CN106504269A - A kind of method for tracking target of many algorithm cooperations based on image classification - Google Patents
A kind of method for tracking target of many algorithm cooperations based on image classification Download PDFInfo
- Publication number
- CN106504269A CN106504269A CN201610917111.8A CN201610917111A CN106504269A CN 106504269 A CN106504269 A CN 106504269A CN 201610917111 A CN201610917111 A CN 201610917111A CN 106504269 A CN106504269 A CN 106504269A
- Authority
- CN
- China
- Prior art keywords
- target
- change
- image
- tracking
- algorithm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- 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
- G06T2207/20212—Image combination
- G06T2207/20224—Image subtraction
Landscapes
- Image Analysis (AREA)
Abstract
A kind of method for tracking target of many algorithm cooperations based on image classification, including step:(1) all of target component is initialized;(2) judge whether current goal is front cross frame:If so, BW SOAMS algorithm keeps track targets are adopted;If it is not, the image being input into after the second frame, then judge to classify to image sequence;(3) corresponding Moving Target Tracking Algorithm is called according to classification;(4) real-time update of target's feature-extraction, sample and object candidate area;(5) position of tracking target is given, next frame is read in, b is continued through and is judged classification tracking until the last frame of video sequence.
Description
Technical field
The present invention relates to target tracking domain, in particular to a kind of target of many algorithm cooperations based on image classification
Tracking.
Background technology
Target following, as a cross-cutting cutting edge technology interdisciplinary, is one of key problem of computer vision.Mesh
Mark tracking technique started to develop from the sixties in last century, till now, had defined a series of method.Due to tracked mesh
The complexity of the multiformity and external environment condition of the change of specimen body, motion target tracking are one and are rich in challenging problem.One
The target tracking algorism of individual robust has to preferably to solve all difficulties that tracking process runs into (such as rotationally-varying, chi
Very little change, illumination variation, block change, context similarity change etc.).Classical target tracking algorism mainly has average drifting to calculate
Method, optical flow method, Kalman filtering method and sparse matrix such as represent at the method, every kind of classical Moving Target Tracking Algorithm respectively have from
Oneself pluses and minuses, and present Moving Target Tracking Algorithm is essentially all to be improved based on above-mentioned classic algorithm or sent out
Open up and come.Although, there is substantial amounts of target tracking algorism in target tracking domain, can also realize the fortune under certain complex environment
The tracking of moving-target.But a good wide target tracking algorism of environment that adapts to of robust performance is still computer vision
The research topic of one great challenge.The present invention is processed to sequence of video images from the source of tracking image sequence,
With according to the difference because playing image change reason, classifying to target to be tracked and background before target following in advance.
Because two classes being can be largely classified into the reason for playing environmental change in object tracking process, that is, track target Self-variation
With the change of tracking environmental, we have found that in actual test and mainly have tracking target because of the factor for playing target Self-variation
Change and change in size is blocked, because playing the main by illumination variation and context similarity change etc. of environmental change.Become according to two kinds
Change image was not divided into before tracking two classes with us because an image change scope:Image overall change and local are interested
Regional change.Person's appropriate algorithm is not then selected to be tracked positioning to video sequence, so for the bar because playing varying environment change
Part goes to track the real-time and robustness that target ensures tracking.
The present invention is have been directed to not using many algorithm fusion Moving Target Tracking Algorithms based on image classification, well reply
With tracking problem caused by factor change, experiments verify that, the present invention is than other epidemic algorithms instantly in real-time and robust
All it is of a relatively high in property.
Sparse matrix definition mathematically is:If in a matrix, most elements is 0, this matrix is called sparse square
Battle array (sparse matrix).Sparse representation method provides new thinking for image processing in recent years.Research shows visual cortex
The expression of complex stimulus uses sparse coding principle, and the sparse representation method based on sparse coding can preferably portray people
Cognitive features of the class visual system to image.Sparse coding is carried out to substantial amounts of image feature information using sparse matrix, while
Dimensionality reduction compression is carried out to the vector information of sparse coding, algorithm is greatly reduced and is obtained amount of calculation, improve the operational efficiency of algorithm.
Mean shift algorithm is a kind of nonparametric technique based on density gradient, finds target location by interative computation,
Realize target following.The advantage of algorithm is that principle is simple, operational efficiency is high therefore wide to blocking certain adaptive ability
General applies in target tracking domain.Meanwhile, mean shift algorithm is by calculating probability distribution density due to the calculating of itself
To realize target following, therefore, when the size for tracking target is bigger, the more algorithms of the pixel quantity that will be counted and calculate
The bigger efficiency of amount of calculation is lower.Simultaneously as average drift algorithm uses the mode based on color characteristic, quick to illumination
Sense, can cause tracking drift in the case where background is similar to color of object.
Shortcoming is had according to mean shift algorithm, and combines the pretreatment of the image sequence of early stage of the present invention and classified, can be with
Mean shift algorithm is cleverly avoided to tracking shortcoming the problems such as illumination and similar background, which is taken full advantage of to blocking
Robustness, and efficient operation efficiency.
Content of the invention
The purpose of the application there are provided a kind of method for tracking target of many algorithm cooperations based on image classification, including
Step:(1) all of target component is initialized;(2) judge whether current goal is front cross frame:If so, calculated using BW-SOAMS
Method tracks target;If it is not, being input into t (t>2) two field picture, then judge to classify to image sequence;(3) adjusted according to classification
Use corresponding Moving Target Tracking Algorithm;(4) real-time update of target's feature-extraction, sample and object candidate area;(5) give
Go out to track the position of target, read in next frame, continue through b and judge classification tracking until the last frame of video sequence.
Preferably, image sequence is carried out judging that the method that is classified is in the step 2:
P=N2/N1
Wherein, the total pixel number of image is defined as N, the pixel change number of self-defined threshold binarization treatment after frames differencing
For N1, the pixel change number in target three times regional extent is N2, solves zones of different pixel change number respectively and accounts for global pixel
Proportion P.
Preferably, step 2 specific classification standard is:Target local is divided into by proportion P to image sequence interested
Regional change and image overall change, wherein, the change of target local region of interest is called and incorporates the yardstick for improving background weighting
The Moving Target Tracking Algorithm of direction-adaptive;The Moving Target Tracking Algorithm based on compression tracking is called in image overall change.
Preferably, the influence factor of the image of described image global change is illumination variation, context similarity and background
Fuzzy.
Preferably, change in size, rotationally-varying of the influence factor of target local region of interest change for target
And block change.
According to the pluses and minuses of the sparse matrix present invention according to image sequence early stage with treatment classification to being divided into the overall situation
The video segment of image sequence carries out target following, sufficiently make use of sparse matrix to represent the advantage of algorithm, reduces target
The dimension of following calculation, greatly reduces because of the excessive caused real-time performance of tracking of the amount of calculation of algorithm caused by image overall change
Ungratified problem.Simultaneously because rarefaction representation algorithm exist when blocking occurs in target following or the adjacent frame position movement of target compared with
Tracking drift can be caused when big, and this kind of change is drawn by the image classification process through early stage in the algorithm of the present invention
The localized variation of image has been divided into it, has dexterously avoided the shortcoming of the algorithm, the computational efficiency for having given full play to its algorithm is high
Advantage.Make algorithm that there is in object tracking process the robustness of higher operational efficiency and Geng Gao.
It should be appreciated that aforementioned description substantially and follow-up description in detail are exemplary illustration and explanation, should not
It is used as the restriction to claimed content of the invention.
Description of the drawings
With reference to the accompanying drawing that encloses, the present invention more purpose, function and advantages are by by the as follows of embodiment of the present invention
Description is illustrated, wherein:
The flow chart that Fig. 1 shows the method for tracking target of many algorithm cooperations based on image classification according to the present invention;
Fig. 2 is shown according to the of the invention method for tracking target cooperated based on many algorithms of image classification based on image
The sequence of video images figure of global change;
Fig. 3 is shown according to the of the invention target tracking algorism cooperated based on many algorithms of image classification based on image
The sequence of video images figure of localized variation;
Fig. 4 shows three times of the target of the target tracking algorism of many algorithm cooperations based on image classification according to the present invention
Area image change pixel count and image overall pixel count percentage change figure.
Specific embodiment
By reference to one exemplary embodiment, the purpose of the present invention and function and the side for realizing these purposes and function
Method will be illustrated.However, the present invention is not limited to one exemplary embodiment disclosed below;Can by multi-form come
Which is realized.The essence of description is only to aid in the detail of the various equivalent modifications Integrated Understanding present invention.
Hereinafter, embodiments of the invention will be described with reference to the drawings.In the accompanying drawings, identical reference represents identical
Or similar part, or same or like step.
As shown in figure 1, the target of many algorithm cooperations based on image classification of the present inventionTrackingStep is:
Step 101:Initialize all of target component;
Step 102:Judge whether current goal is front cross frame:If so, step 106 is entered;If it is not, being input into t (t>
2) two field picture, then judge to classify to image sequence;
According to one embodiment of present invention, the method for judging to be classified is carried out to image sequence in the step 102
For:
P=N2/N1
Wherein, the total pixel number of image is defined as N, the pixel change number of self-defined threshold binarization treatment after frames differencing
For N1, the pixel change number in target three times regional extent is N2, solves zones of different pixel change number respectively and accounts for global pixel
Proportion P.
According to one embodiment of present invention, step 2 specific classification standard is:By proportion P to image sequence point
It is that the change of target local region of interest and image overall change, wherein, the change of target local region of interest is called to incorporate and changed
Enter the adaptive Moving Target Tracking Algorithm of dimension of background weighting;The fortune based on compression tracking is called in image overall change
Tracking of maneuvering target algorithm.
According to one embodiment of present invention, the influence factor of the image of described image global change is illumination variation, the back of the body
Scape similarity and blurred background.
According to one embodiment of present invention, chi of the influence factor of the target local region of interest change for target
Very little change, rotationally-varying and block change.
Step 103:Corresponding Moving Target Tracking Algorithm is called according to classification;
Step 104:The real-time update of target's feature-extraction, sample and object candidate area;
Step 105:Be given tracking target position, read in next frame, continue through step 102 judge classification tracking until
The last frame of video sequence.
Step 106:Using BW-SOAMS algorithm keeps track targets.
Fig. 2 and Fig. 3 is respectively the base of the target tracking algorism of many algorithm cooperations based on image classification according to the present invention
In image overall change and the sequence of video images figure of localized variation;
As illustrated, step is:
A. pretreatment classification is carried out to image sequence;
The reason for according to causing pixel to change, statistical computation is carried out to the amount of pixels for changing, solve change pixel and account for the overall situation
The proportion of pixel;
B. by the catch cropping pin difference method of adjacent image frame, the image of frame after the recovery is obtained;
C. the binary image for carrying out self-defined threshold value to image-region is processed;
Wherein, image sequence is divided into by two classes according to the difference of the order of magnitude:
One class is the change based on image overall;Change is mainly by illumination variation, context similarity and blurred background
Cause;
Another kind of is change based on image local area-of-interest, and change is mainly by change in size, the rotation of target
Change and block what change caused.
Target is tracked according to Selection and call corresponding target tracking algorism is processed.
The pixel change of statistics target three times region and image overall, the pixel count combined using local and the overall situation are become
Change to count to efficiently avoid and changed the phenomenon for being categorized as global change for acutely causing mistake by local pixel.
Image sequence pretreatment classification is by the frame difference operation between adjacent two field picture, obtains the image of frame after the recovery.So
The image binaryzation for carrying out self-defined threshold value (threshold value is 0.2 in inventive algorithm) afterwards to image-region is processed.By testing
Card, the proportion for accounting for global pixel using change pixel are classified to image sequence, it is ensured that certain classification accuracy, but
Be, when target local region of interest pixel a large amount of changes (mainly seriously blocked by target itself or target size
Occur what serious change caused) when, by the way of the above-mentioned proportion that global pixel is simply accounted for using change pixel, image will occur
Classification error, mistake are divided into global change.Therefore, the pixel of present invention statistics target three times region and image overall
Change, the pixel count change statistics combined using local and the overall situation can be prevented effectively from be changed by local pixel and acutely cause mistake
It is mistakenly classified as the phenomenon of global change.
Assume image total pixel number we be defined as N, change of the image by the self-defined threshold binarization treatment of frame after the recovery
Change pixel count is N1, and in the range of the three times area of target area, pixel change number is N2, solves change pixel respectively and accounts for global picture
The proportion of element.The present invention carries out verification experimental verification to the representational image sequence fragment of 20 the international standards organization standard generalized markup data bases, schemes
Photo section and statistical conditions such as table 1 below.
The present invention changes number N2 by the pixel of comparison object three times surface area and changes number N1 in image overall pixel, fixed
Adopted P is N2/N1, and by being compared statistics to above-mentioned 20 sections of sequence of pictures, the span of P is 1 to 100%, and the present invention is with regard to P
Relation of the value between the accuracy rate of image classification make corresponding curve, observe when P values from 1% to 100% when scheme
As classification accurate between relation curve, the curve between accuracy rate and ratio is as shown in Figure 4.
By the relation curve between accuracy rate and ratio P, it can be seen that the standard of image classification when P takes 10% or 11%
Really 100%, therefore we choose 10% as categorised demarcation line.The present invention is judged in global method using local is planned as a whole
Image local change and global change.Effectively avoid because localized target region due to target itself situation (as:Large area
Or all block, change in size, rotationally-varying) acute variation and the probability for being categorized as global change of mistake, can be with standard
Image is divided into two classes really.The high moving target of suitable robustness is not selected us for pixel change principle is caused
Track algorithm is tracked to moving target, so as to not only improve the operation efficiency of algorithm but also improve the robustness of algorithm.
Sequence of video images is accurately divided into by two classes by the classification of draw above shape, i.e., based on regarding that image overall changes
Frequency image sequence and the sequence of video images changed based on image local, it is corresponding suitable to be selected according to different video sequences
Moving Target Tracking Algorithm.
Dimension self adaptation mean shift algorithm using based on background weighting of the invention is (referred to as:BW-SOAMS) conduct
The corresponding track algorithm of local region of interest image sequence.Background weighting algorithm is fused in SOAMS algorithms, is effectively kept away
Exempted from by local region of interest context similarity change, blurred background change and background environment illumination slowly varying
Tracking drifting problem caused by the change interested of local is classified as during identification and classification.While ensure that algorithm real-time, protect again
The high robust of target following is demonstrate,proved.
The invention discloses a kind of mobile target in complex background track algorithm.Due to target itself and environmental change
, during causing tracking, there are all difficulties in multiformity.Different Moving Target Tracking Algorithms respectively has excellent for different problems
Shortcoming, the target tracking algorism for also not having a kind of pervasive robustness good up to now.Become for object variations and background environment
Change, it is proposed that a kind of target tracking algorism of many calculation cooperations based on image classification.The change of target itself causes local to feel emerging
The problem of interesting regional change mainly by the size of target object, rotate and block and cause.It is complete that background environment change causes
Office's variation issue is mainly caused by illumination, context similarity and blurred background.The algorithm is first to image sequence pretreatment point
Class, then selects to be suitable for tracking drifting problem caused by the environmental change.Experiments verify that, the present invention is than other popular mesh
Mark track algorithm has more preferable robustness.
Compared with the prior art, the target tracking algorism of many algorithm cooperations based on image classification proposed by the present invention has
Beneficial effect is embodied in:
(1) present invention is processed to sequence of video images from the source of tracking image sequence, with according to because playing image
The difference of reason of changes, classifies to target to be tracked and background before target following in advance.
(2) because can be largely classified into two classes the reason for playing environmental change in object tracking process, that is, target itself is tracked
Change and the change of tracking environmental, we have found that in actual test and mainly have tracking mesh because of the factor for playing target Self-variation
Target blocks change and change in size, because playing the main by illumination variation and context similarity change etc. of environmental change.According to two
Kind of change image was not divided into before tracking two classes with us because of image change scope:Feel in image overall change and local
Interest regional change.Then do not select person's appropriate algorithm that positioning is tracked to video sequence, be so directed to because playing varying environment change
Condition go track target ensure tracking real-time and robustness.
In conjunction with the explanation and practice of the present invention for disclosing here, the other embodiment of the present invention is for those skilled in the art
All will be readily apparent and understand.Illustrate and embodiment be to be considered only as exemplary, the present invention true scope and purport equal
It is defined in the claims.
Claims (5)
1. the method for tracking target that a kind of many algorithms based on image classification cooperate, including step:
(1) all of target component is initialized;
(2) judge whether current goal is front cross frame:If so, BW-SOAMS algorithm keeps track targets are adopted;If it is not, i.e. input the
Image after two frames, then judge to classify to image sequence;
(3) corresponding Moving Target Tracking Algorithm is called according to classification;
(4) real-time update of target's feature-extraction, sample and object candidate area;
(5) position of tracking target is given, next frame is read in, b is continued through and is judged that classification tracking is last until video sequence
One frame.
2. method according to claim 1, it is characterised in that:Carrying out judgement in the step 2 to image sequence is carried out point
The method of class is:
P=N2/N1
Wherein, the total pixel number of image is defined as N, and after frames differencing, the pixel change number of self-defined threshold binarization treatment is
N1, the pixel change number in target three times regional extent is N2, solves zones of different pixel change number respectively and accounts for global pixel
Proportion P.
3. method according to claim 2, it is characterised in that:Step 2 specific classification standard is:By proportion P pair
Image sequence is divided into the change of target local region of interest and image overall changes, and wherein, target local region of interest changes
Call and incorporate the adaptive Moving Target Tracking Algorithm of the dimension for improving background weighting;Image overall change is called based on pressure
The Moving Target Tracking Algorithm of contracting tracking.
4. method according to claim 3, it is characterised in that:The influence factor of the image of described image global change is light
According to change, context similarity and blurred background.
5. method according to claim 3, it is characterised in that:The influence factor of the target local region of interest change
For the change in size of target, rotationally-varying and block change.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610917111.8A CN106504269B (en) | 2016-10-20 | 2016-10-20 | A kind of method for tracking target of more algorithms cooperation based on image classification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610917111.8A CN106504269B (en) | 2016-10-20 | 2016-10-20 | A kind of method for tracking target of more algorithms cooperation based on image classification |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106504269A true CN106504269A (en) | 2017-03-15 |
CN106504269B CN106504269B (en) | 2019-02-19 |
Family
ID=58318139
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610917111.8A Active CN106504269B (en) | 2016-10-20 | 2016-10-20 | A kind of method for tracking target of more algorithms cooperation based on image classification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106504269B (en) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101739551A (en) * | 2009-02-11 | 2010-06-16 | 北京智安邦科技有限公司 | Method and system for identifying moving objects |
CN101901354A (en) * | 2010-07-09 | 2010-12-01 | 浙江大学 | Method for detecting and tracking multi targets at real time in monitoring videotape based on characteristic point classification |
CN102236901A (en) * | 2011-06-30 | 2011-11-09 | 南京大学 | Method for tracking target based on graph theory cluster and color invariant space |
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 |
CN105654505A (en) * | 2015-12-18 | 2016-06-08 | 中山大学 | Collaborative tracking algorithm based on super-pixel and system thereof |
CN105913455A (en) * | 2016-04-11 | 2016-08-31 | 南京理工大学 | Local image enhancement-based object tracking method |
-
2016
- 2016-10-20 CN CN201610917111.8A patent/CN106504269B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101739551A (en) * | 2009-02-11 | 2010-06-16 | 北京智安邦科技有限公司 | Method and system for identifying moving objects |
CN101901354A (en) * | 2010-07-09 | 2010-12-01 | 浙江大学 | Method for detecting and tracking multi targets at real time in monitoring videotape based on characteristic point classification |
CN102236901A (en) * | 2011-06-30 | 2011-11-09 | 南京大学 | Method for tracking target based on graph theory cluster and color invariant space |
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 |
CN105654505A (en) * | 2015-12-18 | 2016-06-08 | 中山大学 | Collaborative tracking algorithm based on super-pixel and system thereof |
CN105913455A (en) * | 2016-04-11 | 2016-08-31 | 南京理工大学 | Local image enhancement-based object tracking method |
Also Published As
Publication number | Publication date |
---|---|
CN106504269B (en) | 2019-02-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Le et al. | Video salient object detection using spatiotemporal deep features | |
Jiang et al. | Crowd counting and density estimation by trellis encoder-decoder networks | |
Stallkamp et al. | Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition | |
CN110503613B (en) | Single image-oriented rain removing method based on cascade cavity convolution neural network | |
Le et al. | Deeply Supervised 3D Recurrent FCN for Salient Object Detection in Videos. | |
CN107633226A (en) | A kind of human action Tracking Recognition method and system | |
Xu et al. | Video salient object detection via robust seeds extraction and multi-graphs manifold propagation | |
CN111986180B (en) | Face forged video detection method based on multi-correlation frame attention mechanism | |
CN105512618B (en) | Video tracing method | |
CN103093198A (en) | Crowd density monitoring method and device | |
CN104866843B (en) | A kind of masked method for detecting human face towards monitor video | |
CN110163057B (en) | Object detection method, device, equipment and computer readable medium | |
CN110909741A (en) | Vehicle re-identification method based on background segmentation | |
Wang et al. | Background extraction based on joint gaussian conditional random fields | |
Chen et al. | Salbinet360: Saliency prediction on 360 images with local-global bifurcated deep network | |
Tsai et al. | Joint detection, re-identification, and LSTM in multi-object tracking | |
Gehrig et al. | A real-time multi-cue framework for determining optical flow confidence | |
Shivakumara et al. | A new iterative-midpoint-method for video character gap filling | |
Dai et al. | Effective moving shadow detection using statistical discriminant model | |
Luo et al. | Rain-like layer removal from hot-rolled steel strip based on attentive dual residual generative adversarial network | |
CN106504269A (en) | A kind of method for tracking target of many algorithm cooperations based on image classification | |
Zhang et al. | Global guidance-based integration network for salient object detection in low-light images | |
Wang et al. | Multi-scale features fused network with multi-level supervised path for crowd counting | |
Chuprov et al. | Are industrial ml image classifiers robust to withstand adversarial attacks on videos? | |
Yan et al. | SDCNet: size divide and conquer network for salient object detection |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | 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 |