CN109543615A - A kind of double learning model method for tracking target based on multi-stage characteristics - Google Patents

A kind of double learning model method for tracking target based on multi-stage characteristics Download PDF

Info

Publication number
CN109543615A
CN109543615A CN201811405327.1A CN201811405327A CN109543615A CN 109543615 A CN109543615 A CN 109543615A CN 201811405327 A CN201811405327 A CN 201811405327A CN 109543615 A CN109543615 A CN 109543615A
Authority
CN
China
Prior art keywords
target
layer
response
feature
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.)
Granted
Application number
CN201811405327.1A
Other languages
Chinese (zh)
Other versions
CN109543615B (en
Inventor
张建明
金晓康
李旭东
陆朝铨
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.)
Changsha University of Science and Technology
Original Assignee
Changsha 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 Changsha University of Science and Technology filed Critical Changsha University of Science and Technology
Priority to CN201811405327.1A priority Critical patent/CN109543615B/en
Publication of CN109543615A publication Critical patent/CN109543615A/en
Application granted granted Critical
Publication of CN109543615B publication Critical patent/CN109543615B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
    • G06V20/42—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items of sport video content
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/94—Hardware or software architectures specially adapted for image or video understanding

Abstract

The invention belongs to mode identification technology more particularly to a kind of double learning model method for tracking target based on multi-stage characteristics.This method is specific as follows: A1, reading target video sequence, obtains the target position in t frame image；A2, it is cut out in t+1 frame image and includes the candidate region of target position and carry out feature extraction；A3, training correlation filter model: A4, future position: A5, sorter model re-detection obtain the final position of target in t+1 frame image；The update of A6, dual model, real-time online update correlation filter model and sorter model.This method effectively increases the accuracy rate and success rate of tracking, has well adapted to quick movement, has blocked, the influence of the disturbing factors such as illumination.

Description

A kind of double learning model method for tracking target based on multi-stage characteristics
Technical field
The invention belongs to mode identification technology more particularly to a kind of double learning model targets based on multi-stage characteristics with Track method.
Background technique
Target following is that a basis and important research direction, basic thought are to pass through sequence in computer vision field The video information of column image establishes model, and correlation according to the time and spatially determines the posture and fortune of interested target Dynamic rail mark.Currently, target following technology is all widely used on civilian, military, as video monitoring, human-computer interaction, nobody Drive and the tracking of guided missile intercept etc., but still there is many factors that can not be overcome, as illumination variation, dimensional variation, It blocks, quickly movement, complex background and rotation etc..
In recent years, target following achieves significant progress, many competitive algorithms is proposed, according to apparent mould The difference of type is broadly divided into production and discriminate.Production track algorithm mainly models prospect, passes through minimum weight Structure error searches for candidate region, finds optimal matching position in the current frame, utilizes on-line study new mechanism object module. The problem of tracking problem is converted two classification by discriminate track algorithm, by acquiring one group of positive and negative samples, training in each frame Classifier with discriminating power, to maximumlly distinguish target and background.The performance of discriminate track algorithm relies primarily on In the method quality of feature extraction, the superiority and inferiority of classifier and the viability of online updating classifier mechanism.In order to overcome target Appearance with time change, indicate that target is particularly important using suitable Feature Descriptor, such as color histogram, Haar-like, SURF, HOG, subspace expression, super-pixel etc., even manifold set.
Currently, computer vision field has started the upsurge of deep learning.Depth convolutional neural networks (CNN) are strong due to it Big character representation ability, shows excellent performance, such as image classification, target detection and region of interest in many tasks The detection etc. in domain.Contain multiple convolutional layers in convolutional neural networks, pond layer and softmax layers.Wherein these convolutional layers have There is very strong discriminating power, while remaining the information of space and structuring.Depth convolution feature combination correlation filtering prediction bits Set be target following a hot spot, high-rise convolution feature includes semantic information abundant, and the convolution feature of low layer mentions Higher spatial resolution and edge detail information have been supplied, has been played an important role for accurately positioning target.
In discriminate track algorithm, the method for tracking target based on discriminate correlation filtering (DCF) is increasingly becoming research heat Point achieves good result on the standard data set of target following.It trains a correlation filter to be predicted, obtains The classification score of one target, and efficient meter is realized to all space cycle training samples using discrete Fourier transform It calculates, ensure that the real-time of tracking.Therefore, multiple track algorithms combined based on CNN and DCF are proposed in recent years.These are calculated Method relies on the expression ability powerful by convolution feature, achieves excellent tracking effect, while not needing that the time is spent to exist Line updates depth model, substantially increases the real-time of algorithm.
However, merging depth convolution feature in a DCF frame still remains limitation:
(1) in the Fusion Features stage of multilayer, high-rise convolution feature often has bigger weight, because of high-rise letter Breath has richer semantic information, and compared with the feature of low layer, the effect played is more obvious, therefore this is reasonable.But It is to mislead semantic information, and error to be exaggerated by online updating due to being easy to be interfered by various factors in tracking process Target drift is caused even to lose target.Therefore, the method for layered characteristic fusion does not explore effective pass between feature completely System, while the filter of multilayer is merged using weight, is not effectively utilized the information of the response diagram of multiple filters, is made At information waste.
(2) in the process of movement, in the case where seriously being blocked or quickly being moved, correlation filtering is difficult target Such challenge is adapted to, therefore interference information can be introduced into the continuous update of correlation filter by tracker, cause error Accumulation, while causing tracking drift even failure.In this case, single learning model cannot effectively be competent at tracking , need to intervene new correction mechanism.
Summary of the invention
