CN104200487A - Target tracking method based on ORB characteristics point matching - Google Patents

Abstract

The invention relates to a target tracking method based on ORB characteristics point matching. The target tracking method comprises: adopting an ORB (Oriented FAST and Rotated BRIEF) characteristic point set to represent a target object, and in a tracking process of the target object, iteratively updating the ORB characteristic point set by continuously adding new characteristic points and cutting outliers. A nearest neighborhood search algorithm is used between two continuous frames to find out matched characteristic point pairs, and then a motion transformation model of the tracking target is worked out according to the characteristic point pairs by using method that combines random sampling accordance with a multiple transformation model. The target tracking method based on ORB characteristics point matching is better in robustness and higher in efficiency, when being used for processing scenes with complex appearance change, disorder background, illumination variation and shielded target object.

Description

A kind of method for tracking target based on ORB Feature Points Matching

Technical field

The present invention relates to computer vision tracking technique field, be specifically related to the method for tracking target based on ORB Feature Points Matching.

Background technology

Target following is the important subject of computer vision field, the convenience of the develop rapidly based on high-performance computer, inexpensive video camera and growing automated video are analyzed demand, and the mankind have produced very big interest to probing into of visual target tracking algorithm.In recent years, vision is followed the tracks of in a lot of fields application very widely, comprises motion gesture identification, pedestrian's tracking, intelligent monitoring, man-machine interaction, intelligent transportation detection and automobile navigation etc.Be accompanied by the perfect of the development of digital home's industry and digital home's industrial chain, video tracking progressively gos deep in the application of Smart Home monitoring, on the basis of the multi-service application integration support platform of Facing Digital family, for the user of digital home, provide the characteristic digital home interaction services such as household and community security protection.

Although Chinese scholars has been studied, find plurality of target tracking, face several major issues.The first, the appearance change of target object.In actual scene, the outward appearance of target is subject to the impact of factors such as illumination variation, the change of posture form, camera lens shake, target sizes change and constantly changes.The second, complicated background interference.Under the scenes such as traffic detection, automobile navigation, pedestrian's tracking, target following is because being subject to complicated background interference to have huge challenge.In actual complex environment, have much noise, and background is mixed and disorderly, the color of target object is also easily obscured with the color of background mutually.The 3rd, occlusion issue.In actual video scene, it is the most common common situation that target object is blocked even completely by partial occlusion, investigates a kind of robustness of method for tracking target, just must test the processing power of the method to occlusion issue.The 4th, real time problems.Before three kinds of problems be all in the robustness of investigating target tracking algorism, and target following can to meet the demand of real-time processing be also the difficulties of target tracking domain.

In the research and practice process to the method, the present inventor finds: because unique point descriptor all has good robustness to processing the deformation of noise, detection error and geometrical optics, the present invention proposes a kind of method for tracking target based on ORB Feature Points Matching, thereby realizes efficient target tracking.

Summary of the invention

The invention provides a kind of method for tracking target based on ORB Feature Points Matching, the robustness of processing the scenes such as complicated mode of appearance variation, background clutter, illumination variation and target object block is good, and efficiency is high.

The invention provides a kind of method for tracking target based on ORB Feature Points Matching, comprise the following steps:

(1) first the original image frame of input is extracted to ORB (oFAST and rBRIEF) feature point set;

(2) in descriptor space, match and produce the feature set of present frame with the ORB feature point set of previous frame target object constantly;

(3) according to the position at previous frame target place use random sampling one to make peace method that multiple transform model combines, calculate the motion converter of tracking target;

(4) finally by constantly adding new unique point and cutting outlier to upgrade, produce the target signature collection that final present frame is constantly suitable.

Described ORB unique point descriptor represents by a string of binary characters, and each of character string, by choosing at random the big or small generation mutually that partners of image block that two sizes are S * S near unique point, is got the string of binary characters of n=256 bit length here.More particularly, the image block P that we are S * S a size defines with string of binary characters, and definition is as shown in formula (1):

$\mathrm{\&tau;}(p;x,y)=\left\{\begin{array}{ccc}1,& \mathrm{if}& p\left(x\right)<p\left(y\right),\\ 0,& & \mathrm{otherwise}.\end{array}\right.---\left(1\right)$

Wherein, p (x) and p (y) are respectively that P is at the gray-scale value of pixel x and y.By choosing one group of specific <x, y> point is to conduct set, just the character string that has produced like this n position as BRIEF descriptor, defines as formula (2):

$B_n\left(p\right)=\underset{1\&le;i\&le;n}{\mathrm{\&Sigma;}}{2}^{i-1}\mathrm{\&tau;}(p;x_i,y_i).---\left(2\right)$

The ORB unique point descriptor D that described given two length are n position _iand D _j, similarity ρ (D between the two _i, D _j) by the very efficient hamming distance of computing machine, represented.So, its similarity measurement is defined as follows:

$\mathrm{\&rho;}(D_i,D_j)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}[D_i\left(k\right)-D_j\left(k\right)],---\left(3\right)$

Wherein, D _iand D (k) _j(k) be respectively ORB unique point descriptor D _iand D _jk position, and the worthwhile x of function δ [x] equals 1 at 0 o'clock, otherwise equals 0.

As new ORB feature point set S _tafter present frame t is extracted, we attempt finding out its subset belong to S _tand the feature point set of its each element and the target object of previous frame t-1 matching degree is high.

Suppose we are according to the feature point set of the target object of previous frame t-1 and in whole descriptor space, use proximity search algorithm to find out Optimum Matching be defined as follows:

whether the motion of target object is suitable, and we are with relatively having stipulated a kind of distance metric and a kind of conditional probability with former frame target object position.The formula of evaluating a model M is defined as follows:

Wherein, N _c(M) represent the number of consistent unique point, Pr (M) represents the probability of candidate family, and Cp (M) represents the complexity of model M.The concrete condition of these three factors of influence will be described below.

A given specific model M and from unique point set the minimum number meeting the demands of choosing at random individual matching characteristic point, we can calculate the unknown parameter of transformation model M with comparalive ease.Then unique point set all data be all used to test, those can be applicable to well transformation model M formation a consistent point set C (M), and the size of point set C (M) is N _c(M).Here the complexity Cp (M) that we arrange model M is

In order to assess the possibility Pr (M) of given candidate family M, first we defined a distance metric according to the position of previous frame target object, defines as formula (6):

$\mathrm{dist}\left(M\right|\mathrm{\&gamma;})={\&Integral;}_0^1\left|\right|\mathrm{\&gamma;}\left(u\right)-M\left(\mathrm{\&gamma;}\left(u\right)\right)\left|\right|\mathrm{du},---\left(6\right)$

Wherein, γ (u): [0,1] → R ²for the standardization border of previous frame t-1 target object, and M (γ (u)) is the target object border under model M.If the border of tracked target object is discrete, polygon can be used in its border so represent, distance metric is:

$\mathrm{dist}\left(M\right|Q_{t-1})=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}\left|\right|q_{t-1}^k-M\left(q_{t-1}^k\right)\left|\right|.---\left(7\right)$

Next, the probability P r (M) that we can define certain special exercise transformation model M is as formula (8):

$\mathrm{Pr}\left(M\right)=e^{-\mathrm{\&lambda;dist}\left(M\right|Q_{t-1})},---\left(8\right)$

Wherein, λ means the constant parameter of attenuation rate.

Described ORB feature point set is by dynamically adding the renewal that iterates of the method for new unique point and cutting outlier.We use a kind of simple method for the energy value of each unique point maintenance in set.After Feature Points Matching and transformation model are estimated, the t frame teacher's of i unique point in the feature point set of target object energy value upgraded, defined as formula (9):

${V_i}^t=\left\{\begin{array}{cc}{V_i}^{t-1}+\mathrm{\&alpha;},& \mathrm{matched},\mathrm{consensus}\\ {V_i}^{t-1}-\mathrm{\&beta;},& \mathrm{matched},\mathrm{outlier}\\ {V_i}^{t-1},& \mathrm{not\; matched}\end{array}\right.---\left(9\right)$

Formula (9) if represent a unique point with the transformation model of estimation can with the target object feature point set of previous frame t-1 in a bit match, illustrate that so this unique point is consistent, and its energy value is assigned this energy value of previous frame t-1, add an increment α.In like manner contrary situation, if can not match by estimative transformation model, illustrates that this unique point is that peel off, abnormal so, and its energy value will be endowed a decrement β.And the point not mating, its energy value remains unchanged.

According to their energy value, can sort to unique point.So, only have M the unique point that energy value is the highest to be retained, and other unique point because being considered to for tracking target object, its low energy value is unsettled, is dropped.When new unique point is added into the feature point set of target object, their zero energy value is set to all intermediate values of set.Follow the tracks of while starting most, all unique points energy value when the first frame is identical.

Technique scheme can be found out, because adopting ORB feature point set, the embodiment of the present invention represents the characteristics of image of target object, in target object tracing process, adopt and constantly add the method for new unique point and cutting outlier to come iteration to upgrade ORB feature point set.And between two continuous frames, adopt proximity search algorithm to find out the unique point pair of coupling, next according to unique point, adopt random sampling one to make peace method that multiple transform model combines is calculated the motion converter model of tracking target., therefore can process the target following under the scenes such as complicated mode of appearance variation, background clutter, illumination variation and target object block.

Accompanying drawing explanation

In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, to the accompanying drawing of required use in embodiment or description of the Prior Art be briefly described below, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skills, do not paying under the prerequisite of creative work, can also obtain according to these accompanying drawings other accompanying drawing.

Fig. 1 is the process flow diagram of the method for tracking target based on ORB Feature Points Matching in the embodiment of the present invention;

Fig. 2 is changing and blocking the tracking effect figure under situation at target object mode of appearance in the embodiment of the present invention;

Fig. 3 is the ORB characteristic point analysis figure to target occlusion video sequence in the embodiment of the present invention.

Embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is clearly and completely described, obviously, described embodiment is only the present invention's part embodiment, rather than whole embodiment.Embodiment based in the present invention, those of ordinary skills, not making all other embodiment that obtain under creative work prerequisite, belong to the scope of protection of the invention.

The inventive method flow process as shown in Figure 1, I _tthe original image frame that represents the t moment of input, first this method extracts ORB feature point set S _t, then in descriptor space with the ORB feature point set of previous frame t-1 target object constantly match and produce the feature set of present frame then according to the position at previous frame target place use random sampling one to make peace method that multiple transform model combines, calculate the motion converter model T of tracking target _t, finally by constantly adding new unique point and cutting outlier to upgrade, produce the target signature collection that final present frame t is constantly suitable

After ORB Feature Points Matching, the unique point pair of the some based on mating between two continuous frames, we can estimate and the most suitable transformation model of moving target object.Random sampling unification algorism (RANdom SAmple Consensus) is a kind of iterative algorithm of being estimated the parameter of certain mathematical model by the one group of observation data that contains outlier.First we extract and mate SIFT unique point or the SURF unique point of target, then utilizes the unique point of coupling to this method being used for estimating the motion model of target object between two continuous frames.Random sampling unification algorism produces a right model of the abundant unique point of energy matching by the mode of iteration.

Yet RANSAC algorithm requires the unique point pair of prior model of supposition and minimal amount, this point is easy to be breached in target following.Therefore, we use RAMOSAC algorithm to carry out model estimation, and RAMOSAC has combined two kinds of methods of multiple motion model of RAMSAC and different complexities.Multiple transform model is comprised of four kinds of different transformation models, and its detailed formation below 3.2.1 joint is introduced.

(1) multiple transform model representation

In order to describe better the movement locus of target object, we select to represent with table four kinds of listed transformation models of 3-2, and they are respectively the translation model M that contains two unknown parameters _t, the similarity transformation model M that contains four unknown parameters (rotation parameter, zooming parameter and translation parameters) _s, the affine Transform Model M that contains six unknown parameters _aand the projection transformation model M that contains eight unknown parameters _p(3 * 3 matrixes).And four right minimal amount difference N of unique point that model parameters are required more than calculating _min=1,2,3 and 4.

Table 1 conversion motion model

(2) estimation of transformation model

In order to evaluate one according to model M ∈ { M _t, M _s, M _a, M _pestimating to such an extent that whether candidate transformation T is suitable for the motion of describing target object, we are with relatively having stipulated a kind of distance metric and a kind of conditional probability with former frame target object position.The formula of evaluating a model M is defined as follows:

Wherein, N _c(M) represent the number of consistent unique point, Pr (M) represents the probability of candidate family, and Cp (M) represents the complexity of model M.The concrete condition of these three factors of influence will be described below.

A given specific model M and from unique point set the minimum number meeting the demands of choosing at random individual matching characteristic point, we can calculate the unknown parameter of transformation model M with comparalive ease.Then unique point set all data be all used to test, those can be applicable to well transformation model M formation a consistent point set C (M), and the size of point set C (M) is N _c(M).Here the complexity Cp (M) that we arrange model M is

In order to assess the possibility Pr (M) of given candidate family M, first we defined a distance metric according to the position of previous frame target object, defines as formula:

$\mathrm{dist}\left(M\right|\mathrm{\&gamma;})={\&Integral;}_0^1\left|\right|\mathrm{\&gamma;}\left(u\right)-M\left(\mathrm{\&gamma;}\left(u\right)\right)\left|\right|\mathrm{du},$

Wherein, γ (u): [0,1] → R ²for the standardization border of previous frame t-1 target object, and M (γ (u)) is the target object border under model M.If the border of tracked target object is discrete, polygon can be used in its border so represent, distance metric is:

$\mathrm{dist}\left(M\right|Q_{t-1})=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}\left|\right|q_{t-1}^k-M\left(q_{t-1}^k\right)\left|\right|.$

Next, the probability P r (M) that we can define certain special exercise transformation model M is as formula:

$\mathrm{Pr}\left(M\right)=e^{-\mathrm{\&lambda;dist}\left(M\right|Q_{t-1})},$

Wherein, λ means the constant parameter of attenuation rate.

Similar to RANSAC algorithm, RAMOSAC algorithm execution step is as follows:

1) from table 1, in listed transformation model, choose at random a model M;

2) then from ORB feature point set choosing at random a size is oRB unique point subset S _r;

3) according to subset S _rcalculate the parameter of model M, produce consistent feature point set C (M) and calculate Score (M);

4) iteration step 1) to 3) fixing number of times.According to Score (M) maximal value, find out optimal model;

5) use least square method to recalculate model parameter according to the most consistent point set of applicable model.

The inventive method changes and blocks tracking effect under situation as shown in Figure 2 at target object mode of appearance.Under the scene that target appearance metamorphosis, partial occlusion block even completely and similar object is disturbed, the inventive method robustness is good, can accurately tracking target object, obtain good tracking results.The inventive method is to shown in appended Fig. 3 of ORB characteristic point analysis of target occlusion video sequence, analyzed quantitatively the superiority of the inventive method, by add up frame by frame the ORB unique point of target object coupling and wherein the number of intra-office point prove the robustness of this method.

It should be noted that, the contents such as the information interaction between said apparatus and intrasystem each unit, implementation, due to the inventive method embodiment based on same design, particular content can, referring to the narration in the inventive method embodiment, repeat no more herein.

One of ordinary skill in the art will appreciate that all or part of step in the whole bag of tricks of above-described embodiment is to come the hardware that instruction is relevant to complete by program, this program can be stored in a computer-readable recording medium, storage medium can comprise: ROM (read-only memory) (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), disk or CD etc.

A kind of method for tracking target based on ORB the Feature Points Matching above embodiment of the present invention being provided is described in detail, applied specific case herein principle of the present invention and embodiment are set forth, the explanation of above embodiment is just for helping to understand method of the present invention and core concept thereof; , for one of ordinary skill in the art, according to thought of the present invention, all will change in specific embodiments and applications, in sum, this description should not be construed as limitation of the present invention meanwhile.

Claims (5)

1. the method for tracking target based on ORB Feature Points Matching, is characterized in that, the method includes the steps of:

First original image frame to input extracts ORB feature point set;

In descriptor space, match and produce the feature set of present frame with the ORB feature point set of previous frame target object constantly;

According to the position at previous frame target place use random sampling one to make peace method that multiple transform model combines, calculate the motion converter of tracking target;

Finally by constantly adding new unique point and cutting outlier to upgrade, produce the target signature collection that final present frame is constantly suitable.

2. the method for tracking target based on ORB Feature Points Matching according to claim 1, it is characterized in that: described ORB unique point descriptor is represented by a string of binary characters, each of character string is by choosing at random the big or small generation mutually that partners of image block that two sizes are S * S near unique point, here get the string of binary characters of n=256 bit length, the image block P that is S * S by a size defines with string of binary characters, and definition is as shown in formula (1):

Wherein, p (x) and p (y) are respectively that P is at the gray-scale value of pixel x and y; By choosing one group of specific <x, y> point is to conduct set, just the character string that has produced like this n position as BRIEF descriptor, defines as formula (2):

。

3. the method for tracking target based on ORB Feature Points Matching according to claim 1, is characterized in that: the ORB unique point descriptor D that described given two length are n position _iand D _j, similarity ρ (D between the two _i, D _j) by computing machine, hamming is apart from representing very efficiently, its similarity measurement is defined as follows:

Wherein, D _iand D (k) _j(k) be respectively ORB unique point descriptor D _iand D _jk position, and the worthwhile x of function δ [x] equals 1 at 0 o'clock, otherwise equals 0;

As new ORB feature point set S _tafter present frame t is extracted, we attempt finding out its subset belong to S _tand the feature point set of its each element and the target object of previous frame t-1 matching degree is high;

Set according to the feature point set of the target object of previous frame t-1 and in whole descriptor space, use proximity search algorithm to find out Optimum Matching be defined as follows:

。

4. the method for tracking target based on ORB Feature Points Matching according to claim 1, is characterized in that: one of described evaluation is according to model M ∈ { M _t, M _s, M _a, M _pestimate to such an extent that whether candidate transformation T suitable for the motion of describing target object, with former frame target object position relatively stipulated a kind of distance metric and a kind of conditional probability, the formula of evaluating a model M is defined as follows:

Wherein, N _c(M) represent the number of consistent unique point, Pr (M) represents the probability of candidate family, and Cp (M) represents the complexity of model M.The concrete condition of these three factors of influence will be described below; A given specific model M and from unique point set the minimum number meeting the demands of choosing at random individual matching characteristic point, then unique point set all data be all used to test, those can be applicable to well transformation model M formation a consistent point set C (M), and the size of point set C (M) is N _c(M); Here the complexity Cp (M) that I arrange model M is according to the position of previous frame target object, defined a distance metric, defined as formula (6):

Wherein, γ (u): [0,1] → R ²for the standardization border of previous frame t-1 target object, and M (γ (u)) is the target object border under model M; If the border of tracked target object is discrete, polygon can be used in its border so represent, distance metric is:

The probability P r (M) that defines certain special exercise transformation model M is as formula (8):

Wherein, λ means the constant parameter of attenuation rate.

5. the method for tracking target based on ORB Feature Points Matching according to claim 1, it is characterized in that: described ORB feature point set is by dynamically adding the renewal that iterates of the method for new unique point and cutting outlier, after Feature Points Matching and transformation model are estimated, the t frame teacher's of i unique point in the feature point set of target object energy value upgraded, defined as formula (9):

Formula (9) if represent a unique point with the transformation model of estimation can with the target object feature point set of previous frame t-1 in a bit match, illustrate that so this unique point is consistent, and its energy value is assigned this energy value of previous frame t-1, add an increment α; In like manner contrary situation, if can not match by estimative transformation model, illustrates that this unique point is that peel off, abnormal so, and its energy value will be endowed a decrement β, and the point not mating, its energy value remains unchanged.