CN108550161A - A kind of dimension self-adaption core correlation filtering fast-moving target tracking method - Google Patents
A kind of dimension self-adaption core correlation filtering fast-moving target tracking method Download PDFInfo
- Publication number
- CN108550161A CN108550161A CN201810228008.1A CN201810228008A CN108550161A CN 108550161 A CN108550161 A CN 108550161A CN 201810228008 A CN201810228008 A CN 201810228008A CN 108550161 A CN108550161 A CN 108550161A
- Authority
- CN
- China
- Prior art keywords
- target
- image block
- window
- adaption
- size
- 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.)
- Granted
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
- G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
Abstract
The present invention relates to a kind of dimension self-adaption core correlation filtering fast-moving target tracking methods, by introducing two-dimensional variable bandwidth Gaussian window with the separation of background before promoting in KCF frames and adding size estimation module, estimate the dimensional variation of target by image block characteristics point matching result, can be combined with displacement prediction on this basis to judge target whether in blocking.Once target is in blocking, then it is assumed that scale free converts and without template renewal at this time, and then strengthens the accuracy of tracking in this way.It is an advantage of the invention that calculating simply, quickly, operation efficiency can be significantly improved on the basis of compatible dimension self-adaption.
Description
Technical field
The present invention relates to core correlation filtering (Kernel Correlation Filter) target following technologies, more particularly to
A kind of quick nuclear phase pass filtered target tracking of dimension self-adaption, belongs to Video Analysis Technology field.
Background technology
Target following is one of key problem of video analysis, is had in fields such as human-computer interaction, video monitoring, augmented realities
It is widely applied.Although being made great progress in the research for carrying out the technology in the past few decades, since there are light in application
According to variation, rigid deformation, many factors such as quickly movement, partial occlusion, background be complicated, seeks fast and stable, is suitble to object change
The tracking of change is a challenge always.
In recent years, the tracking based on detection (tracking-by-detection) is a kind of typical target following pattern.
This tracing mode integrates on-line study and template renewal, and new location information is obtained from detection.But this pattern
It needs to collect many sample forms in target neighborhood, and it is greatly overlapping that these templates, which have, it is very high to cause
Redundant computation.For this purpose, researcher proposes based on cycle nuclear structure (CSK:Circulant Structure with
Kernels track algorithm) effectively increases operation efficiency, and is further developed on this basis based on core correlation filtering
(KCF:Kernel Correlation Filter) target following technology.KCF technologies are obtained by introducing multiple features channel concept
It obtained target following robustness and accuracy is promoted, becoming a kind of at present has the target following skill for representing meaning and application prospect
Art.
Currently, becoming one of the hot spot of concern to the improvement of KCF target followings, primarily focuses on and further increase its operation
Efficiency and the adaptability that target scale is changed.For example, it is a kind of for real-time tracking adaptive color feature [Danelljan M,
Khan F S,Felsberg M,et al.Adaptive Color Attributes for Real-Time Visual
Tracking[C].IEEE Conference on Computer Vision and Pattern Recognition.IEEE
Computer Society,2014:1090-1097.].Color attribute is divided into 11 classes by this method, is in time selected than more significant
Color, and using a kind of dimension reduction method similar to PCA by feature vector from 11 tie up drop to 2 dimensions, improve accuracy, still
It is not ideal enough to the tracking effect of multiscale target.The integrated dimension self-adaption nuclear phase of another feature based closes filter tracking
[Li Y,Zhu J.A Scale Adaptive Kernel Correlation Filter Tracker with Feature
Integration[C].European Conference on Computer Vision.Springer,Cham,2014:254-
265.], although this method can solve the problems, such as the dimensional variation of target, due to wherein including 7 predefined dimension scales, often
Secondary all to need to carry out bilinear interpolation to image block to be sized, complexity is very high, influences very much efficiency.In conclusion existing
The improved method of KCF, it is difficult to take into account dimension self-adaption and operation efficiency.
Invention content
It is an object of the invention to:In view of the defects existing in the prior art, a kind of dimension self-adaption core correlation filtering is proposed
Fast-moving target tracking method introduces bandwidth varying Gaussian window with the separation of background before promoting in KCF frames and adds scale and estimate
Module is counted, the dimensional variation of target is estimated by image block characteristics point matching result.
In order to reach object above, the present invention provides a kind of dimension self-adaption core correlation filtering fast-moving target tracking sides
Method, by introducing two-dimensional variable bandwidth Gaussian window with the separation of background before promoting in KCF frames and adding size estimation module,
Estimate the dimensional variation of target by image block characteristics point matching result, includes the following steps:
Step 1. reads first frame image, and carries out initial training:
Step 1.1. gives initial target frame (u, v, w, h), determines that target's center is pos1 (u, v), target sizes are
Target_sz (w, h) is intercepted the image block patch1 of W × H sizes with target's center, determines dimensional Gaussian window gauss;
Step 1.2. extracts f-HOG features to patch1 and adds dimensional Gaussian window to it, then carries out Fourier transformation, obtains
Image block characteristics after to processing
Step 1.3. carries out Fourier transformation according to image block characteristics and obtains frequency domain Gaussian kernel autocorrelation matrixRidge is found out to return
Return dual spaces learning parameterIf initial reference template is respectively with parameterized template
Step 2. reads next frame image, into detection module:
Step 2.1. intercepts the image block patch2 of W × H sizes centered on pos1 (u, v), and f-HOG is extracted to patch2
Feature and to feature Hanning window cos_w processing, then carries out Fourier transformation to get image block characteristics after processing
Step 2.2. willIt substitutes into Gaussian kernel cross-correlation and carries out Fourier transformation, obtain frequency-domain kernel
Cross-correlation matrixThe ridge regression response score of all possible displacements is obtained simultaneously, and maximum score position is that target is pre-
Location sets pos2 (u', v');
Step 3. enters adaptive scale estimation module:Hanning window weight w is distributed to each location of pixels of patch1i, right
It adds target_sz (w, h) size two-value mask to obtain interested part, and extracts the angles Shi-Tomasi of this part
Point and generation random point, these points are expressed as { a1,a2,Λ,an, it filters out wherein weights and is less than threshold value T1Point, obtain { a1,
a2,Λ,am};Using matching remaining characteristic point { a before and after Lucas Kanade in patch2 to optical flow method1,a2,Λ,am,
It filters out less than normalized crosscorrelation matching intermediate value and the point pair higher than reversed light stream matching error intermediate value, leaves reliable point to collection
Close { a1,a2,Λ,akAnd { b1,b2,Λ,bk};Scale is obtained by calculating distance proportion of arbitrary two characteristic point in consecutive frame
Distributed collectionAnd w is carried out to this setiWeighted average obtains weighted scale valueFinal scale is being averaged for weighted scale and intermediate value scale, i.e.,
Step 4. combines the offset vector m between pos1 (u, v) and pos2 (u', v'), is reliably put between with consecutive frame
Relative displacement vector obtain absolute displacement M={ li|li=| | bi-ai-scale*m||};If the displacement is less than threshold value T2, then
Normal point pair is considered, otherwise it is assumed that being abnormal point pair;
Step 5. counts the quantity of normal point pair and abnormal point pair, if the number of abnormal point pair is less than of normal point pair
Number, then update tracking box size, i.e. target_sz (w, h)=target_sz (w, h) * scale sets interpolation coefficient at this time
Factor is 0.02;Otherwise, then it is assumed that for target in blocking, selection does not update tracking box size target_sz (w, h), and
If interpolation coefficient is 0;
Pos1 (u, v) is updated to the value of pos2 (u', v') by step 6., and W × H sizes are intercepted centered on pos1 (u, v)
Image block patch1, patch1 is extracted and f-HOG features and adds Gaussian window to it, then carries out Fourier transformation to get place
Image block characteristics after reasonIt willIt substitutes into gaussian kernel function and carries out
Fourier transformation obtains frequency domain Gaussian kernel autocorrelation matrixRidge regression dual spaces learning parameter is found out using the resultAnd
The variances sigma marked according to target_sz sizes update Gauss ';
Step 7. linear interpolation updates reference template model_xf and learning parameter model_ α, i.e.,
Model_xf=(1-factor) * model_xf+factor*xf
Step 8. judges whether present frame is last frame, if so, terminating, is otherwise transferred to step 2.
The present invention the technical solution that further limits be:In the step 1, target's center region of search size window_
Sz (W, H) is 2.5 times of initial target boundary rectangle sizes.
Further, the dimensional Gaussian window function, as shown in (1) formula
Gauss_w=G (m, n, σw,σh)=g (m, σw)*g(h,σh)' (1)
In formula, m, n are the width and height of image block characteristics, and bandwidth σ is calculated separately in horizontal and vertical direction, that is to say, that
By target size (w, h) and ratio of window_sz (W, the H) size between both horizontally and vertically as calculating variance yields
Standard, i.e.,
The Gaussian kernel correlation function as shown in (2) formula,
Wherein, x and x' is respectively treated image block, σ 0.5,For the Element-Level multiplication of matrix,It indicates
Form of the parameter in Fourier;
Fhog (x) function representations extract f-HOG features to image block x, and wherein cell is 4, direction number 9.
The form of solution of the ridge regression in dual spaces as shown in (3) formula,
Wherein,Letter is marked for ridge regression
Number, i.e., using image block characteristics center as the two-dimensional Gaussian function of peak position, varianceλ is regularization coefficient,
Indicate Element-Level division;
The ridge regression responds scoring function as shown in (4) formula,
Wherein kzxFor core cross-correlation matrix, α is learning parameter.
Cos_w=h (m) the * h (n) ' for two-dimentional Hanning window function, wherein m, n be respectively image block characteristics width and
It is high.
Further, the step 1 and step 6 replace original Hanning window with bandwidth varying Gaussian window:
Hanning window is limited to the size of window_sz, once this region, which determines, just remains constant, and Gaussian window can be with
The bandwidth of control distribution is adjusted to promote the separation of foreground and background by variance.If m, n are the width and height of image block characteristics,
So here ratio by target size (w', h') and characteristic size between both horizontally and vertically as calculating variance yields
Standard, i.e.,That is, here by target size (w, h) and window_sz (W, H) sizes horizontal and
Ratio between vertical direction is as the standard for calculating variance yields, i.e.,
A kind of dimension self-adaption core correlation filtering fast-moving target tracking method proposed by the present invention, introduces in KCF frames
Bandwidth varying Gaussian window is with the separation of background before promoting and adds size estimation module, by image block characteristics point matching result come
Estimate the dimensional variation of target.In calculation amount, if there is n pixel in image block, core relevant operation is only needed in Fourier
Product calculation is carried out, computation complexity is O (nlogn), and for size estimation module, run time is concentrated mainly on feature and carries
It takes and is matched with Optical-flow Feature point, computation complexity is O (kn), and k is characterized a number, and is much smaller than n.As it can be seen that in KCF frames
On the basis of frame, the calculating cost that size estimation module has only paid very little is added.
In addition, can be combined with displacement prediction on this basis to judge target whether in blocking.Once at target
In blocking, then it is assumed that scale free converts and without template renewal at this time, and then strengthens the accurate of tracking in this way
Property.
A kind of dimension self-adaption core correlation filtering fast-moving target tracking method proposed by the present invention, on the basis of existing KCF
On, the Gaussian window for introducing two-dimentional BREATHABLE BANDWIDTH replaces original Hanning window, is distributed by adjusting variance reed time controll, big in target
The separation of foreground and background is better achieved when small variation, is matched in combination with sparse key point to estimate position and the ruler of target
Degree, to extend the ability that KCF handles dimensional variation.Simultaneously as this Scale Estimation Method of the present invention belongs to interpolation
Formula operates, and will not change the process flow of former KCF, may be equally suitable for the mutation of primary correlation filtering.
Compared with prior art, the present invention can not only improve target following of the target in complex scene and cosmetic variation
Robustness realizes efficiently and accurately processing target dimensional variation, while calculation processing is simpler, quick, can be in compatible ruler
Operation efficiency is significantly improved on the basis of degree is adaptive.
Description of the drawings
The present invention will be further described below with reference to the drawings.
Fig. 1 is that the quick dimension self-adaption nuclear phase of the present invention closes the flow diagram of filter tracking.
Fig. 2 (a) is the first frame image of the Singer1 example definition graphs of the present invention.
Fig. 2 (b) is the second frame image of the Singer1 example definition graphs of the present invention.
Fig. 3 is the Gauss function schematic diagram of the present invention.
Fig. 4 (a) is the first frame image block of the Shi-Tomasi corners Matching result figures of the present invention.
Fig. 4 (b) is the second frame image block of the Shi-Tomasi corners Matching result figures of the present invention.
Fig. 5 (a) is the accuracy of the mean curve graph of the tracking performance comparison diagram of the present invention.
Fig. 5 (b) is the average success rate curve graph of the tracking performance comparison diagram of the present invention.
Fig. 6 (a) is the part tracking result design sketch carScale video sequence design sketch of the present invention.
Fig. 6 (b) is the part tracking result design sketch car4 video sequence design sketch of the present invention.
Fig. 6 (c) is the part tracking result design sketch singer1 video sequence design sketch of the present invention.
Fig. 6 (d) is the part tracking result design sketch walking2 video sequence design sketch of the present invention.
Fig. 6 (e) is the part tracking result design sketch jogging-1 video sequence design sketch of the present invention.
Fig. 6 (f) is the part tracking result design sketch jogging-2 video sequence design sketch of the present invention.
Specific implementation mode
To keep the purpose of the present invention, implementation and advantage relatively sharp, below in conjunction with the accompanying drawings to the technical side of the present invention
Case is described in detail:
A kind of dimension self-adaption core correlation filtering fast-moving target tracking method provided by the invention, flow as shown in Figure 1,
By figure dotted line frame as it can be seen that the present invention does not make complete destructive transformation to original frame, and replaced with bandwidth varying Gaussian window
The Hanning window of trained part has been changed, and has increased size estimation module and has modified corresponding model modification strategy.
Function and parameter declaration:
Target's center region of search size window_sz (W, H) is 2.5 times of initial target boundary rectangle sizes.
Dimensional Gaussian window function is provided by formula (1)
Gauss_w=G (m, n, σw,σh)=g (m, σw)*g(h,σh)' (1)
Wherein m, n are the width and height of image block characteristics, and bandwidth σ is calculated separately in horizontal and vertical direction, that is to say, that
By target size (w, h) and ratio of window_sz (W, the H) size between both horizontally and vertically as calculating variance yields
Standard, i.e.,
Gaussian kernel correlation function is indicated by formula (2)
Wherein x and x' is respectively treated image block, σ 0.5,For the Element-Level multiplication of matrix,Expression parameter exists
The form of Fourier still continues to use this representation below.
Fhog (x) function representations extract f-HOG features to image block x, and wherein cell is 4, direction number 9.
Formula (3) indicates the form of solution of the ridge regression in dual spaces
WhereinLetter is marked for ridge regression
Number, i.e., using image block characteristics center as the two-dimensional Gaussian function of peak position, varianceλ is regularization coefficient,
Indicate Element-Level division.
Cos_w=h (m) * h (n) ' are two-dimentional Hanning window function, and wherein m, n are respectively the width and height of image block characteristics.
Formula (4) indicates that ridge regression responds scoring function
Wherein kzxFor core cross-correlation matrix, α is learning parameter.
This method specifically includes following steps:
Step 1 reads first frame image, and carries out initial training:
Given initial target frame (u, v, w, h), wherein target's center is pos1 (u, v) and target sizes are target_sz
(w, h) is intercepted the image block patch1 of W × H sizes centered on pos1 (u, v), dimensional Gaussian is determined with the criterion in formula (1)
Window gauss.F-HOG features are extracted to patch1 and add dimensional Gaussian window to it, after then carrying out Fourier transformation to get processing
Image block characteristicsIt willSubstitution formula (2) simultaneously carries out Fourier transformation and obtains
Frequency domain Gaussian kernel autocorrelation matrixResult is substituted into formula (3) and finds out ridge regression dual spaces learning parameterIf initial reference
Template is respectively with parameterized template
Step 2 reads next frame, the image block patch2 of W × H sizes is intercepted centered on pos1 (u, v), to patch2
It extracts f-HOG features and to feature Hanning window cos_w processing, then carries out Fourier transformation to get image block characteristics after processingIt willSubstitution formula (2) simultaneously carries out Fourier transformation, obtains frequency domain
Core cross-correlation matrixSimultaneously willSubstitution formula (4) responds score with the ridge regression for obtaining all possible displacements,
Maximum score position is target predicted position pos2 (u', v').
Step 3 distributes Hanning window weight w to each location of pixels of patch1i, itself plus target_sz (w, h) size are covered
Film extracts the Shi-Tomasi angle points of this part and generates random point to obtain interested part, these points indicate
For { a1,a2,Λ,an, it filters out wherein weights and is less than threshold value T1Point, obtain { a1,a2,Λ,am}.Before Lucas Kanade
Backward optical flow method matches remaining characteristic point { a in patch21,a2,Λ,am, it filters out and matches intermediate value less than normalized crosscorrelation
With the point pair higher than reversed light stream matching error intermediate value, reliable point is left to set { a1,a2,Λ,akAnd { b1,b2,Λ,
bk}.Size distribution set is obtained by calculating distance proportion of arbitrary two characteristic point in consecutive frameAnd w is carried out to this setiWeighted average obtains weighted scale valueFinal scale is being averaged for weighted scale and intermediate value scale, i.e.,
Step 4, in conjunction with the offset vector m between pos1 (u, v) and pos2 (u', v'), reliably put between with consecutive frame
Relative displacement vector obtain absolute displacement M={ li|li=| | bi-ai- scale*m | | }, if this displacement is less than threshold value T2,
Normal point pair is then considered, otherwise it is assumed that being abnormal point pair.
Step 5, statistics normally with the quantity of abnormal point pair, if the number of abnormal point pair be less than normal point pair number, more
New tracking box size, i.e. target_sz (w, h)=target_sz (w, h) * scale, at this moment set interpolation coefficient factor as
0.02.Otherwise, then it is assumed that during target is for blocking, do not update target_sz (w, h), and set interpolation coefficient as 0.
Step 6, the value that pos1 (u, v) is updated to pos2 (u', v') intercept W × H sizes centered on pos1 (u, v)
Image block patch1, dimensional Gaussian window gauss is determined with the criterion in formula (1).F-HOG features are extracted to patch1 and to it
Add Gaussian window, then carries out Fourier transformation to get image block characteristics after processingIt willSubstitution formula (2) simultaneously carries out Fourier transformation and obtains frequency domain Gaussian kernel autocorrelation matrixResult is substituted into formula (3)
Find out ridge regression dual spaces learning parameterAnd according to target_sz sizes update Gauss label variances sigma '.
Step 7, linear interpolation update reference template model_xf and learning parameter model_ α, i.e.,
Model_xf=(1-factor) * model_xf+factor*xf
Step 8 judges whether present frame is last frame, if so, terminating, otherwise goes to step 2.
Based on foregoing invention content, below we illustrated with example Singer1 (Fig. 2):
Given initial target frame (47,99,43,145), wherein initial target center are pos1 (47,99) and target sizes
Target_sz is 43 × 145, and target's center region of search window_sz is 43 × 145.
For training module, the image block patch1 of window_sz sizes is intercepted centered on pos1 from the 1st frame image
(43 × 145) are in ratio both horizontally and vertically according to target_sz (43 × 145) and window_sz (107 × 362)
Variance (σw=0.40, σh=0.40) feature sizes Gaussian window gauss is determined.Simultaneously to patch1 extraction f-HOG features (26 × 90)
Gaussian window weighting is carried out to it, then carries out Fourier transformation to get image block characteristics after processingIt willIt substitutes into and calculates Gaussian kernel auto-correlation function and carry out Fourier's change
It gets in returnAnd then substitute into the solution formula of ridge regression dual spacesFind out learning parameterIf initial reference template
It is respectively with parameterized template
Fig. 3 is bandwidth varying Gauss function schematic diagram.As seen from the figure, Gaussian function
Peak position determined that and bandwidth is by variances sigma by desired value μ2It determines.σ describes the dispersion degree of data distribution, and σ is bigger, number
More disperse according to distribution, σ is smaller, and data distribution is more concentrated.That is σ is bigger, and curve bandwidth is bigger, conversely, bandwidth is smaller.It follows
When ring obtains sample, former algorithm is to add Hanning windowExtraction, but Hanning window cannot
Change and change the size of extraction with target sizes.It when target scale change is small, can make in next training, introduce more
Background information enters positive sample;When target scale becomes larger, positive sample can only include a part for target.Both of these case all can
Cause the inaccuracy of tracking.Gauss function can be according to variances sigma2Variation and in time adjust bandwidth, with present frame mesh
Bandwidth of the ratio of size and tracing area size as variance control dimensional Gaussian window is marked, to be indirectly controlled positive sample
Size.At the same time, the variance of Gauss label also makes corresponding adjustment with the scale of target.
For detection module, intercepted centered on pos1 (47,99) in the 2nd frame image window_sz sizes (43 ×
145) image block patch2 (107 × 362), and to patch2 extraction f-HOG features (26 × 90) and carry out Hanning window weighting
Processing.Then carry out Fourier transformation to getIt willIt substitutes into
It calculates Gaussian kernel cross-correlation and carries out Fourier transformation, obtainAnd then it substitutes intoTo obtain centered on pos1
The regions window_sz in all possible displacements ridge regression respond score, maximum score position is target prediction position
Set pos2 (47,99).
In size estimation module, Hanning window weight w is distributed to each location of pixels of patch1i, to itself plus target_sz
(43 × 145) size two-value mask extracts Shi-Tomasi angle points and the life of this part to obtain interested part
At random point, these points are expressed as { a1,a2,Λ,an, the point that weights are less than threshold value (0.86) is filtered out, { a is obtained1,a2,Λ,
am, as shown in Fig. 4 (a).Using matching remaining characteristic point { a before and after Lucas Kanade in patch2 to optical flow method1,a2,
Λ,am, as shown in Fig. 4 (b), filters out less than normalized crosscorrelation matching intermediate value (0.9994) and missed higher than reversed light stream matching
The point pair of poor intermediate value (0.0065) leaves reliable point to set { a1,a2,Λ,akAnd { b1,b2,Λ,bk}.Appointed by calculating
Distance proportion of two characteristic points of anticipating in consecutive frame obtains size distribution setAnd to this
Set carries out wiWeighted average obtains weighted scale value(0.9986), final ruler
Degree is being averaged for weighted scale and intermediate value scale (0.9981), i.e.,Knot
The offset vector m (0,0) between pos1 (47,99) and pos2 (47,99) is closed, the relative displacement between is reliably put with consecutive frame
Vector obtains absolute displacement M={ li|li=| | bi-ai- scale*m | | }, if this displacement is less than threshold value (5pixel),
Normal point pair is considered, otherwise it is assumed that being abnormal point pair.Statistics normally with the quantity of abnormal point pair, if the number of abnormal point pair
(0) be less than normal point pair number (13), update tracking box size, i.e. target_sz=target_sz*scale,
(42.9292,144.7613), if interpolation coefficient factor is 0.02.Otherwise, then it is assumed that during target is for blocking, do not update
Target_sz, if difference coefficient is 0.Pos1 (47,99) is updated to the value (47,99) of pos2, is recycled.
To verify the effect of the method for the present invention, following confirmatory experiment has been carried out
The hardware and software emulation environment that the present invention tests are as shown in table 1:
CPU | Intel(R)Core(TM)i5-2450CPU@2.5GHz |
Memory | 4G |
Operating system | Window 7 |
Development environment | MATLAB2012a、Visual Studio 2010 |
Programming language | MATLAB、C/C++ |
1 hardware of table and software environment
The present invention it is generally acknowledged using 50 on OTB (Online Tracking Benchmark) platform, regarding of having marked
Frequency sequence is tested, various situations in their simulation of real scenes, including illumination variation, change of scale, part or tight
It blocks again, deformation etc..
On OTB test platforms, there are mainly two types of evaluation criterions:Accuracy (Precision Plot) and success rate
(Success Plot)。
During tracking, the target location of algorithm estimation is referred to as predicted value, and the target location manually marked is referred to as
The case where actual value, difference is less than given threshold value between the two, accounts for the percentage of total frame, referred to as accuracy, in general threshold value
It is set as 20 pixels.Given threshold value is different, and the accuracy generally yielded is also different, and the accuracy of all situations is combined
Get up to be fitted to a curve.
For convenience's sake, algorithm keeps track result is denoted as BT, real goal frame is denoted as BG, area statistics function is denoted as
Area (), then tracking coverage rate is:In the video frame, if coverage rate is more than given threshold value,
Then target is successfully tracked, and success tracking frame accounts for the percentage of total frame, and referred to as success rate, in general threshold value is set as
0.5.Given threshold value is different, and the success rate generally yielded is also different, and the success rate of all situations is combined can
It is depicted as a curve.
In conclusion accuracy and success rate are higher, the performance of target following is better.
And the standard for evaluating tracking efficiency uses processing frame number (frame/second) per second, the frame number of algorithm processing per second is more,
Illustrate that real-time performance of tracking is better.
As shown in figure 5, being shown according to the operation result of 50 video sequences, the method for the present invention is in bat and averagely
Success rate all achieves good effect.Specifically, compared to primal algorithm, the present invention improves in bat
6.08%, average success rate improves 17.50%, while in operational efficiency, the method for the present invention can reach 53.4629 frames/
The average speed of second.And SAMF is consistent with our goal of the invention, but the speed of service only has about 7 frames/second.Although and CN speed
On quickly, but tracking box cannot change with target sizes, to the ineffective of multiscale target tracking.
Fig. 6 is the part tracking effect figure of operation result of the present invention.Red is the tracking box of primal algorithm, and green is this
The tracking box of inventive method.As seen from the figure, tracking box changes with the variation of target sizes, and in jogging, when encountering
When blocking, primal algorithm tracking failure, and the method for the present invention can continue to track.
It should be noted that the foregoing is merely the specific embodiment of the present invention, it is not intended to limit the invention, this
Data set and attack mode used are only limitted to the present embodiment in embodiment, all within the spirits and principles of the present invention, made by
Any modification, equivalent substitution, improvement and etc. should all be included in the protection scope of the present invention.
Claims (7)
1. a kind of dimension self-adaption core correlation filtering fast-moving target tracking method, it is characterised in that:By being introduced in KCF frames
Two-dimensional variable bandwidth Gaussian window passes through image block characteristics point and matches knot to promote the separation of preceding background and add size estimation module
Fruit estimates the dimensional variation of target, includes the following steps:
Step 1. reads first frame image, and carries out initial training:
Step 1.1. gives initial target frame (u, v, w, h), determines that target's center is pos1 (u, v), target sizes target_
Sz (w, h) is intercepted the image block patch1 of W × H sizes with target's center, determines dimensional Gaussian window gauss;
Step 1.2. extracts f-HOG features to patch1 and adds dimensional Gaussian window to it, then carries out Fourier transformation, obtains everywhere
Image block characteristics after reason
Step 1.3. carries out Fourier transformation according to image block characteristics and obtains frequency domain Gaussian kernel autocorrelation matrixFind out ridge regression pair
Even space learning parameterIf initial reference template and parameterized template are respectively model_xf=xf,
Step 2. reads next frame image, into detection module:
Step 2.1. intercepts the image block patch2 of W × H sizes centered on pos1 (u, v), and f-HOG features are extracted to patch2
And to feature Hanning window cos_w processing, Fourier transformation is then carried out to get image block characteristics after processing
Step 2.2. willIt substitutes into Gaussian kernel cross-correlation and carries out Fourier transformation, obtain frequency-domain kernel cross-correlation
MatrixThe ridge regression response score of all possible displacements is obtained simultaneously, and maximum score position is target predicted position
pos2(u',v');
Step 3. enters adaptive scale estimation module:Hanning window weight w is distributed to each location of pixels of patch1i, to itself plus
Target_sz (w, h) size two-value mask to obtain interested part, and extract this part Shi-Tomasi angle points with
And random point is generated, these points are expressed as { a1,a2,Λ,an, it filters out wherein weights and is less than threshold value T1Point, obtain { a1,a2,
Λ,am};Using matching remaining characteristic point { a before and after Lucas Kanade in patch2 to optical flow method1,a2,Λ,am, filter
Except intermediate value and higher than the point pair of reversed light stream matching error intermediate value is matched less than normalized crosscorrelation, reliable point is left to set
{a1,a2,Λ,akAnd { b1,b2,Λ,bk};Scale point is obtained by calculating distance proportion of arbitrary two characteristic point in consecutive frame
Cloth setAnd w is carried out to this setiWeighted average obtains weighted scale valueFinal scale is being averaged for weighted scale and intermediate value scale, i.e.,
Step 4. combines the offset vector m between pos1 (u, v) and pos2 (u', v'), and the phase between is reliably put with consecutive frame
Contraposition, which is shifted to, measures absolute displacement M={ li|li=| | bi-ai-scale*m||};If the displacement is less than threshold value T2, then it is assumed that
It is normal point pair, otherwise it is assumed that being abnormal point pair;
Step 5. counts the quantity of normal point pair and abnormal point pair, if the number of abnormal point pair is less than the number of normal point pair,
Update tracking box size, i.e. target_sz (w, h)=target_sz (w, h) * scale, set at this time interpolation coefficient factor as
0.02;Otherwise, then it is assumed that for target in blocking, selection does not update tracking box size target_sz (w, h), and sets interpolation system
Number is 0;
Pos1 (u, v) is updated to the value of pos2 (u', v') by step 6., and the figure of W × H sizes is intercepted centered on pos1 (u, v)
As block patch1, f-HOG features are extracted to patch1 and add Gaussian window to it, after then carrying out Fourier transformation to get processing
Image block characteristicsIt willIt substitutes into gaussian kernel function and carries out Fourier
Transformation, obtains frequency domain Gaussian kernel autocorrelation matrixRidge regression dual spaces learning parameter is found out using the resultAnd according to
The variances sigma of target_sz sizes update Gauss label ';
Step 7. linear interpolation updates reference template model_xf and learning parameter model_ α, i.e.,
Model_xf=(1-factor) * model_xf+factor*xf
Step 8. judges whether present frame is last frame, if so, terminating, is otherwise transferred to step 2.
2. dimension self-adaption core correlation filtering fast-moving target tracking method according to claim 1, it is characterised in that:It is described
In step 1, target's center region of search size window_sz (W, H) is 2.5 times of initial target boundary rectangle sizes.
3. dimension self-adaption core correlation filtering fast-moving target tracking method according to claim 1, it is characterised in that:It is described
Dimensional Gaussian window function, as shown in (1) formula
Gauss_w=G (m, n, σw,σh)=g (m, σw)*g(h,σh)' (1)
In formula, m, n are the width and height of image block characteristics, and bandwidth σ is calculated separately in horizontal and vertical direction, that is to say, that by mesh
Dimensioning (w, h) is with ratio of window_sz (W, the H) size between both horizontally and vertically as the mark for calculating variance yields
Standard, i.e.,
4. dimension self-adaption core correlation filtering fast-moving target tracking method according to claim 1, it is characterised in that:It is described
Gaussian kernel correlation function as shown in (2) formula,
Wherein, x and x' is respectively treated image block, σ 0.5,For the Element-Level multiplication of matrix,Indicate that expression parameter exists
The form of Fourier;
Fhog (x) function representations extract f-HOG features to image block x, and wherein cell is 4, direction number 9.
5. dimension self-adaption core correlation filtering fast-moving target tracking method according to claim 1, it is characterised in that:It is described
The form of solution of the ridge regression in dual spaces as shown in (3) formula,
Wherein,For ridge regression labeling function, i.e.,
Using image block characteristics center as the two-dimensional Gaussian function of peak position, varianceλ is regularization coefficient,It indicates
Element-Level division;
The ridge regression responds scoring function as shown in (4) formula,
Wherein kzxFor core cross-correlation matrix, α is learning parameter.
6. dimension self-adaption core correlation filtering fast-moving target tracking method according to claim 1, it is characterised in that:It is described
Cos_w=h (m) * h (n) ' are two-dimentional Hanning window function, and wherein m, n are respectively the width and height of image block characteristics.
7. dimension self-adaption core correlation filtering fast-moving target tracking method according to claim 1, it is characterised in that:It is described
Step 1 and step 6 replace original Hanning window with bandwidth varying Gaussian window:
Hanning window is limited to the size of window_sz, once this region, which determines, just remains constant, and Gaussian window can pass through
Variance come adjust control distribution bandwidth to promote the separation of foreground and background;If m, n are the width and height of image block characteristics, by mesh
The ratio of dimensioning (w', h') and characteristic size between both horizontally and vertically is as the standard for calculating variance yields, i.e.,
That is, the ratio of target size (w, h) and window_sz (W, H) size between both horizontally and vertically is made
To calculate the standard of variance yields, i.e.,
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810228008.1A CN108550161B (en) | 2018-03-20 | 2018-03-20 | Scale self-adaptive kernel-dependent filtering rapid target tracking method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810228008.1A CN108550161B (en) | 2018-03-20 | 2018-03-20 | Scale self-adaptive kernel-dependent filtering rapid target tracking method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108550161A true CN108550161A (en) | 2018-09-18 |
CN108550161B CN108550161B (en) | 2021-09-14 |
Family
ID=63516650
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810228008.1A Active CN108550161B (en) | 2018-03-20 | 2018-03-20 | Scale self-adaptive kernel-dependent filtering rapid target tracking method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108550161B (en) |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109859243A (en) * | 2019-01-18 | 2019-06-07 | 昆明理工大学 | A kind of motion target tracking method based on dimension self-adaption block particle |
CN109858455A (en) * | 2019-02-18 | 2019-06-07 | 南京航空航天大学 | A kind of piecemeal detection scale adaptive tracking method for circular target |
CN109886280A (en) * | 2019-02-21 | 2019-06-14 | 西安微电子技术研究所 | A kind of heterologous image object matching process based on core correlation filtering |
CN109977971A (en) * | 2019-03-29 | 2019-07-05 | 苏州大学 | Dimension self-adaption Target Tracking System based on mean shift Yu core correlation filtering |
CN109993052A (en) * | 2018-12-26 | 2019-07-09 | 上海航天控制技术研究所 | The method for tracking target and system of dimension self-adaption under a kind of complex scene |
CN109993777A (en) * | 2019-04-04 | 2019-07-09 | 杭州电子科技大学 | A kind of method for tracking target and system based on double-template adaptive threshold |
CN110032709A (en) * | 2019-01-24 | 2019-07-19 | 太原理工大学 | A kind of positioning and estimation method for abnormal point in geographical coordinate conversion |
CN110033006A (en) * | 2019-04-04 | 2019-07-19 | 中设设计集团股份有限公司 | Vehicle detecting and tracking method based on color characteristic Nonlinear Dimension Reduction |
CN110097575A (en) * | 2019-04-28 | 2019-08-06 | 电子科技大学 | A kind of method for tracking target based on local feature and scale pond |
CN110175649A (en) * | 2019-05-28 | 2019-08-27 | 南京信息工程大学 | It is a kind of about the quick multiscale estimatiL method for tracking target detected again |
CN110276785A (en) * | 2019-06-24 | 2019-09-24 | 电子科技大学 | One kind is anti-to block infrared object tracking method |
CN110298868A (en) * | 2019-06-26 | 2019-10-01 | 北京工业大学 | A kind of multiscale target tracking of high real-time |
CN110349177A (en) * | 2019-07-03 | 2019-10-18 | 广州多益网络股份有限公司 | A kind of the face key point-tracking method and system of successive frame video flowing |
CN110826403A (en) * | 2019-09-27 | 2020-02-21 | 深圳云天励飞技术有限公司 | Tracking target determination method and related equipment |
CN111161321A (en) * | 2019-12-31 | 2020-05-15 | 北京理工大学重庆创新中心 | Relevant filtering tracking method based on background perception |
CN111354022A (en) * | 2020-02-20 | 2020-06-30 | 中科星图股份有限公司 | Target tracking method and system based on kernel correlation filtering |
CN111563915A (en) * | 2020-05-06 | 2020-08-21 | 南京航空航天大学 | KCF target tracking method integrating motion information detection and Radon transformation |
CN111814590A (en) * | 2020-06-18 | 2020-10-23 | 浙江大华技术股份有限公司 | Personnel safety state monitoring method, equipment and computer readable storage medium |
CN114119970A (en) * | 2022-01-29 | 2022-03-01 | 中科视语(北京)科技有限公司 | Target tracking method and device |
CN115631359A (en) * | 2022-11-17 | 2023-01-20 | 诡谷子人工智能科技(深圳)有限公司 | Image data processing method and device for machine vision recognition |
CN115797647A (en) * | 2022-11-14 | 2023-03-14 | 西安电子科技大学广州研究院 | Target stable tracking method under embedded open environment |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106557774A (en) * | 2015-09-29 | 2017-04-05 | 南京信息工程大学 | The method for real time tracking of multichannel core correlation filtering |
CN107240122A (en) * | 2017-06-15 | 2017-10-10 | 国家新闻出版广电总局广播科学研究院 | Video target tracking method based on space and time continuous correlation filtering |
-
2018
- 2018-03-20 CN CN201810228008.1A patent/CN108550161B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106557774A (en) * | 2015-09-29 | 2017-04-05 | 南京信息工程大学 | The method for real time tracking of multichannel core correlation filtering |
CN107240122A (en) * | 2017-06-15 | 2017-10-10 | 国家新闻出版广电总局广播科学研究院 | Video target tracking method based on space and time continuous correlation filtering |
Non-Patent Citations (2)
Title |
---|
ANDR´ES SOL´IS MONTERO等: "Scalable Kernel Correlation Filter with Sparse Feature Integration", 《2015ICCVW》 * |
李麒骥: "尺度自适应的核相关滤波跟踪器", 《计算机应用》 * |
Cited By (37)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109993052A (en) * | 2018-12-26 | 2019-07-09 | 上海航天控制技术研究所 | The method for tracking target and system of dimension self-adaption under a kind of complex scene |
CN109993052B (en) * | 2018-12-26 | 2021-04-13 | 上海航天控制技术研究所 | Scale-adaptive target tracking method and system under complex scene |
CN109859243A (en) * | 2019-01-18 | 2019-06-07 | 昆明理工大学 | A kind of motion target tracking method based on dimension self-adaption block particle |
CN110032709A (en) * | 2019-01-24 | 2019-07-19 | 太原理工大学 | A kind of positioning and estimation method for abnormal point in geographical coordinate conversion |
CN110032709B (en) * | 2019-01-24 | 2023-04-14 | 太原理工大学 | Positioning and estimation method for abnormal point in geographic coordinate conversion |
CN109858455A (en) * | 2019-02-18 | 2019-06-07 | 南京航空航天大学 | A kind of piecemeal detection scale adaptive tracking method for circular target |
CN109886280A (en) * | 2019-02-21 | 2019-06-14 | 西安微电子技术研究所 | A kind of heterologous image object matching process based on core correlation filtering |
CN109886280B (en) * | 2019-02-21 | 2022-05-31 | 西安微电子技术研究所 | Heterogeneous image target matching method based on kernel correlation filtering |
CN109977971A (en) * | 2019-03-29 | 2019-07-05 | 苏州大学 | Dimension self-adaption Target Tracking System based on mean shift Yu core correlation filtering |
CN109993777A (en) * | 2019-04-04 | 2019-07-09 | 杭州电子科技大学 | A kind of method for tracking target and system based on double-template adaptive threshold |
CN110033006A (en) * | 2019-04-04 | 2019-07-19 | 中设设计集团股份有限公司 | Vehicle detecting and tracking method based on color characteristic Nonlinear Dimension Reduction |
CN109993777B (en) * | 2019-04-04 | 2021-06-29 | 杭州电子科技大学 | Target tracking method and system based on dual-template adaptive threshold |
CN110097575A (en) * | 2019-04-28 | 2019-08-06 | 电子科技大学 | A kind of method for tracking target based on local feature and scale pond |
CN110097575B (en) * | 2019-04-28 | 2022-07-01 | 电子科技大学 | Target tracking method based on local features and scale pool |
CN110175649A (en) * | 2019-05-28 | 2019-08-27 | 南京信息工程大学 | It is a kind of about the quick multiscale estimatiL method for tracking target detected again |
CN110175649B (en) * | 2019-05-28 | 2022-06-07 | 南京信息工程大学 | Rapid multi-scale estimation target tracking method for re-detection |
CN110276785A (en) * | 2019-06-24 | 2019-09-24 | 电子科技大学 | One kind is anti-to block infrared object tracking method |
CN110276785B (en) * | 2019-06-24 | 2023-03-31 | 电子科技大学 | Anti-shielding infrared target tracking method |
CN110298868B (en) * | 2019-06-26 | 2021-06-25 | 北京工业大学 | High-instantaneity multi-scale target tracking method |
CN110298868A (en) * | 2019-06-26 | 2019-10-01 | 北京工业大学 | A kind of multiscale target tracking of high real-time |
CN110349177B (en) * | 2019-07-03 | 2021-08-03 | 广州多益网络股份有限公司 | Method and system for tracking key points of human face of continuous frame video stream |
CN110349177A (en) * | 2019-07-03 | 2019-10-18 | 广州多益网络股份有限公司 | A kind of the face key point-tracking method and system of successive frame video flowing |
CN110826403A (en) * | 2019-09-27 | 2020-02-21 | 深圳云天励飞技术有限公司 | Tracking target determination method and related equipment |
CN110826403B (en) * | 2019-09-27 | 2020-11-24 | 深圳云天励飞技术有限公司 | Tracking target determination method and related equipment |
CN111161321A (en) * | 2019-12-31 | 2020-05-15 | 北京理工大学重庆创新中心 | Relevant filtering tracking method based on background perception |
CN111161321B (en) * | 2019-12-31 | 2023-07-04 | 北京理工大学重庆创新中心 | Relevant filtering tracking method based on background perception |
CN111354022A (en) * | 2020-02-20 | 2020-06-30 | 中科星图股份有限公司 | Target tracking method and system based on kernel correlation filtering |
CN111354022B (en) * | 2020-02-20 | 2023-08-22 | 中科星图股份有限公司 | Target Tracking Method and System Based on Kernel Correlation Filtering |
CN111563915A (en) * | 2020-05-06 | 2020-08-21 | 南京航空航天大学 | KCF target tracking method integrating motion information detection and Radon transformation |
CN111563915B (en) * | 2020-05-06 | 2023-04-14 | 南京航空航天大学 | KCF target tracking method integrating motion information detection and Radon transformation |
CN111814590A (en) * | 2020-06-18 | 2020-10-23 | 浙江大华技术股份有限公司 | Personnel safety state monitoring method, equipment and computer readable storage medium |
CN111814590B (en) * | 2020-06-18 | 2023-09-29 | 浙江大华技术股份有限公司 | Personnel safety state monitoring method, equipment and computer readable storage medium |
CN114119970B (en) * | 2022-01-29 | 2022-05-03 | 中科视语(北京)科技有限公司 | Target tracking method and device |
CN114119970A (en) * | 2022-01-29 | 2022-03-01 | 中科视语(北京)科技有限公司 | Target tracking method and device |
CN115797647A (en) * | 2022-11-14 | 2023-03-14 | 西安电子科技大学广州研究院 | Target stable tracking method under embedded open environment |
CN115797647B (en) * | 2022-11-14 | 2023-09-08 | 西安电子科技大学广州研究院 | Target stable tracking method under embedded open environment |
CN115631359A (en) * | 2022-11-17 | 2023-01-20 | 诡谷子人工智能科技(深圳)有限公司 | Image data processing method and device for machine vision recognition |
Also Published As
Publication number | Publication date |
---|---|
CN108550161B (en) | 2021-09-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108550161A (en) | A kind of dimension self-adaption core correlation filtering fast-moving target tracking method | |
CN105405151B (en) | Anti-shelter target tracking based on particle filter and weighting Surf | |
CN104574445B (en) | A kind of method for tracking target | |
CN105069472B (en) | A kind of vehicle checking method adaptive based on convolutional neural networks | |
CN103514441B (en) | Facial feature point locating tracking method based on mobile platform | |
CN108665481A (en) | Multilayer depth characteristic fusion it is adaptive resist block infrared object tracking method | |
CN108346159A (en) | A kind of visual target tracking method based on tracking-study-detection | |
CN110135500A (en) | Method for tracking target under a kind of more scenes based on adaptive depth characteristic filter | |
CN103886325B (en) | Cyclic matrix video tracking method with partition | |
CN107403175A (en) | Visual tracking method and Visual Tracking System under a kind of movement background | |
CN101394546B (en) | Video target profile tracing method and device | |
CN105678811A (en) | Motion-detection-based human body abnormal behavior detection method | |
CN102142085B (en) | Robust tracking method for moving flame target in forest region monitoring video | |
CN104933738B (en) | A kind of visual saliency map generation method detected based on partial structurtes with contrast | |
CN100531405C (en) | Target tracking method of sports video | |
CN111582349B (en) | Improved target tracking algorithm based on YOLOv3 and kernel correlation filtering | |
CN113706581B (en) | Target tracking method based on residual channel attention and multi-level classification regression | |
CN111340758B (en) | Novel efficient iris image quality evaluation method based on deep neural network | |
CN110991397B (en) | Travel direction determining method and related equipment | |
CN109242883A (en) | Optical remote sensing video target tracking method based on depth S R-KCF filtering | |
CN109448307A (en) | A kind of recognition methods of fire disaster target and device | |
CN109376736A (en) | A kind of small video target detection method based on depth convolutional neural networks | |
CN112149665A (en) | High-performance multi-scale target detection method based on deep learning | |
CN110309789A (en) | Video monitoring human face clarity evaluation method and device based on deep learning | |
CN104680536A (en) | Method for detecting SAR image change by utilizing improved non-local average algorithm |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |