CN106815576A - Target tracking method based on consecutive hours sky confidence map and semi-supervised extreme learning machine - Google Patents
Target tracking method based on consecutive hours sky confidence map and semi-supervised extreme learning machine Download PDFInfo
- Publication number
- CN106815576A CN106815576A CN201710047829.0A CN201710047829A CN106815576A CN 106815576 A CN106815576 A CN 106815576A CN 201710047829 A CN201710047829 A CN 201710047829A CN 106815576 A CN106815576 A CN 106815576A
- Authority
- CN
- China
- Prior art keywords
- target
- sample
- trail
- followed
- semi
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
-
- 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
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of target tracking method based on consecutive hours sky confidence map and semi-supervised extreme learning machine, the method is in time continuous in view of video frame image, target location to be followed the trail of will not also undergo mutation simultaneously, other video frame image is also spatially continuous, spatial continuity is embodied in target and target ambient background has certain particular kind of relationship, when great changes will take place for the outward appearance of target, this relation can help distinguish between target and background region to be followed the trail of.The present invention is directed to deformation and occlusion issue, fully take into account the information that real goal can be provided, fully excavate the distribution similarity for having exemplar and unlabeled exemplars, improve the precision followed the trail of, propose a kind of semi-supervised method for tracing based on extreme learning machine for excavating and having exemplar and unlabeled exemplars distribution similarity, above two method is combined in a tracking framework for coupling, the present invention realizes that a kind of robustness is good, high robust tracking.
Description
Technical field
Target tracking method the present invention relates to be based on consecutive hours sky confidence map and semi-supervised extreme learning machine, belongs to intelligence
Information processing and target tracking technical field.
Background technology
Target following is indispensable link in most vision systems.(such as video in specific scene application
The fields such as monitoring), automation, the quick, target tracking of high robust attract attention.Video monitoring, Vehicle Detection, intelligence machine
The aspects such as people, sea floor object detection tracking have broad application prospects.
Target tracking is an extremely important part in computer vision field, and moving objects in video track algorithm is logical
The information of the video image for analyzing each frame in sequence of video images to be followed the trail of is crossed, data mining is carried out in video, learn mesh
Mark behavior simultaneously carries out substantial amounts of motion capture, and a series for the treatment of is carried out to information, obtains and marks tracked target and exist
Corresponding position in video image.The complexity for blocking deformation, background between object, the change of illumination light and shade, real-time and strong
Strong property difference etc. is tracing process problem demanding prompt solution.Classical method for tracing such as Meanshift, particle filter etc. is depended on and regarded
The abundant degree of contained target information in frequency, in actual sequence of video images, the information that target can be provided is fairly limited,
Cause the tracking target that can not stablize, such as there is deformation to block in scene, these classic algorithms are even more helpless.
That is subject matter present in prior art：(1) real-time and strong during being followed the trail of in video scene to be followed the trail of
Strong property is poor, and target time space position information is deficient, the unobvious problem of target signature；(2) when scene has shelter and target to be followed the trail of
In the case of deforming upon, especially occur that whole target is blocked and target to be followed the trail of occurs the situation of huge deformation,
The problem that the target followed the trail of can be caused to lose.
The content of the invention
It is an object of the invention to provide a kind of target tracking based on consecutive hours sky confidence map and semi-supervised extreme learning machine
Method, to make up the deficiencies in the prior art.
The present invention is in time continuous in view of video frame image, and the continuity of time is embodied in neighbouring interframe and waits to chase after
Track object variations will not be very big, while target location to be followed the trail of will not also undergo mutation；At the same time video frame image is in space
On be also continuous, spatial continuity is embodied in target and target ambient background and there is certain particular kind of relationship, when the outward appearance of target
When great changes will take place, this relation can help distinguish between target and background region to be followed the trail of, and propose to utilize the vacant letter of consecutive hours
The method for tracing that graphics is practised overcomes real-time and robustness is poor, target time space position information scarcity and target signature are obvious etc.
Problem.For deformation and occlusion issue, the information that real goal can be provided is fully taken into account, fully excavating has exemplar
With the distribution similarity of unlabeled exemplars, the precision followed the trail of is improved, it is proposed that one kind is excavated exemplar and unlabeled exemplars
The semi-supervised method for tracing based on extreme learning machine of distribution similarity, above two method is combined in a tracking for coupling
In framework, realize that a kind of robustness is good, high robust tracking.
To reach above-mentioned purpose, concrete technical scheme that the present invention takes is through the following steps that realize：
Step one, it is specific treat tracing monitoring scene in gather n frames target video A={ I to be followed the trail of1,…,Ii,…In,
Wherein IiThe i-th frame sequence of video images to be followed the trail of is represented, video to be followed the trail of is pre-processed using image filtering denoising, contrast enhancing
SequenceReduce noise and protrusion interested treats trace regions；
Step 2, in t frames sequence of video images I to be followed the trail oftMiddle use rectangular window chooses target O to be followed the trail of, it is determined that
Target's center position o*, O represents fresh target presence in the scene, and o represents new target location, defines a two dimension mesh to be followed the trail of
Target confidence map MODEL Ct(o)；It is by target area expansion twice formation local background region representation to be followed the trail of It is interior
Intensity locations feature w (k) at coordinate position k are extracted, intensity locations feature set is constituted
The brightness of image at I (k) denotation coordinations position k,Denotation coordination o*Neighborhood；Set up the priori of t frames target to be followed the trail of
Model P (w (k) | O), t frame space-time models are gone out in prediction on such basis
Step 3, in target's center position to be followed the trail of region overlap sampling, obtain N1Individual region unit image is used as just
Sample and N2Individual region unit image extracts positive and negative sample data feature x as negative samplej, the class label for remembering positive sample is 1, is born
The class label of sample is 0, yj∈{1,0}；Foundation has mark sample setWith without mark sample set XuGroup
Into training sample set X={ Xs,Xu}={ (xj,yj), j=1 ..., N1+N2；
Step 4, the training sample set X obtained with step 3 train semi-supervised extreme learning machine network model；
Step 5, in It+1In, the t frame space-time models tried to achieve using step 2Model modification is carried out, is calculated
The space-time model of t+1 framesUsing the t+1 frame space-time models tried to achieveConvolution It+1Obtain the space-time confidence map of fresh target
Ct+1O (), maximizes Ct+1O () determines the target location o in t+1 frames；
Step 6, judge whether target is blocked, if target is not blocked, into step 5, conversely, into step 7；
Step 7, in It+1In, by ItIn the o that has tried to achieve*It is target location, in target location o*Region, with target
Region rectangle window size overlap sampling, obtains N number of region unit image as candidate target, extracts candidate target data characteristicsSet up target image block test sample collection to be followed the trail ofBy test sample collection input step four
The semi-supervised extreme learning machine network for completing is trained, t+1 frames test output T is obtained, online semi-supervised extreme learning machine is maximized
Maximum classification response position, obtains target location o in t+1 frames；
Step 8, carry out online semi-supervised extreme learning machine network model to maximum classification response result and update threshold value sentencing
It is fixed, if online semi-supervised extreme learning machine model need not update, into step 5, otherwise enter step 9；
Step 9, there is labeled data collection by what step 3 was obtainedThe test specimens obtained with step 7
This collectionAs without labeled data collection Xu=Xt+1, carry out step 4, the semi-supervised limit of re -training
Habit machine network model；
Circulating repetition above-mentioned steps, until follow the trail of completing whole video sequence.
Further, the step 3：In target's center position o to be followed the trail of*Region, with target area rectangular window
Size overlap sampling, j-th sampled point be to the Euclidean distance of target's center positionWhenWhen, sampling obtains N1It is individual
Region unit image as positive sample, whenWhen, sampling obtains N2Individual region unit image is used as negative sample, r1、r2And r3
It is respectively sample radius；Extract positive and negative sample data feature xj, target image block training sample set to be followed the trail of is set up, (N is collected altogether1
+N2) individual target image block is used as training sample set X={ (xj,yj), j=1 ..., N1+N2, remembering the class label of positive sample is
1, the class label of negative sample is 0, yj∈{1,0}；The sample order that training sample is concentrated is upset and reset, is taken before coming most
A certain proportion of sample (usual ratio is relatively low) conduct in face has marked sample set Xs, take remaining sample (usual ratio is higher) work
Not mark sample set Xu, and X={ Xs,Xu}。
The step 4：Input weights are set using random fashion and hidden layer is biased, if representing hidden layer knot with (a, b)
Input weight a and threshold value b that point is obtained, training sample are have labeled data collectionWithout labeled data collectionWherein XsAnd XuRepresent input sample, YsIt is and XsCorresponding output sample；The mapping function of hidden layer is G
X (), mapping function form can be expressed as G (x)=1/ (1+e-x), output weight is represented with β, h (xi)=[G (a1,b1,
xi),…,G(aL,bL,xi)]s×mI-th hidden layer output matrix is represented, the nodes of hidden layer are m, eiRepresent i-th input
The learning error (residual error) of node
The object function of semi-supervised extreme learning machine is：
fi=h (xi) β, i=1 ..., s+u
Wherein ciPunishment parameter is represented, λ represents balance parameter, and L is the Tula obtained by label data and without label data
This operation result of pula, F is the output matrix of network, and Tr is mark computing；
Represent that semi-supervised extreme learning machine object function is with matrix form：
WhereinIt is that preceding s rows are equal to Ys, the rear null output label sample of u rows, C is that preceding s diagonal entry is CiIt is remaining
It is zero diagonal matrix；
Local derviation is asked to obtain β above formula：
It is zero to make local derviation, and solution obtains output weight beta and is：
When there is label data to be more than hidden layer nodal point number
When there is label data to be less than hidden layer nodal point number
Wherein, HTIt is the transposed matrix of matrix H.
The method whether target is blocked of judging described in step 6 is the result to confidence mapCarry out occlusion threshold
th1Judge, ifWhen, show that blocking occurs in target, th1The critical value that expression is blocked, can become according to the difference of scene
Change, when this algorithm is applied to different scenes, artificially adjust th1Value, can fluctuate, when target quilt in certain scope under normal circumstances
When blocking,Can decline rapidly, the value after will fall rapidly upon is defined as th1Value, judges whether target is blocked with this.
It is to maximum that whether the online semi-supervised extreme learning machine network model of judgement described in step 8 needs the method for updating
Classification response result TmaxIt is updated threshold value th2Judge, if Tmax>th2When, show online semi-supervised extreme learning machine network mould
Type is not required to update, to update threshold value th2Judge whether network model needs to update.
Beneficial effects of the present invention：The vacant letter graphics of consecutive hours is practised method for tracing and semi-supervised extreme learning machine by the present invention
Method for tracing is combined, and solves in tracing process that real-time and robustness are poor, target time space position information is deficient, target signature
Not obvious and deformation blocks to cause to follow the trail of the problem that target is lost.The present invention is particular by calculating consecutive hours sky confidence map
Carry out occlusion threshold judgement, obtained it is a kind of judge target whether enter occlusion area method, effectively solve target screening
The decision problem of gear, and carry out semi-supervised limit study by calculating semi-supervised extreme learning machine network output maximum response
Machine network model update threshold determination, obtained it is a kind of judge the network model whether need renewal method, efficiently solve net
The problem of network model generalization difference.Present invention greatly enhances the precision followed the trail of, realize that a kind of robustness is good, high robust
Tracing process.
Brief description of the drawings
Fig. 1 is overall trace flow schematic diagram of the invention.
Fig. 2 is the zone marker figure of target to be followed the trail of in specific embodiment.
Fig. 3 is based on consecutive hours sky confidence map target tracking method block diagram in specific embodiment.
Fig. 4 is the basic framework figure of semi-supervised extreme learning machine network.
Fig. 5 be specific embodiment in be based on semi-supervised extreme learning machine target tracking method block diagram.
To follow the trail of effect example under circumstance of occlusion in specific embodiment, figure (a) is that have interesting target to be followed the trail of to Fig. 6
Frame of video, (b), (c), (d), (e), (f) are respectively the frame of video being tracked to the interesting target after (a) frame.
Specific embodiment
To make the purpose of the present invention, embodiment and advantage relatively sharp, below in conjunction with the accompanying drawings and by specific embodiment
To further illustrate the present invention.
Particular flow sheet of the invention is as shown in Figure 1.
It is specific using one section of corridor monitor video caviar (384*288 pixels, 25 frame per second) of classics in the present embodiment
As video to be followed the trail of.
Step one, pre-process video sequence to be followed the trail of using image filtering denoising, contrast enhancing, reduction noise and protrusion
It is interested to treat trace regions；Specifically include following steps：
Step 1-1, the corridor monitor video caviar to one section of classics are defined as A, carry out sub-frame processing 200 frames of acquisition and treat
Follow the trail of sequence of video images, i.e. A={ I1,…,Ii,…I200, wherein IiRepresent that corridor monitoring video the i-th frames of caviar are waited to chase after
Track video image；
Step 1-2, to the 200 frame video image sequenceIt is filtered denoising, the enhanced pretreatment of contrast.
Step 2, in t=1 frames sequence of video images I to be followed the trail ofT=1In choose target O to be followed the trail of, determine target's center
Position o*, O represents fresh target presence in the scene, and o represents new target location, defines a confidence for two dimension target to be followed the trail of
Graph model Ct(o)；The prior model P (w (k) | O) of t frames target to be followed the trail of is set up, t frame space-time models are gone out in prediction on such basisAs schemed
Shown in 3；Specifically include following steps：
Step 2-1, in IT=1In choose interesting target O to be followed the trail of, target area rectangle using rectangular window W by user
The width of window is w and height is h, and o represents new target location, and target area is expanded into twice forms local background's region representation
ForAs shown in Figure 2；Intensity locations feature w (k) at coordinate position k are extracted in local background region, intensity locations are constituted
Feature setThe brightness of image at I (k) denotation coordinations position k,Denotation coordination o*
Neighborhood；
Step 2-2, tracing problem is converted into calculates interesting target position confidence map problem to be followed the trail of：
Wherein CtO () represents the confidence graph model of t frames, represent new target location o and old target location o*Relative position
And direction, new target location is nearer apart from old target location, and the value of the confidence is bigger；Represent space-time mould
Type, the relative position and direction of description fresh target o and local background area coordinate points k, P (w (k) | O) prior model is represented, retouch
State old target locationIntensity and relative position direction with local background area coordinate points k, simulate interesting target O to be followed the trail of
Rudimentary profile information；
Step 2-3, calculating t=1 frames IT=1Confidence mapSimultaneously maximum is obtained to put
Letter value
Step 2-4, the prior model for calculating t=1 framesWhereinIt is chi
Degree parameter；
Step 2-5, by calculating t=1 frames IT=1Confidence map MODEL CtO () and prior model P (w (k), O) calculate the
The space-time model of the interesting target of t=1 frames
Wherein F represents FFT, F-1Represent fast fourier inverse transformation.
Step 3, in target's center position to be followed the trail of region overlap sampling, obtain N1Individual region unit image is used as just
Sample and N2Individual region unit image extracts positive and negative sample data feature x as negative samplej, setting up has mark sample setWith without mark sample set XuComposition training sample set X={ (xj,yj), j=1 ..., N1+N2, such as Fig. 5
It is shown, specifically include following steps：
Step 3-1, in center o*Region, with target area rectangular window size overlap sampling, j-th sampling
Euclidean distance of the point to target's center position beWhenWhen, sampling obtains 45 region unit images as positive sample,
WhenWhen, sampling obtains 31 region unit images as negative sample, sample radius r1、r2And r3It is respectively to set
(the unit of parameter 5,10 and 20：Pixel)；
Step 3-2, the positive and negative sample data feature x of extractionj, target image block training sample set to be followed the trail of is set up, collect altogether
76 target image blocks are used as training sample set X={ (xj,yj), j=1 ..., 76, the class label for remembering positive sample is 1, is born
The class label of sample is 0, yj∈{1,0}；
Step 3-3, by training sample concentrate sample order upset and reset, take 50 that come foremost sample make
To have marked sample set Xs, remaining 26 samples are taken as not marking sample set Xu, and X={ Xs,Xu}。
Step 4, the training sample set X obtained with step 3 train online semi-supervised extreme learning machine network model, specifically
Comprise the following steps：
Step 4-1, semi-supervised extreme learning machine are a kind of BP network models of single hidden layer as shown in figure 4, whole
Network model is divided into three layers, including：Input layer, hidden layer and output layer, input weights and hidden layer are set using random fashion
Biasing, independently of training sample, algorithm structure simple computation efficiency high, if representing the input that hidden layer node is obtained with (a, b)
Weight a and threshold value b, training sample is have labeled data collectionWithout labeled data collectionWherein
XsAnd XuRepresent input sample, YsIt is and XsCorresponding output sample；The mapping function of hidden layer is G (x), mapping function form
G (x)=1/ (1+e can be expressed as-x), output weight is represented with β, h (xi)=[G (a1,b1,xi),…,G(a2000,b2000,
xi)]50×2000I-th hidden layer output matrix is represented, the nodes of hidden layer are 2000, eiRepresent i-th of input node
Practise error (residual error)；
Step 4-2, the object function of the semi-supervised extreme learning machine that need to be trained are：
fi=h (xi) β, i=1 ..., s+u
Wherein ciPunishment parameter is represented, λ represents balance parameter, and L is the Tula obtained by label data and without label data
This operation result of pula, F is the output matrix of network, TrIt is mark computing；
Step 4-3, represent that semi-supervised extreme learning machine object function is with matrix form：
WhereinIt is that preceding 50 row is equal to Ys, the null output label sample of 26 rows afterwards.C is that preceding 50 diagonal entries are
CiRemaining as zero diagonal matrix；
Step 4-4, local derviation is asked to obtain β above formula：
Step 4-5, local derviation is made to be zero, solution obtains output weight beta and is：
When there is label data to be more than hidden layer nodal point number
When there is label data to be less than hidden layer nodal point number
Wherein, HTIt is the transposed matrix of matrix H, so far semi-supervised extreme learning machine network model training is finished.
Step 5, in t+1 frames, the t frame space-time models tried to achieve using step 2Model modification is carried out, is calculated
Obtain the space-time model of t+1 framesUsing the t+1 frame space-time models tried to achieveConvolved image It+1Obtain fresh target when
Empty confidence map Ct+1O (), maximizes the confidence map C for trying to achievet+1O () determines the target location o in t+1 frames, as shown in figure 3, tool
Body is comprised the following steps：
Step 5-1, in It+1In, with o*The local background region of twice target sizes is taken for target location
Intensity locations feature is extracted in the region, intensity locations feature set is constituted
Step 5-2, the space-time model of t frames interesting target to be followed the trail of update：
Wherein ρ is learning rate,It is interesting target space-time model to be followed the trail of that t frames are calculated,In frequency domain
It is expressed as：
WhereinIt isTime domain Fourier transform.Time domain filtering FwIt is expressed as：
Fw=ρ/(ejw-(1-ρ))
Wherein j is imaginary unit；
Step 5-3, calculating t+1 frames interesting target confidence map to be followed the trail of：
Step 5-4, in t+1 frames interesting target position o be maximize t+1 frames confidence map：
O=arg maxCt+1(o)
Maximum the value of the confidence is
Step 6, judge whether target is blocked, specifically include following steps：
Step 6-1, the maximum the value of the confidence obtained to step 5-4Carry out occlusion threshold th1Judge, ifWhen,
Show that blocking occurs in target, judge whether target is blocked with this.th1The critical value that expression is blocked, can be according to the difference of scene
And change, when this algorithm is applied to different scenes, artificially adjust th1Value, can fluctuate in certain scope under normal circumstances, work as mesh
When mark is blocked,Can decline rapidly, the value after will fall rapidly upon is defined as th1Value；Defined in this concrete scheme
If step 6-2,When, show that blocking does not occur in target, step 5-1 is carried out, otherwise carry out step 7-
1。
Step 7, in t+1 frames, tracked overlap sampling at target's center position in t frames, extract candidate target
Data characteristics, sets up target image block test sample to be followed the trail of, and test sample is input into the above-mentioned online semi-supervised pole for training
Be prediction new target location by maximum classification response position in test sample in limit learning machine, as shown in figure 5, specifically include with
Lower step：
Step 7-1, to t+1 frame video images, with o*It is target location, in center o*Region, with target area
Domain rectangular window size overlap sampling, j-th sampled point to o*Euclidean distance beWhenWhen, sampling obtains 232
Region unit image is test data as candidate target, and the sample data of extraction is characterized asAnd remember that test set isSample radius r1The parameter of setting is 20 (units：Pixel)；
Step 7-2, test are output as：
T=H*β
Wherein β is the output weight that t frames are calculated, H*It is the hidden layer output matrix of test,
Step 7-3, in t+1 frames, interesting target position o to be followed the trail of be maximize the semi-supervised extreme learning machine of t+1 frames
Maximum classification response position：
O=arg max T
Maximum classification response value is Tmax。
Step 8, carry out online semi-supervised extreme learning machine network model to maximum classification response result and update threshold value sentencing
It is fixed, comprise the following steps that：
Step 8-1, to maximum classification response result TmaxCarry out the renewal threshold value th of semi-supervised extreme learning machine2Judge, if
Tmax>th2When, show that online semi-supervised extreme learning machine network model is not required to update, judge that the online semi-supervised limit learns with this
Whether machine model needs to update, th2The critical value for updating is represented, th defined in this concrete scheme2=0.
If step 8-2, Tmax>When 0, show that online semi-supervised extreme learning machine network model is not required to update, carry out step 5-
1, otherwise carry out step 9.
The online semi-supervised extreme learning machine network model of step 9, re -training, as shown in figure 5, specific as follows：By step
What 3-3 was obtained has labeled data collectionThe test set obtained with step 7-1
As without labeled data collection Xu=Xt+1, carry out the online semi-supervised extreme learning machine network model of step 4-1 re -trainings.
Circulating repetition above-mentioned steps, until completing the tracking for entirely treating tracing monitoring video sequence.
Tracing monitoring video is treated to above-mentioned, respectively with particle filter, Meanshift and the inventive method traceability
The comparing of energy, the results are shown in Table 1, it can be seen that no matter the inventive method center bias contribution or deviation mean square deviation result
Particle filter and Meanshift methods are superior to, robustness and robustness to target tracking is realized.
Table 1 is contrasted for display particle filter, Meanshift and the inventive method tracking performance
| Particle filter | Meanshift | The inventive method | |
| Center deviation | 75.4796 | 22.9740 | 10.1834 |
| Deviation mean square deviation | 47.8903 | 12.2607 | 7.9702 |
Fig. 6 is tracking effect example under circumstance of occlusion in specific embodiment, it can be seen that serious twice having successively gone through
Circumstance of occlusion under, can accurately still follow the trail of target, further demonstrate the robustness and robustness of the inventive method.
The above is the preferred embodiment of the present invention, it should be pointed out that：For those skilled in the art
For, under the premise without departing from the principles of the invention, some improvements and modifications can also be made, these improvements and modifications should also be regarded
It is protection scope of the present invention.
Claims (5)
1. it is a kind of based on consecutive hours sky confidence map and semi-supervised extreme learning machine target tracking method, it is characterised in that including
Following steps：
Step one, it is specific treat tracing monitoring scene in gather n frames target video A={ I to be followed the trail of1,…,Ii,…In, wherein Ii
The i-th frame sequence of video images to be followed the trail of is represented, video sequence to be followed the trail of is pre-processed using image filtering denoising, contrast enhancingReduce noise and protrusion interested treats trace regions；
Step 2, in t frames sequence of video images I to be followed the trail oftMiddle use rectangular window chooses target O to be followed the trail of, in determining target
Heart position o*, O represent fresh target presence in the scene, and o represents new target location, one two dimension target O's to be followed the trail of of definition
Confidence map MODEL Ct(o)；It is by target area expansion twice formation local background region representation to be followed the trail of Interior extraction
Intensity locations feature w (k) at coordinate position k, constitute intensity locations feature setI(k)
The brightness of image at the k of denotation coordination position,The neighborhood of denotation coordination o*.Set up the prior model of t frames target to be followed the trail of
P (w (k) | O), t frame space-time models are gone out in prediction on such basis
Step 3, in target's center position to be followed the trail of region overlap sampling, obtain N1Individual region unit image as positive sample and
N2Individual region unit image extracts positive and negative sample data feature x as negative samplej, the class label for remembering positive sample is 1, negative sample
Class label is 0, yj∈{1,0}；Foundation has mark sample setWith without mark sample set XuComposition training
Sample set X={ Xs,Xu}={ (xj,yj), j=1 ..., N1+N2；
Step 4, the training sample set X obtained with step 3 train semi-supervised extreme learning machine network model；
Step 5, in It+1In, the t frame space-time models tried to achieve using step 2Model modification is carried out, t+1 is calculated
The space-time model of frameUsing the t+1 frame space-time models tried to achieveConvolution It+1Obtain the space-time confidence map C of fresh targett+1
O (), maximizes Ct+1O () determines the target location o in t+1 frames；
Step 6, judge whether target is blocked, if target is not blocked, into step 5, conversely, into step 7；
Step 7, in It+1In, by ItIn the o* that has tried to achieve be target location, in target location o* regions, with target area
Rectangular window size overlap sampling, obtains N number of region unit image as candidate target, extracts candidate target data characteristicsBuild
Found target image block test sample collection to be followed the trail ofTest sample collection input step four has been trained
Into semi-supervised extreme learning machine network, obtain t+1 frames test output T, maximize the maximum classification of semi-supervised extreme learning machine and ring
Position is answered, target location o in t+1 frames is obtained；
Step 8, semi-supervised extreme learning machine network model is carried out to maximum classification response result update threshold determination, if half supervises
Superintending and directing extreme learning machine model need not update, and into step 5, otherwise enter step 9；
Step 9, there is labeled data collection by what step 3 was obtainedThe test sample collection obtained with step 7As without labeled data collection Xu=Xt+1, carry out step 4, the semi-supervised extreme learning machine of re -training
Network model；
Circulating repetition above-mentioned steps, until follow the trail of completing whole video sequence.
2. target tracking method as claimed in claim 1, it is characterised in that the step 3 is specially：In target to be followed the trail of
Center o*Region, with target area rectangular window size overlap sampling, j-th sampled point to target's center position
Euclidean distance isWhenWhen, sampling obtains N1Individual region unit image as positive sample, whenWhen, sampling
Obtain N2Individual region unit image is used as negative sample, r1、r2And r3It is respectively sample radius.Extract positive and negative sample data feature xj, build
Target image block training sample set to be followed the trail of is found, (N is collected altogether1+N2) individual target image block is used as training sample set X={ (xj,
yj), j=1 ..., N1+N2, the class label for remembering positive sample is 1, and the class label of negative sample is 0, yj∈{1,0}；Will training
Sample order in sample set is upset and is reset, and takes and comes a certain proportion of sample of foremost as having marked sample set Xs,
Remaining sample is taken as not marking sample set Xu, and X={ Xs,Xu}。
3. target tracking method as claimed in claim 1, it is characterised in that the step 4 is specially：Using random fashion
Input weights and hidden layer biasing are set, if representing input weight a and threshold value b that hidden layer node is obtained with (a, b), sample is trained
This is have labeled data collectionWithout labeled data collectionWherein XsAnd XuRepresent input sample,
YsIt is and XsCorresponding output sample；The mapping function of hidden layer is G (x), and mapping function form can be expressed as G (x)=1/ (1
+e-x), output weight is represented with β, h (xi)=[G (a1,b1,xi),…,G(aL,bL,xi)]1×mRepresent i-th hidden layer output square
Battle array, the nodes of hidden layer are m, eiRepresent i-th learning error of input node；
The object function of semi-supervised extreme learning machine is：
fi=h (xi) β, i=1 ..., s+u
Wherein ciPunishment parameter is represented, λ represents balance parameter, and L is the figure Laplce obtained by label data and without label data
Operation result, F is the output matrix of network, and Tr is mark computing；
Represent that semi-supervised extreme learning machine object function is with matrix form：
WhereinIt is that preceding s rows are equal to Ys, the rear null output label sample of u rows, C is that preceding s diagonal entry is CiRemaining as zero
Diagonal matrix；
Local derviation is asked to obtain β above formula：
It is zero to make local derviation, and solution obtains output weight beta and is：
When there is label data to be more than hidden layer nodal point number
When there is label data to be less than hidden layer nodal point number
Wherein, HTIt is the transposed matrix of matrix H.
