CN108898621A - A kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window - Google Patents
A kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window Download PDFInfo
- Publication number
- CN108898621A CN108898621A CN201810662407.9A CN201810662407A CN108898621A CN 108898621 A CN108898621 A CN 108898621A CN 201810662407 A CN201810662407 A CN 201810662407A CN 108898621 A CN108898621 A CN 108898621A
- Authority
- CN
- China
- Prior art keywords
- target
- window
- suggests
- perception
- detection
- 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
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
A kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window, can in the correlation filtering frame based on CNN according to the stability self-adaptive of detection selecting scale estimation model and target re-detection mode, improve algorithm move, block in size estimation, quickly, in terms of robustness.The target suggestion window generated by EdgeBoxes is all that the target for having high similarity with object instance suggests window, referred to as example perception target suggestion window based on what is obtained after apparent similarity and spatial weighting sequence.The target of example perception suggests that window is further directed to optimal location by the correlation filter based on CNN, therefrom choose it is most significant it is guided after example perception target suggest window, size estimation or re-detection as target is as a result, the dimensional variation and target that can efficiently solve during tracking are lost.On standard data set, the method for proposition obtains very high performance indicator.
Description
Technical field
The present invention relates to computer vision techniques, suggest the correlation of window more particularly, to a kind of Case-based Reasoning perception target
Filter tracking method.
Background technique
The mankind to extraneous video have very high visual ability, brain can to the specific objective in particular video frequency into
Row quickly and accurately tracks.Computer will imitate the visual ability of the mankind it is necessary to reach people in speed and precision
The level of class.Target following is under the jurisdiction of computer vision, is the basic content of visual perception, its speed and precision determines
The real-time and accuracy of visual perception.Modern target following technology is widely used in video monitoring, human-computer interaction, virtually shows
The fields such as reality, robot navigation, video compress, public safety, if computer, which has, is similar to the mankind in particular video frequency
The tracking ability of specific objective saves a large amount of time cost, warp then can replace the mankind in the field that is widely applied
Cost of helping and human cost.Therefore, the target following technology in computer vision is furtherd investigate, its precision is continuously improved, speed
Degree is constantly promoted, and is had great theoretical and practical significance.Target following solve a basic problem be：It is regarded first at one
The target to be tracked is initialized in the initial frame of frequency, then in next each video frame, determines tracked target
State：That is the position of target and scale.Target following is a challenging research topic, during tracking, can be met
Rotation inside and outside to illumination variation, dimensional variation, plane deformation, is blocked, background interference, quickly movement, camera shake, low resolution
Various challenges, these challenging objectives tracing tasks such as rate, the disengaging visual field become more difficult, design one kind and all compare all kinds of challenges
Target tracking algorism compared with robust is that the emphasis of target tracking domain is also difficult point.
Currently, target tracking algorism is usually divided into two classes with apparent model by researcher both at home and abroad：Production model with
Discriminative model.Production model models target area in present frame, and next frame is found and the most like region of model,
The region by tracking target predicted state, such method has Kalman filtering, particle filter, average drifting etc., this kind of side
Method only considers the foreground information of target, does not account for the background information of target, and tracking accuracy is not high.Discriminative model is by target
Tracking is considered as a classification problem, extracts target information and background information in the current frame to train classifier, in the next frame
Target is separated from background with trained classifier, to obtain the predicted state of tracked target, such methods
Target and background information is considered simultaneously, there is apparent advantage in existing tracking.In recent years, it is based on correlation filtering
Method for tracking target respectively due to its speed advantage becomes the main flow direction studied at present, the method based on correlation filtering belongs to
The scope of discriminative model.
Method based on correlation filtering becomes one of the research hotspot of target tracking domain in recent years, such method is due to it
Speed advantage all yields good result in standard data set and every match.2012,F.Henriques is mentioned
The core correlation filtering for having gone out high speed, opens the new era of correlation filtering.Correlation filtering is a kind of based on detection
Tracking, in the study stage, target and surrounding background area construct training sample by cycle spinning, by this
A little training samples revert to a Gaussian function by ridge regression, to learn to a correlation filter；In detection-phase, learn
The candidate target of the correlation filter and region of search that arrive carries out related operation, selects the maximum candidate target of response as tracking
As a result.Due to entirely realizing that process is transformed into frequency domain, the convolution of high time complexity by Fast Fourier Transform (FFT) by time-domain
Operation is converted to the point multiplication operation of low time complexity, so speed is quickly.
Although correlation filtering speed is quickly, there are problems that the following aspects：It can not carry out size estimation；It deposits
In boundary effect；The peak value of response is not prominent enough；The inadequate robust of feature used；Search range is small to be easily lost target.?
In terms of size estimation, 2014, Martin Danelljan etc. proposed DSST, in addition to the 2D of one estimated location of training is related
Outside filter, the 1D correlation filter of also additional one estimation scale of training, to a certain extent can scale to target into
Row ART network.In order to solve the Boundary Effect problem of correlation filtering, 2015, Martin Danelljan etc. was proposed
SRDCF, due to punishing the introducing of weight, was entirely realized by alleviating boundary effect to the punishment weight increased to boundary
Journey is carried out in time-domain, greatly reduces the speed of service；2017, Matthias Mueller etc. proposed CACF, will be upper and lower
The sample that text is constituted is used to train correlation filter together with original target sample, significantly improves while guaranteeing real-time
Precision.In order to enhance the separating capacity of correlation filter, 2017, Rui Yao etc. proposed RCF, on the basis of L2 norm
L1 norm is introduced, the peak value of response diagram is highlighted, enhances the ability that correlation filter distinguishes target and background.In feature Shandong
In terms of stick, 2014, Martin Danelljan etc. added additional color characteristic on the basis of original HOG feature
(Color Naming) goes characterization target, improves the robustness of algorithm；2016, Luca Bertinetto etc. was proposed
Staple, color histogram and HOG feature are effectively fused together, and the robust of algorithm is improved while guaranteeing real-time
Property；2015, Chao Ma etc. proposed HCF, and the depth characteristic of different layers is incorporated multiple correlation filter collaborations and completes tracking
Task, since the introducing of depth characteristic is so that the precision of algorithm is significantly promoted, using GPU, speed is reachable
10FPS.Correlation filtering search range is small be easily lost target aiming at the problem that, 2015, Zhibin Hong etc. was proposed
MUSTer is carried out short distance tracking based on correlation filtering, is carried out based on Feature Points Matching with the long-range re-detection after losing, Chao Ma
Etc. proposing LCT, after location estimation and size estimation that target is carried out using correlation filtering, on-line training is based on SVM classifier
Re-detection is carried out, due to the introducing of re-detection function, so that tracking performance is significantly promoted.In addition, 2016, Martin
Correlation filtering is expanded to continuous domain from discrete domain and proposes CCOT by Danelljan etc., and tracking accuracy is significantly promoted, and speed is big
Amplitude decline, only 0.3FPS or so, subsequent Martin Danelljan etc., which improves CCOT, proposes ECO, and precision mentions
Speed is also promoted while liter.
Summary of the invention
The purpose of the present invention is to provide can in the correlation filtering frame based on CNN it is adaptive according to the stability of detection
Ground selecting scale estimation model and target re-detection mode are answered, algorithm is improved and moves, blocks, background in size estimation, quickly
The robustness of interference etc. can efficiently solve one kind that dimensional variation and target during tracking are lost and be based in fact
Example perception target suggests the correlation filtering tracking of window.
The present invention includes the following steps：
1) for the i-th frame video, the Initial state estimation of target is carried out by the correlation filter based on CNN, obtains phase
Close filter response figure；
In step 1), for the i-th frame video, estimated by the original state that the correlation filter based on CNN carries out target
The specific method of meter can be：Target is detected in correlation filtering region of search based on the correlation filter of CNN and obtains correlation filtering
Response diagram, original state of the state as target corresponding to maximum value, is denoted as (l in response diagramic,siC), wherein liC is indicated
Position, siC indicates scale.
2) to correlation filtering response diagram defined in step 1), the stability of detection is calculated, determines mould locating for present frame
Formula；
In step 2), described to calculate the stability of detection to correlation filtering response diagram defined in step 1), determination is worked as
The specific method of mode locating for previous frame can be：The stability of detection is codetermined by two indices, and one is in response in figure
Maximum value is denoted as Rmax, another is in response to figure peak sidelobe ratio, is denoted as PSR, and defined formula (1) is as follows：
Wherein, RμAnd RσRespectively the mean value and variance of response diagram, mode determining method are as follows：WhenOr
PersonWhen, start target re-detection mode, otherwise start size estimation mode, whereinFor RmaxMean value,For the mean value of PSR；The mode can be size estimation or target re-detection；λ=0.72.
3) target being extracted in the search range corresponding to the mode that step 2) determines and suggesting window, target is suggested into window
Suggest window as candidate target；Since EdgeBoxes suggests in method that window generates that recall rate is high, speed in target
Fastly, therefore target is generated based on EdgeBoxes and suggests window；
4) window is suggested to target obtained in step 3), based on the apparent similarity and spatial weighting between example set
It is ranked up, the target for selecting example perception suggests window, equipped with having M example, E={ e in example set E1,e2,…,eM,
Suggest thering is N number of target to suggest window, P={ p in window collection P in target1,p2,…,pN}；Each target suggest window with it is each
The apparent similarity calculating method of a example is respectively：
Color similarity：I-th of target suggests window piWith j-th of example ejBetween color similarityDefinition
For the COS distance between their color histogram vectors, defined formula (2) is as follows：
Wherein, his () indicates that target suggests the color histogram vector of window or example；
Shape similarity：I-th of target suggests window piWith j-th of example ejBetween shape similarityDefinition
For the COS distance between their HOG feature vectors, defined formula (3) is as follows：
Wherein, hog () indicates that target suggests the HOG feature vector of window or example；
Spatial weighting：Space weight wiIt is defined as i-th of target and suggests window piWith dbjective state o according to a preliminary estimatecBetween
Jie Lade distance, defined formula (4) is as follows：
Wherein, box () indicates that target suggests the bounding box of window or dbjective state according to a preliminary estimate.
Since above-mentioned color similarity and shape similarity are two complementary apparent similarities, i-th of target is built
Discuss the ballot score v of windowiDefined formula (5) is as follows：
Based on viWindow sequence is suggested to target, chooses the forward target that sorts and suggests that window is built as example perception target
Discuss window；
It is described based on apparent including color and shape etc. between example set in step 4)；Parameter θ=0.7.
5) window is suggested to the target of the perception of each example obtained in step 4), by the correlation filter based on CNN
Guidance to after optimum state, therefrom select response it is maximum it is guided after target suggest further shape of the window as target
State is denoted asWherein,Indicate position,Indicate scale；At this point,Corresponding maximum response and peak value
Secondary lobe ratio is denoted asAnd PSRp；
6) according to the detection stability being calculated in step 2), based on obtained in step 1)With in step 5)
It obtainsIt carries out adaptive targets positioning and adaptive model updates.
In step 6), the detection stability can be by RmaxIt is determined jointly with PSR；
The self-adapting objective locating method can be：Under re-detection mode, ifAndIt so indicates to be successfully detected target, the dbjective state of present frame isOtherwise, present frame
Dbjective state isUnder size estimation mode, the dbjective state of present frame is determined by the biggish dbjective state of response；
It is described to be from detection patternOr
The adaptive model update method can be：Under re-detection mode, indicate that target is lost, more without model
Newly；Under size estimation mode, when detecting high stability, model modification is carried out with learning rate η；Otherwise, learning rate is reduced
Model modification is carried out for c η；The re-detection mode can beOrThe detection stability
It is more a height ofAnd
The parameter lambda=0.72, μ=0.90, δ=1.00, η=0.01, c=0.7.
