CN105513092A - Template characteristic selection method for target tracking - Google Patents

Abstract

The invention, which belongs to the technical field of visual tracking, discloses a template characteristic selection method for target tracking. With the method, a problem of template tracking drifting caused by a non target matching point in the template during the existing long-range target tracking process can be solved. The method comprises: an optical flow vector of an original matching feature point pair in a template is calculated; polar coordinate processing is carried out on the calculated optical flow vector; probability density statistics is carried out on the processed optical flow vector in polar coordinate space to generate an angular distribution histogram; and median filtering is carried out on the angular distribution histogram to obtain a target matching feature point, so that a non-target feature point in the target template is effectively removed.

Description

A kind of template characteristic system of selection for target following

Technical field

The present invention relates to a kind of Visual Tracking, particularly relate to the optical flow field analysis of template characteristic point in target following, sports ground analysis and probability density distribution analysis, belong to image procossing and area of pattern recognition.

Background technology

It is the important ingredient of image procossing and area of pattern recognition that template is followed the tracks of, and is widely used in intelligent monitoring, in the vision tracking module of the various system such as man-machine interaction and unmanned platform navigation.Its basic process can be described as: when without any priori, with rectangular window selected certain object (i.e. template) in the first two field picture, then follows the tracks of this object in subsequent video sequence.

At present, researcher has carried out large quantifier elimination to template track algorithm, and the LK optical flow method wherein based on template characteristic coupling is one of most important technological means.Its techniqueflow is: first carry out feature extraction to To Template, obtains the unique point in template; Then utilize LK optical flow method to follow the tracks of each unique point in next frame image, obtain the matching characteristic point of each unique point, form matching characteristic point pair thus; Sports ground finally by matching characteristic point between two two field pictures (matching characteristic point to) tries to achieve the geometric transform relation between two two field pictures, and final position and the size obtaining target in this two field picture, complete a tracking cycle.Constantly repeat above-mentioned tracking cycle can follow the tracks of by the realize target in whole video sequence.The hypothesis of this technological means is: although each unique point is tracked separately coupling by LK optical flow method in template, as a part for tracked target, in template, all unique points follow Movement consistency.

But the shortcoming of the Moulding board method of existing feature based coupling is: along with constantly carrying out of tracking cycle, template is easily drifted about, and finally causes track rejection.Its main cause is, although most of region tracked target covered in template, but background pixel is inevitably introduced into, and then the intersection point of pixel in background image and background and objective contour is also often extracted as unique point, and is followed the tracks of by LK optical flow method.Therefore, there is non-targeted feature point pairs in the feature point pairs mated between two two field pictures obtained.But these non-targeted unique points do not belong to a part for target, do not meet Movement consistency hypothesis, therefore interference is caused to solving of geometric transformation, the estimation of template size in next frame and position is caused a deviation.Along with the continuous repetition of tracking cycle, the non-targeted unique point in template is constantly introduced into, and template deviation constantly accumulates, and finally causes the drift of template.

Summary of the invention

For solving the template drift problem caused because non-targeted unique point is introduced in object tracking process, the present invention proposes a kind of template characteristic system of selection for target following, by carrying out the probability density analysis based on directional spreding to the optical flow field of the matching characteristic point between two two field pictures to formation, reject the non-targeted unique point in template, thus the template drift in containment object tracking process.

The concrete steps of the method are:

The light stream vector of original match feature point pairs in step one, calculation template

If there is N number of original match feature point pairs in template, follow the tracks of original match unique point, for the unique point in the (n+1)th two field picture according to LK optical flow method $P^{n+1}=\left\{p_i^{n+1}\right|p_i^{n+1}=\left(\begin{array}{cc}{x}_i^{n+1}& y_i^{n+1}\end{array}\right),i=1,2,\mathrm{...},N\}$ With characteristic of correspondence point in the n-th two field picture the light stream vector d of each original match feature point pairs in (1) calculation template according to the following equation _i(n|n+1), 1≤i≤N:

$d_i\left(n\right|n+1)=\left(\begin{array}{cc}{x}_i^{n+1}-x_i^n,& y_i^{n+1}-y_i^n,\end{array}\right),1\≤i\≤N---\left(1\right)$

Wherein, n is image frame number; with be respectively at the n-th two field picture I _nwith the (n+1)th two field picture I _n+1i-th unique point of middle correspondence, with be respectively unique point and unique point coordinate in image coordinate system;

The light stream vector that in step 2, employing polar coordinate representation template, original match point is right

According to following formula (2)

Obtain the polar coordinate representation (ρ of the light stream vector of each original match feature point pairs _i(n|n+1), θ _i(n|n+1)); Wherein ρ _i(n|n+1) be light stream vector d _i(n|n+1) length, θ _i(n|n+1) be light stream vector d _i(n|n+1) deflection;

