CN108805897A - Improved moving target detection VIBE algorithm - Google Patents
Improved moving target detection VIBE algorithm Download PDFInfo
- Publication number
- CN108805897A CN108805897A CN201810498273.1A CN201810498273A CN108805897A CN 108805897 A CN108805897 A CN 108805897A CN 201810498273 A CN201810498273 A CN 201810498273A CN 108805897 A CN108805897 A CN 108805897A
- Authority
- CN
- China
- Prior art keywords
- pixel
- frame
- ghost
- value
- background model
- 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
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 95
- 238000001514 detection method Methods 0.000 title claims abstract description 63
- 238000000034 method Methods 0.000 claims abstract description 65
- 230000008859 change Effects 0.000 claims abstract description 5
- 230000003044 adaptive effect Effects 0.000 claims description 31
- 238000005516 engineering process Methods 0.000 claims description 13
- 239000011159 matrix material Substances 0.000 claims description 9
- 230000008030 elimination Effects 0.000 claims description 5
- 238000003379 elimination reaction Methods 0.000 claims description 5
- 230000003068 static effect Effects 0.000 abstract description 34
- 238000000605 extraction Methods 0.000 abstract description 15
- 230000008569 process Effects 0.000 abstract description 13
- 230000000694 effects Effects 0.000 abstract description 8
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 238000012360 testing method Methods 0.000 description 29
- 230000004069 differentiation Effects 0.000 description 13
- 238000012545 processing Methods 0.000 description 11
- 238000002474 experimental method Methods 0.000 description 10
- 238000010586 diagram Methods 0.000 description 8
- 230000006872 improvement Effects 0.000 description 8
- 238000013461 design Methods 0.000 description 7
- 238000011160 research Methods 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 4
- 230000008034 disappearance Effects 0.000 description 4
- 241000894007 species Species 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000005021 gait Effects 0.000 description 3
- 238000007689 inspection Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 206010000117 Abnormal behaviour Diseases 0.000 description 1
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000012795 verification Methods 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
- G06T7/215—Motion-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/254—Analysis of motion involving subtraction of images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Aiming at the problems that a ghost phenomenon occurs in subsequent detection when an input first frame contains a moving target in a classic VIBE foreground extraction algorithm, the detection effect of a fixed radius in a model in a complex scene is poor in pixel discrimination and the like, the invention provides an improved moving target detection VIBE algorithm, which distinguishes whether a static area is a ghost area or a static target and describes the difference degree between a current pixel and a sample in a background model, and dynamically adjusts the radius in the discrimination model according to a description word in the pixel discrimination process, so that more foreground points are detected when the scene change degree is small, pixel points with small fluctuation can be prevented from being detected as foreground points when the change degree is large, and noise in a detection result is reduced. The beneficial technical effects are as follows: compared with the original VIBE algorithm, the method can remove the ghost in fewer frames and ensure that the detected moving target is more accurate.
Description
Technical field
The present invention is a kind of image/video treatment technology, specially a kind of improved moving object detection VIBE algorithms.
Background content
Moving object detection[1], i.e., remove to obtain the moving target in sequence of frames of video (without in scene by certain means
Background information), it is relatively common in fields such as video processing, traffic monitoring, social securities.Currently, Gait Recognition[2], target with
Track[3], video abnormal behaviour analysis[4]It is burning hoter etc. multiple research fields, although different research fields has differences, have
It is a little identical, that is, the research object of these research fields is all video information.So how to obtain people's sense from video
The information of interest becomes the top priority of related field.Moving target detecting method[5-7]It can be divided into three classes：Frame difference method, light
Stream method and background subtraction.Frame difference method is the then basis by being made the difference between video consecutive frame (or frame of same intervals)
Threshold value obtains foreground and background.Its clear principle is understandable, and code implements also very simply, and the speed of service can also reach real-time
Requirement, but there are slur phenomenons for the profile of frame difference method extraction, and when closely spaced between two frames, object is in front and back two frame
The region of middle overlapping just will appear hole region after subtracting each other.Optical flow method is to describe sports ground by calculating light stream, further according to
Light stream amplitude threshold carries out foreground extraction.Although its testing result comparison is accurate, generation time paid when light stream is being calculated
Valence is big, therefore is not applied for the processing of real-time video, calculates the influence noises such as can not also avoid shade, block when light stream.The back of the body
Scape relief method needs to build a background model with the first frame of input video or former frames, then with present frame and background model ratio
Compared with the pixel class differentiated in present frame.The advantages of this method is can be by adapting to scene to timely updating for background model
Continuous variation to obtain preferable testing result in complicated scene.Mixed Gaussian algorithm[8](GMM) belong to background to subtract
A kind of common algorithms of division, GMM can be background slow pixel study is changed, and change fast pixel and are considered as foreground, to reach
To the separation of foreground and background.But GMM initialization procedures are long, and parameter Estimation is slow, can not be suitable for the processing of real-time video.
For various problems existing for above-mentioned algorithm, Olivier Barnich et al.[9]A kind of nothing was proposed in 2009
The background subtraction algorithm of parameter Estimation -- VIBE algorithms, the algorithm detection result is relatively good, and speed of service ratio is very fast, and close
It is widely used over year, gradually develops into a kind of general background subtraction algorithm[10].But traditional VIBE is calculated
There are 2 points of deficiencies for method：1) when containing moving target in first frame, detection just will appear " ghost " phenomenon；2) VIBE backgrounds mould
Type differentiates that radius is fixed, and this fixed threshold value cannot well adapt to the foreground detection in dynamic scene.
In order to enable the detection result of VIBE algorithms also can be satisfactory under complex scene, numerous scholars calculate in VIBE
Also respective improvement is made that on the basis of method.
For " ghost " test problems, current algorithm is broadly divided into two classes:The first kind is the movement according to foreground pixel
Detection of attribute " ghost ".For example, Yang etc.[11]It borrows frame difference method and compares difference of some pixel between present frame and previous frame
It is different, the motion state of each pixel is described with molility factor, when its value be 0, illustrate the pixel be static pixel, it should be determined as
" ghost ".Second class is the automatically updating function using background model.For example, Stauffer etc.[12]With each frame to background mould
Type is modified, gradually by background model initializing when " ghost " pixel for introducing replace away.
For the problem for differentiating radius and fixing in VIBE background models, there are also researchs at present.For example, document [13]
When differentiating pixel classification, threshold value is adjusted according to the variance of the sample of current pixel point, achieved the effect that it is certain, but it is big
The variance of amount calculates the execution efficiency for having seriously affected program；Document[14]Then by calculating the background complexity of each frame, according to
Background complexity dynamically adjusts the value for differentiating radius so that certain promotion that the accuracy rate of testing result obtains；Document[15]
Using in neighborhood of pixel points pixel maximum value and minimum value obtain difference adaptively adjust differentiate radius, but also Detection accuracy
It gets a promotion, but when neighborhood territory pixel spot noise is more, testing result will be had adverse effect on.
In summary, it is also necessary to above-mentioned image processing method is improved and perfect.
Invention content
In order to be improved to problem listed by background technology, the present invention provides a kind of improved moving object detection VIBE calculations
Method, the specific method is as follows：
A kind of improved moving object detection VIBE algorithms, on the basis of existing moving object detection VIBE algorithms,
Increase " ghost " minimizing technology and adaptive threshold method.
Furtherly, a kind of improved moving object detection VIBE algorithms, carry out as follows：
Step 1：By the current nth frame image of video input, judge whether present frame is first frame；
If so, utilizing first frame initial background model BGM；
If it is not, then entering step 2；
Step 2：Judge whether present frame is last frame frame number+1；
If so, operation terminates；
If it is not, then entering step 3；
Step 3：According to initialized background model BGM, some pixel in traversal present frame utilizes improved
Adaptive threshold judges that the pixel is background dot or foreground point；
If background dot then enters step 4；
If foreground point then enters step 5；
Step 4：Corresponding position sets value as 0, and according to the pixel position whether in reality in the mask code matrix of corresponding frame
Different more new strategies is made to background model BGM in target boundary rectangle frame, enters step 9；
Step 5：Corresponding position sets value as 255 in the mask code matrix of corresponding frame, and judge the frame whether traverse finish and
Current frame number can be divided exactly by some positive integer；
If it is not, then entering step 3；
If so, entering step 6；
Step 6：Stagnant zone in present frame is obtained, and judges the attribute of stagnant zone；
If stagnant zone is ghost region, 7 are entered step；
If stagnant zone is actual motion target area, 8 are entered step；
Step 7：Local replacement policy update background model BGM accelerates the elimination of ghost, enters step 9；
Step 8：Actual motion target boundary rectangle location information is preserved, enters step 1；
Step 9：Background model BGM updates finish, and enter step 1.
Beneficial technique effect
For " ghost removal " problem, though existing method inhibits the generation or acceleration of " ghost " to a certain extent
The elimination of " ghost ", but is all mostly " do one's bit " in former background model so that background model becomes complicated, too fat to move,
Cause model rate in operation slack-off, it is difficult to be reused in practical applications.Method proposed by the invention can compared with
Ghost is just detected in few frame number, and does not interfere with the real-time of VIBE algorithms, can be used in practice.Further
It says, a kind of " ghost " minimizing technology compared based on contour similarity proposed by the invention, by comparing stagnant zone and correspondence
The profile that is arrived with Canny operator extractions of gray areas between similarity measurements, it is " ghost " region or quiet to distinguish stagnant zone
Then only target is handled it respectively
For " adaptive threshold problem ", though existing method can reach certain detection result, algorithm is complicated
Degree is higher, such as by the method for variance, and a large amount of variance calculates the execution efficiency that certainly will affect algorithm.It is proposed by the invention
Method it is fast by LBP operator calculating speeds the advantages that, copy the building mode of LBP operators to construct LBP-T describing words so that
VIBE algorithm testing results nor affect on the execution efficiency of algorithm while more accurate.This method by LBP-T describing words come
It describes the difference degree between sample, this difference degree in current pixel and background model and also directly reflects video Scene
Variation degree.The size of radius in discrimination model is dynamically adjusted according to describing word when pixel differentiates so that become in scene
Change degree has more foreground points to be detected when smaller, the pixel that can prevent fluctuation smaller when variation degree is larger is tested
It is foreground point to survey, and reduces the noise in testing result.
The experimental results showed that compared with original VIBE algorithms, the improved sides based on profile similarity system design this paper
Method can remove " ghost " in less frame number；Adaptive threshold based on LBP-T describing words makes the movement mesh detected
Mark is more accurate.
The present invention provides the ghost minimizing technology based on profile similarity system design, and this method is not from increase model complexity
From the aspect of problem, but the content in the actual frame corresponding to static motion target area and ghost region exist compared with
This essence of big difference, respectively and the profile in corresponding actual frame region using Canny operator extractions, because of ghost
Region contour and actual frame corresponding region profile differences are very big, and actual motion objective contour is deposited with actual frame corresponding region profile
In certain similitude to obtain the attribute of static target.After obtaining ghost region, so that it may with where ghost in present frame
The pixel in region reinitializes a local background model and replaces corresponding position in former background model, allow quiet and secluded model more
Add closing to reality, to accelerate the release rate of ghost.
E in Fig. 4) and when f) to be current quiet region be ghost, the ghost profile that is carried respectively with Canny operators and work as
The profile of previous frame corresponding region, it can be found that profile differences between the two are larger.
G in Fig. 4) and when h) to be current quiet region be actual motion target, the practical fortune that is carried respectively with Canny operators
Moving-target two-value map contour and the profile of present frame corresponding region, it can be found that profile between the two has centainly similar
Property.
Fig. 5 is that " ghost " region partial model reinitializes replacement archetype corresponding position.The figure is partial model
One process display diagram of replacement policy.The left side is the background model for containing moving target initialization, and centre is to use ghost
Local background's model that corresponding actual frame region reinitializes, the right are replaced models.It can be found that replaced
Model no longer contains moving target information so that model more closing to reality background, to accelerate the release rate of ghost.
Fig. 6 is the velocity contrast that the present invention is eliminated with original VIBE algorithms ghost, and the left side is frame number when ghost occurs, in
Between be that former VIBE algorithms ghost eliminates the frame number for being, the right is the frame number that ghost of the present invention is eliminated, it can be found that the present invention can
Ghost can be eliminated in less frame number clean.
The present invention provides the adaptive threshold VIBE improvement strategies based on LBP-T describing words, this method in order to not influence
The execution efficiency of entire VIBE algorithms, uses for reference the principle of LBP operators, by acquire the mean value of background model, current pixel it is complete
Portion neighbours and a multi-threshold structure jointly constructs go out the LBP-T describing words of the pixel.Pass through the constructions of LBP-T describing words
Journey, difference between current pixel and background model sampled pixel to be sorted can be described by having shown LBP-T describing words, when this
When species diversity is smaller, then illustrate that scene changes degree is smaller, so that it may suitably to reduce the value for differentiating radius R, enabling
Detect more movable informations, when this differ greatly, then scene changes degree is higher, it is necessary to increase and differentiate radius R
Value, avoiding complex information excessive in scene from being detected as foreground causes noise spot in testing result more.Experiment shows base
It is more accurate in the moving target result that the adaptive threshold of LBP-T describing words enables to VIBE to detect.
Figure 13 is the comparison improved the present invention is based on LBP-T adaptive thresholds VIBE between original VIBE algorithms.The left side
It is the frame actually entered, centre is that original VIBE algorithms detect as a result, the right is testing result of the present invention, it can be found that originally
The result of invention is more accurate than former VIBE testing results.
Figure 16 is that the present invention carries out the accuracy in detection of several different motion object detection methods on data set Canoe
Comparison, contrast curve are as follows.As can be seen from Figure 16 original VIBE algorithms detection result ratio GMM algorithms are good very much,
But original VIBE algorithms have used fixed differentiation radius, do not adapt to the scene of different fluctuating ranges, and improved side
Method uses adaptive threshold, the value of radius can be differentiated with the size adjust automatically of scene changes degree so that detection is accurate
True rate is improved, and whole detection effect is better than traditional VIBE algorithms detection result.
Description of the drawings
Fig. 1 is that the radius threshold of VIBE algorithms judges schematic diagram.
Fig. 2 is the influence of " ghost " to moving target.
Fig. 3 is the schematic diagram of stagnant zone detection.
Fig. 4 is stagnant zone and corresponding grey scale region Canny operator results.
Fig. 5 is that " ghost " region partial model reinitializes replacement archetype corresponding position.The figure is partial model
One process display diagram of replacement policy.The left side is the background model for containing moving target initialization, and centre is to use ghost
Local background's model that corresponding actual frame region reinitializes, the right are replaced models.It can be found that replaced
Model no longer contains moving target information so that model more closing to reality background, to accelerate the release rate of ghost.
Fig. 6 is compared with the method for the present invention removes speed with original VIBE " ghost ".Three width figures below Fig. 6 be the present invention with
The velocity contrast that original VIBE algorithms ghost is eliminated, the left side is frame number when ghost occurs, and centre is that former VIBE algorithms ghost disappears
Except the frame number for being, the right is the frame number that ghost of the present invention is eliminated, it can be found that the present invention can be in less frame number
Ghost is eliminated clean.Three width figures above Fig. 6 are that the present invention is based on LBP-T adaptive thresholds VIBE to improve and original VIBE calculations
Comparison between method.The left side is the frame actually entered, and centre is that original VIBE algorithms detect as a result, the right is inspection of the present invention
It surveys as a result, it can be found that the result of the present invention is more accurate than former VIBE testing results.
Fig. 7 is different the influence for differentiating radius to VIBE testing results.
Fig. 8 is LBP operator building process.
Fig. 9 is the different situations that identical LBP values correspond to local pixel distribution.
Figure 10 is background model mean value.
Figure 11 is the difference statistical result of pixel to be sorted and background model mean value.
Figure 12 is the schematic diagram of multi-threshold structure.
Figure 13 is improvement adaptive threshold provided by the present invention and other methods testing result contrast schematic diagram.
Figure 14 is current quiet region when being ghost, and the ghost profile and present frame carried respectively with Canny operators corresponds to
The profile in region, it can be found that profile differences between the two are larger.
Figure 15 is current quiet region when being actual motion target, the actual motion target two carried respectively with Canny operators
It is worth map contour and the profile of present frame corresponding region, it can be found that there are certain similitudes for profile between the two.
Figure 16 is that the present invention carries out the accuracy in detection of several different motion object detection methods on data set Canoe
Contrast schematic diagram.
Figure 17 is the flow diagram of the present invention.
Specific implementation mode
A kind of improved moving object detection VIBE algorithms, on the basis of existing moving object detection VIBE algorithms,
Increase " ghost " minimizing technology and adaptive threshold method.
Referring to Figure 17, a kind of improved moving object detection VIBE algorithms of the present invention carry out as follows：
Step 1：By the current nth frame image of video input, judge whether present frame is first frame；
If so, utilizing first frame initial background model BGM；
If it is not, then entering step 2；
Step 2：Judge whether present frame is last frame frame number+1；
If so, operation terminates；
If it is not, then entering step 3；
Step 3：According to initialized background model BGM, some pixel in traversal present frame utilizes improved
Adaptive threshold judges that the pixel is background dot or foreground point；
If background dot then enters step 4；
If foreground point then enters step 5；
Step 4：Corresponding position sets value as 0, and according to the pixel position whether in reality in the mask code matrix of corresponding frame
Different more new strategies is made to background model BGM in target boundary rectangle frame, enters step 9；
Step 5：Corresponding position sets value as 255 in the mask code matrix of corresponding frame, and judge the frame whether traverse finish and
Current frame number can be divided exactly by some positive integer；
If it is not, then entering step 3；
If so, entering step 6；
Step 6：Stagnant zone in present frame is obtained, and judges the attribute of stagnant zone；
If stagnant zone is ghost region, 7 are entered step；
If stagnant zone is actual motion target area, 8 are entered step；
Step 7：Local replacement policy update background model BGM accelerates the elimination of ghost, enters step 9；
Step 8：Actual motion target boundary rectangle location information is preserved, enters step 1；
Step 9：Background model BGM updates finish, and enter step 1.
Furtherly, step 1 the specific steps are：
By the current nth frame image of video input, and judge present frame whether be video first frame；
If so, using first frame initial background model BGM, specific initialization strategy is as follows：
The all pixels point of traversal first frame contains any one pixel v (x) in the background model of each pixel
There is N number of sample, is denoted as M (x)={ v1, v2... vk... vn, N values are 20, and the sample in the background model of each pixel is adopted
With randomized policy, n times randomly choose a pixel value in 8 neighborhoods of the pixel as a value in the pixel samples；
When first frame has traversed, background model BGM just initializes completion；
If it is not, then entering step 2.
Furtherly, judge that the pixel is background dot or foreground point, i.e. the detailed process of pixel classifications is as follows：
In theorem in Euclid space, one is defined with v (x) as the center of circle, threshold value R is the circle S of radiusR(v(x))；This circle indicates
With the set to the point less than threshold value R at a distance from center of circle v (x)；
The number that current pixel background model M (x) and v (x) distances are less than R is counted, if more than given threshold value Dmin(value
For 2), then it is assumed that current pixel point and background sample are close, and current pixel is divided into background (pixel value is set as 0)；Otherwise, it just draws
It is divided into foreground point (pixel value is set as 1)；Calculation formula is as follows：
Main three parameters of detection process of VIBE algorithms：Wherein sample set number N, threshold value min are set as N=20, min
=2, and the threshold value R of closely located judgement does not use global fixed value, it should each pixel is arranged respective
Differentiate radius, then proposes the adaptive threshold based on LBP-T describing words.
Furtherly, in step 4, corresponding position sets value as 0, and according to the pixel point in the mask code matrix of corresponding frame
It sets and whether in realistic objective boundary rectangle frame makes different more new strategies to background model BGM：
If the pixel is in realistic objective boundary rectangle frame, pixel does not update its background model；
If the pixel is not in realistic objective boundary rectangle frame, it hasThe probability pixel pixel
Value goes to replace a value of sample M (x) in its background model,Value range between 2 to 128, preferablyTogether
When, also haveThe pixel value of the probability pixel go in the background model sample for replacing its some neighborhood territory pixel point
One value；Then, 9 are entered step；
Furtherly, judge stagnant zone in present frame attribute whether be ghost specific method, i.e., " ghost " detect
Method is：It is the method based on profile similarity system design, including following two steps：
Step 1 calculates the static foreground area in VIBE algorithm testing result video sequences；
Step 2 passes through static foreground area and co-located region corresponding grey scale graph region in present frame respectively
Then Canny operator extraction profiles calculate its similarity on given contour similarity formula, are compared by contour similarity
Obtain the attribute of static foreground area.
Furtherly, the processing of " local replacement policy update background model BGM accelerates the elimination of ghost " described in step 7
Strategy is：When judging that stagnant zone is " ghost " region, it is necessary to accelerate the release rate of " ghost ".
Selection real background region first carries out model initialization with VIBE algorithms again；
Then the pixel model in " ghost " region in archetype is replaced；
Finally obtain the background model of " clean " only comprising background pixel initialization.
Furtherly, the adaptive threshold method in step 3 is specially following three steps based on LBP-T describing words
Suddenly：
Step 1 calculates the mean value of pixel background model to be sorted and obtains 8 neighborhood territory pixels of the pixel and by background
The mean value of model is placed on current pixel position；
Step 2, counts the distribution of the difference rule of multiple pixels to be sorted and background model mean value, and builds multi-threshold
Structure chart；
Step 3 calculates the LBP-T values of current pixel, the radius value when value being used to classify as current pixel.
Differentiation radius R in original VIBE algorithms used in background model more new strategy is changeless, is owned
Pixel completely uses this unique discrimination threshold, fixed to differentiate that radius is unfavorable for background detection, can not adapt to complicated field
The variation of scape, in order to adapt to complex scene variation to improve fixed threshold to be adaptive threshold, this just needs to find one and can retouch
The describing word of video Scene variation degree is stated, is then that a differentiation radius is arranged in each pixel according to describing word, replaces
Fall the fixation discrimination threshold in original VIBE models.
Differentiate the stage in pixel, copy the building mode of LBP operators, the present invention is by whole neighbours of current pixel to be sorted
Pixel builds a LBP-T describing word with the sample value to be compared in background model, which can describe the neighbour of current pixel
Occupy the difference between some sampled pixel in pixel and background model.Because distribution has continuity between pixel, this
Difference also reflects the difference between current that sampled pixel of pixel and background model to be sorted indirectly, when this species diversity is smaller
When, then illustrate that scene changes degree is smaller, so that it may suitably to reduce the value for differentiating radius R, enabling detect more
Movable information, when this differ greatly, then scene changes degree is higher, it is necessary to increase the value for differentiating radius R, avoid field
Excessive complex information, which is detected as foreground, in scape causes noise spot in testing result more.Therefore it can be gone according to the size of describing word
Determine the fixed differentiation radius R in VIBE models.
Referring to Figure 10, furtherly, one the step of for " adaptive threshold ", it is implemented as follows：
Wherein X is current pixel to be sorted, neik(k=1,2...8) is 8 neighborhood territory pixels of current pixel X.N is background
Number of samples in model,NiFor the mean value of background sample.
Referring to Figure 11, two the step of for " adaptive threshold ", specially：The present invention is in multiple video detections, statistics
The differences of multiple pixels and background model mean value to be compared, statistical result such as Figure 11.
Figure it is seen that the difference overwhelming majority of pixel to be sorted and background model mean value be all distributed in section [0,
40], and in section [0,40] subinterval [0,20] occupy entire section more than half.
Based on discussed above, the present invention constructs a multi-threshold structure chart, as shown in figure 12：
Wherein, Thi, i=1,2,3...8 be multiple threshold values, according to clockwise (with LBP formation binary strings sequence consensus)
Sequence is placed, and maximum threshold value is placed on to the highest order of binary string, minimum threshold value is placed on the lowest order of binary string.Thi
Shown in specific value such as Fig. 3 (b).Above-mentioned multi-threshold structure chart is for by the mean value of current pixel neighbor pixel and background model
What difference was compared with the threshold value of corresponding position.
It the step of for " adaptive threshold " three, is implemented as follows：
Work as Δk=| neik-μ|≤ThkWhen, corresponding position is just set as 0 in structure chart, is otherwise just set as 1.
LBP-T describing words can be described as follows with formula：
Wherein (xc, yc) central elements of 3 × 3 neighborhoods is represented, k is the number of neighbours, value 8.C is constant, in experiment
Value is 2.ΔkIndicate (xc, yc) k-th of neighbour and background model mean μ between absolute difference.D (x) is symbol letter
Number：
With R (xc, yc)=η × LBP-T (xc, yc)#
As the new radius in former VIBE models, each pixel has one's own radius value when differentiating, without
It is global fixed value R again.η is the factor, and experiment value is 1/3.
Use this building method：
When describing word LBP-T values are bigger, show the sampled pixel in the neighbor pixel point and background model of current pixel
Point is mutually poorly bigger, and scene changes degree is higher, at this point, threshold value just should set more greatly, avoids the noise spot quilt in scene
Foreground point is detected as,
On the contrary, when describing word LBP-T values are smaller, then show the neighbor pixel point and background model sample picture of current pixel
Relatively, scene changes degree is relatively low for vegetarian refreshments, and threshold value should be then arranged smaller, allow for more pixels in this way
It is detected as foreground.
Furtherly, the detailed step of the ghost minimizing technology of step 6 proposition is：" ghost " minimizing technology is based on wheel
The removing method of wide similarity system design, specially following three steps：
Step 1 calculates the static foreground area in VIBE algorithm testing result video sequences；This step is specific as follows：Just
Beginningization two full 0 matrixes f_c and s_r；Current pixel is detected as foreground point, take out the corresponding position of f_c matrixes value add 1 after
It is stored back to；It is presently processing frame number aliquot P, P=30, every 30 frame does one-time detection, finds and is more than or equal to Q in f_c matrixes,
The position of the point of Q=20, the value that corresponding position is set in s_r matrixes are set as 255；Find qualified company in s_r matrixes
The boundary rectangle frame in logical region, gives up when boundary rectangle frame is small area, retains when boundary rectangle frame is large area；It is described
Small area refers to that rectangle frame of the number of pixels less than 20 is given up, conversely, being large area；The position of rectangle frame is exactly static target institute
It is set to full 0 matrix again in region, and by two matrixes of f_c and s_r.Preferred scheme, P, Q are positive number.
Step 2 passes through static foreground area and co-located region corresponding grey scale graph region in present frame respectively
Then Canny operator extraction profiles calculate its similarity on given contour similarity formula, are compared by contour similarity
Obtain the attribute of static foreground area；This step is specific as follows：After obtaining stagnant zone above, using Canny edge extractings
Contour detecting is all carried out to the corresponding gray scale graph region of stagnant zone and the region：Using calculating binary map contour similarity side
Formula judges the attribute in current quiet region；Attribute refers to that stagnant zone is " ghost " or real static target；Steps are as follows：
1. with the profile of Canny operator extraction stagnant zones, it is denoted as C0；2. according to stagnant zone, corresponding position is found on gray-scale map,
Equally with its profile of Canny operator extractions, it is denoted as C1；3. calculating C0With C1Contour similarity, define bianry image C0、C1Phase
It is as follows like spending：
Wherein size (C0∩C1) it is C0∩C1The product of the length and width of result figure, fgCount(C0∩C1) it is C0∩C1In result figure
The number of foreground point (value be 255), then the attribute SR of stagnant zone can indicate as follows：
Wherein T0For threshold value, value T in experiment0=0.02；After the completion of the attribute of stagnant zone determines, need static to this
Make corresponding processing in region.
Step 3 determines the attribute of stagnant zone, after being ghost or practical static moving target, further according to not belonging to
Property gives static foreground area different processing；This step is specific as follows：
If stagnant zone is the static of actual moving target, in order to inhibit the disappearance of moving target, in VIBE algorithms
When detection current pixel is background dot, judge the point whether in the boundary rectangle frame of stagnant zone.If be reduced by if
This then directly selects and the pixel is allowed not update background model to the size of the updating factor of background model.
If stagnant zone is " ghost " region, it is necessary to accelerate the release rate of " ghost ", select real background region
Again model initialization is carried out with VIBE algorithms, then replaces the pixel model in " ghost " region in archetype；It obtains
The background model of one " clean " only comprising background pixel initialization.
Technology in order to preferably illustrate, more of the invention a little, is now changed an interpretation and is described as follows：
Referring to Figure 17, detailed process frame of the invention is described as follows as shown, changing an angle：
One：The pixel in every frame is clicked through using the adaptive threshold VIBE algorithm improvements strategy based on LBP-T describing words
Row classification, the i.e. pixel are background pixel point or foreground pixel point.Main three ginsengs of detection process of wherein VIBE algorithms
Number：Wherein sample set number N, threshold value #min is set as N=20, #min=2, and the threshold value R of closely located judgement is not used
Global fixed value, it should which respective differentiation radius is arranged to each pixel.With R (xc, yc)=η × LBP-T (xc,
yc)#
As the new radius value in former VIBE models, each pixel has one's own radius value when differentiating, and
No longer it is global fixed value R.η is the factor, and experiment value is 1/3.
Two：When pixel to be sorted is identified as background dot, whether the pixel is first checked in actual motion target
Extraneous rectangular area in；If the pixel is not updated model；If not existing, also according to original update
Factor pair model is updated.I.e. it hasThe pixel value of the probability pixel go to replace its background mould
A value of sample M (x) in type, meanwhile, also haveThe pixel value of the probability pixel go to replace its some neighborhood picture
A value in the background model sample of vegetarian refreshments.
Three：When pixel to be sorted is identified as foreground point, detect whether the frame is disposed and can frame number by just
Integer P (P=30) divides exactly.If ineligible, continue to traverse according to original step；It is eligible, then detect rest point,
There is at least Q (Q=20) number to be detected as the pixel of foreground in P frames.
Four：Region in the corresponding actual frame of stagnant zone and stagnant zone that detects is carried with Canny operators respectively
Contouring calculates the two contour similarity, works as similarity>When 0.02, then it is assumed that the stagnant zone is practical static target, then protects
The zone position information is deposited, the pixel for being detected as background dot does not just update background model in the area；Work as similarity<=
0.02 is, then it is assumed that is ghost region.The background model that a part is reinitialized using the region, replaces original background
The corresponding position numerical value of model.
Operation principle about VIBE algorithms：
VIBE algorithms use background modeling and foreground detection techniques based on Pixel-level, it initializes the back of the body by first frame
Then scape model is made prospect background to the pixel in new each frame and is differentiated, the pixel for being judged as background is also certain
Probability go update background model in sample.
VIBE algorithm frames include mainly three big modules：1) initialization of background model；2) foreground detection；3) background model
Update.
1) initialization of model：VIBE algorithms are that each pixel v (x) of video head frames establishes a background mould
Type, the background model of each pixel have collectively constituted the background model of entire VIBE algorithms.The background model of each pixel
In contain N number of sample, be denoted as M (x)={ v1, v2... vk... vn, usual N values are 20.In the background model of each pixel
Sample use randomized policy, n times randomly choose a pixel value in 8 neighborhoods of the pixel as in the pixel samples
One value.
2) foreground detection：In theorem in Euclid space, one is defined with v (x) as the center of circle, threshold value R is the circle S of radiusR(v(x))。
This circle expression and the set for arriving the point at a distance from center of circle v (x) less than threshold value R, as shown in Figure 1.
The number that current pixel background model M (x) and v (x) distances are less than R is counted, if more than given threshold value Dmin(value
For 2), then it is assumed that relatively, it is proper that current pixel is divided into background for current pixel point and background sample, otherwise, just draws
It is divided into foreground point.Calculation formula is as follows：
3) background model more new strategy：When some pixel is classified as background pixel point, then the pixel just needs
Update more is made to the background model of VIBE algorithms.VIBE algorithms are using the mechanism randomly updated, for the background picture being determined
Vegetarian refreshments v (x), it hasThe pixel value of the probability pixel go to replace sample M (x) in its background model
One value, meanwhile, also haveThe pixel value of the probability pixel go to replace the background of its some neighborhood territory pixel point
A value in model sample.
VIBE algorithm realization principles are simple, and real-time is good, and detection result is good, are widely applied to foreground extraction, movement mesh
The fields such as mark detection, but there is also certain disadvantages for VIBE algorithms itself.First, VIBE algorithm are automatically updated using model
" ghost " needs to consume a large amount of video frame；Second, when pixel differentiates, using fixed threshold, complex scene is caused to go out
Some existing pixels are by flase drop.It will be improved in terms of the two herein.
About improved VIBE algorithms in the present invention
Insufficient for above-mentioned 2 points of VIBE algorithms, the present invention is proposed " ghost " based on profile similarity system design respectively
Minimizing technology, the adaptive threshold method based on LBP-T describing words and the fortune being combined based on super-pixel segmentation and conspicuousness
Moving-target cavity filling algorithm.For " ghost " minimizing technology based on profile similarity system design proposed, it is detected first
Static foreground area in VIBE algorithm testing results, then passes through according to static foreground area and corresponding grey scale graph region respectively
The profile of Canny operator extractions relatively obtains the attribute of static foreground area by contour similarity, finally according to different attribute
Different processing is given to static foreground area；Adaptive threshold method based on LBP-T describing words is then in the pixel classifications stage
The structure form of LBP operators is copied to build the difference journey between a sampled pixel that can be described in current pixel and background model
The describing word of degree makes the fixed differentiation radius in VIBE algorithms into adaptive threshold according to describing word size.
The ghost detection and removal compared based on contour similarity
Ghost case study：" ghost " problem is that generally faced in the foreground extractions algorithm such as VIBE one is more intractable
Problem.The reason of " ghost " is formed is VIBE background models in initialization, has just contained moving target in first frame, has led
It has caused just to introduce when the background model initializing of the pixel of moving target region and has distinguished larger value with real background.
Moving target starts when mobile, it, which just starts the region occupied, will be detected as foreground, but the foreground zone detected
Domain is practically without the moving target being truly present, and here it is so-called " ghost " phenomenons.In Fig. 2 (b) red rectangle frames
Foreground is " ghost "." ghost " presence will certainly influence the acquisition of real foreground target for a long time, also to target detection, gait
Identification, the correlative study of the computer vision fields such as target following produce certain negative effect
Fig. 2 is certain the frame testing result obtained in data set [16] streetLight using VIBE algorithms.By Fig. 2 (b)
It is found that the form of expression and moving target existing for " ghost " and its similar, it is difficult to distinguish.The presence of " ghost " influences " ghost " area
The initialization and update operation of the background model of domain pixel, to produce certain influence to real target detection.
By VIBE Model Backgrounds more new principle it is found that the pixel for being detected as background of " ghost " areas adjacent can delay
Slowly it erodes " ghost ".The pixel of " ghost " areas adjacent just has certain probability update when being identified as background pixel
The sample value of its neighborhood, that is, " ghost " value in sample pattern can slowly be replaced by background value, but renewal process is ratio
Slower.As shown in Fig. 2 (c), when 170 frame, " ghost " region in figure can be just completely removed, it is seen that VIBE algorithms
The process that automatically updates consume a large amount of frame.
Algorithm improvement principle：Original VIBE algorithms need to remove " ghost " region after multiframe clean.This
The purpose of literary innovatory algorithm is just to speed up the disappearance speed of " ghost ", and inhibits the disappearance speed of practical static target, because
It is only temporary static in view of practical static target is also a part for moving target.Such as it is different in monitor video system
Often in detection, practical static target is also required to warn if there is abnormal, it is necessary to ensure that practical static target is detected.
Q (P in P frames is detected first with VIBE algorithms<Q it) is detected as the pixel of foreground, by being obtained after processing
The position of stagnant zone；Then the gray-scale map pair of stagnant zone and the corresponding present frame of stagnant zone is obtained by Canny operators
The profile information for answering region differentiates that the attribute of stagnant zone is (" ghost " or quiet by calculating the similarity of two profile informations
Only target)；Different processing methods finally takes the pixel of stagnant zone according to the attribute of stagnant zone.
Improved method detailed process：It is known that " ghost " region (Ghost) and practical static target region (Object)
Taking the form of in VIBE algorithm testing results is the same, all shows as same position in multiple video frame and is always detected
For foreground pixel point, so, it is necessary first to detect the area for being always detected as foreground point in current VIBE algorithms output result
Domain, that is, stagnant zone, detecting step are as follows：
1) two full 0 matrixes f_c and s_r are initialized；
2) current pixel is detected as foreground point, take out the corresponding position of f_c matrixes value add 1 after be stored back to；
3) be presently processing frame number aliquot P (P=30, every 30 frame do one-time detection), find in f_c matrixes be more than etc.
In the position of the point of Q (Q=20), the value that corresponding position is set in s_r matrixes is 255；
4) the boundary rectangle frame (small area rectangle frame is given up) of qualified connected region in s_r matrixes, rectangle are found
The position of frame is exactly static target region, is marked in original graph.As shown in Fig. 3 (c).
5) two matrixes of f_c and s_r are set to full 0 matrix again, doing initialization for cycle next time prepares.
After obtaining stagnant zone above, use Canny edge extractings corresponding to stagnant zone and the region herein
Gray scale graph region all carries out contour detecting, and the results are shown in Figure 4 for contour detecting.
Content in the actual frame corresponding with practical static target region of " ghost " region is different, " ghost " region
It is typically all background area in corresponding actual frame, as shown in Fig. 4 (b), the general pixel distribution in background area is smoother, and real
It is exactly static moving target true that border static target is corresponding, as shown in Fig. 4 (d), actual moving target and week
There are profile, the profile is also just more similar in shape to the profile of current stagnant zone binary map in the region enclosed.It is based on
Such inspiration judges the attribute in current quiet region (" ghost " still using calculating binary map contour similarity mode herein
Real static target).Steps are as follows for calculating：
1) profile for using Canny operator extraction stagnant zones, is denoted as C0；
2) according to stagnant zone, corresponding position is found on gray-scale map, equally with its profile of Canny operator extractions, is denoted as
C1；
3) C is calculated0With C1Contour similarity, define bianry image C0、C1Similarity it is as follows：
Wherein size (C0∩C1) it is C0∩C1The product of the length and width of result figure, fgCount(C0∩C1) it is C0∩C1In result figure
The number of foreground point (value be 255), then the attribute SR of stagnant zone can indicate as follows：
Wherein T0For threshold value, value T in experiment0=0.02.
After the completion of the attribute of stagnant zone determines, need to make corresponding processing to the stagnant zone.
If stagnant zone is the static of actual moving target, in order to inhibit the disappearance of moving target, in VIBE algorithms
When detection current pixel is background dot, judge the point whether in the boundary rectangle frame of stagnant zone.If be reduced by if
This directly selects the pixel is allowed not update background model herein to the size of the updating factor of background model.
If stagnant zone is " ghost " region, it is necessary to accelerate the release rate of " ghost ", select real background region
Again model initialization is carried out with VIBE algorithms, then replaces the pixel model in " ghost " region in archetype.Such as Fig. 5
(c) shown in, the background model of " clean " only comprising background pixel initialization is obtained.
The verification of improvement effect is analyzed：Herein by " ghost " minimizing technology compared based on contour similarity and original of proposition
Slowly update goes " ghost " method to be made that comparison is real in two groups of videos of data set PETS2006 and streetLight to beginning VIBE
It tests, experimental result is as shown in Figure 6.
" ghost " region in Fig. 6 (a), (d) red rectangle frame, (b), (e) be original VIBE algorithms slowly update " ghost
The frame number of shadow " consumption, (c), (f) be improved " ghost " removal consumption herein frame number, it is obvious that " ghost " is slow in video 1
The frame consumed when updating out slowly is 120 frames, and improved method only needs 53 frames " ghost " just almost to disappear herein；" ghost in video 2
The frame that shadow " consumes when slowly updating out is 170 frames, and improved method only needs 74 frames " ghost " just almost to disappear herein, tests
Prove improved method herein to the removal of " ghost " more quickly and effectively.
Adaptive threshold based on LBP-T describing words
It is analyzed about Threshold：The selection of threshold value is extremely important to the result of foreground detection, and threshold value has selected small will
Lead to the noise spot there are many containing in the result of foreground detection, and threshold value selection is big, may result in target and larger area occurs
Missing inspection.
Fig. 7 is provided with two different differentiation radiuses to VIBE algorithms, has been two groups of Experimental comparisons respectively.From Fig. 7 (a),
Fig. 7 (c) two images can be seen that when differentiating the larger of radius setting, and the noise in scene can will not be detected；But
It is to find out from Fig. 7 (b), Fig. 7 (d), differentiates that radius setting is big, the profile of moving target will be missed, and cause profile unclear
It is clear, there is incompleteness.
Conclusion as can be drawn from Figure 7, it is exactly best that the differentiation radius of VIBE algorithms, which is not fixed, it should with field
The transformation adjustment appropriate of scape can obtain best testing result and be only preferably.
3.2.2 algorithm improvement principle
Differentiation radius R in original VIBE algorithms used in background model more new strategy is changeless, is owned
Pixel completely uses this unique discrimination threshold, fixed to differentiate that radius is unfavorable for background detection, can not adapt to complicated field
The variation of scape, in order to adapt to complex scene variation to improve fixed threshold to be adaptive threshold, this just needs to find one and can retouch
The describing word of video Scene variation degree is stated, is then that a differentiation radius is arranged in each pixel according to describing word, replaces
Fall the fixation discrimination threshold in original VIBE models.
Differentiate the stage in pixel, copies LBP[17]The building mode of operator, herein by all adjacent of current pixel to be sorted
It occupies pixel and builds a LBP-T describing word with the sample value to be compared in background model, which can describe current pixel
Difference in neighbor pixel and background model between some sampled pixel.Because distribution has continuity between pixel, this
Species diversity also reflects the difference between current that sampled pixel of pixel and background model to be sorted indirectly, when this species diversity compared with
Hour, then illustrate that scene changes degree is smaller, so that it may suitably to reduce the value for differentiating radius R, enabling detect more
More movable informations, when this differ greatly, then scene changes degree is higher, it is necessary to increase the value for differentiating radius R, avoid
Excessive complex information, which is detected as foreground, in scene causes noise spot in testing result more.Therefore it can be according to the size of describing word
Remove to replace the fixed differentiation radius R in VIBE models.
Improved method detailed process：It calculates simply, the characteristic that analysis etc. in real time be LBP operators can be carried out to image.Such as
Shown in Fig. 8, traditional LBP is defined in 3 × 3 neighborhood of pixel using intermediate pixel as threshold value, with neighbor pixel value with intermediate
The pixel value of position is compared, if neighbours' value is more than center pixel value, the position where the neighbours is marked as 1, otherwise
It is just 0.In this way, 8 pixels in 3 × 3 neighborhoods will obtain eight bit through comparing, then according to clockwise
A binary string is sequentially formed, which is converted into the LBP values that decimal number is exactly the pixel.
Above-mentioned building process shows that LBP operators describe some pixel and the relationship of surrounding pixel.But LBP is calculated
The value of son can not directly indicate this difference degree between neighbor pixel and intermediate pixel, as shown in Figure 9.Fig. 9 shows
(a) as the LBP values of (d), still the gap between (a) and neighborhood territory pixel and intermediate pixel (d) is different, (d)
Intermediate pixel and neighborhood territory pixel are approximate, and still the intermediate pixel of (a) is really more much bigger than neighborhood territory pixel difference.Assuming that Fig. 9 (a)
Intermediate pixel is sample average of some pixel when carrying out pixel differentiation in the background model of corresponding position in VIBE algorithms, and
Its 8 neighbor pixel is all neighbours of current pixel, it is known that, all neighbours of current pixel are very big with sample value gap,
The complexity of scene changes is higher.Distribution in image between pixel is not arbitrary, and is existed between part and part
Certain regularity.Current pixel has certain similitude with its neighbor pixel in spatial distribution.Current pixel it is big
Part neighborhood territory pixel all in background model pixel value gap it is smaller when, also illustrate current pixel and background model gap
It is smaller, it is lower that side reflects scene changes degree, it should suitably reduce discrimination threshold so that more foreground pixel points
It is detected；When most of neighborhood territory pixel and background model pixel value gap it is bigger when, current pixel and background model are poor
Away from it is larger, side reflects that scene changes degree is higher, then should suitably increase differentiation radius, prevents to fluctuate some smaller
Pixel is detected as foreground point, reduces the noise in testing result.
It is therefore desirable to find a kind of describing word can describe current pixel 8 neighbor pixels and background model sample value it
Between difference degree, i.e. the describing word of scene variation degree.
The definition of LBP operators, the present invention is copied to build LBP-T (threshold value) describing word.Steps are as follows for construction method：
(1) background model mean value computation
Wherein X is current pixel to be sorted, neik(k=1,2...8) is 8 neighborhood territory pixels of current pixel X.N is background
Number of samples in model,NiFor the mean value of background sample.
(2) multi-threshold structure chart is built
Herein in multiple video detections, the difference of multiple pixels and background model mean value to be compared has been counted, has been counted
As a result such as figure below.
It can be seen from figure 11 that the difference overwhelming majority of pixel to be sorted and background model mean value be all distributed in section [0,
40], and in section [0,40] subinterval [0,20] occupy entire section more than half.
A multi-threshold structure chart is constructed herein：In fig. 12, Thi, i=1,2,3...8 be multiple threshold values, according to suitable
Hour hands (forming binary string sequence consensus with LBP) sequence is placed, and maximum threshold value is placed on to the highest order of binary string, minimum
Threshold value be placed on the lowest order of binary string.ThiShown in specific value such as Figure 12 (b).Above-mentioned multi-threshold structure chart is for inciting somebody to action
Current pixel neighbor pixel is compared with the equal value difference of background model with the threshold value of corresponding position.
(3) LBP-T describing words are calculated
Work as Δk=| neik-μ|≤ThkWhen, corresponding position is just set as 0 in structure chart, is otherwise just set as 1.
LBP-T describing words can be described as follows with formula：
Wherein (xc, yc) central elements of 3 × 3 neighborhoods is represented, k is the number of neighbours, value 8.C is constant, in experiment
Value is 2.ΔkIndicate (xc, yc) k-th of neighbour and background model mean μ between absolute difference.D (x) is symbol letter
Number：
With R (xc, yc)=η × LBP-T (xc, yc) (6)
As the new radius value in former VIBE models, each pixel has one's own radius value when differentiating, and
No longer it is global fixed value R.η is the factor, and experiment value is 1/3.
Above-mentioned building method shows to show the neighbor pixel point and background of current pixel when describing word LBP-T values are bigger
Sampled pixel point in model is mutually poorly bigger, and scene changes degree is higher, at this point, threshold value just should set more greatly, avoids
Noise spot in scene is detected as foreground point, on the contrary, when describing word LBP-T values are smaller, then shows the neighbours of current pixel
Relatively, scene changes degree is relatively low, and threshold value should be then arranged smaller for pixel and background model sampled pixel point,
More pixels are allowed in this way is detected as foreground.
Results contrast and analysis：Three videos of example video are carried in data set pedestrians, canoe, opencv
In, select same frame to input, in GMM algorithms, original VIBE algorithms and proposed in this paper based on the adaptive of LBP-T describing words
Respective testing result is respectively obtained on three algorithms of threshold value VIBE innovatory algorithms and is compared.Figure 13 presents 3 groups of experiments
As a result comparison, the image of (a) row, which is that data set pedestrians, canoe, opencv are included respectively successively from top to bottom, to be shown
Example video in the 520th, 958,435 frames input figure.(b) row image is inspection of the mixed Gaussian GMM algorithms on corresponding input frame
Survey result, it can be seen that have in GMM testing results many foreground points by flase drop be background dot.(c) row image is that original VIBE is calculated
Testing result of the method on corresponding input frame, hence it is evident that better than the testing result of GMM algorithm very much.(d) row image is to carry herein
Testing result of the VIBE innovatory algorithms of the adaptive threshold based on LBP-T describing words gone out on corresponding input frame, experiment
Prove, with improved adaptive threshold value, obtained foreground detection result can be more preferable, than the detection of original VIBE algorithms result more
Full, the foreground pixel point contained is more, and foreground is truer.
In conclusion for " ghost " problem of classics VIBE foreground extraction algorithms, the present invention proposes a kind of based on wheel
" ghost " detection method of wide similarity system design, this method automatically update " ghost " with VIBE and compare, and the frame number used is less,
" ghost " removes faster；Differentiate that radius is bad to complex background detection result for fixed in VIBE background models, it is proposed that
A kind of VIBE improved methods of the adaptive threshold of LBP-T describing words so that the result obtained after improvement is than original detection knot
Fruit will get well.Improved method is that the researchs such as subsequent Gait Recognition, abnormality detection are made that feasible place mat work.
6 bibliography
1.Hu,W.,et al.,Moving object detection using tensor-based low-rank
and saliently fused-sparse decomposition.IEEE Transactions on Image
Processing,2017.26(2):p.724-737.
2.Man,J.and B.Bhanu,Individual recognition using gait energy
image.IEEE transactions on pattern analysis and machine intelligence,2006.28
(2):p.316-322.
3.Leal-Taixé,L.,et al.,Tracking the trackers:an analysis of the state
of the art in multiple object tracking.arXiv preprint arXiv:1704.02781,2017.
4.Bala,R.and V.Monga,Video Anomaly Detection.Computer Vision and
Imaging in Intelligent Transportation Systems,2017:p.227-256.
5.Bouwmans,T.,F.El Baf,and B.Vachon,Background modeling using mixture
of gaussians for foreground detection-a survey.Recent Patents on Computer
Science,2008.1(3):p.219-237.
6.Bouwmans,T.,Traditional and recent approaches in background
modeling for foreground detection:An overview.Computer Science Review,
2014.11:p.31-66.
7.Kim,K.,et al.,Real-time foreground–background segmentation using
codebook model.Real-time imaging,2005.11(3):p.172-185.
8.Reynolds,D.,Gaussian mixture models.Encyclopedia of biometrics,
2015:p.827-832.
9.Barnich,O.and M.Van Droogenbroeck.ViBe:a powerful random technique
to estimate the background in video sequences.in Acoustics,Speech and Signal
Processing,2009.ICASSP 2009.IEEE International Conference on.2009.IEEE.
10.Barnich,O.and M.Van Droogenbroeck,ViBe:A universal background
subtraction algorithm for video sequences.IEEE Transactions on Image
processing,2011.20(6):p.1709-1724.
11.Li,Y.,W.Chen,and R.Jiang.The integration adjacent frame difference
of improved ViBe for foreground object detection.in 2011 7th International
Conference on Wireless Communications,Networking and Mobile Computing.2011.
12.Stauffer,C.and W.E.L.Grimson,Learning patterns of activity using
real-time tracking.IEEE Transactions on pattern analysis and machine
intelligence,2000.22(8):p.747-757.
13.Brutzer,S.,B.and G.Heidemann.Evaluation of background
subtraction techniques for video surveillance.in Computer Vision and Pattern
Recognition(CVPR),2011 IEEE Conference on.2011.IEEE.
14.Hofmann,M.,P.Tiefenbacher,and G.Rigoll.Background segmentation
with feedback:The pixel-based adaptive segmenter.in Computer Vision and
Pattern Recognition Workshops(CVPRW),2012 IEEE Computer Society Conference
on.2012.IEEE.
15.Su,Y.,A.Li,and K.Jiang,Improved visual background extractor model
for moving objects detecting algorithm.J.Comput.Aided Des.Comput.Graph,
2014.26(2):p.232-240.
16.PETS 2012 Benchmark Data http://pets2012.net/.
17.Wang,X.,T.X.Han,and S.Yan.An HOG-LBP human detector with partial
occlusion handling.in Computer Vision,2009 IEEE 12th International Conference
on.2009.IEEE.
18.Van Droogenbroeck,M.and O.Paquot.Background subtraction:
Experiments and improvements for ViBe.in Computer Vision and Pattern
Recognition Workshops(CVPRW),2012 IEEE Computer Society Conference
on.2012.IEEE.
Claims (4)
1. a kind of improved moving object detection VIBE algorithms, it is characterised in that：In existing moving object detection VIBE algorithms
On the basis of, increase " ghost " minimizing technology and adaptive threshold method.
2. a kind of improved moving object detection VIBE algorithms according to claim 1, which is characterized in that as follows
It carries out：
Step 1：By the current nth frame image of video input, judge whether present frame is first frame；
If so, utilizing first frame initial background model BGM；
If it is not, then entering step 2；
Step 2：Judge whether present frame is last frame frame number+1；
If so, operation terminates；
If it is not, then entering step 3；
Step 3：According to initialized background model BGM, some pixel in traversal present frame utilizes improved adaptive
Threshold value is answered to judge that the pixel is background dot or foreground point；
If background dot then enters step 4；
If foreground point then enters step 5；
Step 4：Corresponding position sets value as 0, and according to the pixel position whether in realistic objective in the mask code matrix of corresponding frame
Different more new strategies is made to background model BGM in boundary rectangle frame, enters step 9；
Step 5：Corresponding position sets value as 255 in the mask code matrix of corresponding frame, and judges whether the frame traverses and finish and currently
Frame number can be divided exactly by some positive integer；
If it is not, then entering step 3；
If so, entering step 6；
Step 6：Stagnant zone in present frame is obtained, and judges the attribute of stagnant zone；
If stagnant zone is ghost region, 7 are entered step；
If stagnant zone is actual motion target area, 8 are entered step；
Step 7：Local replacement policy update background model BGM accelerates the elimination of ghost, enters step 9；
Step 8：Actual motion target boundary rectangle location information is preserved, enters step 1；
Step 9：Background model BGM updates finish, and enter step 1.
3. a kind of improved moving object detection VIBE algorithms according to claim 2, which is characterized in that the tool of step 1
Body step is：By the current nth frame image of video input, and judge present frame whether be video first frame：
If so, using first frame initial background model BGM, initialized：The all pixels point for traversing first frame, to arbitrary
One pixel v (x) contains N number of sample in the background model of each pixel, is denoted as M (x)={ v1, v2... vk,
...vn, N values are 20, and the sample in the background model of each pixel uses randomized policy, n times to randomly choose the pixel
A pixel value in 8 neighborhoods is as a value in the pixel samples；When first frame has traversed, background model BGM is with regard to initial
Change and completes；
If it is not, then entering step 2.
4. a kind of improved moving object detection VIBE algorithms according to claim 3, which is characterized in that in step 3
Adaptive threshold method is based on LBP-T describing words, specially：
Step 1 calculates the mean value of pixel background model to be sorted and obtains 8 neighborhood territory pixels of the pixel and by background model
Mean value be placed on current pixel position；
Step 2, counts the distribution of the difference rule of multiple pixels to be sorted and background model mean value, and builds multi-threshold structure
Figure；
Step 3 calculates the LBP-T values of current pixel, the radius value when value being used to classify as current pixel.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810498273.1A CN108805897B (en) | 2018-05-22 | 2018-05-22 | Improved moving target detection VIBE method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810498273.1A CN108805897B (en) | 2018-05-22 | 2018-05-22 | Improved moving target detection VIBE method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108805897A true CN108805897A (en) | 2018-11-13 |
| CN108805897B CN108805897B (en) | 2022-01-21 |
Family
ID=64091508
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810498273.1A Expired - Fee Related CN108805897B (en) | 2018-05-22 | 2018-05-22 | Improved moving target detection VIBE method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108805897B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109636866A (en) * | 2018-12-13 | 2019-04-16 | 福建工程学院 | The foreground extracting method of stochastic clustering statistics based on resolution decreasing |
| CN109785562A (en) * | 2018-12-29 | 2019-05-21 | 华中光电技术研究所（中国船舶重工集团有限公司第七一七研究所） | A kind of vertical photoelectricity ground based threats warning system and suspicious object recognition methods |
| CN110084129A (en) * | 2019-04-01 | 2019-08-02 | 昆明理工大学 | A kind of river drifting substances real-time detection method based on machine vision |
| CN110288630A (en) * | 2019-06-27 | 2019-09-27 | 浙江工业大学 | A kind of moving target ghost suppressing method of background modeling |
| CN110363197A (en) * | 2019-06-22 | 2019-10-22 | 东北电力大学 | Based on the video area-of-interest exacting method for improving visual background extraction model |
| CN111080562A (en) * | 2019-12-06 | 2020-04-28 | 合肥科大智能机器人技术有限公司 | Substation suspender identification method based on enhanced image contrast |
| CN111104870A (en) * | 2019-11-27 | 2020-05-05 | 珠海欧比特宇航科技股份有限公司 | Motion detection method, device and equipment based on satellite video and storage medium |
| CN111259866A (en) * | 2020-03-06 | 2020-06-09 | 大连科技学院 | Offshore ship target detection method based on improved background difference method |
| CN111553931A (en) * | 2020-04-03 | 2020-08-18 | 中国地质大学（武汉） | ViBe-ID foreground detection method for indoor real-time monitoring |
| CN111985314A (en) * | 2020-07-09 | 2020-11-24 | 东南大学 | ViBe and improved LBP-based smoke detection method |
| CN112084880A (en) * | 2020-08-14 | 2020-12-15 | 江铃汽车股份有限公司 | Image processing method, device, storage medium and equipment |
| CN113129331A (en) * | 2019-12-31 | 2021-07-16 | 中移(成都)信息通信科技有限公司 | Target movement track detection method, device and equipment and computer storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104899839A (en) * | 2015-06-05 | 2015-09-09 | 河海大学 | Ghost quick-inhibition method based on ViBe algorithm |
| CN106157332A (en) * | 2016-07-07 | 2016-11-23 | 合肥工业大学 | A kind of motion inspection optimization method based on ViBe algorithm |
| WO2017028029A1 (en) * | 2015-08-14 | 2017-02-23 | 富士通株式会社 | Background model extracting method and apparatus and image processing device |
| CN106651782A (en) * | 2016-09-26 | 2017-05-10 | 江苏科海智能***有限公司 | ViBe-oriented foreground ghosting removal method |
-
2018
- 2018-05-22 CN CN201810498273.1A patent/CN108805897B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104899839A (en) * | 2015-06-05 | 2015-09-09 | 河海大学 | Ghost quick-inhibition method based on ViBe algorithm |
| WO2017028029A1 (en) * | 2015-08-14 | 2017-02-23 | 富士通株式会社 | Background model extracting method and apparatus and image processing device |
| CN106157332A (en) * | 2016-07-07 | 2016-11-23 | 合肥工业大学 | A kind of motion inspection optimization method based on ViBe algorithm |
| CN106651782A (en) * | 2016-09-26 | 2017-05-10 | 江苏科海智能***有限公司 | ViBe-oriented foreground ghosting removal method |
Non-Patent Citations (3)
| Title |
|---|
| YUMIN TIAN，ET AL: "《Moving Object Detection with Vibe and Texture Feature》", 《PCM2016》 * |
| 吴尔杰: "《监控视频中多目标检测与跟踪研究》", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 王辉，等: "《一种基于阈值的自适应Vibe目标检测算法》", 《计算机科学》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109636866A (en) * | 2018-12-13 | 2019-04-16 | 福建工程学院 | The foreground extracting method of stochastic clustering statistics based on resolution decreasing |
| CN109785562B (en) * | 2018-12-29 | 2023-08-15 | 华中光电技术研究所（中国船舶重工集团有限公司第七一七研究所） | Vertical photoelectric ground threat alert system and suspicious target identification method |
| CN109785562A (en) * | 2018-12-29 | 2019-05-21 | 华中光电技术研究所（中国船舶重工集团有限公司第七一七研究所） | A kind of vertical photoelectricity ground based threats warning system and suspicious object recognition methods |
| CN110084129A (en) * | 2019-04-01 | 2019-08-02 | 昆明理工大学 | A kind of river drifting substances real-time detection method based on machine vision |
| CN110363197B (en) * | 2019-06-22 | 2023-05-23 | 东北电力大学 | Video region of interest extraction method based on improved visual background extraction model |
| CN110363197A (en) * | 2019-06-22 | 2019-10-22 | 东北电力大学 | Based on the video area-of-interest exacting method for improving visual background extraction model |
| CN110288630A (en) * | 2019-06-27 | 2019-09-27 | 浙江工业大学 | A kind of moving target ghost suppressing method of background modeling |
| CN110288630B (en) * | 2019-06-27 | 2021-11-09 | 浙江工业大学 | Moving target ghost suppression method for background modeling |
| CN111104870B (en) * | 2019-11-27 | 2024-02-13 | 珠海欧比特卫星大数据有限公司 | Motion detection method, device, equipment and storage medium based on satellite video |
| CN111104870A (en) * | 2019-11-27 | 2020-05-05 | 珠海欧比特宇航科技股份有限公司 | Motion detection method, device and equipment based on satellite video and storage medium |
| CN111080562A (en) * | 2019-12-06 | 2020-04-28 | 合肥科大智能机器人技术有限公司 | Substation suspender identification method based on enhanced image contrast |
| CN111080562B (en) * | 2019-12-06 | 2022-12-20 | 合肥科大智能机器人技术有限公司 | Substation suspender identification method based on enhanced image contrast |
| CN113129331B (en) * | 2019-12-31 | 2024-01-30 | 中移(成都)信息通信科技有限公司 | Target movement track detection method, device, equipment and computer storage medium |
| CN113129331A (en) * | 2019-12-31 | 2021-07-16 | 中移(成都)信息通信科技有限公司 | Target movement track detection method, device and equipment and computer storage medium |
| CN111259866B (en) * | 2020-03-06 | 2023-07-28 | 大连科技学院 | Marine ship target detection method based on improved background difference method |
| CN111259866A (en) * | 2020-03-06 | 2020-06-09 | 大连科技学院 | Offshore ship target detection method based on improved background difference method |
| CN111553931A (en) * | 2020-04-03 | 2020-08-18 | 中国地质大学（武汉） | ViBe-ID foreground detection method for indoor real-time monitoring |
| CN111985314A (en) * | 2020-07-09 | 2020-11-24 | 东南大学 | ViBe and improved LBP-based smoke detection method |
| CN111985314B (en) * | 2020-07-09 | 2024-04-30 | 东南大学 | Smoke detection method based on ViBe and improved LBP |
| CN112084880A (en) * | 2020-08-14 | 2020-12-15 | 江铃汽车股份有限公司 | Image processing method, device, storage medium and equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108805897B (en) | 2022-01-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108805897A (en) | Improved moving target detection VIBE algorithm | |
| CN105678811B (en) | A kind of human body anomaly detection method based on motion detection | |
| CN105844295B (en) | A kind of video smoke sophisticated category method based on color model and motion feature | |
| US20230289979A1 (en) | A method for video moving object detection based on relative statistical characteristics of image pixels | |
| López-Rubio et al. | Foreground detection in video sequences with probabilistic self-organizing maps | |
| CN105404847B (en) | A kind of residue real-time detection method | |
| CN105184818B (en) | A kind of video monitoring anomaly detection method and its detecting system | |
| KR101414670B1 (en) | Object tracking method in thermal image using online random forest and particle filter | |
| CN106228137A (en) | A kind of ATM abnormal human face detection based on key point location | |
| CN105957356B (en) | A kind of traffic control system and method based on pedestrian's quantity | |
| CN106548488B (en) | A kind of foreground detection method based on background model and inter-frame difference | |
| CN105243356B (en) | A kind of method and device that establishing pedestrian detection model and pedestrian detection method | |
| CN104077757B (en) | Road background extraction and updating method with fusion of real-time traffic state information | |
| CN110298297A (en) | Flame identification method and device | |
| CN107590427B (en) | Method for detecting abnormal events of surveillance video based on space-time interest point noise reduction | |
| CN105913441A (en) | Shadow removal method for improving target detection performance in video | |
| CN106570490A (en) | Pedestrian real-time tracking method based on fast clustering | |
| CN111062273A (en) | Tracing detection and alarm method for left-over articles | |
| CN107992875B (en) | A kind of well-marked target detection method based on image bandpass filtering | |
| CN109685009A (en) | A kind of method of region crowd density video detection | |
| Liu et al. | A image segmentation algorithm based on differential evolution particle swarm optimization fuzzy c-means clustering | |
| Luque et al. | An ART-type network approach for video object detection. | |
| Shi et al. | Moving cast shadow detection in video based on new chromatic criteria and statistical modeling | |
| CN109544605A (en) | Moving shadow detection method based on space-time relation modeling | |
| CN114429597A (en) | Initial fire intelligent recognition system in petrochemical industry based on neural 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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20220121 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |