CN105354863A - Adaptive scale image sequence target tracking method based on feature filtering and fast motion detection template prediction - Google Patents

Abstract

The present invention provides an adaptive scale image sequence target tracking method based on feature filtering and fast motion detection template prediction. The method comprises the following steps of: 1, specifying a to-be-tracked object in an initial frame, transforming the object into a frequency domain by fast Fourier transformation based on a Gaussian kernel, and constructing a feature filter; 2, constructing a fast motion prediction template; 3, constructing an object imaging scale change model; 4, in a second frame, acquiring a predicted position of the object by the fast motion prediction template, acquiring a position of a corresponding matched search box according to the object imaging scale change model by using the predicted position as the center, performing filteration by the feature filter, regarding a position of a maximum response value as an object location, and simultaneously, generating a new feature filter and using the new feature filter as a to-be-used feature filter in a next frame; 5, sequentially repeating the step 4 in subsequent frames until a video or image sequence is finished. The method provided by the present invention is effectively applicable to tracking of a distant and small object in a complicated background and has good tracking performance.

Description

The adaptive scale image sequence method for tracking target of feature based filtering and Fast Moving Detection template prediction

Technical field

The present invention relates to video frequency object tracking field, especially a kind of image sequence method for tracking target.

Background technology

Camera remains target following in process that is remote and fast approaching target in high-speed motion state, is the one in various target following application scenarios.Under remote sight, target object imaging size on sensor device is very little, is often less than 15 × 15 pixels, even only has 5 × 5 pixels.And under such yardstick, the feature of imaging region is very not obvious, be unfavorable for utilizing the features such as object color, texture, structure to follow the tracks of.Under natural scene, the background of destination object is very complicated and random, and prospect may be blocked in addition.Existing track algorithm, under such application scenarios, all can not complete tracing task very well.And this algorithm can be sane tracking target.

In order to improve the robustness to performance of target tracking, adopting the track algorithm of faster handling property can tackle target and moving fast and the change of camera attitude.This algorithm can remain above the processing speed of 300 frames/second online, thus adds the robustness of strong algorithms.

The defect of existing camera image sequence method for tracking target: the distance small target be identified under complex background cannot be detected, very easily with losing target.Move the feature with camera rapid movement for target, original method tracking performance is poorer.

Summary of the invention

In order to overcome existing method for tracking target under complex background, be difficult to detect the poor deficiency of feature, tracking performance to distance small target, the invention provides a kind of be effectively applicable to distance small target under complex background tracking, the good feature based filtering of tracking performance and Fast Moving Detection template prediction image sequence method for tracking target.

The technical solution adopted for the present invention to solve the technical problems is:

An image sequence method for tracking target for the adaptive scale change of feature based filtering and Fast Moving Detection template prediction, comprises the steps:

1) from camera or image sequence, obtain image, in initial frame, specify the object needing to follow the tracks of, object is transformed into frequency domain by the fast fourier transform based on gaussian kernel, construction feature wave filter;

2) rapid movement prediction module is built, for the motion prediction of moving target fast;

3) object images dimensional variation model is built, for prediction in object tracking process, to the prediction of target scale change;

4) in the second frame subsequently, obtain object by rapid movement prediction module and estimate position, estimate centered by position by this, then obtain the position of corresponding match search frame according to object images dimensional variation model;

In match search frame, use the feature filters built to carry out filtering, using maximum response position as target localization, complete the target following of the second frame;

Meanwhile, according to the object of the second frame target following, feature filters is regenerated, as the feature filters that next frame will use;

5) in frame subsequently, step 4 is repeated successively), until video or image sequence terminate.

Technical conceive of the present invention is: based on the object detection method of core wave filter to subjective scales sensitive, can not change in imaging plane mesoscale by adaption object; And due to object and camera motion, make object often have larger jumping etc. in imaging plane, easily cause detected object be in detection block edge or exceed detection block; And expand detection window the calculated amount of fast fourier transform will be caused significantly to increase, reduce handling property and speed.And traditional Kalman filter carries out position prediction, need to consume larger computing power, algorithm complex is 2 × O ³, and adopting Fast Moving Detection template, algorithm complex is C constant amount.Therefore, adopt and can significantly improve the performance of the target tracking algorism based on core wave filter based on the self-adapting detecting window change of object images dimensional variation model and Fast Moving Detection template, and the tracking velocity that maintenance is very fast.

Beneficial effect of the present invention is mainly manifested in: the tracking, the tracking performance that are effectively applicable to distance small target under complex background are better; Effective extraction target area feature, rapid movement template prediction, accurate self-adapting window merges, and improve the robustness of tracking performance, the method may be used for but is not limited to run at high speed automobile to traffic mark recognition and tracking, high-speed aircraft to target following etc.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the image sequence method for tracking target of feature based filtering and Fast Moving Detection template prediction.

Fig. 2 is the schematic diagram of object images dimensional variation model.

Fig. 3 is the schematic diagram of first order motion model.

Fig. 4 is the schematic diagram of Motion mask.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1 ~ Fig. 4, the adaptive scale image sequence method for tracking target of a kind of feature based filtering and Fast Moving Detection template prediction, comprises the steps:

1) from camera or image sequence, obtain image, manually specify or use additive method to obtain the appointment needing tracing object at initial frame.Object is transformed into frequency domain by the fast fourier transform based on gaussian kernel, construction feature wave filter.

2) rapid movement prediction module is built, for the motion prediction of moving target fast.

3) object images dimensional variation model is built, for prediction in object tracking process, to the prediction of target scale change.

4) in the second frame subsequently, obtain object by rapid movement prediction module and estimate position, estimate centered by position by this, then obtain the position of corresponding match search frame according to object images dimensional variation model.

In match search frame, using step 1) feature filters that built carries out filtering, using maximum analog value position as target localization, completes the target following of the second frame.

Meanwhile, according to step 1) identical method, the second frame target following object is regenerated feature filters, as the wave filter that next frame will use.

5) in frame subsequently, step 4 is repeated successively), until video or image sequence terminate.

In the present embodiment, with the object images dimensional variation model of target imaging on the imageing sensor of method representation shown in Fig. 2, in figure, L is the length of target, d ₀for reference position range-to-go (object distance), f is focal length.In the t=0 moment, u ₀for imaging size, O ₀for the photocentre of reference position.In the t=1 moment, u ₁for imaging size, O ₁for photocentre.V is camera motion speed, and a is acceleration of motion, and t is for measuring the moment.In this model, hypothetical target, camera photocentre, camera motion track are axially same.

With this understanding, in video camera imaging plane, the size of target is:

$u_t=\frac{L\×f}{d_0-vt-{\mathrm{at}}^2}=\frac{L\×{\mathrm{n\μ}}_0}{d_0-vt-{\mathrm{at}}^2}=\frac{L\×{\mathrm{n\μ}}_0}{d_1}---\left(1\right)$

Wherein μ ₀for pixel physical size, f=n μ ₀.

Calibrate according to above variation model, obtain video camera Distance geometry target relation on the image plane, and then provide the dimensional variation of tracing object for target following.

According to the first order motion model of such as Fig. 3, construct the Motion mask as Fig. 4, according to the target location of front cross frame, the target azimuth of prediction next frame.Although this Template Location is just omited, the cardinal principle orientation of target can be provided.Because be specifically located through wave filter to obtain, but the method that so calculated amount is little, but for follow-up filtering algorithm provides necessary optimization;

The building process of feature filters: the solution of the ridge regression RidgeRegression of kernelised:

$\mathrm{\α}=\frac{y}{K+\mathrm{\λ}I}$

K＝κ(y,z)

Wherein, α is image y and z relationship map, and K is kernel function, and λ is constant coefficient, and I is vector of unit length.Be shown below, we have employed gaussian kernel, and σ is constant variance.

$K=\exp (-\frac{\left|\right|y-z|{|}^2}{{\mathrm{\σ}}^2})=\exp (-\frac{\left|\right|y|{|}^2+|\left|z\right|{|}^2-2yz}{{\mathrm{\σ}}^2})$

After obtaining two known two field picture z (former frame) and y (present frame), calculate a kind of mapping relations α between two width images.When a rear two field picture x inputs, can be associated in the nuclear space of two width images with α and x, y, and then the response of y on x-y nuclear space can be obtained, and then can by finding peak response point and obtain the location of target on x region, the incidence relation of three can as follows expressed by:

response function, α _ithe mapping relations corresponding to each pixel, y _i, x _ibe each pixel respective value of present frame and next frame image respectively, i refers to each pixel.Ask the position of value maximum, namely obtained the position of target.In order to calculate fast, the frequency domain when calculating all after Fourier transform carries out, and sample window adopts Hamming window to convert.

F in formula is the symbol of Fourier transform, and * F is the conjugation of Fourier transform, F ^-1be the inverse transformation of Fourier transform, I is vector of unit length.

Wherein, z, y, x are the image window that former frame and present frame contain target respectively, and a rear frame is image window to be detected.

The scheme of the present embodiment is applied to camera approximate direct application scenarios close to the target following of target in high-speed motion, obtains stable tracking effect.

Claims (1)

1. an adaptive scale image sequence method for tracking target for feature based filtering and Fast Moving Detection template prediction, is characterized in that: comprise the steps:

1) from camera or image sequence, obtain image, in initial frame, specify the object needing to follow the tracks of, object is transformed into frequency domain by the fast fourier transform based on gaussian kernel, construction feature wave filter;

2) rapid movement prediction module is built, for the motion prediction of moving target fast;

3) object images dimensional variation model is built, for prediction in object tracking process, to the prediction of target scale change;

4) in the second frame subsequently, obtain object by rapid movement prediction module and estimate position, estimate centered by position by this, then obtain the position of corresponding match search frame according to object images dimensional variation model;

In match search frame, use the feature filters built to carry out filtering, using maximum response position as target localization, complete the target following of the second frame;

Meanwhile, according to the object of the second frame target following, feature filters is regenerated, as the feature filters that next frame will use;

5) in frame subsequently, step 4 is repeated successively), until video or image sequence terminate.