CN109584277A - A kind of nuclear phase pass filter tracking method based on binary search - Google Patents
A kind of nuclear phase pass filter tracking method based on binary search Download PDFInfo
- Publication number
- CN109584277A CN109584277A CN201811491175.1A CN201811491175A CN109584277A CN 109584277 A CN109584277 A CN 109584277A CN 201811491175 A CN201811491175 A CN 201811491175A CN 109584277 A CN109584277 A CN 109584277A
- Authority
- CN
- China
- Prior art keywords
- target
- candidate
- scale
- frame
- binary search
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
- G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
Abstract
The invention discloses a kind of, and the nuclear phase based on binary search closes filter tracking method, comprising: S1 tracks target by candidate family first, realizes search for the first time to provide candidate detection position；S2, trace model extract benchmark candidate samples in the position of previous frame tracking result and candidate detection position respectively；S3, thus cyclic shift construction sample carries out the differentiation of classifier again, realizes the binary search of target；S4 finally goes out to track the offset of target, and then determines position of the target in new frame image with the maximum value calculation of all position response figures.The present invention constructs the differentiation that sample carries out classifier again by cyclic shift, realizes and carries out binary search to tracing detection target, carries out real-time tracking to detection target to realize.
Description
Technical field
The present invention relates to unmanned aerial vehicle (UAV) control, embedded development, field of image processings, more particularly to one kind to be based on binary search
Nuclear phase close filter tracking method.
Background technique
Nuclear phase pass filtered target track algorithm (Kernelized Correlation Filters, abbreviation KCF) utilizes and follows
The circulation intensive sampling that ring matrix theory carries out, significantly improves processing speed, huge speed advantage makes KCF algorithm very
It is suitable for real-time modeling method.However KCF algorithm can not be coped on unmanned aerial vehicle platform well because of camera perspective variation, tracking
Target tracks drifting problem caused by quickly moving etc., this greatly reduces the performance of tracking.
Summary of the invention
It is lost for tracking caused by exceeding KCF algorithm detection range when target during unmanned plane tracks quickly moves
Problem is lost, the invention proposes a kind of, and the nuclear phase based on binary search closes filter tracking method.
To achieve the above object, the technical scheme adopted by the invention is that:
A kind of nuclear phase pass filter tracking method based on binary search, comprising:
S1 tracks target by candidate family first, realizes search for the first time to provide candidate detection position；
S2, trace model extract benchmark candidate's sample in the position of previous frame tracking result and candidate detection position respectively
This；
S3, thus cyclic shift construction sample carries out the differentiation of classifier again, realizes the binary search of target；
S4 finally goes out to track the offset of target, and then determines that target exists with the maximum value calculation of all position response figures
Position in new frame image.
Preferably, in step S1, the size M of candidate family training samplec,NcMeet respectively:
Mc=(1+Pc)M (1)
Nc=(1+Pc)N (2)
In formula, PcIt is the constant for extending candidate family size, M, N are the sizes of target to be tracked；
Cosine Window is added to sample characteristics when extracting image pattern, are as follows:
In formula,It is the training sample X of corresponding candidate familycCore auto-correlation,It is the training of corresponding candidate family
Nuclear phase between sample and candidate samples to be detected closes,It is solution of the matrix A in frequency domain, matrixIndicate back heading mark, λ
It is regularization parameter,Refer to the response diagram square of sample image block z；
In a new frame image, image pattern is extracted as candidate family using the position of previous frame image trace result
The couple candidate detection point that the position of obtained response diagram maximum value is provided as candidate family is realized in the input of correlation filter
Positioning is searched for the first time of target.
Preferably, in step S3, trace model more new formula are as follows:
In formula, i and i-1 respectively indicate the index of the i-th frame and the (i-1)-th frame, and θ is the learning rate of model,It is frequency domain tracking
The parameter of model, WithRespectively correspond the coefficient matrix and training sample of frequency domain classifier, T table
Diagram as sequence frame number,WithIt is the model parameter in the (i-1)-th and i moment respectively,It is to be instructed with the sample of the i-th frame
The model parameter practised.
Preferably, in step S4, size estimation is added to trace model, it is assumed that scale when trace model initializes is ST
=(Sx,Sy), Sx、SyThe scale of the x-axis and y-axis when trace model initialization is respectively referred to, the target scale that former frame traces into is
St, the scale pond that definition contains k scale factor is S={ t1,t2,…,tk, so that different detections can be constructed in present frame
Size { tiSt|ti∈ S }, need to obtain the image block of these different scales of the position acquisition of candidate samples in present frame, and pass through
Image block is zoomed to fixed dimension S by bilinear interpolation algorithmT, and response diagram is calculated separately, final scale are as follows:
In formula,It is the sample image block after scaling, scale becomes fixed size ST, corresponding original ruler
Degree is tiSt,It is scale tiStResponse diagram.
Compared with prior art, the beneficial effects of the present invention are:
The present invention starts with from target background modeling, proposes a kind of nuclear phase pass filter tracking method based on binary search,
First time search is first carried out by additional core correlation filter, provides new test point to expand the model of original algorithm detection
It encloses, then reuses core correlation filter in different test points and re-search for target, compensated in the way of searching for twice
Recycle the deficiency of intensive sampling；Meanwhile size estimation is added to improved algorithm to further increase accuracy rate in the present invention.
Certainly, it implements any of the products of the present invention and does not necessarily require achieving all the advantages described above at the same time.
Detailed description of the invention
Fig. 1 is that a kind of nuclear phase based on binary search of the invention closes filter tracking method flow diagram.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, below in conjunction with attached drawing to each reality of the invention
The mode of applying is explained in detail.
Nuclear phase based on binary search closes shown in filter tracking algorithm idea such as step (1)~(17).
Input: T frame image sequence, the initial position P of target1=(x1,y1)；
Output: the position P of t frame targett=(xt,yt)；
(1) the image I of the 1st frame is obtained1In (x1,x2) at sample, training classifier coefficient；
(2) trace model M is initializedtWith candidate family Mc, wherein the size of trace model is ST；
(3) For t=2 ..., T
(4) the image I of t frame is obtainedtIn Pt-1The detection sample at place
(5) response diagram for the candidate family being calculatedAnd obtain candidate test point P '；
(6) For i=1 ..., k
(7) the image I of t frame is obtainedtIt is t in the place P ' scaleiStDetection sampleZoom to ST；
(8) scale t is calculatediStResponse diagram
(9)End for
(10) For i=1 ..., k
(11) the image I of t frame is obtainedtIn Pt-1Place's scale is tiStDetection sampleZoom to ST；
(12) scale t is calculatediStResponse diagram
(13)End for
(14) by all response diagramsWithObtain the position P of current goaltAnd scale；
(15) t frame image I is obtainedtIn PtThe training sample at place calculates classifier coefficient；
(16) trace model M is updatedtWith candidate family Mc；
(17)End for
As shown in Figure 1, a kind of nuclear phase based on binary search closes filter tracking method, including Step1~Step4, specifically
Are as follows:
Step1 tracks target by candidate family first, realizes search for the first time to provide candidate detection position.
Candidate family is realized to provide reliable candidate samples when trace model detection using core correlation filter.In order to
To background modeling, candidate family is more much greater than the size of traditional tracker model, so as to including the biggish space of target
Context graph picture carries out regression training, so that classifier is still able to include correctly target when quickly movement occurs for target.Its
The size M of training samplec,NcIt needs to meet:
Mc=(1+Pc)M (1)
Nc=(1+Pc)N (2)
In formula, PcIt is the constant for extending candidate family size, M, N are the sizes of target to be tracked.Because target
Displacement may be larger, it is therefore desirable to when expanding classifier training in regression matrix virtual value range, be according to trace model
Mark label variances sigma is returned to expand variance yields.
In order to weaken boundary effect, Cosine Window is added to sample characteristics when extracting image pattern.
In formula,It is the training sample X of corresponding candidate familycCore auto-correlation,It is the instruction of corresponding candidate family
The nuclear phase practiced between sample and candidate samples to be detected closes,It is solution of the matrix A in frequency domain, matrixIndicate back heading mark,
λ is regularization parameter,Refer to the response diagram square of sample image block z.
In a new frame image, image pattern is extracted as the correlation of candidate family using the position of previous frame tracking result
The input of filter, to the couple candidate detection point that the position that obtained response diagram is maximized is provided as candidate family, realization pair
The first time of target searches for positioning.Candidate family has divided more positive samples in training, although the accuracy of classification may not
Height, but its model dimension is larger, is able to achieve the coarse positioning to target.
Step2, trace model extract benchmark candidate in the position of previous frame tracking result and candidate detection position respectively
Sample.
Step3, thus cyclic shift construction sample carries out the differentiation of classifier again, realizes the binary search of target.
During tracking, whether trace model or candidate family be adapted to because block, illumination and background it is miscellaneous
The variation of the brings target appearance and context such as random, therefore real-time update is required, but different each other, target appearance
Change not consistent with the step of change in context.
Trace model more new formula is as follows:
In formula, i and i-1 respectively indicate the index of the i-th frame and the (i-1)-th frame, and θ is the learning rate of model,It is frequency domain tracking
The parameter of model, WithRespectively correspond the coefficient matrix and training sample of frequency domain classifier, T table
Diagram as sequence frame number,WithIt is (i-1)-th and i moment total model parameter respectively,It is to be instructed with the sample of the i-th frame
The model parameter practised.
Step4 finally goes out to track the offset of target, and then determines target with the maximum value calculation of all position response figures
Position in new frame image.
Because candidate family mainly provides test point for trace model, moulded dimension is more much larger than the size of target,
Size estimation only is added to trace model to reduce complexity.It might as well assume that scale when trace model initialization is ST=(Sx,
Sy), Sx、SyThe scale of the x-axis and y-axis when trace model initialization is respectively referred to, the target scale that former frame traces into is St, it is fixed
The scale pond that justice contains k scale factor is S={ t1,t2,…,tk, so that different detecting sizes can be constructed in present frame
{tiSt|ti∈ S }, need to obtain the image block of these different scales of the position acquisition of candidate samples in present frame, and pass through two-wire
Image block is all zoomed to fixed dimension S by property interpolation algorithmT, and response diagram is calculated separately, the calculating of final scale is as follows:
Wherein,It is the sample image block after scaling, scale becomes fixed size ST, corresponding original ruler
Degree is tiSt,It is scale tiStResponse diagram.Because the result response diagram that classifier finally differentiates is a matrix, corresponding
Each element is the classification results of each cyclic shift sample, it is only necessary to which the maximum value for finding all scale response diagrams can determine
Corresponding scale.Since the corresponding displacement of tracking target is lain in response matrix, is finally adjusted and be displaced according to corresponding t
Obtain final deviant.
The problem of tracking target scale variation cannot be coped with well due to KCF algorithm, the present invention is to the improved calculation of calcium
Method enters size estimation to further increase accuracy rate.The invention also includes: camera data, benefit are read by Jetson TX2
KCF algorithm is write with QT Creator IDE exploitation environment；KCF algorithm is accelerated using CUDA programming, is sent out by serial ports
It send to winged control module.
Present invention could apply to unmanned plane real-time tracking system or the application scenarios of other demands.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with scope of protection of the claims
Subject to.
Claims (4)
1. a kind of nuclear phase based on binary search closes filter tracking method characterized by comprising
S1 tracks target by candidate family first, realizes search for the first time to provide candidate detection position；
S2, trace model extract benchmark candidate samples in the position of previous frame tracking result and candidate detection position respectively；
S3, thus cyclic shift construction sample carries out the differentiation of classifier again, realizes the binary search of target；
S4 finally goes out to track the offset of target, and then determines target new one with the maximum value calculation of all position response figures
Position in frame image.
2. a kind of nuclear phase based on binary search according to claim 1 closes filter tracking method, which is characterized in that step
In S1, the size M of candidate family training samplec,NcMeet respectively:
Mc=(1+Pc)M (1)
Nc=(1+Pc)N (2)
In formula, PcIt is the constant for extending candidate family size, M, N are the sizes of target to be tracked；
Cosine Window is added to sample characteristics when extracting image pattern, are as follows:
In formula,It is the training sample X of corresponding candidate familycCore auto-correlation,It is the training sample of corresponding candidate family
Nuclear phase between candidate samples to be detected closes,It is solution of the matrix A in frequency domain, matrixIndicate back that heading mark, λ are just
Then change parameter,Refer to the response diagram square of sample image block z；
In a new frame image, image pattern is extracted as the correlation of candidate family using the position of previous frame image trace result
The couple candidate detection point that the position of obtained response diagram maximum value is provided as candidate family is realized to mesh in the input of filter
Target search positioning for the first time.
3. a kind of nuclear phase based on binary search according to claim 2 closes filter tracking method, which is characterized in that step
In S3, trace model more new formula are as follows:
In formula, i and i-1 respectively indicate the index of the i-th frame and the (i-1)-th frame, and θ is the learning rate of model,It is frequency domain trace model
Parameter, WithThe coefficient matrix and training sample of frequency domain classifier are respectively corresponded, T indicates figure
Picture sequence frame number,WithIt is the model parameter in the (i-1)-th and i moment respectively,It is to be gone out with the sample training of the i-th frame
Model parameter.
4. a kind of nuclear phase based on binary search according to claim 3 closes filter tracking method, which is characterized in that step
In S4, size estimation is added to trace model, it is assumed that scale when trace model initializes is ST=(Sx,Sy), Sx、SyIt respectively refers to
The scale of x-axis and y-axis when trace model initializes, the target scale that former frame traces into are St, define and contain k scale
The scale pond of the factor is S={ t1,t2,…,tk, so that different detecting size { t can be constructed in present frameiSt|ti∈ S }, working as
Previous frame needs to obtain the image block of these different scales of the position acquisition of candidate samples, and passes through bilinear interpolation algorithm for image
Block zooms to fixed dimension ST, and response diagram is calculated separately, final scale are as follows:
In formula,It is the sample image block after scaling, scale becomes fixed size ST, corresponding to original scale is
tiSt,It is scale tiStResponse diagram.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811491175.1A CN109584277A (en) | 2018-12-07 | 2018-12-07 | A kind of nuclear phase pass filter tracking method based on binary search |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811491175.1A CN109584277A (en) | 2018-12-07 | 2018-12-07 | A kind of nuclear phase pass filter tracking method based on binary search |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109584277A true CN109584277A (en) | 2019-04-05 |
Family
ID=65927682
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811491175.1A Pending CN109584277A (en) | 2018-12-07 | 2018-12-07 | A kind of nuclear phase pass filter tracking method based on binary search |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109584277A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112396065A (en) * | 2020-10-19 | 2021-02-23 | 北京理工大学 | Scale-adaptive target tracking method and system based on correlation filtering |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102904838A (en) * | 2011-07-25 | 2013-01-30 | 联芯科技有限公司 | Method and device for judging interfering boundary |
