CN110503664A - One kind being based on improved local auto-adaptive sensitivity background modeling method - Google Patents

One kind being based on improved local auto-adaptive sensitivity background modeling method Download PDF

Info

Publication number
CN110503664A
CN110503664A CN201910725944.8A CN201910725944A CN110503664A CN 110503664 A CN110503664 A CN 110503664A CN 201910725944 A CN201910725944 A CN 201910725944A CN 110503664 A CN110503664 A CN 110503664A
Authority
CN
China
Prior art keywords
background
pixel
distance threshold
model
value
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
Application number
CN201910725944.8A
Other languages
Chinese (zh)
Other versions
CN110503664B (en
Inventor
成科扬
孙爽
付艳云
师文喜
刘海强
牟超
李鑫
李鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Jiangsu University
China Electronics Technology Group Corp CETC
Original Assignee
Tsinghua University
Jiangsu University
China Electronics Technology Group Corp CETC
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Tsinghua University, Jiangsu University, China Electronics Technology Group Corp CETC filed Critical Tsinghua University
Priority to CN201910725944.8A priority Critical patent/CN110503664B/en
Publication of CN110503664A publication Critical patent/CN110503664A/en
Application granted granted Critical
Publication of CN110503664B publication Critical patent/CN110503664B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • 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
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/90—Determination of colour characteristics
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

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

The invention discloses one kind to be based on improved local auto-adaptive sensitivity background modeling method.The color space information and local binary parallel pattern (LBSP) feature for obtaining background pixel point using a kind of new Pixel-level method first, establish background model.This method utilizes iteration thought, reduces slowly movement and influence of the short time static target to background model authenticity, obtains more reliable background model.Secondly foreground detection is carried out using unanimity of samples strategy.Finally distance threshold R (x) and learning rate T (x) are obtained using the model modification strategy of adaptive sensitivity.To make the distance threshold R (x) obtained more rationally, distance threshold correction mechanism is proposed for obtaining more appropriate distance threshold.The complexity that current pixel is determined by a kind of novel method, then assigns distance threshold correction mechanism on this basis, adjusts the dynamic of distance threshold more reliable.Background modeling method disclosed by the invention can obtain more accurate foreground target under complex background.

Description

One kind being based on improved local auto-adaptive sensitivity background modeling method
Technical field
The invention belongs to image identification technical field, the intelligence view under the common scenes such as cell, school, square can be applied to In frequency monitoring.
Background technique
The general step of foreground detection based on background modeling algorithm be by static or background there are portion disturbances with Certain way establishes background model, is compared by the sample in the pixel and background model of present frame and extracts prospect mesh Mark, then updates background model.The difficult point of background modeling be how disturbance cancelling information complete extraction target.It proposes at present Algorithm includes the method based on pixel and region class, also there is the background modeling method based on colouring information and textural characteristics, this A little methods have its specific advantage, ensure that real-time, but most of are unable to making an uproar for complete extraction target or false retrieval disturbance Point.
In based on region class modeling, the propositions such as Liu Cuiwei utilize the learning method of online subspace to be used for model modification, 2015, Beaugendre et al. proposed a kind of background modeling method that adaptive region is propagated.Maity in 2017 et al. will The statistical nature extractive technique of block is for detecting prospect.These methods all have the shortcomings that region class modeling, i.e., can not obtain essence True prospect and profile, so it is ineffective.
Olivier Barnich et al. is based on picture after background subtraction method (ViBe) of the proposition based on pixel in 2009 The background modeling method of vegetarian refreshments obtains significant development, and the problem of region class modeling is brought can be effectively treated.2012, Hofmann et al. proposes adaptivenon-uniform sampling pixel-based (PBAS) algorithm.The algorithm proposes on the basis of vibe algorithm The concept of adaptive distance threshold and turnover rate.Pierre-Luc St-Charles in 2015 et al. is proposed based on part (SuBSENSE) algorithm is divided in adaptive sensitivity.This method is characterized in Pixel-level model with the space time information of color and texture Sample, propose a kind of novel Pixel-level feedback scheme, it is adaptive to adjust internal sensitivity to change and update distance Threshold rates.By monitoring partial model fidelity and segmentation noise continuously to instruct these adjustment, allow the dynamic to interval Background motion makes quick response.However this method can not remove static or slow moving target in initiate background model Influence, distance threshold update on it cannot be prevented excessive or too small, cause part missing inspection or false retrieval, therefore the algorithm still has Shortcoming.
Summary of the invention
Goal of the invention: extracting the defect of prospect in order to solve background subtraction method under complex background, in local auto-adaptive spirit On the basis of background modeling (SuBSENSE) algorithm of sensitivity, the invention proposes a kind of sides of new initiate background model Method devises a kind of correction mechanism in the renewal process of adaptive distance threshold, and optimizes the update side of background model Formula enables to extract more accurately prospect, has higher robustness.
Technical solution: the invention proposes one kind to be based on improved local auto-adaptive sensitivity background modeling method, including Following steps:
Step (1): the preceding M frame of selecting video sequence removes slow moving target and in short-term as background model candidate frame Background model is established in the influence of static target, and background model is made of sample；
Step (2): on the basis of step (1) described background model, the tactful detection prospect of unanimity of samples is utilized；
Step (3): according to the model modification Developing Tactics distance threshold of adaptive sensitivity, and current pixel point is obtained Learning rate；
Step (4): pixel is calculated on rgb space the sum of with the cum rights color difference of its neighborhood, is then obtained current The average value of the minimum threshold of distance of pixel and the sample of background model.Background complexity journey is measured by the weighted sum of the two Spend simultaneously corrected range threshold value, final updating color distance threshold value and LBSP characteristic distance threshold value；
Step (5): it instructs to update the sample of background model with current autoadapted learning rate.
Further, the step (1), the method for background model is: the frame for the use of difference step size being n in preceding M frame Between calculus of finite differences, obtained from i-th, i+n and i+2*n frame and the location of pixels of marker motion target, the initial value of frame number i be 1, difference Step-length n=2.Then the rgb value and LBSP characteristic value for obtaining residual pixel position, are deposited as a sample of current pixel Enter background model.Then increase i value in preceding M-N frame inner iteration, repeat aforesaid operations and obtain pixel samples and counted, when certain When the sample size of pixel is N, then the initialization of the background model of the pixel is completed.When sample size is less than N and i+2*n When >=M-N, the insufficient pixel model of sample size is filled by the Pixel Information in subsequent N frame, until sample number is when being N Only.
The background model B (x) obtained by the above method is specific as follows:
B (x)={ B1(x) ..., BK(x) ..., BN(x)}
Wherein, BK(x) k-th of sample in model is indicated, N is total sample number, under normal circumstances N=20；
Further, the step (2), in carrying out the foreground detection stage, calculate separately the Euclidean of pixel and sample away from From and Hamming distance, as the similarity for measuring color and LBSP feature between pixel and sample.Its prospect decision rule is such as Under:
I.e. if being less than adaptive distance threshold R ' (x) in current pixel point I (x) and background model B (x) at a distance from sample Number be less than smallest match quantity min when, which is detected as prospect, on the contrary be background.
Further, in the step (3), by using background dynamics degree Dmin(x) it is updated with point of scintillation counter v (x) Distance threshold R (x).
Wherein, background dynamics degree Dmin(x)=Dmin(x)(1-θ)+dt(x) * θ, θ are learning rate, dtIt (x) is sample in model The minimum value of color and LBSP distance between sheet and pixel.
Point of scintillation counter mechanism v (x) updates as follows:
And XtIt (x) is the exclusive or result of current frame pixel and the segmentation of former frame pixel.
Adaptive distance threshold adjustment mode is by Dmin(x) it is indicated with v (x):
Further, in the step (4), with distance between samples all in background model B (x) and current pixel point I (x) The sum of the average value and current pixel point of minimum value and the cum rights rgb space color difference of 16 neighborhood territory pixel point of surrounding describe current The complex degree of background of pixel, is denoted as C (x).S is the threshold limit value of complex degree of background, RsIt is background complexity under s Distance threshold.It is as follows for the correction mechanism of distance threshold R (x):
As C (x) < s, background is more stable, takes R (x) history minimum value and RsBetween the greater as decision threshold, Change R (x) quickly, prevents decision threshold is too small from causing missing inspection；As C (x) >=s, i.e., The background of current pixel point is complicated, takes R (x) history maximum value and RsSmaller between+α (C (x)-s) as decision threshold, Avoid decision threshold is excessive from causing erroneous detection.R ' (x) has taken into account background stability and dynamic at this time, can obtain better detection Effect.
Last actual color threshold and the calculating of LBSP distance threshold are as follows, whereinWith30 and 3 are taken respectively:
Further, in the step (5), when determining that current pixel is background, record current pixel and sample distance are most Small pixel samples, and this sample is replaced with the learning rate of current pixel point with this Pixel Information, when being determined as prospect, It does not update then.
Beneficial effects of the present invention:
The advantages of maintaining original algorithm based on the background modeling method for improving local auto-adaptive sensitivity, largely On reduce slowly movement and pollution of the static target to background model in short-term, and impart distance threshold R (x) self-recision Mechanism optimizes the more new strategy of model, keeps the target prospect determined more complete accurate.
Detailed description of the invention
Fig. 1 is the nuclear structure schematic diagram of the background modeling algorithm of improved local auto-adaptive sensitivity of the present invention.
Specific embodiment
The present invention will be further explained below with reference to the attached drawings.
As shown in Figure 1, the background modeling algorithm of improved local auto-adaptive sensitivity of the present invention mainly includes background Model initialization, foreground detection, distance threshold are updated to be updated with correction strategy and learning rate.Below in terms of these in detail Illustrate implementation method of the invention.
Background model: the frame differential method for the use of difference step size being n in the preceding M frame (M=120), from i-th, i+n and I+2*n frame obtains and the location of pixels of marker motion target, and the initial value of frame number i is 1, difference step size n=2.Then it obtains surplus The rgb value and LBSP characteristic value of remaining location of pixels are stored in background model as a sample of current pixel.Then preceding M-N frame (N=20) inner iteration increases i value, repeats aforesaid operations and obtains pixel samples and counted, when the sample of certain pixel When quantity is N, then the initialization of the background model of the pixel is completed.When sample size is less than N and i+2*n >=M-N, pass through Pixel Information in subsequent N frame fills the insufficient pixel model of sample size, until when sample number is N.
The background model B (x) obtained by the above method is specific as follows:
B (x)={ B1(x) ..., BK(x) ..., BN(x)}
Wherein, BK(x) k-th of sample in model is indicated, N is total sample number.
This method combines temporal information and spatial information, can remove slowly movement and the influence of static target in short-term is built Vertical background model, enables model sample to embody true background.
Foreground detection: foreground and background separation, decision procedure are carried out for the pixel of a new frame video image are as follows:
I.e. if current pixel point It(x) in background model B (x) at a distance from each sample be less than distance threshold R ' (x) when number is less than number of matches min, which is detected as prospect；It otherwise is background.Wherein detection color value is similar L1 distance is used when spending, detection LBSP characteristic similarity uses Hamming distance.R (x) is also classified into color distance threshold value simultaneously With LBSP characteristic distance threshold value.
Distance threshold updates and correction strategy: by using background dynamics degree Dmin(x) more with point of scintillation counter v (x) New distance threshold R (x).
Wherein, background dynamics degree Dmin(x)=Dmin(x)(1-θ)+dt(x) * θ, θ are learning rate, dtIt (x) is sample in model The minimum value of color and LBSP distance between sheet and pixel.
Point of scintillation counter mechanism v (x) updates as follows:
XtIt (x) is the exclusive or result of current frame pixel and the segmentation of former frame pixel.
Adaptive distance threshold R (x) adjustment mode is by Dmin(x) it is indicated with v (x):
With the average value between samples all in background model B (x) and current pixel point I (x) apart from minimum value and currently The background complexity journey of the sum of cum rights rgb space color difference of pixel and 16 neighborhood territory pixel point of surrounding description current pixel point Degree, is denoted as C (x).S is the threshold limit value of complex degree of background, RsBe background complexity be s under distance threshold.
Wherein, the cum rights block of pixels that color background complexity is shown as table 1 is denoted as C (O) obtained by calculating.Central pixel point Difference weight with neighborhood A, B, C is respectivelyWithIts object is to improve the close neighborhood territory pixel point of distance center pixel To the descriptive power of background complexity, influence of the farther away neighborhood territory pixel point of distance center pixel to background complexity is reduced Power.C (O) calculation is as follows:
In formula, Ai、BiAnd CiRespectively indicate pixel value of the pixel in each channel RGB in 1 pixel region block of table, XoCentered on The pixel value of pixel O.
The cum rights block of pixels of the color complexity of the predefined measurement current pixel point of the present invention of table 1
C4 C3 C6
B3 A0 B1
C0 A1 O A3 C1
B0 A2 B2
C7 C2 C5
If current pixel point mutates, if point of scintillation flashes suddenly or stop flashing, then its background complexity is quick It increasing or reducing, the variation of above-mentioned distance threshold R (x) is unable to satisfy its speed, therefore the threshold value R (x) that adjusts the distance is modified, Working method is shown below:
Wherein, Weighted Threshold α is set as 6, the corresponding distance threshold R of complexity threshold limit value ssIt is 1.2.
As C (x) < s, background is more stable, takes R (x) history minimum value and RsBetween the greater as decision threshold, Change R (x) quickly, prevents decision threshold is too small from causing missing inspection；As C (x) >=s, that is, work as The background of preceding pixel point is complicated, takes R (x) history maximum value and RsSmaller between+α (C (x)-s) keeps away as decision threshold Exempt from that decision threshold is excessive to cause erroneous detection.R ' (x) has taken into account background stability and dynamic at this time, can obtain preferably detection effect Fruit.
Last actual color threshold and the calculating of LBSP distance threshold are as follows, wherein initial valueWithIt takes respectively 30 and 3:
It updates Background learning rate: when determining that current pixel is background, calculating and record current pixel and sample distance is minimum Sample, and this sample is replaced with the turnover rate of 1/T (x), when being determined as prospect, then not updated.The wherein update of T (x) Mode are as follows:
Wherein, St(x) be pixel in present frame testing result.
The series of detailed descriptions listed above only for feasible embodiment of the invention specifically Protection scope bright, that they are not intended to limit the invention, it is all without departing from equivalent implementations made by technical spirit of the present invention Or change should all be included in the protection scope of the present invention.

Claims (5)

1. one kind is based on improved local auto-adaptive sensitivity background modeling method, which comprises the steps of:
Step (1): the preceding M frame of selecting video sequence removes slow moving target and static in short-term as background model candidate frame Background model is established in the influence of target, and background model is made of sample；
Step (2): on the basis of step (1) described background model, prospect is detected using the principle of unanimity of samples；
Step (3): model modification Developing Tactics distance threshold and learning rate according to adaptive sensitivity；
Step (4): after obtaining distance threshold, distance threshold amendment is carried out；It is adjacent with it on rgb space that pixel is calculated first Then the sum of the cum rights color difference in domain obtains being averaged for the minimum threshold of distance of the sample of current pixel point and background model Value；Complex degree of background and corrected range threshold value, final updating color distance threshold value drawn game are measured by the weighted sum of the two Portion's two-value parallel pattern LBSP characteristic distance threshold value；
Step (5): it instructs to update the sample of background model with current autoadapted learning rate.
2. according to claim 1 a kind of based on improved local auto-adaptive sensitivity background modeling method, feature exists In establishing the method for background model in the step (1):
The frame differential method for the use of difference step size being n in preceding M frame obtains simultaneously marker motion target from i-th, i+n and i+2*n frame Location of pixels, the initial value of frame number i is 1, difference step size n=2；Then rgb value and the LBSP for obtaining residual pixel position are special Value indicative is stored in background model as a sample of current pixel；Then in preceding M-N frame inner iteration increase i value, repetition It states operation to obtain pixel samples and counted, when the sample size of certain pixel is N, then completes the background mould of the pixel The initialization of type.When sample size is less than N and i+2*n >=M-N, sample size is filled by the Pixel Information in subsequent N frame Insufficient pixel model, until when sample number is N；
The background model B (x) obtained by the above method is specific as follows:
B (x)={ B1(x) ..., BK(x) ..., BN(x)}
Wherein, BK(x) k-th of sample in model is indicated, N is total sample number.
3. according to claim 1 a kind of based on improved local auto-adaptive sensitivity background modeling method, feature exists In the realization of the step (3) is using background dynamics degree Dmin(x) and point of scintillation counter v (x) updates distance threshold R (x) and learning rate is obtained, the specific method is as follows:
Background dynamics degree Dmmin(x)=Dmin(x)(1-θ)+dt(x) * θ, θ are learning rate, dtIt (x) is sample and pixel in model The minimum value of color and LBSP distance between point.The update method of point of scintillation counter v (x) is as follows:
XtIt (x) is the exclusive or result of current frame pixel and the segmentation of former frame pixel.
4. according to claim 3 a kind of based on improved local auto-adaptive sensitivity background modeling method, feature exists In distance threshold R (x) update mode is as follows:
5. according to claim 1 a kind of based on improved local auto-adaptive sensitivity background modeling method, feature exists In, in the step (4), measure complex degree of background method and distance threshold correction strategy the specific implementation process is as follows:
With the average value and current pixel between samples all in background model B (x) and current pixel point I (x) apart from minimum value The complex degree of background of the sum of point and the cum rights rgb space color difference of 16 neighborhood territory pixel point of surrounding description current pixel point, note For C (x)；S is the threshold limit value of complex degree of background, RsBe background complexity be s under distance threshold, for distance threshold R (x) correction mechanism is as follows:
As C (x) < s, background is more stable, takes R (x) history minimum value and RsBetween the greater as decision threshold, work as background Complexity changes R (x) quickly, prevents decision threshold is too small from causing missing inspection；As C (x) >=s, i.e., current picture The background of vegetarian refreshments is complicated, takes R (x) history maximum value and RsSmaller between+α (C (x)-s) avoids sentencing as decision threshold Determine that threshold value is excessive to cause erroneous detection；R ' (x) has taken into account background stability and dynamic at this time, so as to obtain preferably detection effect Fruit；
Last actual color threshold and the calculating of LBSP distance threshold are as follows, whereinWith30 and 3 are taken respectively:
CN201910725944.8A 2019-08-07 2019-08-07 Improved local adaptive sensitivity-based background modeling method Active CN110503664B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910725944.8A CN110503664B (en) 2019-08-07 2019-08-07 Improved local adaptive sensitivity-based background modeling method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910725944.8A CN110503664B (en) 2019-08-07 2019-08-07 Improved local adaptive sensitivity-based background modeling method

Publications (2)

Publication Number Publication Date
CN110503664A true CN110503664A (en) 2019-11-26
CN110503664B CN110503664B (en) 2023-03-24

Family

ID=68588038

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910725944.8A Active CN110503664B (en) 2019-08-07 2019-08-07 Improved local adaptive sensitivity-based background modeling method

Country Status (1)

Country Link
CN (1) CN110503664B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111047654A (en) * 2019-12-06 2020-04-21 衢州学院 High-definition high-speed video background modeling method based on color information
CN111062974A (en) * 2019-11-27 2020-04-24 中国电力科学研究院有限公司 Method and system for extracting foreground target by removing ghost

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103810667A (en) * 2012-11-09 2014-05-21 通用电气航空系统有限责任公司 Spectral scene simplification through background substraction
CN108537771A (en) * 2018-01-30 2018-09-14 西安电子科技大学昆山创新研究院 MC-SILTP moving target detecting methods based on HSV
CN109544694A (en) * 2018-11-16 2019-03-29 重庆邮电大学 A kind of augmented reality system actual situation hybrid modeling method based on deep learning

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103810667A (en) * 2012-11-09 2014-05-21 通用电气航空系统有限责任公司 Spectral scene simplification through background substraction
CN108537771A (en) * 2018-01-30 2018-09-14 西安电子科技大学昆山创新研究院 MC-SILTP moving target detecting methods based on HSV
CN109544694A (en) * 2018-11-16 2019-03-29 重庆邮电大学 A kind of augmented reality system actual situation hybrid modeling method based on deep learning

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111062974A (en) * 2019-11-27 2020-04-24 中国电力科学研究院有限公司 Method and system for extracting foreground target by removing ghost
CN111062974B (en) * 2019-11-27 2022-02-01 中国电力科学研究院有限公司 Method and system for extracting foreground target by removing ghost
CN111047654A (en) * 2019-12-06 2020-04-21 衢州学院 High-definition high-speed video background modeling method based on color information

Also Published As

Publication number Publication date
CN110503664B (en) 2023-03-24

Similar Documents

Publication Publication Date Title
CN109583425B (en) Remote sensing image ship integrated recognition method based on deep learning
CN110211097B (en) Crack image detection method based on fast R-CNN parameter migration
CN109613006A (en) A kind of fabric defect detection method based on end-to-end neural network
CN105512640B (en) A kind of people flow rate statistical method based on video sequence
CN103971386B (en) A kind of foreground detection method under dynamic background scene
CN103413120B (en) Tracking based on object globality and locality identification
CN110119728A (en) Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network
CN110276264B (en) Crowd density estimation method based on foreground segmentation graph
CN103530893B (en) Based on the foreground detection method of background subtraction and movable information under camera shake scene
CN109559302A (en) Pipe video defect inspection method based on convolutional neural networks
CN106548488B (en) A kind of foreground detection method based on background model and inter-frame difference
CN104835179B (en) Based on the adaptive improvement ViBe background modeling methods of dynamic background
CN104299243B (en) Target tracking method based on Hough forests
CN105261037A (en) Moving object detection method capable of automatically adapting to complex scenes
CN105787482A (en) Specific target outline image segmentation method based on depth convolution neural network
CN108805897A (en) Improved moving target detection VIBE algorithm
CN103413149B (en) Method for detecting and identifying static target in complicated background
CN106651906A (en) Test method for motion characteristic of high-voltage breaker based on improved Gaussian mixture model
CN102288166A (en) Video-based multi-model combined surface ship detection method
CN107038416A (en) A kind of pedestrian detection method based on bianry image modified HOG features
CN107833241A (en) To real-time vision object detection method of the ambient lighting change with robustness
CN108550163A (en) Moving target detecting method in a kind of complex background scene
CN106991686A (en) A kind of level set contour tracing method based on super-pixel optical flow field
CN107169417A (en) Strengthened based on multinuclear and the RGBD images of conspicuousness fusion cooperate with conspicuousness detection method
CN106327488A (en) Adaptive foreground detection method and detection device

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