The present invention can according to the stability self-adaptive of detection selecting scale be estimated in the correlation filtering frame based on CNN
Meter mode and target re-detection mode, improve algorithm move, block in size estimation, quickly, in terms of Shandong
Stick.By EdgeBoxes generate target suggest window based on apparent similarity and spatial weighting sequence after obtain be all with
There is object instance the target of high similarity to suggest window, and referred to as example perception target suggests window.The mesh of these examples perception
Mark suggests that window by the correlation filter based on CNN is further directed to optimal location, therefrom choose it is most significant it is guided after
Example perception target suggest window, as target size estimation or re-detection as a result, can efficiently solve in this way with
Dimensional variation and target during track are lost.On standard data set, method proposed by the present invention obtains very high performance
Index.
Detailed description of the invention
Fig. 1 is the overall flow figure of the embodiment of the present invention.
Fig. 2 is the embodiment of the present invention and qualitative results ratio of other several method for tracking target on Girl2 and Lemming
Compared with figure.
Specific embodiment
It elaborates with reference to the accompanying drawings and examples to method of the invention.
Referring to Fig. 1, the embodiment of the present invention includes following steps：
A. for the i-th frame video, the Initial state estimation of target is carried out by the correlation filter based on CNN, obtains phase
Close filter response figure.The specific implementation process is as follows：Target is detected in correlation filtering region of search based on the correlation filter of CNN
Correlation filtering response diagram is obtained, original state of the state corresponding to maximum value as target, is denoted as in response diagramIts
In,Indicate position,Indicate scale.
B. to response diagram defined in step A, the stability of detection is calculated, determines mode locating for present frame：Scale is estimated
Meter or target re-detection.Detailed process is as follows：The stability of detection is codetermined by two indices, and one is in response in figure
Maximum value is denoted as Rmax, another is in response to figure peak sidelobe ratio, is denoted as PSR, and defined formula (1) is as follows：
Wherein RμAnd RσThe respectively mean value and variance of response diagram.Mode determining method is as follows：WhenOrWhen, start target re-detection mode, otherwise starts size estimation mode, whereinFor RmaxMean value,For the mean value of PSR.
C. target is extracted in the search range corresponding to the mode that step B is determined and suggests window, these targets are suggested
Window all suggests window as candidate target.Due to EdgeBoxes target suggest in method that window generates recall rate it is high,
Speed is fast, so generating target based on EdgeBoxes here suggests window.
D. window is suggested to target obtained in step C, based on apparent (including color and shape) between example set
Similarity and spatial weighting are ranked up, and the target for selecting example perception suggests window.Equipped with there is M example in example set E,
E={ e1,e2,…,eM, suggest thering is N number of target to suggest window, P={ p in window collection P in target1,p2,…,pN}.Each mesh
Mark suggests window and apparent (including color and shape) similarity calculating method of each example is respectively：
Color similarity：I-th of target suggests window piWith j-th of example ejBetween color similarityDefinition
For the COS distance between their color histogram vectors, defined formula (2) is as follows：
Wherein his () indicates that target suggests the color histogram vector of window or example.
Shape similarity：I-th of target suggests window piWith j-th of example ejBetween shape similarityDefinition
For the COS distance between their HOG feature vectors, defined formula (3) is as follows：
Wherein hog () indicates that target suggests the HOG feature vector of window or example.
Spatial weighting：Space weight wiIt is defined as i-th of target and suggests window piWith dbjective state o according to a preliminary estimatecBetween
Jie Lade distance, defined formula (4) is as follows：
Wherein box () indicates that target suggests the bounding box of window or dbjective state according to a preliminary estimate.
Since above color similarity and shape similarity are two complementary apparent similarities, i-th of target
It is recommended that the ballot score v of windowiDefined formula (5) is as follows：
Based on viWindow sequence is suggested to target.It chooses the forward target that sorts and suggests that window is built as example perception target
Discuss window.
E. window is suggested to the target of the perception of each example obtained in step D, is drawn by the correlation filter based on CNN
After being directed at optimum state, therefrom select response it is maximum it is guided after target suggest further shape of the window as target
State is denoted asWherein,Indicate position,Indicate scale.At this point,Corresponding maximum response and peak value
Secondary lobe ratio is denoted asAnd PSRp。
F. according to the detection stability being calculated in step B (by RmaxDetermined jointly with PSR), based on being obtained in step A
'sWith obtained in step ECarry out adaptive target positioning and model modification.Adaptive targets positioning side
Method is as follows：Re-detection mode (Or) under, ifAndIt so indicates to be successfully detected target, the dbjective state of present frame isOtherwise, present frame
Dbjective state isUnder size estimation mode, the dbjective state of present frame is determined by the biggish dbjective state of response.
Adaptive model update method is as follows：Re-detection mode (Or) under, indicate that target is lost
It loses, without model modification；Under size estimation mode, when detection high stability (And) when, model modification is carried out with learning rate η；Otherwise, reducing learning rate is that c η carries out model modification.
Parameter lambda=0.72 of definition in step B.
Parameter θ=0.7 defined in step D.
Parameter lambda=0.72, μ=0.90, δ=1.00, η=0.01, c=0.7 of definition in step F.
Fig. 2 is the embodiment of the present invention and qualitative results ratio of other several method for tracking target on Girl2 and Lemming
Compared with figure, the method for the present invention can trace into target well always.
Table 1
| Tracking | DeepCFIAP | MCPF | DeepSRDCF | HCF | HDT | CREST | SINT | SiamFC | KCFDPT | MUSTer |
| Precision (%) | 89.3 | 87.3 | 85.1 | 83.7 | 84.8 | 83.8 | 78.9 | 77.0 | 74.9 | 77.4 |
| Success rate (%) | 65.5 | 62.8 | 63.5 | 56.2 | 56.4 | 62.3 | 59.2 | 57.2 | 54.7 | 57.5 |
Table 1 is the precision and success rate of the present invention with comparison of other several method for tracking target in OTB100 data set,
Wherein DeepCFIAP is method of the invention.
DeepSRDCF corresponds to method (M.Danelljan, F.S.Khan, the and of M.Danelljann et al. proposition
M.Felsberg.2015.Convolutional Features for Correlation Filter Based Visual
Tracking.In Proceedings of the 2015IEEE International Conference on Computer
Vision Workshops.621–629.)；
HCF corresponds to method (C.Ma, J.B.Huang, X.Yang, the and of C.Ma et al. proposition
M.H.Yang.2015.Hierarchical Convolutional Features for Visual Tracking.In
Proceedings of the 2015IEEE International Conference on Computer Vision.3074-
3082.)；
HDT correspond to Y.Qi et al. proposition method (Y.Qi, S.Zhang, L.Qin, H.Yao, Q.Huang, J.Lim,
and M.H.Yang.2016.Hedged Deep Tracking.In Proceedings of the 2016IEEE
Conference on Computer Vision and Pattern Recognition.4303-4311.)；
SINT corresponds to method (R.Tao, E.Gavves, the and of R.Tao et al. proposition
A.W.M.Smeulders.2016.Siamese Instance Search for Tracking.In Proceedings of
the 2016IEEE Conference on Computer Vision and Pattern Recognition.1420-
1429.)；
SiamFC correspond to L.Bertinetto et al. proposition method (L.Bertinetto, J.Valmadre,
J.F.Henriques,A.Vedaldi,and P.H.S.Torr.2016.Fully-Convolutional Siamese
Networks for Object Tracking.In Proceedings of the 14th European Conference
on Computer Vision Workshops.850–865.)；
CREST correspond to Y.Song et al. proposition method (Y.Song, C.Ma, L.Gong, J.Zhang, R.W.H.Lau,
and M.H.Yang.2017.CREST:Convolutional Residual Learning for Visual
Tracking.In Proceedings of the 2017IEEE International Conference on
ComputerVision.2574–2583.)；
MCPF corresponds to method (T.Zhang, C.Xu, the and M.H.Yang.2017.Multi- of T.Zhang et al. proposition
task Correlation Particle Filter for Robust Object Tracking.In Proceedings of
the 2017IEEE Conference on Computer Vision and Pattern Recognition.4819–
4827.)；
KCFDPT corresponds to method (D.Huang, L.Luo, Z.Chen, M.Wen, the and of D.Huang et al. proposition
C.Zhang.2017.Applying Detection Proposals to Visual Tracking for Scale and
Aspect Ratio Adaptability.International Journal of Computer Vision 122,3
(2017),524–541.)；
MUSTer correspond to Z.Hong et al. proposition method (Z.Hong, Zhe Chen, C.Wang, X.Mei,
D.Prokhorov,and D.Tao.2015.MUlti-Store Tracker(MUSTer):A Cognitive Psychology
Inspired Approach to Object Tracking.In Proceedings of the 2015IEEE
Conference on Computer Vision and Pattern Recognition.749–758.)。
Claims (10)
1. the correlation filtering tracking that a kind of Case-based Reasoning perception target suggests window, it is characterised in that include the following steps：
1) for the i-th frame video, the Initial state estimation of target is carried out by the correlation filter based on CNN, obtains related filter
Wave response diagram；
2) to correlation filtering response diagram defined in step 1), the stability of detection is calculated, determines mode locating for present frame；
3) extract target in search range corresponding to the mode determined in step 2) and suggest window, using target suggestion window as
Candidate target suggests window；Since recall rate height, speed are fast in the method that target suggests window generation by EdgeBoxes, because
This is based on EdgeBoxes and generates target suggestion window；
4) window is suggested to target obtained in step 3), based on the apparent similarity and spatial weighting progress between example set
Sequence, the target for selecting example perception suggests window, equipped with having M example, E={ e in example set E1,e2,…,eM, in mesh
Mark suggests having N number of target to suggest window, P={ p in window collection P1,p2,…,pN}；Each target suggests window and each reality
Example apparent similarity calculating method be respectively：
Color similarity：I-th of target suggests window piWith j-th of example ejBetween color similarityIt is defined as them
COS distance between color histogram vector, defined formula (2) are as follows：
Wherein, his () indicates that target suggests the color histogram vector of window or example；
Shape similarity：I-th of target suggests window piWith j-th of example ejBetween shape similarityIt is defined as it
COS distance between HOG feature vector, defined formula (3) it is as follows：
Wherein, hog () indicates that target suggests the HOG feature vector of window or example；
Spatial weighting：Space weight wiIt is defined as i-th of target and suggests window piWith dbjective state o according to a preliminary estimatecBetween outstanding person
Ladd distance, defined formula (4) are as follows：
Wherein, box () indicates that target suggests the bounding box of window or dbjective state according to a preliminary estimate；
Since above-mentioned color similarity and shape similarity are two complementary apparent similarities, i-th of target suggests window
The ballot score v of mouthiDefined formula (5) is as follows：
Based on viWindow sequence is suggested to target, chooses the forward target that sorts and suggests that window suggests window as example perception target
Mouthful；
5) window is suggested to the target of the perception of each example obtained in step 4), is guided by the correlation filter based on CNN
To optimum state, therefrom select response it is maximum it is guided after target suggest further state of the window as target,
It is denoted asWherein,Indicate position,Indicate scale；At this point,By corresponding maximum response and peak value
Valve ratio is denoted asWith
6) according to the detection stability being calculated in step 2), based on obtained in step 1)It is obtained in step 5)
'sIt carries out adaptive targets positioning and adaptive model updates.
2. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as described in claim 1, feature exists
In in step 1), for the i-th frame video, the tool of the Initial state estimation of target is carried out by the correlation filter based on CNN
Body method is：Target is detected in correlation filtering region of search based on the correlation filter of CNN and obtains correlation filtering response diagram, is rung
Should original state of the state as target corresponding to maximum value in figure, be denoted asWherein,Indicate position,Indicate ruler
Degree.
3. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as described in claim 1, feature exists
It is described that the stability of detection is calculated to correlation filtering response diagram defined in step 1) in step 2), determine present frame institute
The specific method of the mode at place is：The stability of detection is codetermined by two indices, a maximum value being in response in figure, note
For Rmax, another is in response to figure peak sidelobe ratio, is denoted as PSR, and defined formula (1) is as follows：
Wherein, RμAnd RσRespectively the mean value and variance of response diagram, mode determining method are as follows：WhenOrWhen, start target re-detection mode, otherwise start size estimation mode, whereinFor RmaxMean value,For the mean value of PSR.
4. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as described in claim 1, feature exists
In in step 2), the mode is size estimation or target re-detection.
5. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as claimed in claim 3, feature exists
In λ=0.72.
6. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as described in claim 1, feature exists
It is described based between example set apparently including color and shape in step 4).
7. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as described in claim 1, feature exists
In in step 4), parameter θ=0.7.
8. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as described in claim 1, feature exists
In in step 6), the detection stability is by RmaxIt is determined jointly with PSR.
9. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as described in claim 1, feature exists
In in step 6), the self-adapting objective locating method is：Under re-detection mode, ifAndIt so indicates to be successfully detected target, the dbjective state of present frame isOtherwise, present frame
Dbjective state isUnder size estimation mode, the dbjective state of present frame is determined by the biggish dbjective state of response；
It is described to be from detection patternOr
The adaptive model update method is：Under re-detection mode, indicate that target is lost, without model modification；In ruler
It spends under estimation model, when detecting high stability, model modification is carried out with learning rate η；Otherwise, reduce learning rate be c η into
Row model modification；The re-detection mode can beOrThe detection high stability isAnd
10. a kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window as claimed in claim 9, feature exists
In the parameter lambda=0.72, μ=0.90, δ=1.00, η=0.01, c=0.7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810662407.9A CN108898621B (en) | 2018-06-25 | 2018-06-25 | Related filtering tracking method based on instance perception target suggestion window |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810662407.9A CN108898621B (en) | 2018-06-25 | 2018-06-25 | Related filtering tracking method based on instance perception target suggestion window |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108898621A true CN108898621A (en) | 2018-11-27 |
| CN108898621B CN108898621B (en) | 2021-01-12 |
Family
ID=64346270
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810662407.9A Active CN108898621B (en) | 2018-06-25 | 2018-06-25 | Related filtering tracking method based on instance perception target suggestion window |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108898621B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110070562A (en) * | 2019-04-02 | 2019-07-30 | 西北工业大学 | A kind of context-sensitive depth targets tracking |
| CN112233143A (en) * | 2020-12-14 | 2021-01-15 | 浙江大华技术股份有限公司 | Target tracking method, device and computer readable storage medium |
| CN113361329A (en) * | 2021-05-11 | 2021-09-07 | 浙江大学 | Robust single-target tracking method based on example feature perception |
| CN113808171A (en) * | 2021-09-27 | 2021-12-17 | 山东工商学院 | Unmanned aerial vehicle visual tracking method based on dynamic feature selection of feature weight pool |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8144931B1 (en) * | 2004-04-22 | 2012-03-27 | Hartman Richard L | Real time correlator system and method |
| CN104574445A (en) * | 2015-01-23 | 2015-04-29 | 北京航空航天大学 | Target tracking method and device |
-
2018
- 2018-06-25 CN CN201810662407.9A patent/CN108898621B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8144931B1 (en) * | 2004-04-22 | 2012-03-27 | Hartman Richard L | Real time correlator system and method |
| CN104574445A (en) * | 2015-01-23 | 2015-04-29 | 北京航空航天大学 | Target tracking method and device |
Non-Patent Citations (2)
| Title |
|---|
| J.A.FERNANDEZ，B.V.K.V.KUMAR: ""Zero-aliasing correlation filters"", 《2013 8TH INTERNATIONAL SYMPOSIUM ON IMAGE AND SIGNAL PROCESSING AND ANALYSIS (ISPA),TRIESTE,2013》 * |
| 王暐，王春平，李军，张伟: ""特征融合和模型自适应更新相结合的相关滤波器目标跟踪"", 《光学 精密工程》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110070562A (en) * | 2019-04-02 | 2019-07-30 | 西北工业大学 | A kind of context-sensitive depth targets tracking |
| CN112233143A (en) * | 2020-12-14 | 2021-01-15 | 浙江大华技术股份有限公司 | Target tracking method, device and computer readable storage medium |
| CN112233143B (en) * | 2020-12-14 | 2021-05-11 | 浙江大华技术股份有限公司 | Target tracking method, device and computer readable storage medium |
| CN113361329A (en) * | 2021-05-11 | 2021-09-07 | 浙江大学 | Robust single-target tracking method based on example feature perception |
| CN113361329B (en) * | 2021-05-11 | 2022-05-06 | 浙江大学 | Robust single-target tracking method based on example feature perception |
| CN113808171A (en) * | 2021-09-27 | 2021-12-17 | 山东工商学院 | Unmanned aerial vehicle visual tracking method based on dynamic feature selection of feature weight pool |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108898621B (en) | 2021-01-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108898621A (en) | A kind of Case-based Reasoning perception target suggests the correlation filtering tracking of window | |
| CN104574445B (en) | A kind of method for tracking target | |
| CN105335986B (en) | Method for tracking target based on characteristic matching and MeanShift algorithm | |
| CN106296742B (en) | A kind of matched online method for tracking target of binding characteristic point | |
| CN110837768B (en) | Online detection and identification method for rare animal protection | |
| CN103886619B (en) | A kind of method for tracking target merging multiple dimensioned super-pixel | |
| CN105654139A (en) | Real-time online multi-target tracking method adopting temporal dynamic appearance model | |
| CN102598057A (en) | Method and system for automatic object detection and subsequent object tracking in accordance with the object shape | |
| CN104835178A (en) | Low SNR(Signal to Noise Ratio) motion small target tracking and identification method | |
| CN107194408A (en) | A kind of method for tracking target of the sparse coordination model of mixed block | |
| CN111582349B (en) | Improved target tracking algorithm based on YOLOv3 and kernel correlation filtering | |
| CN104036523A (en) | Improved mean shift target tracking method based on surf features | |
| CN109993770B (en) | Target tracking method for adaptive space-time learning and state recognition | |
| CN104484890A (en) | Video target tracking method based on compound sparse model | |
| CN109410249B (en) | Self-adaptive target tracking method combining depth characteristic and hand-drawn characteristic | |
| CN108830170A (en) | A kind of end-to-end method for tracking target indicated based on layered characteristic | |
| CN107230219A (en) | A kind of target person in monocular robot is found and follower method | |
| CN111476077A (en) | Multi-view gait recognition method based on deep learning | |
| CN106127112A (en) | Data Dimensionality Reduction based on DLLE model and feature understanding method | |
| CN108089695A (en) | A kind of method and apparatus for controlling movable equipment | |
| CN110189362A (en) | The efficient method for tracking target of network is fought from coding based on multiple-limb | |
| CN110531618A (en) | Closed loop based on effective key frame detects robot self-localization error cancelling method | |
| CN115719343A (en) | Thick plate T-shaped joint multi-pass welding position autonomous decision-making method based on analytic hierarchy process | |
| CN115272405A (en) | Robust online learning ship tracking method based on twin network | |
| CN112287906B (en) | Template matching tracking method and system based on depth feature fusion |
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 |