(1) technical problems to be solved
For existing technical problem, the present invention provides a kind of double learning model target followings based on multi-stage characteristics Method, this method effectively increase the accuracy rate and success rate of tracking, well adapted to quick movement, block, illumination etc. it is dry Disturb the influence of factor.
(2) technical solution
In order to achieve the above object, the main technical schemes that the present invention uses include:
A kind of double learning model method for tracking target based on multi-stage characteristics, for target video sequence, given first In frame in the case where target original state, the state of target is estimated in video sequence below, is specifically comprised the following steps:
A1, target video sequence is read, obtains the target position in t frame image, t=1；
A2, cut out in t+1 frame image include the target position candidate region, the candidate region is the 2.5 times of regions in t frame image centered on target position；
The low-level features for extracting the depth characteristic of convolutional layer and the candidate region in the candidate region, will be each described The characteristic pattern of the output of depth characteristic and the output of low-level features as the multichannel convolutive neural network of target in convolutional layerM, N is respectively the width and height of characteristic pattern, and D is channel number；
A3, training correlation filter model:
The candidate region is divided into several cell fritters, establishes Gaussian function label Y for every cell fritter:
Pass through the Gaussian function label Y and characteristic pattern Xl, construct each layer, Mei Yite in multichannel convolutive neural network Levy the correlation filter in channel；
The characteristic pattern of each layer, each feature channel is obtained according to the correlation filter；
The relevant response figure of each layer of characteristic pattern and each is obtained according to the characteristic pattern of each layer, each feature channel Position in the relevant response figure of layer characteristic pattern where maximum response；
A4, future position:
The maximum response of the maximum response of current layer and preceding layer is weighted to obtain the position of preceding layer maximum response It sets, by iteration, obtains final response diagram；
Maximum response is found in final response diagram, using the position of maximum response as the predicted position of target；
A5, sorter model re-detection obtain the HPSR of t+1 frame image according to the relevant response figure of each layer of characteristic pattern Index；
When HPSR index is greater than or equal to given threshold θ, final position of the predicted position of the target as target；
When HPSR index is less than given threshold θ, sorter model is enabled, the result and target that sorter model is obtained Predicted position combine, obtain the final position of t+1 frame image object；
The update of A6, dual model, real-time online update correlation filter model and sorter model, are used for t+2 frame figure The target position of picture determines.
Further, the depth characteristic in the step A2 includes advanced features and mid-level features；
The characteristic pattern for extracting convolutional layer Conv3-4, convolutional layer Conv4-4, convolutional layer Conv5-4, as advanced features；
The characteristic pattern for extracting convolutional layer Conv1-2, convolutional layer Conv2-2, as mid-level features.
Further, the low-level features in the step A2 include HOG feature, Gray feature and Color Name feature, Three kinds of above-mentioned features are linked togather, as low-level features.
Further, before extracting depth characteristic and low-level features, the image of the candidate region is subjected to single precision Processing and resampling, calculate mean value, then go mean value to image after normalization；
Feature extraction is carried out to the candidate region using the network model of VGGNet-19 pre-training, extracts different convolution The depth characteristic of layer.
Further, the step A3 includes:
A31, the candidate region is divided into several cell fritters, establishes Gaussian function label Y for every cell fritter:
Wherein, Y (m, n) indicates the label at (m, n), m ∈ { 0,1 ..., M-1 }, n ∈ { 0,1 ..., N-1 }；σ is Parameter,
Wherein, w and h respectively indicates the width and height of target position in t frame, and σ ' indicates that the output factor, value are 0.1, cell_size indicates the side length of cell fritter in t+1 frame image, value 4；
A32, the correlation filter W for constructing each layer in multichannel convolutive neural network, each feature channell d:
Wherein, λ is the regularization parameter of correlation filter,WithRespectively XlWith the discrete Fourier transform of Y, For XlComplex conjugate；
A33, according to the correlation filter Wl dObtain the characteristic pattern of each layer, each feature channel Obtain l layers of relevant response figure ElWith relevant response figure ElIn maximum response:
Wherein,It is operated for inverse Fourier transform,ForDiscrete Fourier transform, relevant response figure ElIt is big Small is M × N；
The response diagram of whole multi-stage characteristics is denoted as set { E1,E2...,El}。
Further, the step A4 includes:
A41, the maximum response of the maximum response of current layer and preceding layer is weighted to obtain preceding layer maximum response Position final response diagram is obtained by iteration；
El-1(m, n)=αl-1El-1(m,n)+αlEl(m,n) (5)
Wherein, ElThere is maximum response in the response diagram position where (m, n) in the characteristic pattern of l grades of (m, n) expression Value, αlFor l grades of weight；
Obtain final response diagram E；
A42, maximum response is found by final response diagram E, position the center p of current tracking targett=(xt, yt)；
(xt,yt)=argm,nmaxE(m,n) (6)。
Further,
Wherein, max (El) it is ElIn maximum value, μlAnd σlL layers of the response diagram respectively in t+1 frame image Average value and standard deviation, βiFor l layers of relevant response figure ElWeight coefficient.
Further, in the step A5, sorter model is enabled, in the target p that step 42 obtainst=(xt,yt) around Acquire candidate sample set xSVM, svm classifier is completed using updating:
Wherein, scores is the classification score of classifier,It is respectively the parameter and biasing of classifier with b；
The position for the maximum scores being calculated is determined as the position that sorter model obtains, then with correlation filtering The result p of device modelt=(xt,yt) it is weighted synthesis, obtain final result.
Further, the step A6 includes:
A61, correlation filter model is updated, updates the molecule of t+1 frameAnd denominatorTo update t+1 frame Correlation filter Wl d,
Wherein,η is Learning rate；
A62, sorter model is updated, positive and negative sample is collected with the window of target sizes near the target position of t+1 frame This, carries out online SVM training after extracting feature；
When given training dataset is G={ (xSVM,j,ySVM,j), j=1 ... r }, r is the number of sample, xSVM,jFor Training sample, ySVM,jFor sample label, if the encirclement frame area of sample is m, it is n, their weight that the target of t+1 frame, which surrounds frame, Folded rate s=(m ∩ n)/(m ∪ n), is positive sample as s>0.5, and when s<0.1 is negative sample, trained objective function are as follows:
Determination after the completion of update, for t+2 frame target position.
(3) beneficial effect
The beneficial effects of the present invention are:
1, double learning model method for tracking target provided by the invention based on multi-stage characteristics, effectively increase the standard of tracking True rate and success rate have well adapted to quick movement, have blocked, the influence of the disturbing factors such as illumination.
2, the invention proposes the correlation filtering model of multi-stage characteristics fusion, target is obtained by way of recursion layer by layer Position.
3, the present invention obtains different response diagrams using the correlation filter of different stage, and each response diagram is calculated PSR, and be weighted, obtain the evaluation index to current location confidence level.
4, the invention proposes online updating sorter models, and by acquiring positive and negative samples around target, training has The classifier of discriminating power online detects target again, obtains the knot of comprehensive correlation filter model behind new position Fruit obtains final target position.
Detailed description of the invention
Fig. 1 is main algorithm flow chart of the invention；
Fig. 2 is the visualization schematic diagram of multi-stage characteristics in the present invention；
Fig. 3 is fluctuation situation of the HPSR on woman sequence image in the present invention；
Fig. 4 is the target following schematic diagram in the embodiment of the present invention.
Specific embodiment
In order to preferably explain the present invention, in order to understand, with reference to the accompanying drawing, by specific embodiment, to this hair It is bright to be described in detail.
A kind of double learning model method for tracking target based on multi-stage characteristics, for target video sequence, given first In frame in the case where target original state, the state of target is estimated in video sequence below, is specifically comprised the following steps:
A1, target video sequence is read, obtains the target position in t frame image, t=1；
A2, cut out in t+1 frame image include the target position candidate region, the candidate region is the 2.5 times of regions in t frame image centered on target position.The image of the candidate region is subjected to single precision processing and is adopted again Sample calculates mean value, then goes mean value to image after normalization.
The low-level features for extracting the depth characteristic of convolutional layer and the candidate region in the candidate region, will be each described The characteristic pattern of the output of depth characteristic and the output of low-level features as the multichannel convolutive neural network of target in convolutional layerM, N is respectively the width and height of characteristic pattern, and D is channel number.
Wherein, feature extraction is carried out to the candidate region using the network model of VGGNet-19 pre-training, extracted different The depth characteristic of convolutional layer.The characteristic pattern for extracting convolutional layer Conv3-4, convolutional layer Conv4-4, convolutional layer Conv5-4, as more Advanced features in grade feature；The characteristic pattern for extracting convolutional layer Conv1-2, convolutional layer Conv2-2, in multi-stage characteristics Grade feature.
Feature extraction, the feature are carried out again to the candidate region of target using preset multiple manual feature operators For HOG feature, Gray feature and Color Name feature, three kinds of above-mentioned features are linked togather, as in multi-stage characteristics Low-level features.
A3, training correlation filter model:
A31, the candidate region is divided into several cell fritters, establishes Gaussian function label Y for every cell fritter:
Wherein, Y (m, n) indicates the label at (m, n), m ∈ { 0,1 ..., M-1 }, n ∈ { 0,1 ..., N-1 }；σ is Parameter,
Wherein, w and h respectively indicates the width and height of target position in t frame, and σ ' indicates that the output factor, value are 0.1, cell_size indicates the side length of cell fritter in t+1 frame image, value 4；
A32, the correlation filter W for constructing each layer in multichannel convolutive neural network, each feature channell d:
Wherein, λ is the regularization parameter of correlation filter,WithRespectively XlWith the discrete Fourier transform of Y, For XlComplex conjugate；
A33, according to the correlation filter Wl dObtain the characteristic pattern of each layer, each feature channel Obtain l layers of relevant response figure ElWith relevant response figure ElIn maximum response:
Wherein,It is operated for inverse Fourier transform,ForDiscrete Fourier transform, relevant response figure ElIt is big Small is M × N；
The response diagram of whole multi-stage characteristics is denoted as set { E1,E2...,El}。
A4, future position.
A41, the maximum response of the maximum response of current layer and preceding layer is weighted to obtain preceding layer maximum response Position final response diagram is obtained by iteration；
El-1(m, n)=αl-1El-1(m,n)+αlEl(m,n) (5)
Wherein, ElThere is maximum response in the response diagram position where (m, n) in the characteristic pattern of l grades of (m, n) expression Value, αlFor l grades of weight；
Obtain final response diagram E；
A42, maximum response is found by final response diagram E, position the center p of current tracking targett=(xt, yt), the predicted position as target:
(xt,yt)=argm,nmaxE(m,n) (6)。
A5, sorter model re-detection obtain the HPSR of t+1 frame image according to the relevant response figure of each layer of characteristic pattern Index:
Wherein, max (El) it is ElIn maximum value, μlAnd σlL grades of response diagram is flat respectively in t frame image Mean value and standard deviation, βiFor l grades of related corresponding figure ElWeight coefficient.
When HPSR index is greater than or equal to given threshold θ, final position of the predicted position of the target as target；
When HPSR index is less than given threshold θ, sorter model is enabled, the result and target that sorter model is obtained Predicted position combine, obtain the final position of t+1 frame image object.
In the target p that step A42 is obtainedt=(xt,yt) the candidate sample set x of surrounding acquisitionSVM, SVM is completed using updating Classification:
Wherein, scores is the classification score of classifier,It is respectively the parameter and biasing of classifier with b.
The position for the maximum scores being calculated is determined as the position that sorter model obtains, then with correlation filtering The result p of device modelt=(xt,yt) it is weighted synthesis, obtain final result.
The update of A6, dual model, real-time online update correlation filter model and sorter model, are used for t+2 frame figure The target position of picture determines.
A61, correlation filter model is updated, updates the molecule of t+1 frameAnd denominatorTo update t+1 frame Correlation filter
Wherein,η is Learning rate；
A62, sorter model is updated, positive and negative sample is collected with the window of target sizes near the target position of t+1 frame This, carries out online SVM training after extracting feature；
When given training dataset is G={ (xSVM,j,ySVM,j), j=1 ... r }, r is the number of sample, xSVM,jFor Training sample, ySVM,jFor sample label, if the encirclement frame area of sample is m, it is n, their weight that the target of t+1 frame, which surrounds frame, Folded rate s=(m ∩ n)/(m ∪ n), is positive sample as s>0.5, and when s<0.1 is negative sample, trained objective function are as follows:
Determination after the completion of update, for t+2 frame target position.
The technical principle of the invention is described above in combination with a specific embodiment, these descriptions are intended merely to explain of the invention Principle shall not be construed in any way as a limitation of the scope of protection of the invention.Based on explaining herein, those skilled in the art It can associate with other specific embodiments of the invention without creative labor, these modes fall within this hair Within bright protection scope.

Claims (10)

1. a kind of double learning model method for tracking target based on multi-stage characteristics, which is characterized in that it is directed to target video sequence, In given first frame in the case where target original state, the state of target is estimated in video sequence below, specifically include as Lower step:
A1, target video sequence is read, obtains the target position in t frame image, t=1；
A2, cut out in t+1 frame image include the target position candidate region；
The low-level features for extracting the depth characteristic of convolutional layer and the candidate region in the candidate region, by each convolution The characteristic pattern of the output of depth characteristic and the output of low-level features as the multichannel convolutive neural network of target in layerM, N is respectively the width and height of characteristic pattern, and D is channel number；
A3, training correlation filter model:
The candidate region is divided into several cell fritters, establishes Gaussian function label Y for every cell fritter:
Pass through the Gaussian function label Y and characteristic pattern Xl, it is logical to construct each layer, each feature in multichannel convolutive neural network The correlation filter in road；
The characteristic pattern of each layer, each feature channel is obtained according to the correlation filter；
According to the characteristic pattern of each layer, each feature channel obtain each layer of characteristic pattern relevant response figure and each layer of spy Levy the position in the relevant response figure of figure where maximum response；
A4, future position:
The maximum response of the maximum response of current layer and preceding layer is weighted to obtain the position of preceding layer maximum response, is led to Iteration is crossed, final response diagram is obtained；
Maximum response is found in final response diagram, using the position of maximum response as the predicted position of target；
A5, sorter model re-detection refer to according to the HPSR that the relevant response figure of each layer of characteristic pattern obtains t+1 frame image Mark；
When HPSR index is greater than or equal to given threshold θ, final position of the predicted position of the target as target；
When HPSR index be less than given threshold θ when, enable sorter model, by sorter model obtain result and target it is pre- Location, which is set, to be combined, and the final position of t+1 frame image object is obtained；
The update of A6, dual model, real-time online updates correlation filter model and sorter model, for t+2 frame image Target position determines.
2. double learning model method for tracking target according to claim 1, which is characterized in that the candidate region is t 2.5 times of regions in frame image centered on target position.
3. double learning model method for tracking target according to claim 1, which is characterized in that the depth in the step A2 Feature includes advanced features and mid-level features；
The characteristic pattern for extracting convolutional layer Conv3-4, convolutional layer Conv4-4, convolutional layer Conv5-4, as advanced features；
The characteristic pattern for extracting convolutional layer Conv1-2, convolutional layer Conv2-2, as mid-level features.
4. double learning model method for tracking target according to claim 1, which is characterized in that rudimentary in the step A2 Feature includes HOG feature, Gray feature and Color Name feature, three kinds of above-mentioned features is linked togather, as rudimentary spy Sign.
5. double learning model method for tracking target according to claim 1, which is characterized in that extracting depth characteristic and low Before grade feature, the image of the candidate region is subjected to single precision processing and resampling, mean value is calculated after normalization, then right Image goes mean value；
Feature extraction is carried out to the candidate region using the network model of VGGNet-19 pre-training, extracts different convolutional layers Depth characteristic.
6. double learning model method for tracking target according to claim 1, which is characterized in that the step A3 includes:
A31, the candidate region is divided into several cell fritters, establishes Gaussian function label Y for every cell fritter:
Wherein, Y (m, n) indicates the label at (m, n), m ∈ { 0,1 ..., M-1 }, n ∈ { 0,1 ..., N-1 }；σ is parameter,
Wherein, w and h respectively indicates the width and height of target position in t frame, and σ ' indicates the output factor, value 0.1, Cell_size indicates the side length of cell fritter in t+1 frame image, value 4；
A32, the correlation filter W for constructing each layer in multichannel convolutive neural network, each feature channell d:
Wherein, λ is the regularization parameter of correlation filter,WithRespectively XlWith the discrete Fourier transform of Y,For Xl Complex conjugate；
A33, according to the correlation filter Wl dObtain the characteristic pattern of each layer, each feature channel Obtain l layers of relevant response figure ElWith relevant response figure ElIn maximum response:
Wherein,It is operated for inverse Fourier transform,ForDiscrete Fourier transform, relevant response figure ElSize be M ×N；
The response diagram of whole multi-stage characteristics is denoted as set { E1,E2...,El}。
7. double learning model method for tracking target according to claim 1, which is characterized in that the step A4 includes:
A41, the maximum response of the maximum response of current layer and preceding layer is weighted to obtain the position of preceding layer maximum response It sets, by iteration, obtains final response diagram；
El-1(m, n)=αl-1El-1(m,n)+αlEl(m,n)(5)
Wherein, ElThere is maximum response, α in the response diagram position where (m, n) in the characteristic pattern of l grades of (m, n) expressionl For l grades of weight；
Obtain final response diagram E；
A42, maximum response is found by final response diagram E, position the center p of current tracking targett=(xt,yt)；
(xt,yt)=argm,nmax E(m,n) (6)。
8. double learning model method for tracking target according to claim 1, which is characterized in that
Wherein, max (El) it is ElIn maximum value, μlAnd σlThe average value of l layers of response diagram respectively in t+1 frame image And standard deviation, βiFor l layers of relevant response figure ElWeight coefficient.
9. double learning model method for tracking target according to claim 7, which is characterized in that in the step A5, enable Sorter model, in the target p that step 42 obtainst=(xt,yt) the candidate sample set x of surrounding acquisitionSVM, completed using updating Svm classifier:
Wherein, scores is the classification score of classifier,It is respectively the parameter and biasing of classifier with b；
The position for the maximum scores being calculated is determined as the position that sorter model obtains, then with correlation filter mould The result p of typet=(xt,yt) it is weighted synthesis, obtain final result.
10. double learning model method for tracking target according to claim 9, which is characterized in that the step A6 includes:
A61, correlation filter model is updated, updates the molecule of t+1 frameAnd denominatorTo update the correlation of t+1 frame Filter Wl d,
Wherein,η is study Rate；
A62, sorter model is updated, positive negative sample is collected with the window of target sizes near the target position of t+1 frame, is mentioned Online SVM training is carried out after taking feature；
When given training dataset is G={ (xSVM,j,ySVM,j), j=1 ... r }, r is the number of sample, xSVM,jFor training Sample, ySVM,jFor sample label, if the encirclement frame area of sample is m, it is n, their Duplication that the target of t+1 frame, which surrounds frame, S=(m ∩ n)/(m ∪ n) is positive sample as s>0.5, is negative sample, trained objective function when s<0.1 are as follows:
Determination after the completion of update, for t+2 frame target position.
CN201811405327.1A 2018-11-23 2018-11-23 Double-learning-model target tracking method based on multi-level features Active CN109543615B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811405327.1A CN109543615B (en) 2018-11-23 2018-11-23 Double-learning-model target tracking method based on multi-level features

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811405327.1A CN109543615B (en) 2018-11-23 2018-11-23 Double-learning-model target tracking method based on multi-level features

Publications (2)

Publication Number Publication Date
CN109543615A true CN109543615A (en) 2019-03-29
CN109543615B CN109543615B (en) 2022-10-28

Family

ID=65849621

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811405327.1A Active CN109543615B (en) 2018-11-23 2018-11-23 Double-learning-model target tracking method based on multi-level features

Country Status (1)

Country Link
CN (1) CN109543615B (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993776A (en) * 2019-04-04 2019-07-09 杭州电子科技大学 A kind of correlation filtering method for tracking target and system based on multistage template
CN110246155A (en) * 2019-05-17 2019-09-17 华中科技大学 One kind being based on the alternate anti-shelter target tracking of model and system
CN110929848A (en) * 2019-11-18 2020-03-27 安徽大学 Training and tracking method based on multi-challenge perception learning model
CN111640138A (en) * 2020-05-28 2020-09-08 济南博观智能科技有限公司 Target tracking method, device, equipment and storage medium
WO2020228522A1 (en) * 2019-05-10 2020-11-19 腾讯科技（深圳）有限公司 Target tracking method and apparatus, storage medium and electronic device
WO2021007984A1 (en) * 2019-07-18 2021-01-21 深圳大学 Target tracking method and apparatus based on tsk fuzzy classifier, and storage medium

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101814187A (en) * 2009-12-29 2010-08-25 天津市亚安科技电子有限公司 Video tracking method based on multi-stage characteristics
KR20180069312A (en) * 2016-12-15 2018-06-25 한국전자통신연구원 Method for tracking of object using light field video and apparatus thereof
CN108346159A (en) * 2018-01-28 2018-07-31 北京工业大学 A kind of visual target tracking method based on tracking-study-detection
CN108520529A (en) * 2018-03-30 2018-09-11 上海交通大学 Visible light based on convolutional neural networks and infrared video method for tracking target
CN108734151A (en) * 2018-06-14 2018-11-02 厦门大学 Robust long-range method for tracking target based on correlation filtering and the twin network of depth

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101814187A (en) * 2009-12-29 2010-08-25 天津市亚安科技电子有限公司 Video tracking method based on multi-stage characteristics
KR20180069312A (en) * 2016-12-15 2018-06-25 한국전자통신연구원 Method for tracking of object using light field video and apparatus thereof
CN108346159A (en) * 2018-01-28 2018-07-31 北京工业大学 A kind of visual target tracking method based on tracking-study-detection
CN108520529A (en) * 2018-03-30 2018-09-11 上海交通大学 Visible light based on convolutional neural networks and infrared video method for tracking target
CN108734151A (en) * 2018-06-14 2018-11-02 厦门大学 Robust long-range method for tracking target based on correlation filtering and the twin network of depth

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
JUNHUA YAN ET AL.: "Real-time unmanned aerial vehicle tracking of fast moving small target on ground", 《J. OF ELECTRONIC IMAGING》 *
栗宝俊: "基于多表观模型的长期视觉目标跟踪算法研究", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993776A (en) * 2019-04-04 2019-07-09 杭州电子科技大学 A kind of correlation filtering method for tracking target and system based on multistage template
WO2020228522A1 (en) * 2019-05-10 2020-11-19 腾讯科技（深圳）有限公司 Target tracking method and apparatus, storage medium and electronic device
JP2022516055A (en) * 2019-05-10 2022-02-24 テンセント・テクノロジー・（シェンジェン）・カンパニー・リミテッド Goal tracking methods, computer programs, and electronic devices
JP7125562B2 (en) 2019-05-10 2022-08-24 テンセント・テクノロジー・（シェンジェン）・カンパニー・リミテッド Target tracking method, computer program, and electronic device
US11610321B2 (en) 2019-05-10 2023-03-21 Tencent Technology (Shenzhen) Company Limited Target tracking method and apparatus, storage medium, and electronic device
CN110246155A (en) * 2019-05-17 2019-09-17 华中科技大学 One kind being based on the alternate anti-shelter target tracking of model and system
CN110246155B (en) * 2019-05-17 2021-05-18 华中科技大学 Anti-occlusion target tracking method and system based on model alternation
WO2021007984A1 (en) * 2019-07-18 2021-01-21 深圳大学 Target tracking method and apparatus based on tsk fuzzy classifier, and storage medium
CN110929848A (en) * 2019-11-18 2020-03-27 安徽大学 Training and tracking method based on multi-challenge perception learning model
CN110929848B (en) * 2019-11-18 2023-03-31 安徽大学 Training and tracking method based on multi-challenge perception learning model
CN111640138A (en) * 2020-05-28 2020-09-08 济南博观智能科技有限公司 Target tracking method, device, equipment and storage medium
CN111640138B (en) * 2020-05-28 2023-10-27 济南博观智能科技有限公司 Target tracking method, device, equipment and storage medium

Also Published As

Publication number Publication date
CN109543615B (en) 2022-10-28

Similar Documents

Publication Publication Date Title
CN109543615A (en) A kind of double learning model method for tracking target based on multi-stage characteristics
CN112418117B (en) Small target detection method based on unmanned aerial vehicle image
CN107748873B (en) A kind of multimodal method for tracking target merging background information
CN110084151B (en) Video abnormal behavior discrimination method based on non-local network deep learning
CN108665481B (en) Self-adaptive anti-blocking infrared target tracking method based on multi-layer depth feature fusion
CN110197502B (en) Multi-target tracking method and system based on identity re-identification
CN107145862B (en) Multi-feature matching multi-target tracking method based on Hough forest
CN107767405A (en) A kind of nuclear phase for merging convolutional neural networks closes filtered target tracking
CN107481264A (en) A kind of video target tracking method of adaptive scale
CN107230267B (en) Intelligence In Baogang Kindergarten based on face recognition algorithms is registered method
CN109800624A (en) A kind of multi-object tracking method identified again based on pedestrian
CN106570490B (en) A kind of pedestrian&#39;s method for real time tracking based on quick clustering
CN111640138B (en) Target tracking method, device, equipment and storage medium
CN106355604A (en) Target image tracking method and system
CN108647695A (en) Soft image conspicuousness detection method based on covariance convolutional neural networks
CN110263712A (en) A kind of coarse-fine pedestrian detection method based on region candidate
CN110458160A (en) A kind of unmanned boat waterborne target recognizer based on depth-compression neural network
CN110728694B (en) Long-time visual target tracking method based on continuous learning
CN110163891A (en) Long term object tracking based on sensor model
CN108682022A (en) Based on the visual tracking method and system to anti-migration network
CN109344842A (en) A kind of pedestrian&#39;s recognition methods again based on semantic region expression
Xiao et al. Traffic sign detection based on histograms of oriented gradients and boolean convolutional neural networks
CN111460881A (en) Traffic sign countermeasure sample detection method and classification device based on neighbor discrimination
CN111462173B (en) Visual tracking method based on twin network discrimination feature learning
CN106874867A (en) A kind of face self-adapting detecting and tracking for merging the colour of skin and profile screening

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant