CN107316316A - The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features - Google Patents

The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features Download PDF

Info

Publication number
CN107316316A
CN107316316A CN201710355503.4A CN201710355503A CN107316316A CN 107316316 A CN107316316 A CN 107316316A CN 201710355503 A CN201710355503 A CN 201710355503A CN 107316316 A CN107316316 A CN 107316316A
Authority
CN
China
Prior art keywords
target
correlation
illustrative plates
characteristic spectrum
collection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201710355503.4A
Other languages
Chinese (zh)
Inventor
练智超
刘忠耿
李杨
濮柯佳
肖亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Science and Technology
Original Assignee
Nanjing University of Science and Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing University of Science and Technology filed Critical Nanjing University of Science and Technology
Priority to CN201710355503.4A priority Critical patent/CN107316316A/en
Publication of CN107316316A publication Critical patent/CN107316316A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • G06T7/248—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving reference images or patches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/262—Analysis of motion using transform domain methods, e.g. Fourier domain methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/269—Analysis of motion using gradient-based methods
    • 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
    • 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
    • G06T2207/20004—Adaptive image processing
    • 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]
    • 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
    • G06T2207/20081—Training; Learning

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a kind of method for tracking target adaptively merged based on multiple features with nuclear phase pass filtering technique.Method and step is：The target location tracked according to former frame and yardstick, obtain the candidate region of target motion；The gradient orientation histogram and color characteristic of candidate region are extracted, two kinds of features are merged, and carries out FFT, obtains calculating core cross-correlation after characteristic spectrum；Determine that target, in the position of present frame and yardstick, obtains target area；The gradient orientation histogram and color characteristic of target area are extracted, two kinds of features are merged, and carries out FFT, obtains calculating core auto-correlation after characteristic spectrum；The adaptive target of design is corresponding, training position filtering device and scaling filter model；Characteristic spectrum and correlation filter are updated using linear interpolation method.Invention enhances the discriminating power of model, the robustness of target following of the target in complex scene and cosmetic variation is improved, computation complexity is reduced, improves the real-time of tracking.

Description

The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features
Technical field
It is particularly a kind of adaptively to be merged and core correlation filtering skill based on multiple features the present invention relates to computer vision field The method for tracking target of art.
Background technology
Target following is the important research content in computer vision field, and target tracking is mainly according to target in video In the first frame or former frames position, the track that estimation postorder sequence target occurs.At present, target following technology mainly has two Major class：
(1) production method：This method is mainly the appearance features that target is described with generation model, in postorder sequence Find most like with target appearance, that is to say, that minimize reconstructed error by searching for candidate target.Compare representative Algorithm have sparse coding, online density estimation and principal component analysis (PCA) etc..Production method is conceived to target appearance Portray, ignore background information, therefore drift is easily produced when target appearance changes acutely or is blocked, track failure.
(2) discriminate method：This method mainly trains two graders, Ran Hou with online machine learning techniques Target detection is carried out with the grader in postorder sequence, target following is completed.In recent years, various machine learning algorithms are employed In discriminate method, wherein more representational have many case-based learning methods, boosting and structure SVM etc..Discriminate Method is because significantly distinguish the information of background and prospect, and discriminating power is strong, expressively more robust, gradually in target tracking domain Occupy dominant position.It is noted that major part deep learning method for tracking target also belongs to discriminate framework at present.
But, traditional discriminate method has an important defect, i.e., in order to strengthen discriminating power, generally require a large amount of Training sample, while also having aggravated computation burden so that these discriminate methods are struggled in the real-time of tracking.
The content of the invention
Present invention aims at providing, a kind of discriminating power is strong, the target following robustness in complex scene and cosmetic variation High adaptively merges the method for tracking target that filtering technique is closed with nuclear phase based on multiple features, is calculated with being reduced in being handled in frequency domain Complexity, improves the real-time of target following.
The technical solution for realizing the object of the invention is：One kind is adaptively merged and core correlation filtering skill based on multiple features The method for tracking target of art, including following steps：
Step 1, t two field pictures are inputted, step 6 are entered if t=1, otherwise into next step；
Step 2, the target location p tracked according to t-1 framest-1With yardstick st-1, obtain the candidate region z of target motiont；
Step 3, candidate region z is extractedtGradient orientation histogram and color characteristic, two kinds of features are merged, so After carry out FFT, obtain characteristic spectrumWherein ^ represents DFT；
Step 4, according to the characteristic spectrum of target former frameCalculate core cross-correlation
Step 5, the corresponding position of maximum in the corresponding collection of illustrative plates of output of difference test position wave filter and scaling filter, Determine position p of the target in present frametWith yardstick st；
Step 6, according to the target location p of t framestWith yardstick st, obtain target area x；
Step 7, target area x is extractedtGradient orientation histogram and color characteristic, two kinds of features are merged, so After carry out FFT, obtain characteristic spectrum
Step 8, according to characteristic spectrumCalculate core auto-correlation
Step 9, adaptive target response collection of illustrative plates is designedTrain position filtering device and scaling filter model；
Step 10, step 11 is entered if t=1, otherwise into step 12；
Step 11, characteristic spectrum is updated using linear interpolation methodAnd correlation filterAnd enter step 12；
Step 12, target following result is exported, makes t=t+1 be then back to the tracking that step 1 carries out next two field picture.
Further, step 3 and being merged two kinds of features described in step 7, specific as follows：
(3.1) according to size 4M*4N image-region I, gradient orientation histogram, the unit of use are extracted in 9 directions Size 4*4, then after principal component analysis dimensionality reduction, obtains the size M*N of 31 dimensions characteristic spectrum；
(3.2) size 4M*4N image-region I is zoomed into M*N, extracts the color characteristic of 11 dimensions；
(3.3) Fusion Features for extracting (3.1) and (3.2), obtain the size M*N of 42 dimensions characteristic spectrum.
Further, the calculating core auto-correlation described in the calculating core cross-correlation and step 8 described in step 4, specific as follows：
(4.1) Gaussian kernel is used, formula is as follows：
Wherein, k (x, x ') is expressed as the Gaussian kernel of two characteristic spectrum x and x ' calculating, and exp () is expressed as e index letter Number, σ is the standard deviation of Gaussian function, and value is 0.5, | | | |2It is expressed as 2 normal forms of vector or matrix；
(4.2) calculate nuclear phase to close, formula is as follows：
Wherein, kxx′Represent that characteristic spectrum x and x ' nuclear phase is closed, exp () is e index function, σ is the standard of Gaussian function Difference, value is 0.5, | | | |2For vector or matrix 2 normal forms,For the inverse transformation of DFT, * is multiple Conjugation, ^ is DFT,It is multiplied for two matrix corresponding elements.
Further, in the corresponding collection of illustrative plates of the output of difference test position wave filter and scaling filter described in step 5 most It is worth corresponding position greatly, determines position p of the target in present frametWith yardstick st, it is specific as follows：
(5.1) from t two field pictures, with position pt-1With yardstick st-1Extract the candidate region z of location estimationt,trans；
(5.2) candidate region z is extractedt,transCharacteristic spectrum
(5.3) position filtering device correlation output response collection of illustrative plates f is calculated using formula belowt,trans：
Wherein, ftThe corresponding collection of illustrative plates of output of position filtering device is expressed as,It is characterized collection of illustrative platesWithCore cross-correlation,Position filtering device obtain and updated is trained for former frame,For the inverse transformation of DFT, ^ is DFT, ⊙ is that two matrix corresponding elements are multiplied；
The target location p that (5.4) t frames are detectedtTo export corresponding collection of illustrative plates ft,transThe corresponding position of maximum；
(5.5) from t two field pictures, with position ptWith yardstick st-1Extract the candidate region z of size estimationt,sacle, build Yardstick pyramid；
(5.6) scaling filter correlation output response collection of illustrative plates f is calculatedt,sacle；
The target scale s that (5.7) t frames are detectedtTo export corresponding collection of illustrative plates ft,sacleThe corresponding yardstick of maximum.
Further, adaptive target response collection of illustrative plates is designed described in step 9Train position filtering device and yardstick filter Ripple device model, it is specific as follows：
(9.1) in t two field pictures, from away from m position of being sampled in the setting range of previous frame target location；
(9.2) the correlation filtering response collection of illustrative plates of m position is calculated, the maximum of each collection of illustrative plates is taken；
(9.3) target response collection of illustrative plates is filled using maximumCorresponding m position, remaining position is filled with Gauss interpolation Put；
(9.4) training pattern formula is as follows：
Wherein,The correlation filter model tried to achieve is represented,It is characterized collection of illustrative platesCore auto-correlation, ^ be discrete Fourier Leaf transformation,It is multiplied for two matrix corresponding elements, ξ and λ are regularization parameter, and value is respectively 0.01 and 0.001.
Further, the use linear interpolation method described in step 11 updates characteristic spectrumAnd correlation filterIt is public Formula is as follows：
Wherein,WithThe respectively characteristic spectrum and correlation filter of former frame, η is learning rate, and value is 0.02。
Compared with prior art, its remarkable advantage is the present invention：(1) gradient orientation histogram and color characteristic are combined, Wherein gradient orientation histogram feature reflects the structural information of target, and color characteristic focuses on the appearance information of target, and two Complementary characteristic fusion is planted, the discriminating power of model is effectively enhanced, improves the stability of tracking；(2) adaptive chi is used Method of estimation is spent, this method realizes that quickly, size estimation is accurate, can be incorporated into any discriminate track algorithm framework；(3) Using adaptive target response designing technique, it combines the appearance information and movable information of target, devises one more Real target response so that the correlation filter of training effectively prevent detection mistake.
Brief description of the drawings
Fig. 1 closes the flow of the method for tracking target of filtering technique for nuclear phase of the present invention based on adaptive multiple features fusion Figure.
Fig. 2 gradient orientation histograms merge schematic diagram with color characteristic.
Fig. 3 is adaptive scale method of estimation schematic diagram.
Fig. 4 is that adaptive targets respond designing technique schematic diagram.
Fig. 5 evaluation result figures on standard vision track file for the present invention, wherein (a) is the standard of OTB50 data sets Exactness is drawn, and (b) is that the accuracy of OTB50 data sets is drawn, and (c) is that the degree of accuracy of OTB100 data sets is drawn, and (d) is The accuracy of OTB100 data sets is drawn.
Fig. 6 is actual video target following result figure of the present invention, wherein (a) is that Human tests are regarded on OTB100 data sets Frequency result figure, (b) is CarScale test videos result figure on OTB100 data sets, and (c) is Jogging on OTB50 data sets Test video result figure, (d) is Jogging test videos result figure on OTB50 data sets.
Embodiment
Below in conjunction with the accompanying drawings and specific embodiment is described in further detail to the present invention.The present invention is adaptive based on multiple features The method for tracking target that filtering technique is closed with nuclear phase should be merged, this method is broadly divided into four big steps, and the first step carries for various features Take fusion；Second step target detection, including location estimation and size estimation；Target location and chi of 3rd step according to current detection Degree, training pattern；4th step, using simple linear interpolation method more new model.With reference to Fig. 1, comprise the following steps that：
Step 1, t two field pictures are inputted, step 6 are entered if t=1, otherwise into next step；
Step 2, the target location p tracked according to t-1 framest-1With yardstick st-1, obtain the candidate region z of target motiont；
Step 3, candidate region z is extractedtGradient orientation histogram and color characteristic, two kinds of features are merged, so After carry out FFT, obtain characteristic spectrumWherein ^ represents DFT；
Step 4, according to the characteristic spectrum of target former frameCalculate core cross-correlation
Step 5, the corresponding position of maximum in the corresponding collection of illustrative plates of output of difference test position wave filter and scaling filter, Determine position p of the target in present frametWith yardstick st；
Step 6, according to the target location p of t framestWith yardstick st, obtain target area x；
Step 7, target area x is extractedtGradient orientation histogram and color characteristic, two kinds of features are merged, so After carry out FFT, obtain characteristic spectrum
Step 8, according to characteristic spectrumCalculate core auto-correlation
Step 9, adaptive target response collection of illustrative plates is designedTrain position filtering device and scaling filter model；
Step 10, step 11 is entered if t=1, otherwise into step 12；
Step 11, characteristic spectrum is updated using linear interpolation methodAnd correlation filterAnd enter step 12；
Step 12, target following result is exported, makes t=t+1 be then back to the tracking that step 1 carries out next two field picture.
As shown in Fig. 2 giving the method for tracking target that the nuclear phase based on adaptive multiple features fusion closes filtering technique Multiple features fusion mechanism schematic diagram.Nuclear phase of the present invention based on adaptive multiple features fusion close the target of filtering technique with Track method, it is characterised in that various features fusion method, in recent years, gradient orientation histogram are led in Object Detecting and Tracking Domain is done well, and it mainly reflects the structural information of target.Color characteristic is in image retrieval, target detection, the neck such as target identification Domain effect is protruded, and is combined the discriminating power for enhancing this method with histogram of gradients, described in step 3 and step 7 by two kinds of spies Levy and merged, it is specific as follows：
(3.1) according to size 4M*4N image-region I, gradient orientation histogram, the unit of use are extracted in 9 directions Size 4*4, then after principal component analysis dimensionality reduction, obtains the size M*N of 31 dimensions characteristic spectrum；
(3.2) size 4M*4N image-region I is zoomed into M*N, extracts the color characteristic of 11 dimensions；
(3.3) Fusion Features for extracting (3.1) and (3.2), obtain the size M*N of 42 dimensions characteristic spectrum.
The calculating core auto-correlation described in calculating core cross-correlation and step 8 described in step 4, it is specific as follows：
(4.1) Gaussian kernel is used, formula is as follows：
Wherein, k (x, x ') is expressed as the Gaussian kernel of two characteristic spectrum x and x ' calculating, and exp () is expressed as e index letter Number, σ is the standard deviation of Gaussian function, and value is 0.5, | | | |2It is expressed as 2 normal forms of vector or matrix；
(4.2) calculate nuclear phase to close, formula is as follows：
Wherein, kxx′Represent that characteristic spectrum x and x ' nuclear phase is closed, exp () is e index function, σ is the standard of Gaussian function Difference, value is 0.5, | | | |2For vector or matrix 2 normal forms,For the inverse transformation of DFT, * is multiple Conjugation, ^ is DFT, and ⊙ is that two matrix corresponding elements are multiplied.
As shown in figure 3, giving the method for tracking target that the nuclear phase based on adaptive multiple features fusion closes filtering technique Adaptive scale method of estimation schematic diagram.Nuclear phase of the present invention based on adaptive multiple features fusion closes the mesh of filtering technique Mark tracking, it is characterised in that adaptive scale method of estimation.The moulded dimension that traditional nuclear phase closes filtering method target is consolidated It is fixed, it is impossible to handle the change of target scale, therefore it is easily caused tracking failure.The present invention proposes a kind of adaptive scale estimation Method, specifically trains an independent scaling filter, by scaling filter relevant response it is maximum when corresponding yardstick come Estimation, this method application FFT is simple and efficient, is desirably integrated into traditional discriminate method for tracking target In, the corresponding position of maximum in the corresponding collection of illustrative plates of the output of difference test position wave filter and scaling filter described in step 5, Determine position p of the target in present frametWith yardstick st, it is specific as follows：
(5.1) from t two field pictures, with position pt-1With yardstick st-1Extract the candidate region z of location estimationt,trans；
(5.2) candidate region z is extractedt,transCharacteristic spectrum
(5.3) position filtering device correlation output response collection of illustrative plates f is calculated using formula belowt,trans：
Wherein, ftThe corresponding collection of illustrative plates of output of position filtering device is expressed as,It is characterized collection of illustrative platesWithCore cross-correlation,Position filtering device obtain and updated is trained for former frame,For the inverse transformation of DFT, ^ For DFT,It is multiplied for two matrix corresponding elements；
The target location p that (5.4) t frames are detectedtTo export corresponding collection of illustrative plates ft,transThe corresponding position of maximum；
(5.5) from t two field pictures, with position ptWith yardstick st-1Extract the candidate region z of size estimationt,sacle, build Yardstick pyramid；
(5.6) scaling filter correlation output response collection of illustrative plates f is calculatedt,sacle；
The target scale s that (5.7) t frames are detectedtTo export corresponding collection of illustrative plates ft,sacleThe corresponding yardstick of maximum.
As shown in figure 4, giving the method for tracking target that the nuclear phase based on adaptive multiple features fusion closes filtering technique Adaptive targets respond design method schematic diagram.Nuclear phase of the present invention based on adaptive multiple features fusion closes filtering technique Method for tracking target, it is characterised in that adaptive target response design method.Traditional nuclear phase closes filtering method target and rung Should be changeless, target response is produced centered on the first frame target location by Gaussian function, so once detecting Stage is made a mistake, due to the update mechanism of model, and mistake can be propagated down always, until tracking failure.The present invention is understanding Certainly the problem, employs a kind of adaptive target response design method so that target response is change, knot in each frame The appearance information and movable information of target have been closed, the robustness of tracking is improved, the adaptive mesh of the design described in step 9 Mark response collection of illustrative platesPosition filtering device and scaling filter model are trained, it is specific as follows：
(9.1) in t two field pictures, from away from m position of being sampled in the setting range of previous frame target location；
(9.2) the correlation filtering response collection of illustrative plates of m position is calculated, the maximum of each collection of illustrative plates is taken；
(9.3) target response collection of illustrative plates is filled using maximumCorresponding m position, remaining position is filled with Gauss interpolation Put；
(9.4) training pattern formula is as follows：
Wherein,The correlation filter model tried to achieve is represented,Represent characteristic spectrumCore auto-correlation, ^ represents discrete Fourier transform, ⊙ represents two matrix corresponding element multiplications, and ξ and λ are regularization parameters, in order to prevent the model of training from crossing plan Close, ξ and λ values are respectively 0.01 and 0.001 in the present invention.
As shown in figure 5, illustrate the present invention follows the trail of evaluation result figure on data set OTB50 and OTB100 in standard vision, Wherein (a) is the degree of accuracy drawing of OTB50 data sets, and (b) is that the accuracy of OTB50 data sets is drawn, and (c) is OTB100 numbers Drawn according to the degree of accuracy of collection, (d) is that the accuracy of OTB100 data sets is drawn.OTB50 data sets have 50 video sequences, altogether Possess 29000 frames, and OTB100 data sets possess 100 video sequences, and 58897 frames are possessed altogether, they have target per frame Mark.Evaluation metricses mainly have two kinds：The degree of accuracy and success rate.In (a) and (c) is drawn in the degree of accuracy, the degree of accuracy is defined as The distance between algorithm test position and target designation position account for the percentage of total evaluation and test frame number no more than the frame number of 20 pixels； Success rate is drawn in (b) and (d), and Duplication refers to that algorithm detection target bounding box is weighed between the two with target designation bounding box Frame number of the percentage more than 50% that folded area (shipping calculation) accounts for the gross area (union) accounts for the percentage of total evaluation and test frame number Than.From evaluation result as can be seen that the present invention does well in target tracking task.
As shown in fig. 6, illustrating the present invention and some outstanding algorithm target tracking results in actual video in recent years Compare figure, wherein (a) is Human test videos result figure on OTB100 data sets, (b) is CarScale on OTB100 data sets Test video result figure, (c) is Jogging test videos result figure on OTB50 data sets, and (d) is on OTB50 data sets Jogging test video result figures.All in all, the present invention follows the trail of effect preferably, due to using color and gradient direction Nogata Figure fusion feature, size measurement mechanism, the corresponding mechanism of adaptive targets, the present invention can be blocked in target, dimensional variation, It is accurate under the conditions of the unfavorable factor such as target distortion and the quick motion of target to follow the trail of target.

Claims (6)

1. a kind of adaptively merge the method for tracking target that filtering technique is closed with nuclear phase based on multiple features, it is characterised in that including Following steps：
Step 1, t two field pictures are inputted, step 6 are entered if t=1, otherwise into next step；
Step 2, the target location p tracked according to t-1 framest-1With yardstick st-1, obtain the candidate region z of target motiont；
Step 3, candidate region z is extractedtGradient orientation histogram and color characteristic, two kinds of features are merged, then carried out FFT, obtains characteristic spectrumWherein ^ represents DFT；
Step 4, according to the characteristic spectrum of target former frameCalculate core cross-correlation
Step 5, the corresponding position of maximum in the corresponding collection of illustrative plates of output of difference test position wave filter and scaling filter, it is determined that Position p of the target in present frametWith yardstick st；
Step 6, according to the target location p of t framestWith yardstick st, obtain target area x；
Step 7, target area x is extractedtGradient orientation histogram and color characteristic, two kinds of features are merged, then carried out FFT, obtains characteristic spectrum
Step 8, according to characteristic spectrumCalculate core auto-correlation
Step 9, adaptive target response collection of illustrative plates is designedTrain position filtering device and scaling filter model；
Step 10, step 11 is entered if t=1, otherwise into step 12；
Step 11, characteristic spectrum is updated using linear interpolation methodAnd correlation filterAnd enter step 12；
Step 12, target following result is exported, makes t=t+1 be then back to the tracking that step 1 carries out next two field picture.
2. according to claim 1 adaptively merge the method for tracking target that filtering technique is closed with nuclear phase based on multiple features, Characterized in that, step 3 and being merged two kinds of features described in step 7, specific as follows：
(3.1) according to size 4M*4N image-region I, gradient orientation histogram, the unit size of use are extracted in 9 directions 4*4, then after principal component analysis dimensionality reduction, obtains the size M*N of 31 dimensions characteristic spectrum；
(3.2) size 4M*4N image-region I is zoomed into M*N, extracts the color characteristic of 11 dimensions；
(3.3) Fusion Features for extracting (3.1) and (3.2), obtain the size M*N of 42 dimensions characteristic spectrum.
3. according to claim 1 adaptively merge the method for tracking target that filtering technique is closed with nuclear phase based on multiple features, Characterized in that, calculate core cross-correlation and the calculating core auto-correlation described in step 8 described in step 4, it is specific as follows：
(4.1) Gaussian kernel is used, formula is as follows：
Wherein, k (x, x ') is expressed as the Gaussian kernel of two characteristic spectrum x and x ' calculating, and exp () is expressed as e index function, and σ is The standard deviation of Gaussian function, value is 0.5, | | | |2It is expressed as 2 normal forms of vector or matrix；
(4.2) calculate nuclear phase to close, formula is as follows：
Wherein, kxx′Represent that characteristic spectrum x and x ' nuclear phase is closed, exp () is e index function, σ is the standard deviation of Gaussian function, Value is 0.5, | | | |2For vector or matrix 2 normal forms,For the inverse transformation of DFT, * is complex conjugate, ^ is DFT,It is multiplied for two matrix corresponding elements.
4. according to claim 1 adaptively merge the method for tracking target that filtering technique is closed with nuclear phase based on multiple features, Characterized in that, maximum pair in the corresponding collection of illustrative plates of the output of difference test position wave filter and scaling filter described in step 5 The position answered, determines position p of the target in present frametWith yardstick st, it is specific as follows：
(5.1) from t two field pictures, with position pt-1With yardstick st-1Extract the candidate region z of location estimationt,trans；
(5.2) candidate region z is extractedt,transCharacteristic spectrum
(5.3) position filtering device correlation output response collection of illustrative plates f is calculated using formula belowt,trans：
Wherein, ftThe corresponding collection of illustrative plates of output of position filtering device is expressed as,It is characterized collection of illustrative platesWithCore cross-correlation,For Former frame trains obtain and updated position filtering device,For the inverse transformation of DFT, ^ is discrete Fourier transform,It is multiplied for two matrix corresponding elements；
The target location p that (5.4) t frames are detectedtTo export corresponding collection of illustrative plates ft,transThe corresponding position of maximum；
(5.5) from t two field pictures, with position ptWith yardstick st-1Extract the candidate region z of size estimationt,sacle, build yardstick Pyramid；
(5.6) scaling filter correlation output response collection of illustrative plates f is calculatedt,sacle；
The target scale s that (5.7) t frames are detectedtTo export corresponding collection of illustrative plates ft,sacleThe corresponding yardstick of maximum.
5. according to claim 1 adaptively merge the method for tracking target that filtering technique is closed with nuclear phase based on multiple features, Characterized in that, designing adaptive target response collection of illustrative plates described in step 9Train position filtering device and scaling filter mould Type, it is specific as follows：
(9.1) in t two field pictures, from away from m position of being sampled in the setting range of previous frame target location；
(9.2) the correlation filtering response collection of illustrative plates of m position is calculated, the maximum of each collection of illustrative plates is taken；
(9.3) target response collection of illustrative plates is filled using maximumCorresponding m position, remaining position is filled with Gauss interpolation；
(9.4) training pattern formula is as follows：
Wherein,The correlation filter model tried to achieve is represented,It is characterized collection of illustrative platesCore auto-correlation, ^ be discrete Fourier become Change,It is multiplied for two matrix corresponding elements, ξ and λ are regularization parameter, and value is respectively 0.01 and 0.001.
6. according to claim 1 adaptively merge the method for tracking target that filtering technique is closed with nuclear phase based on multiple features, Characterized in that, the use linear interpolation method described in step 11 updates characteristic spectrumAnd correlation filterFormula is as follows：
Wherein,WithThe respectively characteristic spectrum and correlation filter of former frame, η is learning rate, and value is 0.02.
CN201710355503.4A 2017-05-19 2017-05-19 The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features Pending CN107316316A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710355503.4A CN107316316A (en) 2017-05-19 2017-05-19 The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710355503.4A CN107316316A (en) 2017-05-19 2017-05-19 The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features

Publications (1)

Publication Number Publication Date
CN107316316A true CN107316316A (en) 2017-11-03

Family

ID=60181486

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710355503.4A Pending CN107316316A (en) 2017-05-19 2017-05-19 The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features

Country Status (1)

Country Link
CN (1) CN107316316A (en)

Cited By (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108053425A (en) * 2017-12-25 2018-05-18 北京航空航天大学 A kind of high speed correlation filtering method for tracking target based on multi-channel feature
CN108257153A (en) * 2017-12-29 2018-07-06 中国电子科技集团公司第二十七研究所 A kind of method for tracking target based on direction gradient statistical nature
CN108288062A (en) * 2017-12-29 2018-07-17 中国电子科技集团公司第二十七研究所 A kind of method for tracking target based on core correlation filtering
CN108364305A (en) * 2018-02-07 2018-08-03 福州大学 Vehicle-mounted pick-up video target tracking method based on modified DSST
CN108647694A (en) * 2018-04-24 2018-10-12 武汉大学 Correlation filtering method for tracking target based on context-aware and automated response
CN108876818A (en) * 2018-06-05 2018-11-23 国网辽宁省电力有限公司信息通信分公司 A kind of method for tracking target based on like physical property and correlation filtering
CN109035290A (en) * 2018-07-16 2018-12-18 南京信息工程大学 A kind of track algorithm updating accretion learning based on high confidence level
CN109035302A (en) * 2018-07-26 2018-12-18 中国人民解放军陆军工程大学 Target tracking algorism based on the perceptually relevant filtering of space-time
CN109034193A (en) * 2018-06-20 2018-12-18 上海理工大学 Multiple features fusion and dimension self-adaption nuclear phase close filter tracking method
CN109064497A (en) * 2018-07-16 2018-12-21 南京信息工程大学 A kind of video tracing method based on color cluster accretion learning
CN109285179A (en) * 2018-07-26 2019-01-29 昆明理工大学 A kind of motion target tracking method based on multi-feature fusion
CN109410246A (en) * 2018-09-25 2019-03-01 深圳市中科视讯智能系统技术有限公司 The method and device of vision tracking based on correlation filtering
CN109461172A (en) * 2018-10-25 2019-03-12 南京理工大学 Manually with the united correlation filtering video adaptive tracking method of depth characteristic
CN109670410A (en) * 2018-11-29 2019-04-23 昆明理工大学 A kind of fusion based on multiple features it is long when motion target tracking method
CN109858415A (en) * 2019-01-21 2019-06-07 东南大学 The nuclear phase followed suitable for mobile robot pedestrian closes filtered target tracking
CN109886996A (en) * 2019-01-15 2019-06-14 东华大学 A kind of visual pursuit optimization method
CN109949342A (en) * 2019-03-15 2019-06-28 中国科学院福建物质结构研究所 The complementary study method for real time tracking of adaptive fusion based on destination probability model
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
CN110211149A (en) * 2018-12-25 2019-09-06 湖州云通科技有限公司 A kind of dimension self-adaption nuclear phase pass filter tracking method based on context-aware
CN110472607A (en) * 2019-08-21 2019-11-19 上海海事大学 A kind of ship tracking method and system
CN110751670A (en) * 2018-07-23 2020-02-04 中国科学院长春光学精密机械与物理研究所 Target tracking method based on fusion
CN108846851B (en) * 2018-04-25 2020-07-28 河北工业职业技术学院 Moving target tracking method and terminal equipment
CN112598711A (en) * 2020-12-25 2021-04-02 南京信息工程大学滨江学院 Hyperspectral target tracking method based on joint spectrum dimensionality reduction and feature fusion
CN113298851A (en) * 2021-07-07 2021-08-24 沈阳航空航天大学 Target image tracking method based on multi-scale and multi-feature
CN113327273A (en) * 2021-06-15 2021-08-31 中国人民解放军火箭军工程大学 Infrared target tracking method based on variable window function correlation filtering

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015163830A1 (en) * 2014-04-22 2015-10-29 Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi Target localization and size estimation via multiple model learning in visual tracking
CN106570486A (en) * 2016-11-09 2017-04-19 华南理工大学 Kernel correlation filtering target tracking method based on feature fusion and Bayesian classification
CN106651913A (en) * 2016-11-29 2017-05-10 开易（北京）科技有限公司 Target tracking method based on correlation filtering and color histogram statistics and ADAS (Advanced Driving Assistance System)

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015163830A1 (en) * 2014-04-22 2015-10-29 Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi Target localization and size estimation via multiple model learning in visual tracking
CN106570486A (en) * 2016-11-09 2017-04-19 华南理工大学 Kernel correlation filtering target tracking method based on feature fusion and Bayesian classification
CN106651913A (en) * 2016-11-29 2017-05-10 开易（北京）科技有限公司 Target tracking method based on correlation filtering and color histogram statistics and ADAS (Advanced Driving Assistance System)

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
张雷: ""复杂场景下实时目标跟踪算法及实现技术研究"", 《中国博士学位论文全文数据库 信息科技辑》 *

Cited By (38)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108053425A (en) * 2017-12-25 2018-05-18 北京航空航天大学 A kind of high speed correlation filtering method for tracking target based on multi-channel feature
CN108257153A (en) * 2017-12-29 2018-07-06 中国电子科技集团公司第二十七研究所 A kind of method for tracking target based on direction gradient statistical nature
CN108288062A (en) * 2017-12-29 2018-07-17 中国电子科技集团公司第二十七研究所 A kind of method for tracking target based on core correlation filtering
CN108257153B (en) * 2017-12-29 2021-09-07 中国电子科技集团公司第二十七研究所 Target tracking method based on direction gradient statistical characteristics
CN108364305A (en) * 2018-02-07 2018-08-03 福州大学 Vehicle-mounted pick-up video target tracking method based on modified DSST
CN108364305B (en) * 2018-02-07 2021-05-18 福州大学 Vehicle-mounted camera video target tracking method based on improved DSST
CN108647694B (en) * 2018-04-24 2021-04-16 武汉大学 Context-aware and adaptive response-based related filtering target tracking method
CN108647694A (en) * 2018-04-24 2018-10-12 武汉大学 Correlation filtering method for tracking target based on context-aware and automated response
CN108846851B (en) * 2018-04-25 2020-07-28 河北工业职业技术学院 Moving target tracking method and terminal equipment
CN108876818A (en) * 2018-06-05 2018-11-23 国网辽宁省电力有限公司信息通信分公司 A kind of method for tracking target based on like physical property and correlation filtering
CN109034193A (en) * 2018-06-20 2018-12-18 上海理工大学 Multiple features fusion and dimension self-adaption nuclear phase close filter tracking method
CN109064497B (en) * 2018-07-16 2021-11-23 南京信息工程大学 Video tracking method based on color clustering supplementary learning
CN109064497A (en) * 2018-07-16 2018-12-21 南京信息工程大学 A kind of video tracing method based on color cluster accretion learning
CN109035290A (en) * 2018-07-16 2018-12-18 南京信息工程大学 A kind of track algorithm updating accretion learning based on high confidence level
CN110751670B (en) * 2018-07-23 2022-10-25 中国科学院长春光学精密机械与物理研究所 Target tracking method based on fusion
CN110751670A (en) * 2018-07-23 2020-02-04 中国科学院长春光学精密机械与物理研究所 Target tracking method based on fusion
CN109285179A (en) * 2018-07-26 2019-01-29 昆明理工大学 A kind of motion target tracking method based on multi-feature fusion
CN109035302A (en) * 2018-07-26 2018-12-18 中国人民解放军陆军工程大学 Target tracking algorism based on the perceptually relevant filtering of space-time
CN109285179B (en) * 2018-07-26 2021-05-14 昆明理工大学 Moving target tracking method based on multi-feature fusion
CN109410246A (en) * 2018-09-25 2019-03-01 深圳市中科视讯智能系统技术有限公司 The method and device of vision tracking based on correlation filtering
CN109410246B (en) * 2018-09-25 2021-06-11 杭州视语智能视觉系统技术有限公司 Visual tracking method and device based on correlation filtering
CN109461172A (en) * 2018-10-25 2019-03-12 南京理工大学 Manually with the united correlation filtering video adaptive tracking method of depth characteristic
CN109670410A (en) * 2018-11-29 2019-04-23 昆明理工大学 A kind of fusion based on multiple features it is long when motion target tracking method
CN110211149A (en) * 2018-12-25 2019-09-06 湖州云通科技有限公司 A kind of dimension self-adaption nuclear phase pass filter tracking method based on context-aware
CN110211149B (en) * 2018-12-25 2022-08-12 湖州云通科技有限公司 Scale self-adaptive kernel correlation filtering tracking method based on background perception
CN109886996A (en) * 2019-01-15 2019-06-14 东华大学 A kind of visual pursuit optimization method
CN109886996B (en) * 2019-01-15 2023-06-06 东华大学 Visual tracking optimization method
CN109858415A (en) * 2019-01-21 2019-06-07 东南大学 The nuclear phase followed suitable for mobile robot pedestrian closes filtered target tracking
CN109949342A (en) * 2019-03-15 2019-06-28 中国科学院福建物质结构研究所 The complementary study method for real time tracking of adaptive fusion based on destination probability model
CN109949342B (en) * 2019-03-15 2022-07-15 中国科学院福建物质结构研究所 Self-adaptive fusion complementary learning real-time tracking method based on target probability model
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
CN110472607A (en) * 2019-08-21 2019-11-19 上海海事大学 A kind of ship tracking method and system
CN112598711A (en) * 2020-12-25 2021-04-02 南京信息工程大学滨江学院 Hyperspectral target tracking method based on joint spectrum dimensionality reduction and feature fusion
CN112598711B (en) * 2020-12-25 2022-12-20 南京信息工程大学滨江学院 Hyperspectral target tracking method based on joint spectrum dimensionality reduction and feature fusion
CN113327273A (en) * 2021-06-15 2021-08-31 中国人民解放军火箭军工程大学 Infrared target tracking method based on variable window function correlation filtering
CN113327273B (en) * 2021-06-15 2023-12-19 中国人民解放军火箭军工程大学 Infrared target tracking method based on variable window function correlation filtering
CN113298851A (en) * 2021-07-07 2021-08-24 沈阳航空航天大学 Target image tracking method based on multi-scale and multi-feature
CN113298851B (en) * 2021-07-07 2023-09-26 沈阳航空航天大学 Target image tracking method based on multi-scale multi-feature

Similar Documents

Publication Publication Date Title
CN107316316A (en) The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features
CN107154024A (en) Dimension self-adaption method for tracking target based on depth characteristic core correlation filter
Li et al. Density map guided object detection in aerial images
CN110009679B (en) Target positioning method based on multi-scale feature convolutional neural network
Uhrig et al. Sparsity invariant cnns
CN106952288B (en) Based on convolution feature and global search detect it is long when block robust tracking method
CN108665481A (en) Multilayer depth characteristic fusion it is adaptive resist block infrared object tracking method
CN107330357A (en) Vision SLAM closed loop detection methods based on deep neural network
CN109461172A (en) Manually with the united correlation filtering video adaptive tracking method of depth characteristic
CN105741316A (en) Robust target tracking method based on deep learning and multi-scale correlation filtering
CN112597985B (en) Crowd counting method based on multi-scale feature fusion
CN106570893A (en) Rapid stable visual tracking method based on correlation filtering
CN108961308B (en) Residual error depth characteristic target tracking method for drift detection
Ji et al. Parallel fully convolutional network for semantic segmentation
CN109993095A (en) A kind of other characteristic aggregation method of frame level towards video object detection
CN110188708A (en) A kind of facial expression recognizing method based on convolutional neural networks
CN110503081A (en) Act of violence detection method, system, equipment and medium based on inter-frame difference
CN110457515A (en) The method for searching three-dimension model of the multi-angle of view neural network of polymerization is captured based on global characteristics
CN102034267A (en) Three-dimensional reconstruction method of target based on attention
CN110826389A (en) Gait recognition method based on attention 3D frequency convolution neural network
Lan et al. Coherence-aware context aggregator for fast video object segmentation
Zhao et al. Self-generated defocus blur detection via dual adversarial discriminators
Xiao et al. MeMu: Metric correlation Siamese network and multi-class negative sampling for visual tracking
CN110348492A (en) A kind of correlation filtering method for tracking target based on contextual information and multiple features fusion
CN109146925A (en) Conspicuousness object detection method under a kind of dynamic scene

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20171103