each measuring point p to be checked of calculating t frame _ttrack ε (p' _t-1, p _t) with the associated similarity of the track of all measuring points to be checked, be 1 or 0 to judge moving target pixel point set according to similar value with static scene point set

$\mathrm{\&kappa;}\left({\mathrm{\&epsiv;}}_p\right)=\left\{\begin{array}{cc}1,& \underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_p-{\mathrm{\&epsiv;}}_i\left|\right|>\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_i-{\mathrm{\&epsiv;}}_j\left|\right|\\ 0,& \underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_p-{\mathrm{\&epsiv;}}_i\left|\right|\&le;\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|\right|{\mathrm{\&epsiv;}}_i-{\mathrm{\&epsiv;}}_j\left|\right|\end{array}\right.---\left(10\right)$

Wherein, ε _prepresent measuring point p to be checked _ttrack ε (p' _t-1, p _t), ε _i, ε _jrepresent respectively the track of each measuring point to be checked in all measuring points to be checked of t frame, as κ (ε _p)=1 o'clock, as κ (ε _p)=0 o'clock,

Step 8: adopt the mode of facing recently cluster, by moving target pixel point set in the movement locus ε (p' of each pixel _t-1, p _t) carry out Cluster merging synthetic image moving target track ε (o _t-1, o _t);

Step 9: to moving target movement locus ε (o _t-1, o _t) judge; If moving target movement locus be nonlinear, this moving target is spurious motion target; If moving target movement locus ε (o _t-1, o _t) be linear, this moving target is real motion target; According to the continuous path of real motion target, complete the video object of taking photo by plane and automatically detect tracking.

2. the video object of taking photo by plane based on track constraint according to claim 1 detects tracking automatically, and the span that it is characterized in that described pad value parameter d is [10,30], threshold value T _fspan be [100,150], threshold value T _bspan be [100,150].

3. the video object of taking photo by plane based on track constraint according to claim 1 detects tracking automatically, and the span that it is characterized in that described β is [0.1,3].