CN109035302A - Target tracking algorithm based on space-time perception correlation filtering - Google Patents
Target tracking algorithm based on space-time perception correlation filtering Download PDFInfo
- Publication number
- CN109035302A CN109035302A CN201810831686.7A CN201810831686A CN109035302A CN 109035302 A CN109035302 A CN 109035302A CN 201810831686 A CN201810831686 A CN 201810831686A CN 109035302 A CN109035302 A CN 109035302A
- Authority
- CN
- China
- Prior art keywords
- target
- image
- scale
- space
- feature
- 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
- 238000001914 filtration Methods 0.000 title claims abstract description 26
- 230000008447 perception Effects 0.000 title abstract description 3
- 238000012549 training Methods 0.000 claims abstract description 21
- 238000000605 extraction Methods 0.000 claims abstract description 14
- 230000004927 fusion Effects 0.000 claims description 31
- 230000004044 response Effects 0.000 claims description 20
- 239000011159 matrix material Substances 0.000 claims description 17
- 241000282326 Felis catus Species 0.000 claims description 15
- 230000009466 transformation Effects 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 230000017105 transposition Effects 0.000 claims description 2
- 238000000034 method Methods 0.000 abstract description 26
- 230000000694 effects Effects 0.000 abstract description 4
- 230000006872 improvement Effects 0.000 abstract description 4
- 230000007774 longterm Effects 0.000 abstract 1
- 230000008569 process Effects 0.000 description 7
- 230000006870 function Effects 0.000 description 5
- 230000007306 turnover Effects 0.000 description 5
- 238000005070 sampling Methods 0.000 description 3
- 230000000903 blocking effect Effects 0.000 description 2
- 230000004438 eyesight Effects 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000021615 conjugation Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000003014 reinforcing effect Effects 0.000 description 1
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
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
The invention discloses a relevant filtering target tracking method based on space-time perception, and relates to the field of image processing target tracking. The method carries out target tracking according to the ideas of reading a target image, extracting target characteristics, training a filter template, extracting multi-scale characteristics of the target and determining the scale and the position in a new target image, optimizes and improves the structural framework of a related filtering target tracking algorithm and the target characteristic extraction mode, and has obvious advantages in the aspects of tracking robustness and precision comparison compared with the traditional algorithm. The method can enhance the robustness and precision of the related filtering target tracking, solve the problem of model drift caused by linear updating of the template, improve the long-term tracking effect of the related filtering target tracking, ensure the real-time performance of the target tracking, and is an important improvement on the prior art.
Description
Technical field
The present invention relates to computer visions and image procossing target following technical field, particularly relate to a kind of based on space-time sense
Know the target tracking algorism of correlation filtering.
Background technique
Target following technology is increasingly becoming the hot issue of research with the development of computer vision technique, in terms of
It has been more and more widely used.But it due to the deformation of target in target following, blocks and the factors such as the interference of background cause
Target following remains a difficulties.Correlation filtering target following in recent years due to its high speed and preferable robustness, by
Extensive concern.
Foreign scholar conducts in-depth research the field: Bolme et al. for the first time by correlation filtering be applied to target with
Track field passes through the correlation of target and filter template by minimum empiric risk come training objective associated filter template
Target position is judged to complete object tracking process.The algorithm is using gray feature training filter, and tracking velocity is high, robustness
Preferably.Circular matrix property is applied to correlation filtering filter template training process, training sampling process by Henriques et al.
It is equivalent to target signature matrix circular shifting function, and then completes the intensive sampling process of filter template training, this method is excellent
Change target signature training sampling process, further improves the robustness of tracking.Danelljan is true in position correlation filter
It sets the goal behind position, by additional unidimensional scale correlation filter, constructs target scale pond come come the best ruler of estimating target
Degree.This method, which is simple and efficient, preferably to be estimated the dimensional variation of target.Matthias Mueller is filtered in correlation
Background structure perception information is added during wave device template training, the robust of filter is further enhanced by background structure information
Property.Feature Chao Ma abundant using convolutional neural networks Deep Semantics abundant information and shallow-layer detailed information, is arrived by deep layer
Shallow-layer by slightly to essence target position determine method, further improve mesh using the powerful ability in feature extraction of neural network
Mark the precision and robustness of tracking.
Above-mentioned algorithm is innovated and has been improved for correlation filtering target tracking algorism, but still is had the following problems: mesh
The precision and robustness for marking tracking need to be further increased;Filter template uses linear real-time update strategy, when target occurs
It will lead to template drift when blocking, while this update mode causes tracking process bigger therefore long to nearest sample dependence
When tracking will increase unstability.
Summary of the invention
In view of this, it is an object of the invention to propose a kind of target tracking algorism based on the perceptually relevant filtering of space-time,
This method can be improved the tracking accuracy and robustness of target tracking algorism, effectively solve in correlation filtering target tracking algorism
Model modification problem.
Based on above-mentioned purpose, present invention provide the technical scheme that
A kind of target tracking algorism based on the perceptually relevant filtering of space-time, this method read more comprising same target one by one
Sequential picture is opened, and following steps are executed to image:
Step 1: for present image, target position in the images and scale size are obtained;
Step 2: according to current resulting position and scale, HOG the and CN fusion feature of target in present image is extracted;
Step 3: the subject fusion feature training perceptually relevant filter template of space-time currently extracted is utilized;
Step 4: with next image for new present image, corresponding in former present image in new present image
In place of target position, on the basis of current acquired scale, HOG and CN fusion of the target in new present image are extracted
Analysis On Multi-scale Features;
Step 5: position and scale of the target in new present image are determined with the perceptually relevant filter template of current space-time
Size;
Step 6: step 2 is repeated to five, until target following terminates.
Optionally, the concrete mode of the step 1 are as follows:
(101) present image is read, judges whether present image is Three Channel Color image, if present image is threeway
Road color image then enables Cl=1, otherwise enable Cl=0;
(102) the position L=[x, y] and scale size S of target in present image are obtainedz=[W, H], wherein W is current
The width of target in image, H are the height of target in present image.
Optionally, the concrete mode of the step 2 are as follows:
(201) in the target location of present image, image I is taken with 5 times of target scale size;
(202) according to image I, 31 dimensions of target, the HOG feature x that cell size is 4 are extractedh;
(203) if Cl=1, then 11 Victoria C N feature x of target are extracted according to image Ic, otherwise, mesh is extracted according to image I
Target 1 ties up gray feature xg;
(204) if Cl=1, then 42 dimension fusion feature x of target are obtained according to the HOG feature of extraction and CN featuret=
cat(3,xh,xc), otherwise, 32 dimension fusion feature x of target are obtained according to image It=cat (3, xh,xg), wherein cat ()
It indicates to be coupled array function, number 3 indicates to be coupled in a manner of the matrix third dimension.
Optionally, the concrete mode of the step 3 are as follows:
(301) according to the subject fusion feature x of extractiont, the perceptually relevant filter template of space-time, which is arranged, is
In formula,Representing matrix Fourier transformation,
T=WH, representing matrix dimension,
U=1 indicates Lagrangian regularization factors,
λ=14 indicate regularization factors,
The two is scalar,Representing matrix conjugate transposition,
Indicate that filter responds target value, wherein u={ 1,2 ... W }, v={ 1,2 ... H },
(302) intermediate quantity is calculated
In formula, ∮-1() indicates Fourier inversion,
U=min (β u, θ), wherein θ=0.1, β=10, min () expression are minimized function;
(303) according to the subject fusion feature x of extractiont, being once again set up the perceptually relevant filter template of space-time is
In formula,
The two is scalar;
(304) updating intermediate quantity is
(305) according to the required accuracy, step (303)~(304) are repeated 0 time or repeatedly, finally obtain training completion when
Empty perceptually relevant filter template
Optionally, the concrete mode of the step 4 are as follows:
(401) correspond to target in place of the position in former present image in new present image, with current
N times of target scale size takes image, obtains target multi-scale image pond Is;
Wherein, n=5am,N takes odd number, is arranged for scale pond
Number, a are scale step-length;
(402) according to target multi-scale image pond Is, special to the HOG that each target extracts 31 dimensions respectively, cell size is 4
Levy xh s;
(403) if Cl=1, then according to target multi-scale image pond Is11 Victoria C N features of each target are extracted respectively
xc s, otherwise, according to target multi-scale image pond Is1 dimension gray feature x of each target is extracted respectivelyg s;
(404) if Cl=1, then according to target multi-scale image pond Is42 dimension fusion features of each target are obtained respectively
xt s=cat (3, xh s,xc s), otherwise, according to target multi-scale image pond Is32 dimension fusion feature x of each target are obtained respectivelyt s
=cat (3, xh s,xg s)。
Optionally, the concrete mode of the step 5 are as follows:
(501) the perceptually relevant filter template of space-time completed according to step 3 trainingSeek target responseIn formula, ()*The complex conjugate of representing matrix;
(502) target response maximum value position r is searched for according to target response rmax, target response maximum value position is corresponding
Target scale is new current goal scalemNFor the corresponding scale of target response maximum value;
(503) according to target response maximum value position rmax, target response maximum value position is new current goal position
L=[x, y].
The present invention is compared to the advantages of background technique:
1, the robustness and precision for focusing on reinforcing related filtered target tracking of this method, solves template and linearly updates
Caused model drifting problem, improves the long time-tracking effect of correlation filtering target following, while ensure that the reality of target following
Shi Xing.
2, this method fusion HOG and CN feature obtains target more comprehensively characteristic present, improves under complex background
Track robustness.
3, this method avoids the line of template using the perceptually relevant filtered target track algorithm training filter template of space-time
Property update caused by model drifting problem, further improve the robustness of track algorithm.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is an algorithm flow chart of the embodiment of the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference
Attached drawing, the present invention is described in further detail.
As shown in Figure 1, a kind of target tracking algorism based on the perceptually relevant filtering of space-time, this method in image sequence by
(image sequence can be the sequence photo of capture apparatus capture to one reading image, be also possible to the sequence extracted from video view
Frequency frame), and following steps are executed to image:
Step 1: position and scale size of the target in present image are obtained;Concrete mode are as follows:
(101) image is read, judges whether present image is Three Channel Color image.If present image is triple channel coloured silk
Chromatic graph picture then Cl=1, otherwise Cl=0;
(102) according to image information, target position L=[x, y] and scale size S in present image are obtainedz=[W,
H], wherein W is the width of target in present image, and H is the height of target in present image.
Step 2: subject fusion feature is extracted.According to currently obtained target position and target scale, target is extracted
HOG and CN fusion feature;Concrete mode are as follows:
(201) in the target location of present image, image I is taken with 5 times of target scale size;
(202) 31 dimensions of target, the HOG feature x that cell size is 4 are extracted according to image Ih;
(203) if Cl=1,11 Victoria C N feature x of target are extracted according to image Ic, otherwise, target is extracted according to image I
1 dimension gray feature xg;
(204) if Cl=1,42 dimension fusion feature x of target are obtained according to the HOG feature of extraction and CN featuret=cat
(3,xh,xc), otherwise, 32 dimension fusion feature x of target are obtained according to image It=cat (3, xh,xg).Wherein, cat () is indicated
It is coupled array function, number 3 indicates to be coupled in a manner of the matrix third dimension;
Step 3: the training perceptually relevant filter template of space-time.When the subject fusion feature training extracted using step 2
Empty perceptually relevant filter template;Concrete mode are as follows:
(301) according to the subject fusion feature x of extractiont, the perceptually relevant filter template of space-time, which is arranged, is
Wherein,Representing matrix Fourier transformation, T=WH representing matrix dimension, u=1 indicate Lagrangian regularization because
Son, λ=14 indicate regularization factors,It is scalar,Representing matrix conjugation turns
It sets,Be filter response target value, u={ 1,2 ... W }, v={ 1,2 ... H },
(302) intermediate quantity is calculated
Wherein, ∮-1() indicates Fourier inversion,U=min (β u, θ), θ=0.1, β
=10, min () expression are minimized function;
(303) according to the subject fusion feature x of extractiont, being once again set up the perceptually relevant filter template of space-time is
Wherein,For scalar,For scalar;
(304) intermediate quantity is calculated again
Wherein,U=min (β u, θ);
(305) according to required precision, step (303)~(304) are repeated, number of repetition is more, and precision is higher, is instructed
Practice the perceptually relevant filter template of space-time completedIn this example, precision and arithmetic speed, can make step (303) in order to balance
~(304) Exactly-once.
The iteration that template turnover rate is used to adjust filter template in traditional correlation filtering object tracking process updates speed
Degree, biggish turnover rate can meet the violent situation of object variations, but lower to the robustness of target occlusion or background interference etc.,
It is easy to cause model to drift about, and lesser turnover rate is higher to the robustness of target occlusion etc., it can be extensive preferably after blocking
Multiple tracking, but it is not able to satisfy the tracking situation of target drastic mechanical deformation, therefore template turnover rate seriously restricts the effect of target following.
Meanwhile the weight that fixed turnover rate template update mode will lead to current sample is larger, historical frames target sample information weight is got over
Next smaller, i.e., filter template information will gradually replace historical information, and especially target initial target information will be lost, in length
When tracking during, the robustness of target following will gradually decrease.In addition linear update is the mould in order to guarantee tracking velocity
Plate approximation updates, and the robustness of target following template can be further decreased in long time-tracking.Therefore this method can be preferably
Correlation filtering target template replacement problem is solved, the precision and robustness of target following are further increased.
Step 4: target Multiscale Fusion feature is extracted.The mesh corresponded in former present image in new present image
In place of marking position, on the basis of current target scale, the Analysis On Multi-scale Features of HOG and the CN fusion of target are extracted;Specifically
Mode are as follows:
(401) correspond to target in new present image in place of the position in former present image, with current goal scale
N times of size obtains target multi-scale image pond Is,
Wherein, n=5 [am],N is scale pond setting number,
Odd number is taken, number setting more multiscale estimatiL is more accurate, and 7, a is taken as in this example indicates scale step-length, is taken as 1.01 in this example;
" n=5 [am] " meaning be, with each of vector m member usually carry out operation, the knot of each secondary operation respectively
Fruit collectively constitutes vector n;
" target multi-scale image pond I is obtained with n times of current goal scale sizes" meaning be, with the every of vector n
One element obtains an image as multiple respectively, and all images collectively constitute image pond Is。
(402) according to target multi-scale image pond Is, 31 dimensions of target, the HOG feature that cell size is 4 are extracted respectively
xh s;
(403) if Cl=1, according to target multi-scale image pond Is11 Victoria C N feature x of target are extracted respectivelyc s, otherwise,
According to target multi-scale image pond Is1 dimension gray feature x of target is extracted respectivelyg s;
(404) if Cl=1, according to target multi-scale image pond Is42 dimension fusion feature x of target are obtained respectivelyt s=
cat(3,xh s,xc s), otherwise, according to target multi-scale image pond Is32 dimension fusion feature x of target are obtained respectivelyt s=cat (3,
xh s,xg s);
Step 5: new target position and scale are determined.The perceptually relevant filter template of space-time with step 3 training is true
Set the goal position and scale size in present image;Concrete mode are as follows:
(501) according to the perceptually relevant filter template of space-time of step 3 trainingSeek target response(·)*The complex conjugate of representing matrix;
(502) target response maximum value position r is searched for according to target response rmax, target response maximum value position is corresponding
Target scale is new target scalemNFor the corresponding scale of target response maximum value;
(503) according to target response maximum value position rmax, target response maximum value position is new target position L=
[x,y];
Step 6: step 2 is repeated to five, until target following terminates.
So far, target following is completed.
This method is according to reading target image, extraction target signature, training filter template, the extraction multiple dimensioned spy of target
It levies, determine that the thinking of scale and position in target new images carries out target following, to the knot of correlation filtering target tracking algorism
Improvement is optimized in structure frame and target's feature-extraction mode so that this method relative to traditional algorithm tracking robustness and
It has a clear superiority in terms of accuracy comparison.Specifically, the present invention obtains target more comprehensively by fusion HOG and CN feature
Characteristic present is filtered to improve the tracking robustness under complex background using the perceptually relevant filtered target track algorithm training of space-time
Wave device template avoids model drifting problem caused by the linear update of template, improves the robustness of track algorithm.
In short, this method can reinforce the robustness and precision of related filtered target tracking, solution template, which linearly updates, to be led
The model drifting problem of cause, improves the long time-tracking effect of correlation filtering target following, while ensure that the real-time of target following
Property, it is to one kind of the prior art in important improvement.
It should be understood by those ordinary skilled in the art that: the discussion of any of the above embodiment is exemplary only, not
It is intended to imply that the scope of the present disclosure (including claim) is limited to these examples.All within the spirits and principles of the present invention,
Any omission made to the above embodiment, modification, equivalent replacement, improvement etc., should be included in protection scope of the present invention it
It is interior.
Claims (6)
1. a kind of target tracking algorism based on the perceptually relevant filtering of space-time, which is characterized in that reading one by one includes same target
Multiple sequential pictures, and to image execute following steps:
Step 1: for present image, target position in the images and scale size are obtained;
Step 2: according to current resulting position and scale, HOG the and CN fusion feature of target in present image is extracted;
Step 3: the subject fusion feature training perceptually relevant filter template of space-time currently extracted is utilized;
Step 4: with next image for new present image, correspond to target in former present image in new present image
In place of position, on the basis of current acquired scale, the more of HOG and CN fusion of the target in new present image are extracted
Scale feature;
Step 5: determine that position and scale of the target in new present image are big with the perceptually relevant filter template of current space-time
It is small;
Step 6: step 2 is repeated to five, until target following terminates.
2. the target tracking algorism according to claim 1 based on the perceptually relevant filtering of space-time, it is characterised in that: the step
Rapid one concrete mode are as follows:
(101) present image is read, judges whether present image is Three Channel Color image, if present image is triple channel coloured silk
Chromatic graph picture then enables Cl=1, otherwise enable Cl=0;
(102) the position L=[x, y] and scale size S of target in present image are obtainedz=[W, H], wherein W is present image
The width of middle target, H are the height of target in present image.
3. the target tracking algorism according to claim 2 based on the perceptually relevant filtering of space-time, it is characterised in that: the step
Rapid two concrete mode are as follows:
(201) in the target location of present image, image I is taken with 5 times of target scale size;
(202) according to image I, 31 dimensions of target, the HOG feature x that cell size is 4 are extractedh;
(203) if Cl=1, then 11 Victoria C N feature x of target are extracted according to image Ic, otherwise, the 1 of target is extracted according to image I
Tie up gray feature xg;
(204) if Cl=1, then 42 dimension fusion feature x of target are obtained according to the HOG feature of extraction and CN featuret=cat (3,
xh,xc), otherwise, 32 dimension fusion feature x of target are obtained according to image It=cat (3, xh,xg), wherein cat () indicates connection
Array function is tied, number 3 indicates to be coupled in a manner of the matrix third dimension.
4. the target tracking algorism according to claim 3 based on the perceptually relevant filtering of space-time, it is characterised in that: the step
Rapid three concrete mode are as follows:
(301) according to the subject fusion feature x of extractiont, the perceptually relevant filter template of space-time, which is arranged, is
In formula,Representing matrix Fourier transformation,
T=WH, representing matrix dimension,
U=1 indicates Lagrangian regularization factors,
λ=14 indicate regularization factors,
The two is scalar,Representing matrix conjugate transposition,
Indicate that filter responds target value, wherein u={ 1,2 ... W }, v={ 1,2 ... H },
(302) intermediate quantity is calculated
In formula, ∮-1() indicates Fourier inversion,
U=min (β u, θ), wherein θ=0.1, β=10, min () expression are minimized function;
(303) according to the subject fusion feature x of extractiont, being once again set up the perceptually relevant filter template of space-time is
In formula,
The two is scalar;
(304) updating intermediate quantity is
(305) according to the required accuracy, step (303)~(304) are repeated 0 time or repeatedly, finally obtains the space-time sense of training completion
Know associated filter template
5. the target tracking algorism according to claim 4 based on the perceptually relevant filtering of space-time, it is characterised in that: the step
Rapid four concrete mode are as follows:
(401) correspond to target in place of the position in former present image in new present image, with current target
N times of scale size takes image, obtains target multi-scale image pond Is;
Wherein, n=5am,N takes odd number, and number is arranged for scale pond
Mesh, a are scale step-length;
(402) according to target multi-scale image pond Is, to the HOG feature that each target extracts 31 dimensions respectively, cell size is 4
xh s;
(403) if Cl=1, then according to target multi-scale image pond Is11 Victoria C N feature x of each target are extracted respectivelyc s, no
Then, according to target multi-scale image pond Is1 dimension gray feature x of each target is extracted respectivelyg s;
(404) if Cl=1, then according to target multi-scale image pond Is42 dimension fusion feature x of each target are obtained respectivelyt s=
cat(3,xh s,xc s), otherwise, according to target multi-scale image pond Is32 dimension fusion feature x of each target are obtained respectivelyt s=
cat(3,xh s,xg s)。
6. the target tracking algorism according to claim 5 based on the perceptually relevant filtering of space-time, it is characterised in that: the step
Rapid five concrete mode are as follows:
(501) the perceptually relevant filter template of space-time completed according to step 3 trainingSeek target responseIn formula, ()*The complex conjugate of representing matrix;
(502) target response maximum value position r is searched for according to target response rmax, the corresponding target of target response maximum value position
Scale is new current goal scalemNFor the corresponding scale of target response maximum value;
(503) according to target response maximum value position rmax, target response maximum value position is new current goal position L=
[x,y]。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810831686.7A CN109035302B (en) | 2018-07-26 | 2018-07-26 | Target tracking algorithm based on space-time perception correlation filtering |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810831686.7A CN109035302B (en) | 2018-07-26 | 2018-07-26 | Target tracking algorithm based on space-time perception correlation filtering |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109035302A true CN109035302A (en) | 2018-12-18 |
CN109035302B CN109035302B (en) | 2021-07-06 |
Family
ID=64646389
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810831686.7A Active CN109035302B (en) | 2018-07-26 | 2018-07-26 | Target tracking algorithm based on space-time perception correlation filtering |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109035302B (en) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1419680A (en) * | 2001-01-26 | 2003-05-21 | 皇家菲利浦电子有限公司 | Spatio-temporal filter unit and image display apparatus comprising such a spatio-temporal filter unit |
CN106651913A (en) * | 2016-11-29 | 2017-05-10 | 开易(北京)科技有限公司 | Target tracking method based on correlation filtering and color histogram statistics and ADAS (Advanced Driving Assistance System) |
CN107316316A (en) * | 2017-05-19 | 2017-11-03 | 南京理工大学 | The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features |
CN107452022A (en) * | 2017-07-20 | 2017-12-08 | 西安电子科技大学 | A kind of video target tracking method |
CN107578423A (en) * | 2017-09-15 | 2018-01-12 | 杭州电子科技大学 | The correlation filtering robust tracking method of multiple features hierarchical fusion |
-
2018
- 2018-07-26 CN CN201810831686.7A patent/CN109035302B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1419680A (en) * | 2001-01-26 | 2003-05-21 | 皇家菲利浦电子有限公司 | Spatio-temporal filter unit and image display apparatus comprising such a spatio-temporal filter unit |
CN106651913A (en) * | 2016-11-29 | 2017-05-10 | 开易(北京)科技有限公司 | Target tracking method based on correlation filtering and color histogram statistics and ADAS (Advanced Driving Assistance System) |
CN107316316A (en) * | 2017-05-19 | 2017-11-03 | 南京理工大学 | The method for tracking target that filtering technique is closed with nuclear phase is adaptively merged based on multiple features |
CN107452022A (en) * | 2017-07-20 | 2017-12-08 | 西安电子科技大学 | A kind of video target tracking method |
CN107578423A (en) * | 2017-09-15 | 2018-01-12 | 杭州电子科技大学 | The correlation filtering robust tracking method of multiple features hierarchical fusion |
Also Published As
Publication number | Publication date |
---|---|
CN109035302B (en) | 2021-07-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110111366A (en) | A kind of end-to-end light stream estimation method based on multistage loss amount | |
CN109376576A (en) | The object detection method for training network from zero based on the intensive connection of alternately update | |
CN105550678A (en) | Human body motion feature extraction method based on global remarkable edge area | |
CN110021037A (en) | A kind of image non-rigid registration method and system based on generation confrontation network | |
CN106780543A (en) | A kind of double framework estimating depths and movement technique based on convolutional neural networks | |
CN109685732A (en) | A kind of depth image high-precision restorative procedure captured based on boundary | |
CN104376568B (en) | Method for processing DICOM (digital imaging and communications in medicine) medical images on basis of formats | |
CN111325750B (en) | Medical image segmentation method based on multi-scale fusion U-shaped chain neural network | |
CN103605637B (en) | Particle image velocimetry vector estimation method for spatial resolution self-adaptation adjustment | |
CN112927279A (en) | Image depth information generation method, device and storage medium | |
CN116342596A (en) | YOLOv5 improved substation equipment nut defect identification detection method | |
CN114140469B (en) | Depth layered image semantic segmentation method based on multi-layer attention | |
CN110298402A (en) | A kind of small target deteection performance optimization method | |
CN108364305A (en) | Vehicle-mounted pick-up video target tracking method based on modified DSST | |
CN114511710A (en) | Image target detection method based on convolutional neural network | |
CN113989683A (en) | Ship detection method for synthesizing synchronous orbit sequence optical image space-time information | |
CN116645592A (en) | Crack detection method based on image processing and storage medium | |
CN110147724A (en) | For detecting text filed method, apparatus, equipment and medium in video | |
CN109035302A (en) | Target tracking algorithm based on space-time perception correlation filtering | |
CN110728660B (en) | Method and device for lesion segmentation based on ischemic stroke MRI detection mark | |
CN116052108A (en) | Transformer-based traffic scene small sample target detection method and device | |
CN117133014A (en) | Live pig face key point detection method | |
CN115205215A (en) | Corneal nerve image segmentation method and system based on Transformer | |
CN110136185A (en) | A kind of monocular depth estimation method and system | |
CN105654527A (en) | Magnetic resonance imaging reconstruction method and device based on structural dictionary learning |
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 |