A kind of video target tracking method flowing shape and projection group based on Grassmann
Technical field
The present invention relates to a kind of modeling method, particularly relate to a kind of video target tracking method flowing shape and projection group based on Grassmann.
Background technology
For video object significantly nonplanar attitudes vibration, the method for tracking target based on Euclidean space often causes tracking skew or the failure of target.Its main cause is that the characteristic describing the apparent different attitude in target area is not present in certain independent vector space, and traditional linear vector space processing mode cannot meet actual needs.Consider that Grassmann stream shape is a kind of entropy flow shape in Lie Group Manifold, there is the characteristic of the spacing of more applicable tolerance data point, consider that the imaging process essence of image is projective transformation process simultaneously, namely obey projective transformation group (SL (3)).
Summary of the invention
The object of the present invention is to provide a kind of video target tracking method flowing shape and projection group based on Grassmann, the method to target under experience significantly nonplanar geometric deformation, illumination variation and experience partial occlusion situation, stable tracking can both be realized, adopt the geometric deformation model of projective transformation design object more accurately, devise bimodulus video frequency object tracking algorithm.
The object of the invention is to be achieved through the following technical solutions:
Flow a shape and projection group's video target tracking method based on Grassmann, described method comprises following process:
Step 1: inputted video image sequence, totalframes is k, and original template size is m*n(unit: pixel); For the first two field picture, manually determine the target area of image, 8 dimensional vectors on projective transformation group
for the projective transformation parameter of lock-on boundary shape, t is present frame, for the first two field picture, and t=1;
Step 2: predict according to following formula
, j=1,2 ... .L.L be sampling population; V is that state moves to the velocity vector of moment t by moment t-1;
;
Step 3: the image-region corresponding to each sampling, calculates the feature basis matrix of corresponding apparent gray matrix
, then according to formula
calculate the weights of each sampling particle:
Wherein
for feature space
feature basis matrix, vector
represent the target that t frame traces into.
with
be two points that Grassmann flows on shape, leading role's degree is therebetween
;
Step 4: pass through formula
accumulate average in calculating sampling, be the estimated state of target
;
Step 5: according to the update strategy in target signature space, regeneration characteristics space
; Update strategy is as follows:
The minimum value of the proper vector of present frame and each proper vector geodesic line distance of feature space is ds, when ds is greater than given max-thresholds thresmax, illustrate that this two field picture is seriously blocked or distortion, now for ensureing the accuracy not regeneration characteristics space of feature-space information;
Otherwise upgrade template set in two kinds of situation:
(1), when in current feature space, the number of vector is less than prespecified quantity, directly the proper vector of present frame is joined in this feature space set;
(2) otherwise, replace the proper vector in feature space with maximum range value by the proper vector of present frame;
Step 6:t=t+1, if
repeat step 2. otherwise tracing process to terminate.
Advantage of the present invention and effect are:
1. the present invention makes full use of Grassmann and flows the nonlinear characteristic in shape space, the apparent change of target is regarded as the motion of the upper point of stream shape, makes full use of the intrinsic geometry characteristic of state space, design particle filter algorithm, can the apparent change of more accurate estimating target;
2. the present invention uses SL (3) group to describe the geometric deformation process of target.Compared with affined transformation, more precisely describe the deformation process of target, more accurately the geometric deformation of target of prediction;
3. design effectively target signature spatial update strategy of the present invention, in target apparent feature space on-line study process, effectively shields abnormal information, guarantees the accuracy of feature space.
Accompanying drawing explanation
Fig. 1 shows the video frequency object tracking concrete steps flowing shape and projection group based on Grassmann;
Fig. 2 shows the result of described algorithm keeps track geometric deformation target;
Fig. 3 shows the result under described algorithm keeps track non-rigid illumination variation;
Fig. 4 shows the result of described algorithm keeps track experience partial occlusion target.
Embodiment
Below in conjunction with embodiment, the present invention is described in detail.
Embodiment 1:
Step 1:
Input geometric deformation sequence of video images 1, totalframes is 400, and original template size is 80*48(unit: pixel).For the first two field picture, manually determine the target area of image, 8 dimensional vectors [0.03 on projective transformation group; 0.01; 0.01; 0.01; 0.02; 0.02; 10; 10] be the projective transformation parameter of lock-on boundary shape, t is present frame, for the first two field picture, and t=1.
Step 2:
Predict according to following formula
, j=1,2 ... .300.300 be sampling population; V is that state moves to the velocity vector of moment t by moment t-1.
Step 3:
The image-region corresponding to each sampling, calculates the feature basis matrix of corresponding apparent gray matrix
, then according to formula
calculate the weights of each sampling particle:
Wherein
for feature space
feature basis matrix, vector
represent the target that t frame traces into.
with
be two points that Grassmann flows on shape, leading role's degree is therebetween
.
Step 4:
Pass through formula
accumulate average in calculating sampling, be the estimated state of target
.
Step 5:
According to the update strategy in target signature space, regeneration characteristics space
.Update strategy is as follows:
The minimum value of the proper vector of present frame and each proper vector geodesic line distance of feature space is ds, when ds is greater than given max-thresholds
time, illustrate that this two field picture is seriously blocked or distortion, now for ensureing the accuracy not regeneration characteristics space of feature-space information.Otherwise upgrade template set in two kinds of situation:
(1), when in current feature space, the number of vector is less than 10, directly the proper vector of present frame is joined in this feature space set.
(2) otherwise, replace the proper vector in feature space with maximum range value by the proper vector of present frame.
Step 6:
T=t+1, if
repeat step 2. otherwise tracing process to terminate.
Accompanying drawing 2 is the tracking results of described algorithm partial frame.
Embodiment 2:
Step 1:
Input illumination variation sequence of video images 2, totalframes is 500, and original template size is 78*62(unit: pixel).For the first two field picture, manually determine the target area of image, 8 dimensional vectors [0.01 on projective transformation group; 0.001; 0.001; 0.03; 0.01; 0.01; 10; 10] be the projective transformation parameter of lock-on boundary shape.T is present frame, for the first two field picture, and t=1.
Step 2:
Predict according to following formula
, j=1,2 ... .300.300 be sampling population; V is that state moves to the velocity vector of moment t by moment t-1.
Step 3:
The image-region corresponding to each sampling, calculates the feature basis matrix of corresponding apparent gray matrix
, then according to formula
calculate the weights of each sampling particle:
Wherein
for feature space
feature basis matrix, vector
represent the target that t frame traces into.
with
be two points that Grassmann flows on shape, leading role's degree is therebetween
.
Step 4:
Pass through formula
accumulate average in calculating sampling, be the estimated state of target
.
Step 5:
According to the update strategy in target signature space, regeneration characteristics space
.Update strategy is as follows:
The minimum value of the proper vector of present frame and each proper vector geodesic line distance of feature space is ds, when ds is greater than given max-thresholds
time, illustrate that this two field picture is seriously blocked or distortion, now for ensureing the accuracy not regeneration characteristics space of feature-space information.Otherwise upgrade template set in two kinds of situation:
(1), when in current feature space, the number of vector is less than 10, directly the proper vector of present frame is joined in this feature space set.
(2) otherwise, replace the proper vector in feature space with maximum range value by the proper vector of present frame.
Step 6:
T=t+1, if
repeat step 2. otherwise tracing process to terminate.
Accompanying drawing 3 is the tracking results of described algorithm partial frame.
Embodiment 3:
Step 1:
Sequence of video images 3 is blocked in importation, and totalframes is 100, and original template size is 112*52(unit: pixel).For the first two field picture, manually determine the target area of image, 8 dimensional vectors [0.05 on projective transformation group; 0.002; 0.001; 0.02; 0.02; 0.02; 10; 10] be the projective transformation parameter of lock-on boundary shape.T is present frame, for the first two field picture, and t=1.
Step 2:
Predict according to following formula
, j=1,2 ... .300.300 be sampling population; V is that state moves to the velocity vector of moment t by moment t-1.
Step 3:
The image-region corresponding to each sampling, calculates the feature basis matrix of corresponding apparent gray matrix
, then according to formula
calculate the weights of each sampling particle:
Wherein
for feature space
feature basis matrix, vector
represent the target that t frame traces into.
with
be two points that Grassmann flows on shape, leading role's degree is therebetween
.
Step 4:
Pass through formula
accumulate average in calculating sampling, be the estimated state of target
.
Step 5:
According to the update strategy in target signature space, regeneration characteristics space
.Update strategy is as follows:
The minimum value of the proper vector of present frame and each proper vector geodesic line distance of feature space is ds, when ds is greater than given max-thresholds
time, illustrate that this two field picture is seriously blocked or distortion, now for ensureing the accuracy not regeneration characteristics space of feature-space information.Otherwise upgrade template set in two kinds of situation:
(1), when in current feature space, the number of vector is less than 10, directly the proper vector of present frame is joined in this feature space set.
(2) otherwise, replace the proper vector in feature space with maximum range value by the proper vector of present frame.
Step 6:
T=t+1, if
repeat step 2. otherwise tracing process to terminate.
Accompanying drawing 4 is the tracking results of described algorithm partial frame.