4. target tracking method as claimed in claim 1, it is characterised in that judge whether target is blocked described in step 6
Method be result to confidence mapCarry out occlusion threshold th1Judge, ifWhen, show that blocking occurs in target, th1
The critical value that expression is blocked, can change according to the difference of scene, when this algorithm is applied to different scenes, artificially adjust th1
Value, can fluctuate in certain scope under normal circumstances, when target is blocked,Can decline rapidly, the value after will fall rapidly upon
It is defined as th1Value, judges whether target is blocked with this.
5. target tracking method as claimed in claim 1, it is characterised in that the semi-supervised limit study of judgement described in step 8
It is to maximum classification response result T that whether machine network model needs the method for updatingmaxIt is updated threshold value th2Judge, if Tmax>th2
When, show that semi-supervised extreme learning machine network model is not required to update, to update threshold value th2Judge whether network model needs to update.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710047829.0A CN106815576B (en) | 2017-01-20 | 2017-01-20 | Target tracking method based on continuous space-time confidence map and semi-supervised extreme learning machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710047829.0A CN106815576B (en) | 2017-01-20 | 2017-01-20 | Target tracking method based on continuous space-time confidence map and semi-supervised extreme learning machine |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106815576A true CN106815576A (en) | 2017-06-09 |
| CN106815576B CN106815576B (en) | 2020-07-07 |
Family
ID=59111417
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710047829.0A Active CN106815576B (en) | 2017-01-20 | 2017-01-20 | Target tracking method based on continuous space-time confidence map and semi-supervised extreme learning machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106815576B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107423762A (en) * | 2017-07-26 | 2017-12-01 | 江南大学 | Semi-supervised fingerprinting localization algorithm based on manifold regularization |
| CN109255321A (en) * | 2018-09-03 | 2019-01-22 | 电子科技大学 | A kind of visual pursuit classifier construction method of combination history and instant messages |
| CN110211104A (en) * | 2019-05-23 | 2019-09-06 | 复旦大学 | A kind of image analysis method and system for pulmonary masses computer aided detection |
| CN110675382A (en) * | 2019-09-24 | 2020-01-10 | 中南大学 | Aluminum electrolysis superheat degree identification method based on CNN-LapseLM |
| CN113378673A (en) * | 2021-05-31 | 2021-09-10 | 中国科学技术大学 | Semi-supervised electroencephalogram signal classification method based on consistency regularization |
| CN113408984A (en) * | 2021-06-21 | 2021-09-17 | 北京思路智园科技有限公司 | Hazardous chemical substance transportation tracking system and method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8254633B1 (en) * | 2009-04-21 | 2012-08-28 | Videomining Corporation | Method and system for finding correspondence between face camera views and behavior camera views |
| CN102663453A (en) * | 2012-05-03 | 2012-09-12 | 西安电子科技大学 | Human motion tracking method based on second generation Bandlet transform and top-speed learning machine |
| CN103942749A (en) * | 2014-02-24 | 2014-07-23 | 西安电子科技大学 | Hyperspectral ground feature classification method based on modified cluster hypothesis and semi-supervised extreme learning machine |
| CN104992453A (en) * | 2015-07-14 | 2015-10-21 | 国家电网公司 | Target tracking method under complicated background based on extreme learning machine |
| CN106296734A (en) * | 2016-08-05 | 2017-01-04 | 合肥工业大学 | Based on extreme learning machine and the target tracking algorism of boosting Multiple Kernel Learning |
-
2017
- 2017-01-20 CN CN201710047829.0A patent/CN106815576B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8254633B1 (en) * | 2009-04-21 | 2012-08-28 | Videomining Corporation | Method and system for finding correspondence between face camera views and behavior camera views |
| CN102663453A (en) * | 2012-05-03 | 2012-09-12 | 西安电子科技大学 | Human motion tracking method based on second generation Bandlet transform and top-speed learning machine |
| CN103942749A (en) * | 2014-02-24 | 2014-07-23 | 西安电子科技大学 | Hyperspectral ground feature classification method based on modified cluster hypothesis and semi-supervised extreme learning machine |
| CN104992453A (en) * | 2015-07-14 | 2015-10-21 | 国家电网公司 | Target tracking method under complicated background based on extreme learning machine |
| CN106296734A (en) * | 2016-08-05 | 2017-01-04 | 合肥工业大学 | Based on extreme learning machine and the target tracking algorism of boosting Multiple Kernel Learning |
Non-Patent Citations (2)
| Title |
|---|
| 何传阳等: "基于智能监控的中小人群异常行为检测", 《计算机应用》 * |
| 朱博等: "基于极限学习机(ELM)的视线落点估计方法", 《东北大学学报(自然科学版)》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107423762A (en) * | 2017-07-26 | 2017-12-01 | 江南大学 | Semi-supervised fingerprinting localization algorithm based on manifold regularization |
| CN109255321A (en) * | 2018-09-03 | 2019-01-22 | 电子科技大学 | A kind of visual pursuit classifier construction method of combination history and instant messages |
| CN109255321B (en) * | 2018-09-03 | 2021-12-10 | 电子科技大学 | Visual tracking classifier construction method combining history and instant information |
| CN110211104A (en) * | 2019-05-23 | 2019-09-06 | 复旦大学 | A kind of image analysis method and system for pulmonary masses computer aided detection |
| CN110211104B (en) * | 2019-05-23 | 2023-01-06 | 复旦大学 | Image analysis method and system for computer-aided detection of lung mass |
| CN110675382A (en) * | 2019-09-24 | 2020-01-10 | 中南大学 | Aluminum electrolysis superheat degree identification method based on CNN-LapseLM |
| CN113378673A (en) * | 2021-05-31 | 2021-09-10 | 中国科学技术大学 | Semi-supervised electroencephalogram signal classification method based on consistency regularization |
| CN113378673B (en) * | 2021-05-31 | 2022-09-06 | 中国科学技术大学 | Semi-supervised electroencephalogram signal classification method based on consistency regularization |
| CN113408984A (en) * | 2021-06-21 | 2021-09-17 | 北京思路智园科技有限公司 | Hazardous chemical substance transportation tracking system and method |
| CN113408984B (en) * | 2021-06-21 | 2024-03-22 | 北京思路智园科技有限公司 | Dangerous chemical transportation tracking system and method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106815576B (en) | 2020-07-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106815576A (en) | Target tracking method based on consecutive hours sky confidence map and semi-supervised extreme learning machine | |
| CN107203781B (en) | End-to-end weak supervision target detection method based on significance guidance | |
| CN106778604B (en) | Pedestrian re-identification method based on matching convolutional neural network | |
| CN103871029B (en) | A kind of image enhaucament and dividing method | |
| CN106997597B (en) | It is a kind of based on have supervision conspicuousness detection method for tracking target | |
| CN109543606A (en) | A kind of face identification method that attention mechanism is added | |
| CN107506722A (en) | One kind is based on depth sparse convolution neutral net face emotion identification method | |
| CN102256065B (en) | Automatic video condensing method based on video monitoring network | |
| CN102722714B (en) | Artificial neural network expanding type learning method based on target tracking | |
| CN108053419A (en) | Inhibited and the jamproof multiscale target tracking of prospect based on background | |
| CN103942557B (en) | A kind of underground coal mine image pre-processing method | |
| CN106650690A (en) | Night vision image scene identification method based on deep convolution-deconvolution neural network | |
| CN106530246A (en) | Image dehazing method and system based on dark channel and non-local prior | |
| CN107341480A (en) | A kind of crowd massing detection method of modified PCCNN neural network models | |
| CN107025420A (en) | The method and apparatus of Human bodys' response in video | |
| CN110110649A (en) | Alternative method for detecting human face based on directional velocity | |
| CN104751466B (en) | A kind of changing object tracking and its system based on conspicuousness | |
| CN107833239B (en) | Optimization matching target tracking method based on weighting model constraint | |
| CN108009493A (en) | Face anti-fraud recognition methods based on action enhancing | |
| CN105825233B (en) | A kind of pedestrian detection method based on on-line study random fern classifier | |
| CN107833241A (en) | To real-time vision object detection method of the ambient lighting change with robustness | |
| CN106611421A (en) | SAR image segmentation method based on feature learning and sketch line constraint | |
| CN107895379A (en) | The innovatory algorithm of foreground extraction in a kind of video monitoring | |
| CN113139501B (en) | Pedestrian multi-attribute identification method combining local area detection and multi-level feature grabbing | |
| CN104021544A (en) | Greenhouse vegetable disease surveillance video key frame extracting method and extracting system |
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 |