CN106530324A - Visual cortex mechanism simulated video object tracking method - Google Patents
Visual cortex mechanism simulated video object tracking method Download PDFInfo
- Publication number
- CN106530324A CN106530324A CN201610921845.3A CN201610921845A CN106530324A CN 106530324 A CN106530324 A CN 106530324A CN 201610921845 A CN201610921845 A CN 201610921845A CN 106530324 A CN106530324 A CN 106530324A
- Authority
- CN
- China
- Prior art keywords
- layers
- block
- sub
- target
- candidate
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a visual cortex mechanism simulated video object tracking method. The method includes offline training and online tracking. In offline training, image sample sets of a tracking object are constructed; an image pyramid is established for samples; a BIM feature is extracted; and a classifier is trained. In online tracking, a detection area of an image to be detected is initialized, an image pyramid of the detection area is established; the BIM feature is extracted from the image pyramid, matching subblocks are obtained in the BIM feature extracting process, and the matching subblocks are used to predict a candidate object area; potential object areas are selected from the candidate area by iteration according to a determination result of the classifier; and an object area of the highest probability is selected from the potential object areas and serves as a final object area. According to the method, BIM is used to established an appearance model of an object, a tracking framework is introduced, and the object can be tracked stably by simulating the visual cortex mechanism.
Description
Technical field
The invention belongs to computer vision and neurobiology field, and in particular to a kind of analog vision cortex mechanism is regarded
Frequency method for tracking target.
Background technology
Biological excitation model (Biologically inspired model, BIM) is raw according to nerve by Serre et al.
Neo-confucian Hubel and Wiesel is to setting up on the basis of visual cortex achievement in research.The feature extraction of BIM is one 4 layering knot
Structure model, 4 Rotating fields are respectively designated as S1, C1, S2 and C2.The V1 layers of the corresponding visual cortex of S1 with C1 layers, extraction can be tolerated and put down
Move the target texture feature of conversion；The category feature of target can be distinguished by the MST areas of the corresponding visual cortex of S2 with C2 floor, extraction.By
In BIM based on neuro physiology achievement in research, compare other artificial features extracted and there is incomparable superiority,
This model is widely used in Computer Vision Task, such as identification of Activity recognition, age, scene Recognition etc..With high-performance
The exploitation of GPU and distributed system, BIM is possibly realized for video frequency object tracking task.
The research of video frequency object tracking is mainly comprised the following steps：Target detection, the foundation of target appearance model and to target
Fast search.The first step needs from background image to be accurately positioned target, and this link is now as an independence
Research direction；Second step is to extract to be capable of the feature of adaptive targets cosmetic variation to set up the display model of target, selected
Feature can be comprising information such as the color of target, shape, textures；3rd step designs efficient search strategy, quickly positions mesh
Mark.Mainly there is following searching algorithm at present：Particle filter, optical flow method, mean shift algorithm etc..
Video frequency object tracking algorithm research makes great progress in recent years, obtains good tracking effect.But simulate
BIM introducing target tracking algorisms can be such that computer realizes more by the rare research of the target tracking algorism of visual cortex mechanism
Plus meet the working mechanism of human vision cortex, reach that degree of accuracy is high, the tracking effect that robustness is good.
The content of the invention
The problem existed for prior art and urgent needss, the invention provides a kind of video target tracking method and be
System, it is intended that setting up the display model of target using BIM and introducing tracking framework so that computer being capable of analog vision
Cortex working mechanism steadily tracks target.
A kind of video target tracking method of analog vision cortex mechanism, including two ranks of off-line training and online tracking
Section；
The off-line training step is comprised the following steps：
(11) build the positive and negative sample set of tracking target image；
(12) align negative sample and set up image pyramid, extract the BIM features of image pyramid, carry out according to BIM features
The classifier training of the first round, the weight size obtained according to classifier training screen feature to reduce intrinsic dimensionality；
(13) classifier training of the second wheel is carried out using the BIM features reduced after intrinsic dimensionality；
The online tracking phase is comprised the following steps：
(21) detection image initialization detection zone is treated, the image pyramid of detection zone is set up；
(22) BIM features are extracted to the image pyramid of detection zone, are sent to the grader that step (13) is trained,
Grader exports the result of determination of detection zone；Multiple matched sub-blocks are obtained during BIM features are extracted, using matching
Block predicts object candidate area；
(23) potential target region is selected according to grader result of determination iteration from candidate region；
(24) target area maximum probability person is chosen from potential target region be final target area.
Further, the specific implementation of step (12) is：
(121) to each Sample Establishing image pyramid and BIM features are extracted, including S1 layer features, S2 layer features, S3
Layer feature and S4 layer features；
The calculation of the S1 layers feature is：Image pyramid is defined as by the result that S1 layers computing unit is calculated
S1 layer features；S1 layer computing units are the wave filter groups that 4 Gabor filters are constituted, and are expressed as：
x0=xcos θ+ysin θ,
y0=-xsin θ+ycos θ
In formula, F (x, y, θ) represents wave filter in the corresponding response value in coordinate (x, y) place, and θ controls the direction of wave filter,
Value is respectively4 direction 4 Gabor filters of correspondence；γ represents the length-width ratio of wave filter, and σ represents filter
The bandwidth of ripple device, λ represent the wavelength of wave filter；
The calculation of the C1 layers feature is：S1 layers feature is defined as into C1 layers by the result of calculation of C1 computing units
Feature；C1 layers computing unit extracts maximum, the office of computing unit setting in the corresponding local neighborhood of S1 layer features to input
Portion's neighborhood is the local pyramid extracted from the corresponding image pyramid of S1 layer features；
The calculation of the S2 layers feature is：C1 layers feature is defined as into S2 by the result of calculation of S2 layer computing units
Layer feature, S2 layers computing unit calculate characteristic sub-block X and each sub-block template in C1 layer pyramidsBetween it is similar
Property degree, it is describedIt is the sub-block randomly selected from the C1 layer features of multiple samples, d is the quantity for extracting sub-block；It is described
Similarity function is defined as：In formula,
The calculation of the C2 layers feature is：S2 layers feature is defined as into C2 by the result of calculation of C2 layer computing units
Layer feature, C2 layers computing unit calculate the similarity response global maximum of each characteristic sub-block in C1 layer pyramids；
(122) screen characteristic sub-block
C2 layers feature is sent into into classifier training, using the feature selection ability of grader, to d sub-block template according to dividing
The respective weights size that the training of class device is obtained is ranked up, and selects k maximum sub-block template to be used for the grader instruction of the second wheel
Practice, k<d.
Further, the specific implementation of step (22) is：
(221) BIM features are extracted to the image pyramid of detection zone, is sent to the classification that step (13) is trained
Device, grader export the result of determination of detection zone；
(222) BIM features are extracted to the image pyramid of detection zone, obtains the corresponding best match of k sub-block template
Sub-block, predicts the candidate region of target using matched sub-block, and concrete implementation mode is：
Make detection zone image pyramid bottom geometric center be (x, y) zero, the center phase of matched sub-block
Coordinate position (dx, dy) of the matched sub-block relative to zero is defined as to the vector of zero, make two sub-block templates it
Between distance be Dij, i, j=1 ..., k, the distance between its Corresponding matching sub-block are D 'ij；Calculate the target scale of present frame
The factorMedian () takes median to all of ratio, improves the robustness of dimension calculation；According to
Each matched sub-block predicts the center of targetBuild with the centre bit of target
The object candidate area at center is set to, the collection being made up of these object candidate areas is collectively referred to as candidate collection.
Further, the specific implementation of step (23) is：
(231) one object candidate area of random selection from candidate collection.
(232) if grader output is judged to target, choose from all candidate regions from current goal candidate region
Nearest candidate region is used as potential target region, and rejects potential target region from candidate collection and away from detection zone
Candidate samples；Otherwise, the center average of the candidate region away from current detection region is calculated, chosen distance this average is most
Near candidate region is used as potential target region, and rejects potential target region from candidate collection and near current detection area
The candidate region in domain；
(233) the potential target extracted region BIM feature selected to step (232), is sent to step (13) and trains
Grader, the result of determination in grader output potential target region；During BIM features are extracted, S2 layers and C2 layers
Calculating may find new matched sub-block, perform the operation similar with step (222) to new matched sub-block, obtain new mesh
New object candidate area is added to candidate collection by mark candidate region；
(234) judge that candidate collection then enters step (24), otherwise return to step (231) whether as empty, if so,.
Further, step (24) choose that correspondence grader result of determination shows most to have from potential target region can
Final target area can be for target area person.
BIM is used for video frequency object tracking by above-mentioned steps by the present invention, compared to existing technology, with following excellent
Point：
Compared to other target's feature-extraction methods, the complete analog vision cortex work of the feature that BIM is extracted more meets life
Thing capture target general characteristic, with robustness it is good the characteristics of；Sub-block in C1 layers is a kind of local feature, with positioning office
The function of portion's shelter target, and C2 layers are characterized in that a kind of global characteristics, although no positioning function but through machine learning but
There is the ability of very strong differentiation target and background.Both features are combined by the present invention, can play respective advantage, both
The speed of tracking can be improved, the precision of tracking is also ensured that.
Description of the drawings
Fig. 1 is video target tracking method flowchart of the present invention；
Fig. 2 is BIM feature extractions flow chart of the present invention.
Specific embodiment
In order that the objects, technical solutions and advantages of the present invention become more apparent, it is below in conjunction with drawings and Examples, right
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, and
It is not used in the restriction present invention.As long as additionally, technical characteristic involved in invention described below each embodiment
Do not constitute conflict each other can just be mutually combined.
In order to help the understanding to technical solution of the present invention, the first technical term to being related to explain：
BIM models：Biological excitation model, including 4 Rotating fields, respectively S1 layers, C1 layers, S2 layers and C2 layers.Wherein, S generations
Simple cell in the table visual cortex, C represent the complex cell in the visual cortex.Sample obtains being vision through S1 layers and C1 layers
Primary features, be then passed through that S2 and C2 layers obtain is vision advanced features.
Image pyramid：Image pyramid be one with pyramid structure arrange image collection, all of figure in set
As all originating from same original image, and it is by the continuous down-sampled acquisition of original image, until reaching specify minimum
Resolution just stops down-sampled.
Fig. 1 is video target tracking method flowchart of the present invention, including off-line training and two stages of online tracking,
Implement step as follows：
(1) off-line training step：
(11) select positive negative sample.
Positive sample represents target, takes target area manually in video initial frame, in order to be able to allow grader identification local to hide
The target of gear, is blocked to target area at random with the square of white, the size value of white blocks for [8,0.5*min (w,
H)], wherein w, h are respectively the width and height of target.In order to improve the identification ability to positive sample, from 101 figures of CALTECH
As selecting the sample similar with target is tracked in storehouse, the multiformity of positive sample is improved.
Negative sample represents non-targeted, extracts from background area in initial video frame, and the area size of negative sample can be big
In the size of target, the size of target is finally normalized to.In order to increase the representativeness of negative sample, it is also possible to from CALTECH
Select and other close class images of target appearance in 101 image libraries.
(12) first round off-line learning.
(121) to Sample Establishing image pyramid, and BIM features are extracted.
Fig. 2 is BIM feature extractions flow chart of the present invention.
S1 layers are specially：
Computing unit in S1 layers is that 4 Gabor filters constitute wave filter group：
x0=xcos θ+ysin θ,
y0=-xsin θ+ycos θ
In formula, F (x, y, θ) represents wave filter in the corresponding response value in coordinate (x, y) place, and θ controls the direction of wave filter,
Value is respectively4 direction 4 Gabor filters of correspondence；γ represents the length-width ratio of wave filter, and σ represents filter
The bandwidth of ripple device, λ represent the wavelength of wave filter, and these three parameters are determined according to the physical characteristics of human eye receptive field, are constant.
S1 layer computing units are arranged in the pyramid structure similar to input picture so that computing unit can be in input
Calculated on all yardsticks of image pyramid and position.The result definition that image pyramid is calculated by S1 layers computing unit
For S1 layer features.
C1 layers are specially：
C1 layers are convergence-levels, and the computing unit in C1 layers is equally arranged as pyramid structure, each computing unit pair
Maximum is extracted in the corresponding local neighborhood of S1 layer features of input.The local neighborhood that computing unit sets is as from S1 layer features pair
The local pyramid scope of the double-layer structure extracted in the image pyramid answered, the pyramidal bottom size in local are 2 Δ s
×2Δs.The region calculated by C1 layer computing units is overlapped, and overlapping widths are Δ s.S1 layers feature is calculated into single by C1
The result of calculation of unit is defined as C1 layer features.
In order to extract ensuing S2 layers feature, needs are randomly selected various sizes of from the C1 layer features of great amount of samples
Square sub-block is used as sub-block templateSize n of sub-block template be empirical value, for example for (4,8,12,16)；Extract
Sub-block may come from C1 layers and export pyramidal different scale layer, and d is the quantity for extracting sub-block, and generally higher than 1000.
S2 layers are specially：
S2 layers calculate sub-block X and each sub-block template P in the C1 layer pyramids of inputiBetween similarity measurement.Phase
It is defined as like degree function：In formula,Value is only
It is only relevant with size n of sub-block.
C2 layers are specially：
C2 layers are also a convergence-level, respond global maximum by the similarity for calculating each characteristic sub-block, obtain
C2 layer features, the result of calculation of C2 layers are the vectors of a d dimension.
(122) screen characteristic sub-block
C2 layers feature sends into classifier training, and grader can select SVM or Adaboost.Using the feature of grader
Selective power, is ranked up according to the respective weights size that classifier training is obtained to d sub-block template, selects maximum k
Display model of the sub-block template as target following.
The present invention is so that SVM classifier is realized as an example.The C2 features extracted are sent into into SVM classifier training, grader is obtained
Discriminant function be：In formula, weights omegai, correspondence support vector weights omegaiAnd
Biasing b is obtained by training.Select k maximum weights omegaiCorresponding sub-block template xi, k takes 100.This 100 sub-blocks will also
For the second wheel training.
(13) second wheel off-line trainings.
The flow process of the second wheel off-line training is basically identical with first round training flow process, after all Sample Establishing image pyramids
S1 is sequentially passed through, the calculating of C1, S2 and C2 layer computing unit, the C2 layer features of output pass to grader and is trained.It is different
It is the sub-block template after step (122) screening that part is the sub-block template used in S2 layers so that the C2 layer feature dimensions of output
Number is greatly reduced.And the grader that the second training in rotation gets will be used for online tracking phase for recognizing target.
(2) online tracking step：
In order to improve the search efficiency of tracker, the present invention adopts an iterative procedures search target.The process of search is such as
Under：Initialized target detection zone.Calculated according to the S2 layers and C2 layers of BIM, find the best match of all sub-block templates, while
Calculate the grader fraction of detection zone.The position of target is predicted using matched sub-block, target candidate position regional ensemble is obtained；Repeatedly
A candidate region is only detected in generation detection candidate region, each iteration, extracts the BIM features of detection zone and calculates grader
Fraction is detected.In iteration, the selection strategy of candidate samples is determined by the grader fraction of last iteration.It is in detection process, right
The sub-block matching that S2 and C2 layers are newly produced in calculating carries out sub-block prediction, and new predicting the outcome is added to candidate samples set,
Simultaneously in Rejection of samples set some insecure samples reducing the number of times of iteration.When in candidate samples set, no sample is needed
When to be detected, iteration ends.The grader fraction of the candidate samples of all detections is ranked up, and determines final tracking knot
Really.
(21) initialize detection zone.
Detection image initialization detection zone is treated, the image pyramid of detection zone is set up.If being currently the first frame
Or previous frame is when cannot position target, detection zone is initialized using target detection.If during continuous tracking, then
Region determined by previous frame tracking result can be used as detection zone.
(22) BIM features are extracted to the image pyramid of detection zone, are sent to the grader that step (13) is trained,
Grader exports the result of determination of detection zone；Multiple matched sub-blocks are obtained during BIM features are extracted, using matching
Block predicts object candidate area.
(221) BIM features are extracted to the image pyramid of detection zone, is sent to the classification that step (13) is trained
Device, grader export the result of determination of detection zone, i.e. grader fraction.
(222) initialize candidate collection
BIM features are extracted to the image pyramid of detection zone, the corresponding best match sub-block of k sub-block template is obtained.
The candidate region of target is predicted using matched sub-block：The geometric center of the image pyramid bottom of detection zone is made to sit for (x, y)
Mark origin, the vector of the center relative coordinates origin of matched sub-block are defined as coordinate position of the matched sub-block relative to zero
(dx, dy), makes the distance between sub-block template be Dij, i, j=1 ..., k, the distance between its Corresponding matching sub-block are D 'ijMeter
Calculate the target scale factor of present frameMedian () takes median to all of ratio, improves yardstick
The robustness of calculating；The center of target is predicted according to each matched sub-blockStructure
The object candidate area built centered on the center of target.The collection being made up of these object candidate areas is collectively referred to as Candidate Set
Close.
(23) potential target region is selected according to grader result of determination iteration from candidate region.
(231) one object candidate area of selection from candidate collection.
(232) judge whether the object candidate area contains target according to grader fraction S (x), if it is determined that result is yes,
The candidate region is then thought near target, otherwise, the candidate region is away from target.For example, assert grader fraction S (x) non-negative
Target is corresponded to then, conversely, being non-targeted.So, in this step, if S (x) >=0, assertive goal closely detection zone
Domain, then choose from the nearest candidate region in current goal candidate region as potential target region from all candidate regions, and
Potential target region and the candidate samples away from detection zone are rejected from candidate collection.If S (x) ＜ 0, assertive goal does not exist
Near current goal candidate region, then calculate the center average of the candidate region away from current detection region, chosen distance
The nearest candidate region of this average as potential target region, and from candidate collection reject potential target region and near work as
The candidate region of front detection zone.
(233) the potential target extracted region BIM feature selected to step (232), is sent to step (13) and trains
Grader, the result of determination in grader output potential target region.During BIM features are extracted, S2 layers and C2 layers
Calculating may find new matched sub-block.The operation similar with step (222) is performed to new matched sub-block, new mesh is obtained
New object candidate area is added to candidate collection by mark candidate region.
(234) judge that candidate collection then enters step (24), otherwise return to step (231) whether as empty, if so,.
(24) it is chosen for target area maximum probability person and is final target area.For example, assert grader fraction S
X () non-negative then corresponds to target, conversely, being non-targeted.So, in this step, if max (S (x)) >=0, select maximum point
The result of class device fraction is used as final tracking result；If max (S (x)) ＜ 0, present frame cannot find target.
As it will be easily appreciated by one skilled in the art that the foregoing is only presently preferred embodiments of the present invention, not to
The present invention, all any modification, equivalent and improvement made within the spirit and principles in the present invention etc. is limited, all should be included
Within protection scope of the present invention.
Claims (5)
1. a kind of video target tracking method of analog vision cortex mechanism, it is characterised in that including off-line training and it is online with
Two stages of track；
The off-line training step is comprised the following steps：
(11) build the positive and negative sample set of tracking target image；
(12) align negative sample and set up image pyramid, extract the BIM features of image pyramid, first is carried out according to BIM features
The classifier training of wheel, the weight size obtained according to classifier training screen feature to reduce intrinsic dimensionality；
(13) classifier training of the second wheel is carried out using the BIM features reduced after intrinsic dimensionality；
The online tracking phase is comprised the following steps：
(21) detection image initialization detection zone is treated, the image pyramid of detection zone is set up；
(22) BIM features are extracted to the image pyramid of detection zone, is sent to the grader that step (13) is trained, classify
Device exports the result of determination of detection zone；Multiple matched sub-blocks are obtained during BIM features are extracted, it is pre- using matched sub-block
Survey object candidate area；
(23) potential target region is selected according to grader result of determination iteration from candidate region；
(24) target area maximum probability person is chosen from potential target region be final target area.
2. the video target tracking method of analog vision cortex mechanism according to claim 1, it is characterised in that the step
Suddenly the specific implementation of (12) is：
(121) BIM features are extracted to each Sample Establishing image pyramid and, it is special including S1 layer features, S2 layer features, S3 layers
Seek peace S4 layer features；
The calculation of the S1 layers feature is：Image pyramid is defined as into S1 layers by the result that S1 layers computing unit is calculated
Feature；S1 layer computing units are the wave filter groups that 4 Gabor filters are constituted, and are expressed as：
y0=-xsin θ+ycos θ
In formula, F (x, y, θ) represents wave filter in the corresponding response value in coordinate (x, y) place, and θ controls the direction of wave filter, value
Respectively4 direction 4 Gabor filters of correspondence；γ represents the length-width ratio of wave filter, and σ represents wave filter
Bandwidth, λ represents the wavelength of wave filter；
The calculation of the C1 layers feature is：S1 layers feature is defined as into C1 layers by the result of calculation of C1 computing units special
Levy；C1 layers computing unit extracts maximum, the local of computing unit setting in the corresponding local neighborhood of S1 layer features to input
Neighborhood is the local pyramid extracted from the corresponding image pyramid of S1 layer features；
The calculation of the S2 layers feature is：C1 layers feature is defined as into S2 layers by the result of calculation of S2 layer computing units special
Levy, S2 layers computing unit calculates characteristic sub-block X and each sub-block template in C1 layer pyramidsBetween similarity measurements,
It is describedIt is the sub-block randomly selected from the C1 layer features of multiple samples, d is the quantity for extracting sub-block；It is described similar
Degree function is defined as：In formula,
The calculation of the C2 layers feature is：S2 layers feature is defined as into C2 layers by the result of calculation of C2 layer computing units special
Levy, C2 layers computing unit calculates the similarity response global maximum of each characteristic sub-block in C1 layer pyramids；
(122) screen characteristic sub-block
C2 layers feature is sent into into classifier training, using the feature selection ability of grader, to d sub-block template according to grader
The respective weights size that training is obtained is ranked up, and selects k maximum sub-block template to be used for the classifier training of the second wheel, k<
d。
3. the video target tracking method of analog vision cortex mechanism according to claim 2, it is characterised in that the step
Suddenly the specific implementation of (22) is：
(221) BIM features are extracted to the image pyramid of detection zone, is sent to the grader that step (13) is trained, point
Class device exports the result of determination of detection zone；
(222) BIM features are extracted to the image pyramid of detection zone, obtain the corresponding best match sub-block of k sub-block template,
The candidate region of target is predicted using matched sub-block, concrete implementation mode is：
The geometric center for making the image pyramid bottom of detection zone is (x, y) zero, and the center of matched sub-block is relative to be sat
The vector of mark origin is defined as coordinate position (dx, dy) of the matched sub-block relative to zero, makes between two sub-block templates
Distance is Dij, i, j=1 ..., k, the distance between its Corresponding matching sub-block are D 'ij；Calculate the target scale factor of present frameMedian () represents that reduced value takes median；The centre bit of target is predicted according to each matched sub-block
PutThe object candidate area centered on the center of target is built, by these
The collection that object candidate area is constituted is collectively referred to as candidate collection.
4. the video target tracking method of analog vision cortex mechanism according to claim 3, it is characterised in that the step
Suddenly the specific implementation of (23) is：
(231) one object candidate area of random selection from candidate collection.
(232) if grader output is judged to target, choose nearest from current goal candidate region from all candidate regions
Candidate region as potential target region, and reject potential target region and the time away from detection zone from candidate collection
Sampling sheet；Otherwise, the center average of the candidate region away from current detection region is calculated, this average of chosen distance is nearest
Candidate region is used as potential target region, and rejects potential target region from candidate collection and near current detection region
Candidate region；
(233) the potential target extracted region BIM feature selected by step (232), be sent to that step (13) trains point
Class device, the result of determination in grader output potential target region；During BIM features are extracted, the calculating of S2 layers and C2 layers
New matched sub-block may be found, the operation similar with step (222) is performed to new matched sub-block, be obtained new target and wait
New object candidate area is added to candidate collection by favored area；
(234) judge that candidate collection then enters step (24), otherwise return to step (231) whether as empty, if so,.
5. the video target tracking method of analog vision cortex mechanism according to claim 3, it is characterised in that the step
Suddenly (24) are chosen correspondence grader result of determination from potential target region and show most possibly to be finally for target area person
Target area.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610921845.3A CN106530324A (en) | 2016-10-21 | 2016-10-21 | Visual cortex mechanism simulated video object tracking method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610921845.3A CN106530324A (en) | 2016-10-21 | 2016-10-21 | Visual cortex mechanism simulated video object tracking method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106530324A true CN106530324A (en) | 2017-03-22 |
Family
ID=58291528
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610921845.3A Pending CN106530324A (en) | 2016-10-21 | 2016-10-21 | Visual cortex mechanism simulated video object tracking method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106530324A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111612819A (en) * | 2020-05-15 | 2020-09-01 | 安徽工程大学 | Moving target tracking method and system based on image pyramid absolute error sum |
| CN112488135A (en) * | 2020-12-28 | 2021-03-12 | 深圳供电局有限公司 | Method for classifying BIM three-dimensional image features of transformer substation |
| CN112699718A (en) * | 2020-04-15 | 2021-04-23 | 南京工程学院 | Scale and illumination self-adaptive structured multi-target tracking method and application thereof |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101807245A (en) * | 2010-03-02 | 2010-08-18 | 天津大学 | Artificial neural network-based multi-source gait feature extraction and identification method |
| CN102902971A (en) * | 2012-08-31 | 2013-01-30 | 电子科技大学 | Method and system for conducting statistics on elevator visitor flow based on intelligent visual perception |
| CN103914834A (en) * | 2014-03-17 | 2014-07-09 | 上海交通大学 | Significant object detection method based on foreground priori and background priori |
| CN105022990A (en) * | 2015-06-29 | 2015-11-04 | 华中科技大学 | Water surface target rapid-detection method based on unmanned vessel application |
| CN105809718A (en) * | 2016-03-14 | 2016-07-27 | 西南交通大学 | Object tracking method with minimum trajectory entropy |
-
2016
- 2016-10-21 CN CN201610921845.3A patent/CN106530324A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101807245A (en) * | 2010-03-02 | 2010-08-18 | 天津大学 | Artificial neural network-based multi-source gait feature extraction and identification method |
| CN102902971A (en) * | 2012-08-31 | 2013-01-30 | 电子科技大学 | Method and system for conducting statistics on elevator visitor flow based on intelligent visual perception |
| CN103914834A (en) * | 2014-03-17 | 2014-07-09 | 上海交通大学 | Significant object detection method based on foreground priori and background priori |
| CN105022990A (en) * | 2015-06-29 | 2015-11-04 | 华中科技大学 | Water surface target rapid-detection method based on unmanned vessel application |
| CN105809718A (en) * | 2016-03-14 | 2016-07-27 | 西南交通大学 | Object tracking method with minimum trajectory entropy |
Non-Patent Citations (4)
| Title |
|---|
| HAIBIN DUAN 等: "Biologically Inspired Model with Feature Selection for Target Recognition Using Biogeography-Based Optimization", 《JOURNAL OF AEROSPACE INFORMATION SYSTEMS》 * |
| MIN LI 等: "Robust visual tracking based on simplified biologically inspired features", 《IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING》 * |
| RUYI XU 等: "Robust patch-based visual tracking using biologically inspired model", 《PROCEEDINGS OF THE 4TH INTERNATIONAL CONFERENCE ON INTERNET MULTIMEDIA COMPUTING AND SERVICE》 * |
| 黄双萍: "通用视觉目标识别的关键技术研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112699718A (en) * | 2020-04-15 | 2021-04-23 | 南京工程学院 | Scale and illumination self-adaptive structured multi-target tracking method and application thereof |
| CN112699718B (en) * | 2020-04-15 | 2024-05-28 | 南京工程学院 | Scale and illumination self-adaptive structured multi-target tracking method and application thereof |
| CN111612819A (en) * | 2020-05-15 | 2020-09-01 | 安徽工程大学 | Moving target tracking method and system based on image pyramid absolute error sum |
| CN112488135A (en) * | 2020-12-28 | 2021-03-12 | 深圳供电局有限公司 | Method for classifying BIM three-dimensional image features of transformer substation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113567984B (en) | Method and system for detecting artificial small target in SAR image | |
| CN106446930B (en) | Robot operative scenario recognition methods based on deep layer convolutional neural networks | |
| CN105809198B (en) | SAR image target recognition method based on depth confidence network | |
| CN106156744B (en) | SAR target detection method based on CFAR detection and deep learning | |
| CN105184309B (en) | Classification of Polarimetric SAR Image based on CNN and SVM | |
| Zhang et al. | Pedestrian detection method based on Faster R-CNN | |
| CN108009509A (en) | Vehicle target detection method | |
| CN102722712B (en) | Multiple-scale high-resolution image object detection method based on continuity | |
| CN101763507B (en) | Face recognition method and face recognition system | |
| CN107967451A (en) | A kind of method for carrying out crowd's counting to static image using multiple dimensioned multitask convolutional neural networks | |
| CN107066559A (en) | A kind of method for searching three-dimension model based on deep learning | |
| CN109034224A (en) | Hyperspectral classification method based on double branching networks | |
| CN105869173A (en) | Stereoscopic vision saliency detection method | |
| CN102509104B (en) | Confidence map-based method for distinguishing and detecting virtual object of augmented reality scene | |
| CN104537647A (en) | Target detection method and device | |
| CN103824093B (en) | It is a kind of based on KFDA and SVM SAR image target's feature-extraction and recognition methods | |
| CN105426919A (en) | Significant guidance and unsupervised feature learning based image classification method | |
| CN103218831A (en) | Video moving target classification and identification method based on outline constraint | |
| CN107480620A (en) | Remote sensing images automatic target recognition method based on heterogeneous characteristic fusion | |
| CN105574545B (en) | The semantic cutting method of street environment image various visual angles and device | |
| Mao et al. | A deep convolutional neural network trained on representative samples for circulating tumor cell detection | |
| CN108564111A (en) | A kind of image classification method based on neighborhood rough set feature selecting | |
| CN102945374A (en) | Method for automatically detecting civil aircraft in high-resolution remote sensing image | |
| CN107341505A (en) | A kind of scene classification method based on saliency Yu Object Bank | |
| CN110334656A (en) | Multi-source Remote Sensing Images Clean water withdraw method and device based on information source probability weight |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170322 |