CN107016691A - Moving target detecting method based on super-pixel feature - Google Patents
Moving target detecting method based on super-pixel feature Download PDFInfo
- Publication number
- CN107016691A CN107016691A CN201710243141.XA CN201710243141A CN107016691A CN 107016691 A CN107016691 A CN 107016691A CN 201710243141 A CN201710243141 A CN 201710243141A CN 107016691 A CN107016691 A CN 107016691A
- Authority
- CN
- China
- Prior art keywords
- super
- pixel
- pixel block
- seed point
- sample
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses the moving target detecting method based on super-pixel feature, including:Super-pixel segmentation is carried out to every two field picture using SLIC0 split plot designs, and extracts the pixel average of each super-pixel block as its super-pixel characteristic value;Using the super-pixel characteristic value on initial seed point position as this super-pixel block sample value, the sample pattern of the super-pixel block on each initial seed point position is built according to preceding N frames super-pixel block sample value;Extract a new two field picture, calculate the Euclidean distance in the sample pattern of super-pixel block between each sample in the super-pixel block and the seed point in each of which initial seed point, if the sum that Euclidean distance is less than distance threshold between certain super-pixel block and sample judges the super-pixel block for foreground blocks less than matching threshold;All prospect super-pixel block just constitute moving object detection result in this two field picture.The inventive method introduces super-pixel feature, obtains good object edge information, it is ensured that the integrality of moving target outward appearance.
Description
Technical field
The invention belongs to technical field of image processing, and in particular to the moving target detecting method based on super-pixel feature,
It can be applied to robot control and field of intelligent monitoring.
Background technology
Along with the development of computer technology, Detection for Moving Target analyzed as video motion in core content,
The hot issue always studied in the last few years.Its purpose is exactly so as to realize by the dependency relation between sequence of frames of video
Quickly and accurately detect the moving target in monitor video.And the moving object detection field under static background, move mesh
Mark detection method is broadly divided into frame differential method, optical flow and background modeling method.
Wherein background modeling method is most-often used method, in recent years in popular printenv background modeling algorithm master
Have, Barnich et al. (Barnich O, Van Droogenbroeck M.ViBE:A powerful random
technique to estimate the background in video sequences[C]//IEEE
International Conference on Acoustics,Speech and Signal Processing.IEEE
Computer Society,2009:ViBe (Visual Background Extractor) algorithm 945-948.) is proposed,
ViBe algorithms have the spatial characteristics of close pixel value using neighbor pixel, are that each background dot stores a sample
Collection, then calculate new pixel value and sample set distance distinguish before, background.ViBe has good robustness and real-time, but
Occur under specified conditions Ghost regions and for dynamic background treatment effect it is undesirable the problems such as.Hofmann et al.
(Hofmann M,Tiefenbacher P,Rigoll G.Background segmentation with feedback:The
Pixel-Based Adaptive Segmenter[J].2012:38-43.) background is modeled by using pixel value, carried
PBAS (Pixel-Based Adaptive Segmenter) algorithm is gone out, has been each pixel institute mainly by background model
The threshold determination current pixel point set up is foreground point or background dot, therefore background model and each threshold value always can be with the back ofs the body
The renewal of scape and update.
But Pixel-level background modeling algorithm has an essential problem, is just the absence of the support of spatial information, reflection
It is the changing features situation of each pixel, so may result in the loss of information, such as moving object detection not exclusively, is detected
As a result situations such as containing a large amount of flase drop points in.
The content of the invention
It is an object of the invention to overcome deficiency of the prior art, there is provided the inspection of the moving target based on super-pixel feature
Survey method, has preferable marginal information while the characteristics of super-pixel number is controllable, good to obtain using SLIC0 partitioning algorithms
Object edge information, it is ensured that the integrality of moving target outward appearance, while can effectively suppress the interference of dynamic background, obtain more
Good robustness.
In order to solve the above technical problems, the invention provides the moving target detecting method based on super-pixel feature, it is special
Levying is, comprises the following steps:
Step S1, for the preceding N frames of video sequence, super-pixel segmentation is carried out to every two field picture using SLIC0 split plot designs, and
The pixel average of each super-pixel block is extracted as its super-pixel characteristic value;
Step S2, the super-pixel characteristic value on initial seed point position evenly distributed when being split using SLIC0 super picture as this
Plain block sample value, the sample pattern of the super-pixel block on each initial seed point position is built according to preceding N frames super-pixel block sample value;
Step S3, extracts a new two field picture, carries out super-pixel segmentation to image using SLIC0 split plot designs, calculates its every
Euclidean in super-pixel block and the seed point in individual initial seed point in the sample pattern of super-pixel block between each sample away from
From if the sum that Euclidean distance is less than distance threshold R between certain super-pixel block and sample judges the super picture less than matching threshold
Plain block is foreground blocks;All prospect super-pixel block just constitute moving object detection result in this two field picture.
Further, the step of SLIC0 super-pixel segmentations are implemented in step S1 is as follows:
Step S11, is first according to previously given super-pixel number, initial seed point is evenly distributed in entire image,
If picture to be split is altogether containing Num pixel, super-pixel segmentation number set in advance is K, can calculate each super picture
The size of element is about Num/K, and the distance between adjacent initial seed point is approximately then
Step S12, calculates the Grad of all pixels point in seed point n*n neighborhoods, and the neighborhood inside gradient is minimum
Pixel be set to new seed point;
Step S13, for the pixel searched in 2S*2S regions, calculates its distance with the seed point, distance
Metric form has color distance measurement and space length measurement, shown in corresponding computational methods such as formula (1):
Wherein,Represent the color distance between current pixel point j and seed point i, l, a, b is respectively that Lab colors are empty
Between in three elements, l represents brightness, and l scope is 0~100 to represent that from black to white a is represented from carmetta to green
Scope, b represents the scope from yellow to blueness.The space length between current pixel point j and seed point i, x are represented, y is
XY coordinates, Di(j) represent distance metric final between current pixel point j and seed point i, joint color distance and space away from
From NsIt is maximum space distance in class, is defined asNcFor maximum color distance, fix normal with one
Number m is replaced;Joint color distance and space length, so as to obtain final distance metric Di(j), as shown in formula (2):
During seed point search surrounding pixel point, if pixel can be searched by multiple seed points, calculate
The pixel and the minimum value of the distance of surrounding seed point, and it is used as the poly- of the pixel with the seed point corresponding to the minimum value
Class center;
Step S14, continuous iteration no longer changes until each pixel cluster centre, so as to divide the image into be each
Super-pixel block.
Further, the pixel average of super-pixel block is as the characteristic value of the super-pixel feature, and characteristic value calculation formula is such as
Shown in formula (3):
Wherein, p (x) is the pixel value of the pixel in each super-pixel block, and SqFor q-th of super-pixel block, #SqFor q
Pixel number, S in individual super-pixel blockq(p) characteristic value then for q-th of super-pixel block.
Further, in step S2, the sample pattern of the super-pixel block on i-th of initial seed point position of structure is represented
As shown in formula (4):
Wherein Mi(p) it is the sample pattern of the super-pixel block on i-th of initial seed point position, is each super-pixel block
N number of sample is built, andThen represent the n-th sample in the sample pattern of the super-pixel block on i-th of initial seed point position
This value, and the characteristic value of the super-pixel block of its sample value then on i-th of initial seed point position is represented.
Further, in step S3, shown in Euclidean distance such as formula (5):
WhereinRepresent the characteristic value of the current super-pixel block on i-th of initial seed point positionWith it
K-th of sample value in sample patternBetween Euclidean distance.
Further, the testing result according to previous frame image, updates distance threshold and sample pattern, then extracts down again
The new image of one frame, repeat step S3 process obtains moving object detection result, constantly repeats this process until being owned
The moving object detection result of two field picture;
Wherein shown in the more new formula such as formula (7) of distance threshold:
Wherein α, β are a coefficients set in advance, and σ (p) is background complexity;
And sample pattern is updated with the δ (p) of probability 1/, update mode such as formula (8) institute of sample pattern update probability
Show:
Wherein, res (p)=1 represents that current super-pixel is detected as prospect, and res (p)=0 represents that current super-pixel is detected
Survey as background, γ is coefficient set in advance, and σ (p) is background complexity;And only when the super-pixel block is background super-pixel
During block, the renewal of background model can be just carried out, i.e., a sample value S is randomly selected from sample patternk(p) and with current super picture
Plain characteristic value St(p) replace;
Shown in background complexity σ (p) measure formulas such as formula (6):
Compared with prior art, the beneficial effect that is reached of the present invention is:The present invention using SLIC0 algorithms there is segmentation to imitate
Fruit preferably, can keep preferable object boundary information.For each super-pixel block, by the pixel value average of wherein pixel
It is used as the characteristic value of the block of pixels so that input feature vector contains Pixel-level information and area information, there can be preferably inspection
Survey result., as fixed feature value position, preceding N frames have been selected by the use of in the super-pixel block where the position of initial seed point
The characteristic value of image initial seed point super-pixel block builds sample pattern.Initial sample pattern is built complete after, used away from
Moving object detection is adaptively completed from threshold value and sample pattern turnover rate, while good result is achieved, detection algorithm
Speed is greatly promoted, and the detection speed of piece image can improve more than hundred times.
Brief description of the drawings
Fig. 1 is the flow chart of the inventive method;
Fig. 2:It is some frames in four videos to scheme (a), and figure (b) is moving object detection true value, schemes (c) to use ViBe
The moving object detection result of algorithm, figure (d) is the moving object detection result using KDE algorithms, and figure (e) is to be calculated using PBAS
The moving object detection result of method, figure (f) is the moving object detection result using the inventive method.
Embodiment
The invention will be further described below in conjunction with the accompanying drawings.Following examples are only used for clearly illustrating the present invention
Technical scheme, and can not be limited the scope of the invention with this.
As shown in figure 1, a kind of moving target detecting method based on super-pixel feature of the present invention, comprises the following steps:
Step S1, for the preceding N frames of video sequence, super-pixel segmentation is carried out to every two field picture using SLIC0 split plot designs, and
The pixel average of each super-pixel block is extracted as its super-pixel characteristic value.
SLIC0, i.e., simple linear iteraction cluster, is that a kind of thought is simple, realizes convenient super-pixel segmentation algorithm.Should
Algorithm is first to assign 5 dimensional feature vectors for each color pixel cell, is the CIELab face after coloured image conversion respectively
The colour space and XY coordinates, then carry out Local Clustering to image pixel using 5 dimensional vector distance metrics of construction and generate
Final segmentation result.
The step of SLIC0 super-pixel segmentations are implemented is as follows:
Step S11, initializes cluster centre first, also referred to as initialization seed point, i.e., according to previously given super-pixel
Number, evenly distributes seed point in entire image.Assuming that picture to be split is altogether containing Num pixel, it is set in advance super
Pixel segmentation number is K, then can calculate the size about Num/K of each super-pixel, and between adjacent initial seed point
Distance (step-length) be approximately then
Step S12, then reselects seed point in seed neighborhood of a point, that is, calculates seed point n*n and (generally choose
N=3) in neighborhood all pixels point Grad, and the minimum pixel of the neighborhood inside gradient is set to new seed point.
Which step S13, then (belongs to for all pixels point distribution class label in the neighborhood around each seed point
Individual cluster centre).Wherein SLIC0 hunting zone has been limited in 2S*2S, is different from the k-means of standard and is searched in whole figure
Rope, can so accelerate the speed of algorithmic statement.
For the pixel searched in 2S*2S regions, its distance with the seed point, the metric form of distance are calculated
There are color distance measurement and space length measurement, shown in corresponding computational methods such as formula (1):
Wherein,Represent the color distance between current pixel point j and seed point i, l, a, b is respectively that Lab colors are empty
Between in three elements, l represents brightness (Luminosity), and l scope is 0~100 to represent that from black to white a is represented from ocean
The red scope to green, b represents the scope from yellow to blueness.Represent the sky between current pixel point j and seed point i
Between distance, x, y is XY coordinates.Di(j) represent distance metric final between current pixel point j and seed point i, joint color away from
From and space length, NsIt is maximum space distance in class, is defined asNcFor maximum color distance, but
Due to its can it is different with different cluster process and different pictures, so with a fixed constant m (span [1,
40] 10 replacements, are typically taken).Joint color distance and space length, so as to obtain final distance metric Di(j), such as formula
(2) shown in:
During seed point search surrounding pixel point, if pixel can be searched by multiple seed points, calculate
The pixel and the minimum value of the distance of surrounding seed point, and it is used as the poly- of the pixel with the seed point corresponding to the minimum value
Class center.
Step S14, it is necessary to carry out continuous iteration optimization until each pixel cluster centre after once search is completed
No longer change, so as to divide the image into as super-pixel block.
The pixel average of super-pixel block after being split using SLIC0 super-pixel segmentations algorithm is used as the super-pixel feature
Characteristic value, shown in characteristic value calculation formula such as formula (3):
Wherein, p (x) is the pixel value of the pixel in each super-pixel block, and SqFor q-th of super-pixel block, #SqFor q
Pixel number, S in individual super-pixel blockq(p) characteristic value then for q-th of super-pixel block.
Super-pixel characteristic value on step S2, former N frames initial seed point position is built just for this super-pixel block sample value
The sample pattern of super-pixel block on beginning sample pattern, each initial seed point position includes N number of sample.
The background sample value of ViBe algorithms is derived from the random value around pixel in 8*8 neighborhoods, constitutes N number of sample.But
The value of sample does not use the sample value scheme of tradition ViBe algorithms in the background model being characterized with super-pixel, because
Super-pixel segmentation has good edge segmentation effect, and the characteristic value of the super-pixel around it is then possible to the far of meeting difference,
And without neighborhood territory pixel point identical characteristic, so the structure of the first frame completion initial back-ground model can not be used.
Super-pixel number and super-pixel number set in advance after the segmentation of SLIC0 super-pixel segmentations algorithm are not necessarily one
Cause, there can be certain discrepancy, so thus characteristic draws and can not directly use each super-pixel characteristic value as sample
This value.
If the super-pixel characteristic value of same position should in the moving object detection under video camera static position, video sequence
This is all similar.In this case, it is special present invention uses the super-pixel on initial seed point position in step S11
Value indicative builds sample pattern as sample value.
It is super on each initial seed point position of N frames before the present embodiment is directly employed for preferably Detection results
Pixel characteristic value is that to build initial sample pattern (or be background model, sample set, in this paper to each super-pixel block sample value
It is same concept in description).The initial sample pattern built is represented as shown in formula (4):
Wherein Mi(p) it is the sample pattern of the super-pixel block on i-th of initial seed point position, is each super-pixel block
N number of sample is built, andThen represent the n-th sample in the sample pattern of the super-pixel block on i-th of initial seed point position
This value, and the characteristic value of the super-pixel block of its sample value then on i-th of initial seed point position is represented.
Step S3, extracts a new two field picture, carries out super-pixel segmentation to every two field picture using SLIC0 split plot designs, and carry
The pixel average of each super-pixel block is taken as its super-pixel characteristic value, calculate super-pixel block in each of which initial seed point with
Euclidean distance in the seed point in the sample set of super-pixel block between each sample, as shown in formula (5):
WhereinRepresent the characteristic value of the current super-pixel block on i-th of initial seed point positionWith it
K-th of sample value in sample patternBetween Euclidean distance.
The distance calculated is compared with distance threshold R set in advance, if less than distance threshold R, matching
Number Jia 1, otherwise continues the distance between calculating and next sample value.After all statistics is completed, number will be matched and pre-
The matching threshold first set is compared, if less than the threshold value, judging the super-pixel block for foreground blocks, on the contrary then be background
Block.After the completion of the super-pixel block traversal on all initial seed point positions in every piece image, by testing result binaryzation
Represent, i.e., foreground blocks are " 1 ", and background block is " 0 ", and all prospect super-pixel block just constitute testing result.
Step S4, according to the testing result of previous frame image, update distance threshold and sample pattern, extracts next frame new
Image, repeat step S3 process obtains moving object detection result, constantly repeats this process, until obtaining all two field pictures
Foreground detection result.
In order to avoid a certain sample is retained in sample pattern for a long time, so that the accuracy of sample pattern is influenceed, so need
Sample pattern is updated, introducing randomly updates mechanism, i.e., after super-pixel block completes to differentiate, if the super-pixel block belongs to
In background block, then there is 1/ δ (p) probability updating sample set, i.e., a sample value S is randomly selected from sample patternk(p) it is used in combination
Current super-pixel characteristic value St(p) replace, so as to update sample set.So, the mechanism ensure that the life of sample in sample pattern
The life cycle exponentially decays, so as to effectively adapt to the change of scene.
In ViBe algorithms, sample pattern, which updates, possesses a fixed δ (p) of turnover rate 1/, and fixed turnover rate can not be with
The change of background and change, in the case of dynamic background, the ripple of such as water, the shake of leaf etc., just occur a large amount of
Background dot is mistakenly detected as the situation of foreground point, thus fixed turnover rate and distance threshold be poorly suited for use in it is dynamic
State background, should suitably adjust turnover rate and distance threshold to ensure good Detection results.
Background dynamics degree is measured set forth herein standard deviation using current super-pixel block sample pattern.For current super
Pixel characteristic value St(p) with N number of sample value in sample pattern, it regard its standard deviation sigma (p) as the super-pixel block background complexity
Measurement, when a certain super-pixel block keeps background block or when foreground blocks, the standard deviation of dynamic background block just than larger, this
Sample can just distinguish dynamic background, and when a certain super-pixel block just becomes background block from foreground blocks or become from background block
When foreground blocks, though standard deviation is difficult to differentiate between dynamic background and non-dynamic background dot, for result influence it is micro- its
It is micro-, it might even be possible to ignore.Shown in the measure formulas of background complexity such as formula (6):
According to above-mentioned dynamic background complexity σ (p) real-time update distance threshold, for the more video of background dynamics
Sequence, should suitably increase distance threshold, so that the influence that dynamic background is caused to testing result is reduced, while the distance threshold is not
It is suitable too high, otherwise it will cause in subsequent frame, prospect super-pixel block will be detected as background block, it is on the contrary then should reduce apart from threshold
Value, while the distance threshold is also not suitable for too low, otherwise will cause in subsequent frame, background super-pixel block will be detected as prospect and surpass
Block of pixels.So ensure that and be constantly within rational excursion.More new formula such as formula (7) institute of distance threshold
Show:
Wherein distance threshold R initial value, which is taken as 20, α, β, to be taken respectively in a coefficient set in advance, the present embodiment
It is worth for 0.16 and 0.01.
And for the δ (p) of sample pattern update probability 1/, it is general to updating using processing mode more similar with distance threshold
Rate carries out self-adaptive processing.And when the super-pixel block is identified as prospect super-pixel block, it should it is appropriate to reduce sample pattern more
New probability, otherwise the update probability of appropriate increase sample pattern, and only when the super-pixel block is background super-pixel block,
The renewal of sample pattern can be carried out.Shown in the update mode of sample pattern update probability such as formula (8):
Wherein, res (p)=1 represents that current super-pixel is detected as prospect, and res (p)=0 represents that current super-pixel is detected
Survey as background, γ be also in pre-set coefficient, the present embodiment value be 0.1.
Embodiment
The present embodiment is in hardware platform AMD A6,4GB RAM, software development environment Windows10, MATLAB2014a
Lower completion test job.25 sample values have been used in this experiment as the sample set of each super-pixel block, and this experiment is selected
The video sequence collection taken is changedetection.net data sets, wherein have chosen snowFall, Blizzard,
Tetra- video sequences of Pedestrian, canoe have carried out experimental verification, and have chosen ViBe algorithms, KDE calculations in the prior art
Method, PBAS algorithms and inventive algorithm have been contrast experiment, and each algorithm experimental result is as shown in Figure 2:
In fig. 2, figure (a) is some frames in four videos, and figure (b) is moving object detection true value, and figure (c) is
For using the moving object detection result of ViBe algorithms, figure (d) is the moving object detection result using KDE algorithms, is schemed (e)
It is the moving object detection result using PBAS algorithms, figure (f) is the moving object detection result using the inventive method.
As seen from Figure 2, compared to Pixel-level background modeling method, the inventive method is due to combining super-pixel segmentation
Algorithm so that there is more preferable outward appearance to describe for it, while the noise spot in background can be ignored substantially, because Pixel-level is too small
Change is for a super-pixel block, and influence is very little.For example in the canoe video sequences of figure (a), ViBe is calculated
Method occurs in that the noise spots of many under dynamic background, while have also appeared flase drop, originally without target be mistakenly detected as
Foreground target (as shown in figure (c)), the PBAS algorithms of the KDE algorithms and figure (e) of scheming (d) detect failure substantially, and of the invention
Method is then substantially without noise spot, although some information have also been the absence of compared to true value, but have been showed among several algorithms
It is best.In the figure frame of (a) pedestrian video sequences the 972nd, the length length behind people is it also avoid due to the characteristic of super-pixel
Smear caused by flase drop.In blizzard video sequences, the testing result of the inventive method and PBAS algorithms is also ratio
It is relatively close to target true value, ViBe algorithms and KDE algorithms have many flase drop and detection leakage phenomenon.In snowFall video sequences
In row, because the super-pixel segmentation result of background is unsatisfactory, so directly resulting in the inventive method there occurs detection leakage phenomenon.Institute
Compared to traditional algorithm and some current algorithms, to combine the inventive method tool for the area information that super-pixel block is included
There is certain advantage, expression effect more preferably, is more beneficial for motion estimate or the target following in later stage etc..
It is demonstrated experimentally that boundary information can preferably be expressed by having combined the inventive method of super-pixel segmentation, compare simultaneously
In the background model algorithm of Pixel-level, because super-pixel segmentation is quite quick, institute in the process of the present invention in speed quickly, can be with
There are tens to arrive the lifting of hundred times, but be an impediment to the segmentation effect of super-pixel, the testing result to partial video sequence has one
Fixed influence.
Described above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, without departing from the technical principles of the invention, some improvement and modification can also be made, these improvement and modification
Also it should be regarded as protection scope of the present invention.
Claims (6)
1. based on the moving target detecting method of super-pixel feature, it is characterized in that, comprise the following steps:
Step S1, for the preceding N frames of video sequence, carries out super-pixel segmentation, and extract using SLIC0 split plot designs to every two field picture
The pixel average of each super-pixel block is used as its super-pixel characteristic value;
Step S2, the super-pixel characteristic value on initial seed point position evenly distributed when being split using SLIC0 is this super-pixel block
Sample value, the sample pattern of the super-pixel block on each initial seed point position is built according to preceding N frames super-pixel block sample value;
Step S3, extracts a new two field picture, super-pixel segmentation is carried out to image using SLIC0 split plot designs, at the beginning of calculating each of which
Euclidean distance in super-pixel block and the seed point in beginning seed point in the sample pattern of super-pixel block between each sample, if
The sum that Euclidean distance is less than distance threshold R between certain super-pixel block and sample then judges the super-pixel block less than matching threshold
For foreground blocks;All prospect super-pixel block just constitute moving object detection result in this two field picture.
2. the moving target detecting method according to claim 1 based on super-pixel feature, it is characterized in that, in step S1
The step of SLIC0 super-pixel segmentations are implemented is as follows:
Step S11, is first according to previously given super-pixel number, initial seed point is evenly distributed in entire image, if treating
Split picture altogether containing Num pixel, super-pixel segmentation number set in advance is K, can calculate each super-pixel
Size is about Num/K, and the distance between adjacent initial seed point is approximately then
Step S12, calculates the Grad of all pixels point in seed point n*n neighborhoods, and the minimum picture of the neighborhood inside gradient
Vegetarian refreshments is set to new seed point;
Step S13, for the pixel searched in 2S*2S regions, calculates its distance with the seed point, the measurement of distance
Mode has color distance measurement and space length measurement, shown in corresponding computational methods such as formula (1):
Wherein,Represent the color distance between current pixel point j and seed point i, l, a, during b is respectively Lab color spaces
Three elements, l represents brightness, and a represents the scope from carmetta to green, and b represents the scope from yellow to blueness.Generation
Space length between table current pixel point j and seed point i, x, y is XY coordinates, Di(j) current pixel point j and seed point are represented
Final distance metric, N between isIt is maximum space distance in class, is defined asNcFor maximum color
Distance, is replaced with a fixed constant m;Joint color distance and space length, so as to obtain final distance metric Di(j),
As shown in formula (2):
During seed point search surrounding pixel point, if pixel can be searched by multiple seed points, the picture is calculated
Vegetarian refreshments and the minimum value of the distance of surrounding seed point, and with the seed point corresponding to the minimum value as in the cluster of the pixel
The heart;
Step S14, continuous iteration no longer changes until each pixel cluster centre, so as to divide the image into as each super picture
Plain block.
3. the moving target detecting method according to claim 1 based on super-pixel feature, it is characterized in that, super-pixel block
Pixel average as the super-pixel feature characteristic value, shown in characteristic value calculation formula such as formula (3):
Wherein, p (x) is the pixel value of the pixel in each super-pixel block, and SqFor q-th of super-pixel block, #SqQ-th to surpass
Pixel number in block of pixels, Sq(p) characteristic value then for q-th of super-pixel block.
4. the moving target detecting method according to claim 1 based on super-pixel feature, it is characterized in that, in step S2,
The sample pattern of the super-pixel block on i-th of initial seed point position built is represented as shown in formula (4):
Wherein Mi(p) it is the sample pattern of the super-pixel block on i-th of initial seed point position, is that each super-pixel block builds N
Individual sample, andThe n-th sample value in the sample pattern of the super-pixel block on i-th of initial seed point position is then represented,
And the characteristic value of the super-pixel block of its sample value then on i-th of initial seed point position is represented.
5. the moving target detecting method according to claim 1 based on super-pixel feature, it is characterized in that, in step S3,
Shown in Euclidean distance such as formula (5):
WhereinRepresent the characteristic value of the current super-pixel block on i-th of initial seed point positionWith its sample mould
K-th of sample value in typeBetween Euclidean distance.
6. the moving target detecting method according to claim 1 based on super-pixel feature, it is characterized in that, according to previous frame
The testing result of image, updates distance threshold and sample pattern, the new image of next frame, repeat step S3 mistake is then extracted again
Journey obtains moving object detection result, constantly repeats this process until obtaining the moving object detection result of all two field pictures;
Wherein shown in the more new formula such as formula (7) of distance threshold:
Wherein α, β are a coefficients set in advance, and σ (p) is background complexity;
And sample pattern is updated with the δ (p) of probability 1/, shown in the update mode such as formula (8) of sample pattern update probability:
Wherein, res (p)=1 represents that current super-pixel is detected as prospect, and res (p)=0 represents that current super-pixel is detected as
Background, γ is coefficient set in advance, and σ (p) is background complexity;And only when the super-pixel block is background super-pixel block
When, the renewal of background model can be just carried out, i.e., a sample value S is randomly selected from sample patternk(p) and current super-pixel is used
Characteristic value St(p) replace;
Shown in background complexity σ (p) measure formulas such as formula (6):
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710243141.XA CN107016691B (en) | 2017-04-14 | 2017-04-14 | Moving target detecting method based on super-pixel feature |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710243141.XA CN107016691B (en) | 2017-04-14 | 2017-04-14 | Moving target detecting method based on super-pixel feature |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107016691A true CN107016691A (en) | 2017-08-04 |
CN107016691B CN107016691B (en) | 2019-09-27 |
Family
ID=59446639
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710243141.XA Active CN107016691B (en) | 2017-04-14 | 2017-04-14 | Moving target detecting method based on super-pixel feature |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107016691B (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108508425A (en) * | 2018-03-26 | 2018-09-07 | 微瞳科技(深圳)有限公司 | Foreground target detection method based on neighborhood characteristics under a kind of radar near-earth ambient noise |
CN108549833A (en) * | 2018-03-07 | 2018-09-18 | 江苏东大金智信息***有限公司 | A kind of target extraction method of accurate robust |
CN108881465A (en) * | 2018-07-03 | 2018-11-23 | 肖鑫茹 | A kind of intelligent monitor system based on big data |
CN109087330A (en) * | 2018-06-08 | 2018-12-25 | 中国人民解放军军事科学院国防科技创新研究院 | It is a kind of based on by slightly to the moving target detecting method of smart image segmentation |
CN109255321A (en) * | 2018-09-03 | 2019-01-22 | 电子科技大学 | A kind of visual pursuit classifier construction method of combination history and instant messages |
CN109448382A (en) * | 2018-12-20 | 2019-03-08 | 天津天地伟业信息***集成有限公司 | A kind of road depth of accumulated water monitoring and pre-alarming method |
CN110108362A (en) * | 2019-04-17 | 2019-08-09 | 江苏理工学院 | The adaptive online test method of color difference and device based on SLIC super-pixel segmentation |
CN110599517A (en) * | 2019-08-30 | 2019-12-20 | 广东工业大学 | Target feature description method based on local feature and global HSV feature combination |
CN110717356A (en) * | 2018-07-11 | 2020-01-21 | ***通信集团浙江有限公司 | Face recognition detection method and system |
CN110910417A (en) * | 2019-10-29 | 2020-03-24 | 西北工业大学 | Weak and small moving target detection method based on super-pixel adjacent frame feature comparison |
CN111311603A (en) * | 2018-12-12 | 2020-06-19 | 北京京东尚科信息技术有限公司 | Method and apparatus for outputting target object number information |
CN113409338A (en) * | 2021-06-24 | 2021-09-17 | 西安交通大学 | Super-pixel method based on probability distribution |
CN113506266A (en) * | 2021-07-09 | 2021-10-15 | 平安科技(深圳)有限公司 | Method, device and equipment for detecting tongue greasy coating and storage medium |
CN115272353A (en) * | 2022-10-07 | 2022-11-01 | 山东盛世恒机械制造有限公司 | Image processing method suitable for crack detection |
CN117274812A (en) * | 2023-10-08 | 2023-12-22 | 北京香田智能科技有限公司 | Tobacco plant counting method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105741277A (en) * | 2016-01-26 | 2016-07-06 | 大连理工大学 | ViBe (Visual Background Extractor) algorithm and SLIC (Simple Linear Iterative Cluster) superpixel based background difference method |
CN105825234A (en) * | 2016-03-16 | 2016-08-03 | 电子科技大学 | Superpixel and background model fused foreground detection method |
CN106408529A (en) * | 2016-08-31 | 2017-02-15 | 浙江宇视科技有限公司 | Shadow removal method and apparatus |
-
2017
- 2017-04-14 CN CN201710243141.XA patent/CN107016691B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105741277A (en) * | 2016-01-26 | 2016-07-06 | 大连理工大学 | ViBe (Visual Background Extractor) algorithm and SLIC (Simple Linear Iterative Cluster) superpixel based background difference method |
CN105825234A (en) * | 2016-03-16 | 2016-08-03 | 电子科技大学 | Superpixel and background model fused foreground detection method |
CN106408529A (en) * | 2016-08-31 | 2017-02-15 | 浙江宇视科技有限公司 | Shadow removal method and apparatus |
Non-Patent Citations (1)
Title |
---|
徐久强等: "面向运动目标检测的ViBe算法改进", 《东北大学学报(自然科学版)》 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108549833A (en) * | 2018-03-07 | 2018-09-18 | 江苏东大金智信息***有限公司 | A kind of target extraction method of accurate robust |
CN108508425A (en) * | 2018-03-26 | 2018-09-07 | 微瞳科技(深圳)有限公司 | Foreground target detection method based on neighborhood characteristics under a kind of radar near-earth ambient noise |
CN108508425B (en) * | 2018-03-26 | 2020-08-04 | 微瞳科技(深圳)有限公司 | Method for detecting foreground target based on neighborhood characteristics under radar near-earth background noise |
CN109087330A (en) * | 2018-06-08 | 2018-12-25 | 中国人民解放军军事科学院国防科技创新研究院 | It is a kind of based on by slightly to the moving target detecting method of smart image segmentation |
CN108881465A (en) * | 2018-07-03 | 2018-11-23 | 肖鑫茹 | A kind of intelligent monitor system based on big data |
CN110717356A (en) * | 2018-07-11 | 2020-01-21 | ***通信集团浙江有限公司 | Face recognition detection method and system |
CN109255321A (en) * | 2018-09-03 | 2019-01-22 | 电子科技大学 | A kind of visual pursuit classifier construction method of combination history and instant messages |
CN109255321B (en) * | 2018-09-03 | 2021-12-10 | 电子科技大学 | Visual tracking classifier construction method combining history and instant information |
CN111311603A (en) * | 2018-12-12 | 2020-06-19 | 北京京东尚科信息技术有限公司 | Method and apparatus for outputting target object number information |
CN109448382A (en) * | 2018-12-20 | 2019-03-08 | 天津天地伟业信息***集成有限公司 | A kind of road depth of accumulated water monitoring and pre-alarming method |
CN110108362A (en) * | 2019-04-17 | 2019-08-09 | 江苏理工学院 | The adaptive online test method of color difference and device based on SLIC super-pixel segmentation |
CN110599517A (en) * | 2019-08-30 | 2019-12-20 | 广东工业大学 | Target feature description method based on local feature and global HSV feature combination |
CN110910417A (en) * | 2019-10-29 | 2020-03-24 | 西北工业大学 | Weak and small moving target detection method based on super-pixel adjacent frame feature comparison |
CN113409338A (en) * | 2021-06-24 | 2021-09-17 | 西安交通大学 | Super-pixel method based on probability distribution |
CN113409338B (en) * | 2021-06-24 | 2023-04-25 | 西安交通大学 | Super-pixel method based on probability distribution |
CN113506266A (en) * | 2021-07-09 | 2021-10-15 | 平安科技(深圳)有限公司 | Method, device and equipment for detecting tongue greasy coating and storage medium |
CN115272353A (en) * | 2022-10-07 | 2022-11-01 | 山东盛世恒机械制造有限公司 | Image processing method suitable for crack detection |
CN117274812A (en) * | 2023-10-08 | 2023-12-22 | 北京香田智能科技有限公司 | Tobacco plant counting method |
CN117274812B (en) * | 2023-10-08 | 2024-02-20 | 北京香田智能科技有限公司 | Tobacco plant counting method |
Also Published As
Publication number | Publication date |
---|---|
CN107016691B (en) | 2019-09-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107016691B (en) | Moving target detecting method based on super-pixel feature | |
CN105261037B (en) | A kind of moving target detecting method of adaptive complex scene | |
CN106845374B (en) | Pedestrian detection method and detection device based on deep learning | |
CN107993245B (en) | Aerospace background multi-target detection and tracking method | |
CN105809716B (en) | Foreground extraction method integrating superpixel and three-dimensional self-organizing background subtraction method | |
US8280165B2 (en) | System and method for segmenting foreground and background in a video | |
JP4629364B2 (en) | A method for adaptively updating a background image representing the background of a scene | |
JP5045371B2 (en) | Foreground / background classification apparatus, method, and program for each pixel of moving image | |
CN106023257B (en) | A kind of method for tracking target based on rotor wing unmanned aerial vehicle platform | |
CN108876820B (en) | Moving target tracking method under shielding condition based on mean shift | |
CN103258332B (en) | A kind of detection method of the moving target of resisting illumination variation | |
Vosters et al. | Background subtraction under sudden illumination changes | |
CN109460764A (en) | A kind of satellite video ship monitoring method of combination brightness and improvement frame differential method | |
CN109191429B (en) | 3D printing nozzle detection method based on machine vision | |
CN109670401B (en) | Action recognition method based on skeletal motion diagram | |
CN109754440A (en) | A kind of shadow region detection method based on full convolutional network and average drifting | |
US9824454B2 (en) | Image processing method and image processing apparatus | |
Pokrajac et al. | Spatiotemporal blocks-based moving objects identification and tracking | |
Santoyo-Morales et al. | Video background subtraction in complex environments | |
CN104933728A (en) | Mixed motion target detection method | |
CN104715480A (en) | Statistical background model based target detection method | |
Zhang et al. | An optical flow based moving objects detection algorithm for the UAV | |
Shiting et al. | Clustering-based shadow edge detection in a single color image | |
CN105741277A (en) | ViBe (Visual Background Extractor) algorithm and SLIC (Simple Linear Iterative Cluster) superpixel based background difference method | |
CN112487926A (en) | Scenic spot feeding behavior identification method based on space-time diagram convolutional network |
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 |