Step 3, carry out probability density statistics in polar coordinate space interior focusing flow vector, obtain the probability density of light stream vector in each interval, and generate angular distribution histogram thus; The histogrammic unit width of described angular distribution is Δ θ, is highly F _b(j), wherein: Δ θ is after the polar coordinate space of light stream vector is divided into J interval, the width in each interval; F _b(j) light stream vector in the interval probability density of jth, j=1,2 ..., J;

Step 4, medium filtering is carried out to angle distribution histogram, meet the unique point of the light stream vector in the interval of medium filtering condition for followed the tracks of target signature point, retained; The unique point not meeting the light stream vector in the interval of value filtering condition is non-targeted unique point, deletes from template.

Beneficial effect:

Suppose according to target signature point Movement consistency, by analyzing the optical flow field travel direction distribution density of original match feature point pairs in template, to reject in template non-targeted point in original match feature point pairs (i.e. background characteristics point and background and target boundary unique point), thus effectively suppress the problem of template drift in object tracking process, greatly improve the accuracy of target tracking algorism.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the described template characteristic system of selection for target following;

Fig. 2 is original match feature-point optical flow polar plot;

Fig. 3 is the polar plot of feature-point optical flow vector;

Fig. 4 is the angular distribution histogram of feature-point optical flow vector;

Fig. 5 is target matching characteristics point and light stream vector figure thereof.

Embodiment

Below in conjunction with drawings and Examples, the present invention is described in further detail.

The present embodiment provides a kind of template characteristic system of selection for target following, for solving the problem of the template drift caused because non-targeted unique point is introduced in object tracking process.

The overall flow of the method is as shown in Figure 1:

The light stream vector of original match feature point pairs in step one, calculation template, obtains optical flow field D (n|n+1);

If the number of original match feature point pairs is N in template, according to the original match unique point that LK optical flow method traces into, to the (n+1)th two field picture I _n+1interior unique point $P^{n+1}=\left\{p_i^{n+1}\right|p_i^{n+1}=\left(\begin{array}{cc}{x}_i^{n+1}& y_i^{n+1}\end{array}\right),i=1,2,\mathrm{...},N\}$ With the n-th two field picture I _nmiddle characteristic of correspondence point the light stream vector d of each original match feature point pairs in (1) calculation template according to the following equation _i(n|n+1), 1≤i≤N:

$\begin{array}{cc}{d}_i\left(n\right|n+1)=\left(\begin{array}{cc}{x}_i^{n+1}-x_i^n,& y_i^{n+1}-y_i^n,\end{array}\right),& 1\≤i\≤N\end{array}---\left(1\right)$

Wherein, n is image frame number; with be respectively at the n-th two field picture I _nwith the (n+1)th two field picture I _n+1i-th unique point of middle correspondence, with be respectively unique point and unique point coordinate in image coordinate system; N is total logarithm of original match unique point in template.

Finally obtain optical flow field D (n|n+1)={ d of original match point in template _i(n|n+1) | 1≤i≤N}, as shown in Figure 2.

Step 2, by optical flow field D (n|n+1) Polar coordinates of original match point in template

According to the light stream vector d of original match feature point pairs each in template _i(n|n+1) following equation is solved:

Obtain the polar coordinate representation (ρ of the light stream vector of each original match feature point pairs _i(n|n+1), θ _i(n|n+1)), as shown in Figure 3.Wherein ρ _i(n|n+1) be light stream vector d _i(n|n+1) length, θ _i(n|n+1) be light stream vector d _i(n|n+1) deflection.

Step 3, carry out probability density statistics and generate angular distribution histogram in the polar coordinate space of light stream vector to it, concrete steps are:

(301) angular range of polar coordinates 0 ~ 360 ° is divided into J interval, wherein each interval width is Δ θ;

(302) statistics falls into the number S of the light stream vector in a jth interval _j, j=1,2 ..., J; Then light stream vector is obtained at a jth interval probability density F by following formula (2) _b(j):

$\begin{array}{cc}{F}_b\left(j\right)=\frac{S_j}{N},& j=1,2,\mathrm{...},J\end{array}$

(303) according to the jth interval probability density F that step (302) calculates _bj (), sets up angular distribution histogram in the distribution of the angular range interior focusing flow vector of 0 ~ 360 °, as shown in Figure 4; Wherein histogram constituent parts width is Δ θ, is highly F _b(j).

Step 4, carry out medium filtering to angle distribution histogram, obtain target matching characteristics point, concrete steps are:

(401) duplicate removal process is carried out to angle distribution histogram, obtain single angular histogram amplitude unique (F _b(j));

(402) single angular histogram amplitude is sorted by size, find its intermediate value median (unique (F _b(j)));

(403) with intermediate value median (unique (F _b(j))) be thresholding, if a jth interval histogram amplitude unique (F _b(j)) meet:

unique(F _b(j))≥median(unique(F _b(j)))，j＝1,2,…,J

Then think that the unique point of the light stream vector be in a jth interval is for followed the tracks of target signature point, is retained; On the contrary, not think that the unique point of the light stream vector met in the interval of this condition is non-targeted unique point, deleted; The unique point finally retained is the target signature point in template, as shown in Figure 5.

In sum, these are only preferred embodiment of the present invention, be not intended to limit protection scope of the present invention.Within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (3)

1. for a template characteristic system of selection for target following, it is characterized in that, concrete steps are:

The light stream vector of original match feature point pairs in step one, calculation template

If there is N number of original match feature point pairs in template, follow the tracks of original match unique point, for the unique point in the (n+1)th two field picture according to LK optical flow method $P^{n+1}=\left\{p_i^{n+1}\right|p_i^{n+1}=\left(\begin{array}{cc}{x}_i^{n+1}& y_i^{n+1}\end{array}\right),i=1,2,\mathrm{...},N\}$ With characteristic of correspondence point in the n-th two field picture $P^n=\left\{p_i^n\right|p_i^n=\left(\begin{array}{cc}{x}_i^n& y_i^n\end{array}\right),i=1,2,\mathrm{...},N\},$ The light stream vector d of each original match feature point pairs in (1) calculation template according to the following equation _i(n|n+1), 1≤i≤N:

$d_i\left(n\right|n+1)=\left(\begin{array}{cc}{x}_i^{n+1}-x_i^n,& y_i^{n+1}-y_i^n\end{array}\right),1\≤i\≤N---\left(1\right)$

Wherein, n is image frame number; with be respectively at the n-th two field picture I _nwith the (n+1)th two field picture I _n+1i-th unique point of middle correspondence, $\left(\begin{array}{cc}{x}_i^n& y_i^n\end{array}\right)$ With $\left(\begin{array}{cc}{x}_i^{n+1}& y_i^{n+1}\end{array}\right)$ Be respectively unique point and unique point coordinate in image coordinate system;

The light stream vector that in step 2, employing polar coordinate representation template, original match point is right

According to following formula (2)

Obtain the polar coordinate representation (ρ of the light stream vector of each original match feature point pairs _i(n|n+1), θ _i(n|n+1)); Wherein ρ _i(n|n+1) be light stream vector d _i(n|n+1) length, θ _i(n|n+1) be light stream vector d _i(n|n+1) deflection;

Step 3, carry out probability density statistics in polar coordinate space interior focusing flow vector, obtain the probability density of light stream vector in each interval, and generate angular distribution histogram thus; The histogrammic unit width of described angular distribution is Δ θ, is highly F _b(j), wherein: Δ θ is after the polar coordinate space of light stream vector is divided into J interval, the width in each interval; F _b(j) light stream vector in the interval probability density of jth, j=1,2 ..., J;

Step 4, medium filtering is carried out to angle distribution histogram, meet the unique point of the light stream vector in the interval of medium filtering condition for followed the tracks of target signature point, retained; The unique point not meeting the light stream vector in the interval of value filtering condition is non-targeted unique point, deletes from template.

2. as claimed in claim 1 for the template characteristic system of selection of target following, it is characterized in that, described step 3 generates the histogrammic concrete steps of angular distribution and is:

(301) angular range of polar coordinates 0 ~ 360 ° is divided into J interval, wherein each interval width is Δ θ;

(302) statistics falls into the number S of the light stream vector in a jth interval _j, j=1,2 ..., J, obtains light stream vector at a jth interval probability density F _b(j):

$F_b\left(j\right)=\frac{S_j}{N},j=1,2,\mathrm{...},J$

(303) according to the probability density F of light stream vector in a jth interval that step (302) calculates _bj (), sets up angular distribution histogram in the distribution of the angular range interior focusing flow vector of 0 ~ 360 °.

3. as claimed in claim 1 or 2 for the template characteristic system of selection of target following, it is characterized in that, in described step 4 to the concrete steps that angle distribution histogram carries out medium filtering be:

(401) duplicate removal process is carried out to angle distribution histogram, obtain single angular histogram amplitude unique (F _b(j));

(402) single angular histogram amplitude is sorted by size, obtain its intermediate value median (unique (F _b(j)));

(403) with intermediate value median (unique (F _b(j))) be thresholding, if a jth interval histogram amplitude unique (F _b(j)) meet:

unique(F _b(j))≥median(unique(F _b(j)))，j＝1，2，…，J

Then think that it meets medium filtering condition, otherwise, think that it does not meet medium filtering condition.