| KR20150048502A (en) * | 2013-10-28 | 2015-05-07 | 삼성전자주식회사 | Method and Apparatus for quadrature mirror filtering |
| CN108010067A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of visual target tracking method based on combination determination strategy |
| CN108061909A (en) * | 2017-11-23 | 2018-05-22 | 深圳开阳电子股份有限公司 | A kind of catching method, device and the satellite navigation receiver of GNSS satellite signal |
| CN108090919A (en) * | 2018-01-02 | 2018-05-29 | 华南理工大学 | Improved kernel correlation filtering tracking method based on super-pixel optical flow and adaptive learning factor |
-
2018
- 2018-12-07 CN CN201811491175.1A patent/CN109584277A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102904838A (en) * | 2011-07-25 | 2013-01-30 | 联芯科技有限公司 | Method and device for judging interfering boundary |
| KR20150048502A (en) * | 2013-10-28 | 2015-05-07 | 삼성전자주식회사 | Method and Apparatus for quadrature mirror filtering |
| CN108061909A (en) * | 2017-11-23 | 2018-05-22 | 深圳开阳电子股份有限公司 | A kind of catching method, device and the satellite navigation receiver of GNSS satellite signal |
| CN108010067A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of visual target tracking method based on combination determination strategy |
| CN108090919A (en) * | 2018-01-02 | 2018-05-29 | 华南理工大学 | Improved kernel correlation filtering tracking method based on super-pixel optical flow and adaptive learning factor |
Non-Patent Citations (2)
| Title |
|---|
| 吴仁坚: ""基于无人机的车辆实时跟踪算法研究"", 《中国优秀博硕士学位论文全文数据库 信息科技辑》 * |
| 贾永红等: "《人工智能视频分析 智能终端的崛起》", 杭州：浙江工商大学出版社 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112396065A (en) * | 2020-10-19 | 2021-02-23 | 北京理工大学 | Scale-adaptive target tracking method and system based on correlation filtering |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bhat et al. | Learning discriminative model prediction for tracking | |
| Feichtenhofer et al. | Detect to track and track to detect | |
| CN104820997B (en) | A kind of method for tracking target based on piecemeal sparse expression Yu HSV Feature Fusion | |
| CN110009060A (en) | A kind of robustness long-term follow method based on correlation filtering and target detection | |
| CN111161412A (en) | Three-dimensional laser mapping method and system | |
| CN101572770B (en) | Method for testing motion available for real-time monitoring and device thereof | |
| Fan et al. | Long-term visual tracking algorithm for UAVs based on kernel correlation filtering and SURF features | |
| CN107798691A (en) | A kind of unmanned plane independent landing terrestrial reference real-time detecting and tracking method of view-based access control model | |
| CN109448023A (en) | A kind of satellite video Small object method for real time tracking of combination space confidence map and track estimation | |
| CN110555423B (en) | Multi-dimensional motion camera-based traffic parameter extraction method for aerial video | |
| CN109711245A (en) | A kind of closed loop detection method based on image candidate region | |
| JP7339314B2 (en) | Deep Smartphone Sensor Fusion for Indoor Positioning and Tracking | |
| CN109584277A (en) | A kind of nuclear phase pass filter tracking method based on binary search | |
| CN113158904A (en) | Twin network target tracking method and device based on double-mask template updating | |
| Husseini | A survey of optical flow techniques for object tracking | |
| CN115909219A (en) | Scene change detection method and system based on video analysis | |
| Lu et al. | Vision-based localization methods under GPS-denied conditions | |
| CN110781803B (en) | Human body posture recognition method based on extended Kalman filter | |
| Wang et al. | Robust object tracking via online principal component–canonical correlation analysis (P3CA) | |
| Lu et al. | A new meanshift target tracking algorithm by combining feature points from gray and depth images | |
| Chen et al. | Exploiting Acceleration of the Target for Visual Object Tracking | |
| Rinosha | Correlation filter based visual tracking with circular shift on local and semi-local domains | |
| Gao et al. | High-order deep infomax-guided deformable transformer network for efficient lane detection | |
| Wang et al. | Robust visual tracking via guided low-rank subspace learning | |
| Ma et al. | Robust SLAM algorithm in dynamic environment using optical flow |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |