CN104599289B - Method for tracking target and device - Google Patents
Method for tracking target and device Download PDFInfo
- Publication number
- CN104599289B CN104599289B CN201410854540.6A CN201410854540A CN104599289B CN 104599289 B CN104599289 B CN 104599289B CN 201410854540 A CN201410854540 A CN 201410854540A CN 104599289 B CN104599289 B CN 104599289B
- Authority
- CN
- China
- Prior art keywords
- vector
- frame image
- core
- weight
- current frame
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000012806 monitoring device Methods 0.000 claims abstract description 5
- 239000013598 vector Substances 0.000 claims description 136
- 230000009977 dual effect Effects 0.000 claims description 37
- 238000004364 calculation method Methods 0.000 claims description 35
- 238000000605 extraction Methods 0.000 claims description 29
- 230000021615 conjugation Effects 0.000 claims description 6
- 230000004927 fusion Effects 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims 2
- 238000012544 monitoring process Methods 0.000 abstract description 5
- 238000005516 engineering process Methods 0.000 abstract description 2
- 238000001514 detection method Methods 0.000 description 10
- 238000010276 construction Methods 0.000 description 7
- 230000003044 adaptive effect Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 4
- UREBDLICKHMUKA-CXSFZGCWSA-N dexamethasone Chemical compound C1CC2=CC(=O)C=C[C@]2(C)[C@]2(F)[C@@H]1[C@@H]1C[C@@H](C)[C@@](C(=O)CO)(O)[C@@]1(C)C[C@@H]2O UREBDLICKHMUKA-CXSFZGCWSA-N 0.000 description 4
- 229940087412 maxidex Drugs 0.000 description 4
- 230000003068 static effect Effects 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 230000001419 dependent effect Effects 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 230000001960 triggered effect Effects 0.000 description 2
- 238000003339 best practice Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000002045 lasting effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
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
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention is suitable for monitoring technology field, provides a kind of method for tracking target and device, and the method includes obtaining current frame image；According to the apparent model that previous frame image provides, target position locating for target is calculated in current frame image；The apparent model of current frame image is updated according to the target position of calculating, and the posture of monitoring device is adjusted, completes the tracking to the target.The method and device of the embodiment of the present invention one has stronger robustness.
Description
Technical field
The invention belongs to monitoring technology field more particularly to a kind of method for tracking target and device.
Background technique
In monitoring field, PTZ intelligence ball machine, Intelligent target following function is the intelligentized important indicator of ball machine,
It has a wide range of applications in safety monitoring.The characteristics of safety monitoring video is that environment is complicated and changeable, and such as change round the clock, picture shake
Dynamic, image quality deviation, illumination variation, target deformation etc., it is desirable that it is long that target tracker carries out duration under such complex environment
The tracking of time is an extremely challenging task.Method for tracking target under complex environment should have following characteristics: 1, certainly
Adapt to environment still be able to accurately describe target with the variation of ambient enviroment, and complete lasting target with
Track；2, tracking device can track in real time target, have to complete at the appointed time during tracking
The processing of tracking, it means that intelligent algorithm must be simple and efficient；3, the device tracked is easy to use, and it is excessive that no setting is required
Parameter, can install and use.
A kind of rotation of existing patent and the tracking of zoom Pan/Tilt/Zoom camera method and device (CN, number of patent application:
201210343868.2), this method provides the method for tracking target that a kind of multiple features combine, which determines according to former frame
Target area, on the image of present frame according at least one feature search with the most like current goal area in target area
Domain.The Fusion Features of the invention are empirically to carry out that linear weighted function is average to multiple features result, pair that can not be adaptive
Plurality of target feature is merged, therefore environment that cannot well adaptively around, is easy to produce target drift and is showed with what is lost
As, while the parameter setting of various features fusion, lead to the practicability for reducing tracking.
Paper " High-Speed Tracking with Kernelized Correlation Filters " proposes one
The superfast method for tracking target of kind, Target Tracking Problem is transformed into operation under frequency domain by this method, to reach quick height
The performance of effect, it is per second that the speed of tracking has reached 172 frames, therefore has very strong practicability, but this method is merely capable of making
Target is tracked with a kind of feature, reduces tracking to the adaptability of environment, in addition tracking uses core letter
Number needs to be arranged reasonable parameter, and regular factor is also required to reasonably be arranged, to reduce the easy-to-use of tracking
Property.
Summary of the invention
The embodiment of the present invention is designed to provide a kind of method for tracking target and device, it is intended to solve existing target with
Track method target is easy the problem of drift.
The embodiments of the present invention are implemented as follows, and a kind of method for tracking target, described method includes following steps:
Obtain current frame image；
According to the apparent model that previous frame image provides, target position locating for target is counted in current frame image
It calculates；
The apparent model of current frame image is updated according to the target position of calculating, and to the appearance of monitoring device
State is adjusted, and completes the tracking to the target.
Further, the apparent model provided according to previous frame image, to locating for target in current frame image
Target position is calculated, and is specifically included:
Extract at least two features that region is specified in current frame image；
The core associated vector under its frequency domain is constructed respectively according to each feature of extraction, and is generated nuclear phase and closed vector set
It closes；
Vector set is closed to the nuclear phase and carries out fusion treatment；
Decision vector is calculated according to fused core associated vector；
The target position is calculated according to the decision vector.
Further, the calculation formula of the fused core associated vector are as follows:
The calculation formula of the decision vector are as follows:
Wherein, wherein belong to any r 1,2..., R),It is M for a lengthrColumn vector；
Rel () expression takes complex vector located real part vector, and Conj () expression takes complex vector located conjugation, IDFT ()
It indicates to solve the Fast Fourier Transform (FFT) under frequency domain, ⊙ indicates vector dot.
Further, described that the apparent model of current frame image is updated according to the target position of calculating, tool
Body includes:
Extract at least two features in the target position of current frame image；
The core associated vector under its frequency domain is constructed respectively according to each feature of extraction, and is generated nuclear phase and closed vector set
It closes；
Calculate the core weight and dual variable of current frame image.
Further, the core weight and dual variable for calculating current frame image, comprising:
It obtains and refers to core weight, refer to core weight calculation dual variable according to described；
The core weight of current frame image is calculated according to the dual variable；
When the core weight for the current frame image being calculated meets preset condition, by the core of the current frame image
Weight and the dual variable are exported as updated apparent model；
When the core weight for the current frame image being calculated is unsatisfactory for preset condition, then will be calculated described in
The core weight of current frame image, which is used as, refers to weight, and calculates again current frame image core weight, until being calculated
Until current frame image core weight meets preset condition.
Further, the calculation formula of the dual variable are as follows:
Wherein,DFT function expression does Fast Fourier Transform (FFT) to a certain vector.
Further, the calculation formula of the core weight are as follows:
Wherein,
The present invention also proposes a kind of target tracker, and described device includes:
Module is obtained, for obtaining current frame image；
Computing module, the apparent model for being provided according to previous frame image, to locating for target in current frame image
Target position is calculated；
Update module, for being updated according to the target position of calculating to the apparent model of current frame image, and
The posture of monitoring device is adjusted, the tracking to the target is completed.
Further, the computing module includes:
First extraction unit, for extracting at least two features for specifying region in current frame image；
First structural unit, for constructing the core associated vector under its frequency domain respectively according to each feature of extraction,
And it generates nuclear phase and closes vector set；
Integrated unit, for calculating decision vector according to fused core associated vector；
First computing unit, for being calculated according to the decision vector the target position.
Further, the calculation formula of the fused core associated vector are as follows:
The calculation formula of the decision vector are as follows:
Wherein, wherein belong to any r 1,2..., R),It is M for a lengthrColumn vector；
Rel () expression takes complex vector located real part vector, and Conj () expression takes complex vector located conjugation, IDFT ()
It indicates to solve the Fast Fourier Transform (FFT) under frequency domain, ⊙ indicates vector dot.
Further, the update module includes:
Second extraction unit, at least two features in the target position for extracting current frame image；
Second structural unit, for constructing the core associated vector under its frequency domain respectively according to each feature of extraction,
And it generates nuclear phase and closes vector set；
Second computing unit, for calculating the core weight and dual variable of current frame image.
Further, second computing unit includes:
Dual variable computation subunit refers to core weight for obtaining, and refers to core weight calculation dual variable according to described；
Core weight calculation subelement, for calculating the core weight of current frame image according to the dual variable；
Subelement is exported, for when the core weight for the current frame image being calculated meets preset condition, by institute
The core weight and the dual variable for stating current frame image are exported as updated apparent model；
Subelement is recycled, for when the core weight for the current frame image being calculated is unsatisfactory for preset condition, then
Using the core weight for the current frame image being calculated as weight is referred to, and current frame image core weight is counted again
It calculates, until current frame image core weight is calculated and meets preset condition.
Further, the calculation formula of the dual variable are as follows:
Wherein,DFT function expression does Fast Fourier Transform (FFT) to a certain vector.
Further, the calculation formula of the core weight are as follows:
Wherein,
The method and device of the embodiment of the present invention accurately merges the feature of plurality of target, such as the color of target, profile, line
The features such as reason, this method are a kind of method of printenv, i.e. user is without adjusting any parameter, with the variation of environment, with
Track algorithm can be adaptive the weights of the above-mentioned various features of adjustment can be found always most suitable with adapting to current environment
Feature target is described, thus have very strong robustness, be not easy by target drift about with losing.In addition, the present invention is implemented
Tracking and the device calculating of example are simple and efficient, due to carrying out operation under frequency domain, having calculation amount small and calculating speed
Fast feature is spent, can be easily carried out under hardware.
Detailed description of the invention
Fig. 1 is the flow chart for the method for tracking target that the embodiment of the present invention one provides；
Fig. 2 is the flow chart that target position is calculated in the method for tracking target of the offer of the embodiment of the present invention one；
Fig. 3 is to solve the flow chart of optimal core weight in method for tracking target that the embodiment of the present invention one provides；
Fig. 4 is that the flow chart for watching model quietly is updated in the method for tracking target of the offer of the embodiment of the present invention one；
Fig. 5 is the structure chart of target tracker provided by Embodiment 2 of the present invention；
Fig. 6 is the structure chart of computing module in target tracker provided by Embodiment 2 of the present invention；
Fig. 7 is the structure chart of update module in target tracker provided by Embodiment 2 of the present invention；
Fig. 8 is the structure chart of the second computing unit in target tracker provided by Embodiment 2 of the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
Embodiment one
The embodiment of the present invention one proposes a kind of method for tracking target.As shown in Figure 1, the method for the embodiment of the present invention one includes
Step:
S1, current frame image is obtained.Assuming that current frame image is the color image I of RGB, the wide height of image is respectively W and H,
The pixel of h row in image, w column is expressed as I (h, w).
S2, according to previous frame image provide apparent model, in current frame image to target position locating for target into
Row calculates.As shown in Fig. 2, specifically including step:
S201, judge whether start-up trace, enter step S202 if without starting, otherwise enter step S203.Assuming that
IsTrigger is enabled to indicate whether triggering tracking, initialization value IsTrigger=false, expression, which is not detected, to be needed to track
Target.
S202, tracking detection trigger, and return step S1 reacquires image.In the case where not triggering tracking, camera
Be it is static, therefore the background of picture be it is static, can be used Detection for Moving Target in picture moving target carry out
Detecting and tracking, when moving target reaches the tracking that certain condition triggers target.Detection for Moving Target is mature, such as
Well known OpenCV just has such detection algorithm: assuming that tracking detection trigger has been detected by the target A indicated with rectangle
={ x, y, w, h }, x, y, w, h respectively indicate the abscissa of top left corner apex of the target A in image I, ordinate, target width
The tracking that will trigger target is indicated then IsTrigger is set as true with the height of target.
S203, judge whether apparent model initializes, S204 is entered step if not, otherwise enters step S205.It enables
IsInitTracker indicate whether initialized target apparent model and its relevant parameter, if IsInitTracker
=false indicates that tracking is just triggered, and apparent model not yet initializes, and needs to make the apparent model needs of target corresponding
Initialization.
S204, apparent model initialization, specifically include step:
A feature) is extracted.The feature in specified region is calculated, the feature of extraction can be color, shape, Texture eigenvalue, obtain
To the set of a feature.It is expressed as belonging to arbitrary r, r { 1,2 ..., R }, it is special
Levy vectorWherein M=w × h indicates that the dimension of feature is (practical
It can be equal to number of pixels).Notice that the dimension of the sample of different characteristic is identical.
B)x0Class weight Y={ y, y, y ..., y }, class weight vector y=(y1, y2..., yM)T, wherein M is indicated
The dimension of feature y, and to index m ∈ { 0,1 ..., M }, the abscissa and ordinate of correspondence image are respectively m%w, m/w, preceding
Person indicates to take the remainder, such as 10%2=0,10%3=1, and the latter indicates round numbers.The building method of weight is according to actual characteristic
Put in order and reasonably constructed, such construction is completed often through two-dimensional Gaussian function.
C apparent model) is established.Establish apparent modeld0=(d00, d01, d02, d03..., d0R)T
Indicate the weight of core associated vector,It is real number value, with above-mentioned y for column vector
It is corresponding.
D)It indicates the dual variable under frequency domain, is column vector, element
For plural number.It solves optimalThe embodiment of the present invention one provides a kind of optimal core weight d=(d of solution0, d1, d2,
d3..., dR)TAnd dual variableOptimization method, this method is using a kind of thermal starting
Method initializes core weight d, i.e., is used as the nuclei originis weight of the embodiment of the present invention one, reality using last optimal core weight
Border application shows that the efficiency for solving optimal core weight can be increased substantially in this way.It is known to be referred to from characteristic extraction step
Determine the feature in regionT=0,1,2 ..., it is shown in Figure 3, solve best practice the step of
It is as follows:
S20441, parameter initialization.
Firstly, initialization core weightd0To refer to core weight dcJust
Initial value.If t is equal to 0, initializeIf t is not equal to 0, d is initialized0
=dt-1。
Then, it constructs and vector set is closed according to the nuclear phase under construction frequency domainWherein any r is belonged to 1,
2..., R }, it is M for a lengthrColumn vector, using linear kernel letter, Polynomial kernel function, Gaussian kernel
One of kernel function of function etc. is constructed.Wherein v be a M dimension column to
Amount, first element value are 1, and other elements value is 0, i.e. v=(1,0,0 ..., 0)T, iteration count symbol c=0.
S20442, fixed dc, calculate
S20443, it is fixedUpdate dc+1。
Firstly, calculating normal vector w=(w0, w1, w2..., wR)T, with core weight vectors d have identical dimensional, element and
It corresponds, calculation formula is as follows:
Secondly, calculating weight residual delta d=(Δ d0, Δ d1, Δ d2, Δ d3..., Δ dR)T,
Finally, updating core weight d, calculation formula is as follows:
It wherein, is updated core weight, the subsequent core weight for before update.
S20444 judges that core weight meets the following conditions:
Satisfaction then goes to step S20445, otherwise c=c+1, and return step S20442, it is noted that TsIndicate optimization algorithm
The condition of stopping, value range are the real number greater than 0, it is preferable that one value T of present examples=0.001.
S20445, outputdc+1。
E) IsInitTracker=true, tracking frame counter t=t+1 are set.
S205, target position is calculated.As shown in figure 4, specifically including step:
S2051, at least two features that region is specified in current frame image are extracted.Calculate the spy in the specified region of image I
Sign, the feature of extraction can be color, shape, and Texture eigenvalue obtains the set of a feature, is expressed as belonging to arbitrary r, r { 1,2 ..., R }, feature vectorWherein M indicates the dimension of feature.Pay attention to the feature of different samples
Dimension be identical.
S2052, the core associated vector under its frequency domain is constructed according to each feature of extraction respectively, and generates nuclear phase and closes vector set
It closes.Construction process is identical as nuclear phase pass vector set is constructed in S204.According to construction frequency domain
Under nuclear phase close vector set
{ 1,2..., R } wherein is belonged to any r, is M for a lengthrColumn vector, using linear kernel letter, more
One of kernel function of item formula kernel function, gaussian kernel function etc. is constructed,
A) linear kernel function:
B) Polynomial kernel function:
C) gaussian kernel function:
Wherein DFT function expression does Fast Fourier Transform (FFT) to a certain vector, and IDFT function representation is under a certain frequency domain
Vector does inverse fast Fourier transform, a, b, and σ is the adjustable parameter of dependent kernels.
S2053, vector set progress fusion treatment is closed to nuclear phase.
Calculation formula are as follows:
S2054, target position is calculated according to decision vector, calculation formula are as follows:
Wherein Rel () expression takes complex vector located real part vector, and Conj () expression takes complex vector located conjugation, IDFT
() indicates to solve the Fast Fourier Transform (FFT) under frequency domain, and ⊙ indicates vector dot.
The v being calculated according to previous stept, confirm the position A of targett, formula is as follows:
Id=MaxIdex (vt)；
MaxIdex indicates vtIn maximum element position subscript index corresponding m.The position of so target is then At={ id%
W, id/w, w, h }.
S206, judge whether target following terminates, be, set false for IsTrigger, and terminate process, otherwise
Enter step S3.
S3, the apparent model of current frame image is updated according to the target position of calculating, and to the appearance of monitoring device
State is adjusted, and completes the tracking to target.
The step of more new model, is as follows:
Firstly, extracting at least two features in the target position of current frame image.Calculate the specified region A of image It=
The feature of { id%w, id/w, w, h }, the feature of extraction can be color, shape, and Texture eigenvalue obtains the collection of a feature
It closes, is expressed as
Secondly, each feature according to extraction constructs the core associated vector under its frequency domain respectively, and generate nuclear phase pass
Vector set.
Secondly, calculating the core weight and dual variable of current frame image, more new modelIt will be pre-
The position of survey is converted into needing the multiplying power of the angle and scaling that rotate, to adjust camera posture.Return step S1.In the step,
It is optimal that method solution shown in Fig. 3 can be used
The method of the embodiment of the present invention one accurately merges the feature of plurality of target, such as the color of target, profile, texture
Feature, this method are a kind of method of printenv, i.e. user is without adjusting any parameter, and with the variation of environment, tracking is calculated
Method can be adaptive the weights of the above-mentioned various features of adjustment can find most suitable spy always to adapt to current environment
Target is described in sign, to have very strong robustness, is not easy to drift about target with losing.In addition, the embodiment of the present invention one
Tracking calculating be simple and efficient, due to carrying out operation under frequency domain, have calculation amount small and the fast spy of calculating speed
Point can be easily carried out under hardware.
Embodiment two
The embodiment of the present invention two proposes a kind of target tracker.As shown in figure 5, the device of the embodiment of the present invention two includes
Obtain module 10, computing module 20 and update module 30.
Module 10 is obtained, for obtaining current frame image；Assuming that current frame image is the color image I of RGB, the wide height of image
The pixel of respectively W and H, the h row in image, w column are expressed as I (h, w).
Computing module 20, the apparent model for being provided according to previous frame image, to locating for target in current frame image
Target position calculated.As shown in fig. 6, computing module 20 includes the first extraction unit 21, the first structural unit 22, fusion
Unit 23 and the first computing unit 24, the working principle of computing module 20 specifically:
Computing module 20 judges whether start-up trace, it is assumed that IsTrigger is enabled to indicate whether triggering tracking, initialization
Value IsTrigger=false indicates that the target for needing to track is not detected.
Tracking detection trigger is carried out if without starting, and returns and obtains the reacquisition image of module 10.Do not trigger with
In the case where track, camera be it is static, therefore the background of picture be it is static, Detection for Moving Target can be used in picture
Moving target carry out detecting and tracking, when moving target reaches the tracking that certain condition triggers target.Moving object detection skill
Art is mature, such as well known OpenCV just has such detection algorithm: assuming that tracking detection trigger has been detected by use
Rectangle indicate target A={ x, y, w, h }, x, y, w, h respectively indicate top left corner apex of the target A in image I abscissa,
The height of ordinate, the width of target and target indicates the tracking that will trigger target then IsTrigger is set as true.
Continue to judge whether apparent model initializes if having been turned on tracking.IsInitTracker is enabled to indicate whether
The apparent model of initialized target and its relevant parameter indicate that tracking is just triggered if IsInitTracker=false,
Apparent model not yet initializes, and needs to need to do corresponding initialization to the apparent model of target.Apparent model initialization is specific
Include:
A feature) is extracted.The feature in specified region is calculated, the feature of extraction can be color, shape, Texture eigenvalue, obtain
To the set of a feature.It is expressed as belonging to arbitrary r, r { 1,2 ..., R }, it is special
Levy vectorWherein M=w × h indicates that the dimension of feature is (practical
It can be equal to number of pixels).Notice that the dimension of the sample of different characteristic is identical.
B)x0Class weight Y={ y, y, y ..., y }, class weight vector y=(y1, y2..., yM)T, wherein M is indicated
The dimension of feature y, and to index m ∈ { 0,1 ..., M }, the abscissa and ordinate of correspondence image are respectively m%w, m/w, preceding
Person indicates to take the remainder, such as 10%2=010%3=1, and the latter indicates round numbers.The building method of weight is according to actual characteristic
It puts in order and is reasonably constructed, such construction is completed often through two-dimensional Gaussian function.
C apparent model) is established.Establish apparent modeld0=(d00, d01, d02, d03..., d0R)T
Indicate the weight of core associated vector,It is real number value for column vectorWith it is above-mentioned
Y is corresponding.
D)It indicates the dual variable under frequency domain, is column vector, element
For plural number.It solves optimalIt solves about method for solving with update module 30 optimalMethod it is identical, in detail
See below the description as described in update module 30.
E) IsInitTracker=true, tracking frame counter t=t+1 are set.
If apparent model initialization has been completed, then the first extraction unit 21 extracts in current frame image and specifies region extremely
Few two features.First extraction unit 21 calculates the feature in the specified region of image I, and the feature of extraction can be color, shape,
Texture eigenvalue obtains the set of a feature, is expressed as belonging to arbitrary r, r
{ 1,2..., R }, feature vectorWherein M indicates the dimension of feature.
Notice that the dimension of the feature of different samples is identical.
First structural unit 22 is used to construct the core associated vector under its frequency domain respectively according to each feature according to extraction,
And it generates nuclear phase and closes vector set.First structural unit 22 is according to the core under construction frequency domain
Associated vector setIts
In { 1,2..., R } is belonged to any r, be a length be MrColumn vector, using linear kernel letter, multinomial
One of kernel function of formula kernel function, gaussian kernel function etc. is constructed,
D) linear kernel function:
E) Polynomial kernel function:
F) gaussian kernel function:
Wherein DFT function expression does Fast Fourier Transform (FFT) to a certain vector, and IDFT function representation is under a certain frequency domain
Vector does inverse fast Fourier transform, a, b, and σ is the adjustable parameter of dependent kernels.
Integrated unit 23 is used to calculate decision vector according to fused core associated vector.First more according to previous frame image
The apparent model and nuclear phase of new target close vector setCalculate fused core
Associated vector calculation formula are as follows:
Calculate decision vector vt, calculation formula are as follows:
Wherein Rel () expression takes complex vector located real part vector, and Conj () expression takes complex vector located conjugation, IDFT
() indicates to solve the Fast Fourier Transform (FFT) under frequency domain, and ⊙ indicates vector point multiplication.
First computing unit 24 is for calculating target position according to decision vector.According to the v being calculatedt, really
Recognize the position A of targett, formula is as follows:
Id=MaxIdex (vt)；
MaxIdex indicates vtIn maximum element position subscript index corresponding m.The position of so target is then At={ id%
W, id/w, w, h }.
Judge whether target following terminates, be, set false for IsTrigger, and stop target following, otherwise by
Update module 30 updates apparent model.
As shown in fig. 7, update module 30 includes the second extraction unit 31, the second structural unit 32 and the second computing unit
33, in which:
Second extraction unit 31 is used to extract at least two features in the target position of current frame image.Calculate image I
Specified region AtThe feature of={ id%w, id/w, w, h }, the feature of extraction can be color, shape, and Texture eigenvalue obtains
To the set of a feature, it is expressed as
Second structural unit 32 is used to construct the core associated vector under its frequency domain respectively according to each feature of extraction,
And it generates nuclear phase and closes vector set.Vector set is closed according to the nuclear phase under construction frequency domainWherein any r is belonged to 1,2...,
R }, it is M for a lengthrColumn vector, using linear kernel letter, Polynomial kernel function, gaussian kernel function etc.
One of kernel function constructed.Wherein v is the column vector of M dimension, the
One element value is 1, and other elements value is 0, i.e. v=(1,0,0 ..., 0)T, iteration count symbol c=0.
Second computing unit 33 is used to calculate the core weight and dual variable of current frame image.Second computing unit 33 uses
A kind of method of thermal starting initializes core weight d, i.e., using last optimal core weight as the embodiment of the present invention two
Nuclei originis weight, practical application show that the efficiency for solving optimal core weight can be increased substantially in this way.
As shown in figure 8, the second computing unit 33 specifically includes dual variable computation subunit 331, core weight calculation son list
Member 332, output subelement 333, circulation subelement 334.
Known second extraction unit 31 is extracted the feature in specified regionT=0,1,
2 ....
Dual variable computation subunit 331, which obtains, refers to core weight, according to reference core weight calculation dual variable.To mutation
It measures computation subunit 331 and initializes core weightd0To refer to core weight dc's
Initial value.If t is equal to 0, initializeIf t is not equal to 0, initially
Change d0=dt-1。
Fixed dc, calculate
Core weight calculation subelement 332 is fixed according to the core weight that dual variable calculates current frame imageUpdate dc +1.Calculate normal vector w=(w0, w1, w2..., wR)T, there is identical dimensional with d, element and d are corresponded, and calculation formula is such as
Under:
Calculate weight residual delta d=(Δ d0, Δ d1, Δ d2, Δ d3..., Δ dR)T,
It obtainsIt is more for updated core weight
Core weight before new.
Output subelement 333 judges whether core weight meets the following conditions:
Satisfaction then exportsComplete the prediction of target position.Otherwise c=c+1, then by 334 weight of circulation subelement
It is new to calculate, it is noted that TsIndicate that the condition that optimization algorithm stops, value range are the real number greater than 0, it is preferable that present example
Two value Ts=0.001.
After update module 30 updates apparent model, the target position that computing module 20 is predicted is converted into the angle for needing to rotate
The multiplying power of degree and scaling, to adjust camera posture.
The device of the embodiment of the present invention two accurately merges the feature of plurality of target, such as the color of target, profile, texture
Feature, the method that the present apparatus uses is a kind of method of printenv, i.e., user is without adjusting any parameter, with the change of environment
Change, the weight for the above-mentioned various features of adjustment that track algorithm can be adaptive can be found most always with adapting to current environment
Target is described in suitable feature, to have very strong robustness, is not easy to drift about target with losing.In addition, of the invention
The tracking device calculating of embodiment two is simple and efficient, due to carrying out operation under frequency domain, having calculation amount small and calculating speed
Fast feature is spent, can be easily carried out under hardware.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. a kind of method for tracking target, which is characterized in that described method includes following steps:
Current frame image is obtained,
Start-up trace, after initializing apparent model, according to the apparent model that previous frame image provides, to mesh in current frame image
The locating target position of mark is calculated,
The apparent model of current frame image is updated according to the target position of calculating, and to the posture of monitoring device into
Row adjustment, completes the tracking to the target；
It is described that the apparent model of current frame image is updated according to the target position of calculating, it specifically includes:
At least two features in the target position of current frame image are extracted,
The core associated vector under its frequency domain is constructed respectively according to each feature of extraction, and is generated nuclear phase and closed vector set,
Calculate the core weight and dual variable of current frame image；
The core weight and dual variable for calculating current frame image, comprising:
It obtains and refers to core weight, refer to core weight calculation dual variable according to described,
The core weight of current frame image is calculated according to the dual variable,
When the core weight for the current frame image being calculated meets preset condition, by the core weight of the current frame image
And the dual variable is exported as updated apparent model,
It is when the core weight for the current frame image being calculated is unsatisfactory for preset condition, then described current by what is be calculated
The core weight of frame image, which is used as, refers to weight, and calculates again current frame image core weight, until being calculated current
Until frame image core weight meets preset condition.
2. the method as described in claim 1, which is characterized in that the apparent model provided according to previous frame image is being worked as
Target position locating for target is calculated in prior image frame, is specifically included:
Extract at least two features that region is specified in current frame image；
The core associated vector under its frequency domain is constructed respectively according to each feature of extraction, and is generated nuclear phase and closed vector set；
Vector set is closed to the nuclear phase and carries out fusion treatment；
Decision vector is calculated according to fused core associated vector；
The target position is calculated according to the decision vector.
3. method according to claim 2, which is characterized in that the calculation formula of the fused core associated vector are as follows:
The calculation formula of the decision vector are as follows:
Wherein, wherein belonging to { 1,2..., R } to any r,The weight of core associated vector is respectively indicated,For frequency domain
Under core associated vector,It is M for a lengthrColumn vector,For the dual variable under frequency domain；
Rel () expression takes complex vector located real part vector, and Conj () expression takes complex vector located conjugation, and IDFT () is indicated
The Fast Fourier Transform (FFT) under frequency domain is solved, ⊙ indicates vector dot.
4. the method as described in claim 1, it is characterised in that: the calculation formula of the dual variable are as follows:
Wherein,V is the column vector of M dimension, and first element value is 1, other elements value
It is 0, i.e. v=(1,0,0 ..., 0)T, iteration count symbol c=0, y are class weight vector,Respectively indicate nuclear phase pass
The weight of vector,It is M for a lengthrColumn vector, DFT function expression fast Fourier is done to a certain vector
Transformation.
5. the method as described in claim 1, which is characterized in that the calculation formula of the core weight are as follows:
Wherein,
Wherein,For updated core weight,
For the core weight before update, Δ drFor weight residual error, w0、wr、wiRespectively normal vector, drIndicate the weight of core associated vector,The dual variable under frequency domain is respectively indicated,For the core associated vector under frequency domain,It is M for a lengthr
Column vector, ⊙ indicate vector dot.
6. a kind of target tracker, which is characterized in that described device includes:
Module is obtained, for obtaining current frame image,
Computing module is used for start-up trace, after initializing apparent model, according to the apparent model that previous frame image provides, is working as
Target position locating for target is calculated in prior image frame,
Update module, for being updated according to the target position of calculating to the apparent model of current frame image, and to prison
The posture of control equipment is adjusted, and completes the tracking to the target；
The update module includes:
Second extraction unit, at least two features in the target position for extracting current frame image,
Second structural unit constructs the core associated vector under its frequency domain for each feature according to extraction respectively, and raw
It is nucleated associated vector set,
Second computing unit, for calculating the core weight and dual variable of current frame image；
Second computing unit includes:
Dual variable computation subunit refers to core weight for obtaining, and refers to core weight calculation dual variable according to described,
Core weight calculation subelement, for calculating the core weight of current frame image according to the dual variable,
Subelement is exported, for working as by described in when the core weight for the current frame image being calculated meets preset condition
The core weight and the dual variable of prior image frame are exported as updated apparent model, are recycled subelement, are used for
When the core weight for the current frame image being calculated is unsatisfactory for preset condition, then the present frame figure that will be calculated
The core weight of picture, which is used as, refers to weight, and calculates again current frame image core weight, until present frame figure is calculated
Until meeting preset condition as core weight.
7. device as claimed in claim 6, which is characterized in that the computing module includes:
First extraction unit, for extracting at least two features for specifying region in current frame image；
First structural unit constructs the core associated vector under its frequency domain for each feature according to extraction respectively, and raw
It is nucleated associated vector set；
Integrated unit, for calculating decision vector according to fused core associated vector；
First computing unit, for being calculated according to the decision vector the target position.
8. device as claimed in claim 7, which is characterized in that the calculation formula of the fused core associated vector are as follows:
The calculation formula of the decision vector are as follows:
Wherein, wherein belonging to { 1,2..., R } to any r,The weight of core associated vector is respectively indicated,For frequency domain
Under core associated vector,It is M for a lengthrColumn vector,For the dual variable under frequency domain；
Rel () expression takes complex vector located real part vector, and Conj () expression takes complex vector located conjugation, and IDFT () is indicated
The Fast Fourier Transform (FFT) under frequency domain is solved, ⊙ indicates vector dot.
9. device as claimed in claim 6, it is characterised in that: the calculation formula of the dual variable are as follows:
Wherein,V is the column vector of M dimension, and first element value is 1, other elements value
It is 0, i.e. v=(1,0,0 ..., 0)T, iteration count symbol c=0, y are class weight vector,Respectively indicate nuclear phase pass
The weight of vector,It is M for a lengthrColumn vector, DFT function expression fast Fourier is done to a certain vector
Transformation.
10. device as claimed in claim 6, which is characterized in that the calculation formula of the core weight are as follows:
Wherein,
Wherein,For updated core weight,
For the core weight before update, Δ drFor weight residual error, w0、wr、wiThe respectively ordinate of correspondence image, drIndicate core associated vector
Weight,The dual variable under frequency domain is respectively indicated,For the core associated vector under frequency domain,It is long for one
Degree is MrColumn vector, ⊙ indicate vector dot.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410854540.6A CN104599289B (en) | 2014-12-31 | 2014-12-31 | Method for tracking target and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410854540.6A CN104599289B (en) | 2014-12-31 | 2014-12-31 | Method for tracking target and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104599289A CN104599289A (en) | 2015-05-06 |
| CN104599289B true CN104599289B (en) | 2018-12-07 |
Family
ID=53125039
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410854540.6A Active CN104599289B (en) | 2014-12-31 | 2014-12-31 | Method for tracking target and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104599289B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105303541A (en) * | 2015-08-14 | 2016-02-03 | 深圳市瀚海基因生物科技有限公司 | Single-molecule image correction method |
| CN108961308B (en) * | 2018-06-01 | 2021-07-02 | 南京信息工程大学 | Residual error depth characteristic target tracking method for drift detection |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103310466A (en) * | 2013-06-28 | 2013-09-18 | 安科智慧城市技术（中国）有限公司 | Single target tracking method and achievement device thereof |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8296248B2 (en) * | 2009-06-30 | 2012-10-23 | Mitsubishi Electric Research Laboratories, Inc. | Method for clustering samples with weakly supervised kernel mean shift matrices |
| CN103679744B (en) * | 2012-09-17 | 2017-09-12 | 浙江大华技术股份有限公司 | A kind of rotation and the method and device of zoom Pan/Tilt/Zoom camera tracking |
| CN104200237B (en) * | 2014-08-22 | 2019-01-11 | 浙江生辉照明有限公司 | One kind being based on the High-Speed Automatic multi-object tracking method of coring correlation filtering |
| CN104243833B (en) * | 2014-09-30 | 2017-12-08 | 精宸智云（武汉）科技有限公司 | Camera posture adjusting method and device |
-
2014
- 2014-12-31 CN CN201410854540.6A patent/CN104599289B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103310466A (en) * | 2013-06-28 | 2013-09-18 | 安科智慧城市技术（中国）有限公司 | Single target tracking method and achievement device thereof |
Non-Patent Citations (2)
| Title |
|---|
| High-Speed Tracking with Kernelized Correlation Filters;Joao F. Henriques et al.;《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》;20140711;第588页第5.2节 * |
| 基于表观模型的视觉跟踪系统研究与实现;李天平;《中国优秀硕士学位论文全文数据库 信息科技辑》;20131015;第26页第3.2节，第41页图4.1 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104599289A (en) | 2015-05-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105022982B (en) | Hand motion recognition method and apparatus | |
| US20200026910A1 (en) | Gesture identification, control, and neural network training methods and apparatuses, and electronic devices | |
| US20200074165A1 (en) | Image analysis using neural networks for pose and action identification | |
| Paragios et al. | Unifying boundary and region-based information for geodesic active tracking | |
| Chen et al. | A passive auto-focus camera control system | |
| CN105279769B (en) | A kind of level particle filter tracking method for combining multiple features | |
| Xu et al. | Video saliency detection via graph clustering with motion energy and spatiotemporal objectness | |
| CN109165589A (en) | Vehicle based on deep learning recognition methods and device again | |
| CN104537381B (en) | A kind of fuzzy image recognition method based on fuzzy invariant features | |
| Papa et al. | Projections onto convex sets through particle swarm optimization and its application for remote sensing image restoration | |
| CN104599289B (en) | Method for tracking target and device | |
| Li et al. | A maximum a posteriori estimation framework for robust high dynamic range video synthesis | |
| CN108021852A (en) | A kind of demographic method, passenger number statistical system and electronic equipment | |
| Wang et al. | Low-altitude infrared small target detection based on fully convolutional regression network and graph matching | |
| Li et al. | A closer look tells more: a facial distortion based liveness detection for face authentication | |
| Bakour et al. | Soft-CSRNet: real-time dilated convolutional neural networks for crowd counting with drones | |
| Buddubariki et al. | Multiple object tracking by improved KLT tracker over SURF features | |
| Chiranjeevi et al. | Spatially correlated background subtraction, based on adaptive background maintenance | |
| Caseiro et al. | Foreground segmentation via background modeling on Riemannian manifolds | |
| CN112257492A (en) | Real-time intrusion detection and tracking method for multiple cameras | |
| Ding et al. | Video saliency detection by 3D convolutional neural networks | |
| CN113657186B (en) | Feature extraction method and device based on pedestrian re-recognition and storage medium | |
| Yang et al. | OPTIMAL SPATIAL ADAPTATION FOR LOCAL REGION-BASED ACTIVE CONTOURS-An Intersection of Confidence Intervals Approach | |
| Zhang et al. | Robust object tracking based on sparse representation | |
| CN104182990B (en) | A kind of Realtime sequence images motion target area acquisition methods |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information | ||
| CB02 | Change of applicant information |
Address after: 518034 room 1306, Press Plaza, Shennan Avenue, Futian District, Shenzhen, Guangdong Applicant after: ANKE ROBOT CO.,LTD. Address before: 518034 room 1306, Press Plaza, Shennan Avenue, Futian District, Shenzhen, Guangdong Applicant before: ANKE SMART CITY TECHNOLOGY (PRC) Co.,Ltd. |
|
| TA01 | Transfer of patent application right |
Effective date of registration: 20181018 Address after: 211100 89 Shengli Road, Jiangning economic and Technological Development Zone, Nanjing, Jiangsu Applicant after: NANJING CHIEBOT ROBOT TECHNOLOGY CO.,LTD. Address before: 518034 room 1306, Press Plaza, Shennan Avenue, Futian District, Shenzhen, Guangdong Applicant before: ANKE ROBOT CO.,LTD. |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| PE01 | Entry into force of the registration of the contract for pledge of patent right | ||
| PE01 | Entry into force of the registration of the contract for pledge of patent right |
Denomination of invention: Method and device for particle filter target tracking based on pipeline parallel processing technique Effective date of registration: 20191112 Granted publication date: 20181207 Pledgee: Nanjing Bank Co.,Ltd. Chengnan Branch Pledgor: NANJING CHIEBOT ROBOT TECHNOLOGY CO.,LTD. Registration number: Y2019320000272 |
|
| PC01 | Cancellation of the registration of the contract for pledge of patent right | ||
| PC01 | Cancellation of the registration of the contract for pledge of patent right |
Date of cancellation: 20210309 Granted publication date: 20181207 Pledgee: Nanjing Bank Co.,Ltd. Chengnan Branch Pledgor: NANJING CHIEBOT ROBOT TECHNOLOGY Co.,Ltd. Registration number: Y2019320000272 |