CN108549833A - A kind of target extraction method of accurate robust - Google Patents
A kind of target extraction method of accurate robust Download PDFInfo
- Publication number
- CN108549833A CN108549833A CN201810184857.1A CN201810184857A CN108549833A CN 108549833 A CN108549833 A CN 108549833A CN 201810184857 A CN201810184857 A CN 201810184857A CN 108549833 A CN108549833 A CN 108549833A
- Authority
- CN
- China
- Prior art keywords
- pixel
- super
- target
- foreground
- video frame
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-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
- G06T7/12—Edge-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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/49—Segmenting video sequences, i.e. computational techniques such as parsing or cutting the sequence, low-level clustering or determining units such as shots or scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20112—Image segmentation details
- G06T2207/20156—Automatic seed setting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20172—Image enhancement details
- G06T2207/20192—Edge enhancement; Edge preservation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of target extraction methods of accurate robust, are a kind of methods for extracting foreground target in video frame, realize the external Objective extraction in static or background close to static state.This method includes three parts:1) super-pixel figure is generated;2) the foreground target detection based on super-pixel;3) foreground target refines.Compared with existing other methods, advantage of the invention is that:Target can accurately be extracted and identify boundary, efficiency high robust is good.
Description
Technical field
The present invention relates to a kind of target extraction methods of accurate robust, more particularly to are carried out under the fixed scene of video camera
The method of detection and the extraction of external object, belongs to video analysis field.
Background technology
Objective extraction is usually to need means to be used in video related application, and the purpose is to automatically extract in the video frame
Interested target, Objective extraction are commonly used in the Video Applications such as video monitoring, video compress and target identification.At these
In, video camera fixes irremovable or rotation, background be almost it is static constant, will be by when there is exterior object to enter visual field
It is detected and extracts as " foreground target ".So-called background " almost static constant " is considered based on following:1) in video
Background it is static constant, the movement of any pixel is not present;2) in scene there may be some relatively regular object of which movement,
For example the leaf in background is shaken due to wind, camera pedestal is due to shaking etc. caused by wind or other vibrations;
3) sunray migrates under outdoor scene and the slow scene brightness of generation changes, etc..
A kind of common target extraction method is the background subtraction based on mixed Gauss model, and this method is suitable for environment
Change slow scene.The basic thought of background subtraction is:This pixel quilt if some pixel is larger with respect to background deviation
It is considered foreground.The shortcomings that background subtraction, is:1) foreground object being extracted is very coarse, sometimes can only detect and carry
Take out a part for target;2) when pixel magnitude is very much that efficiency of algorithm is very low greatly;3) very sensitive to ambient noise, that is, it is easy
It is interfered by ambient noise.
Invention content
Technical problem to be solved by the invention is to provide a kind of target extraction method of accurate robust, the method it is basic
Thought is:First, each video frame is expressed as sparse form, i.e., each frame is converted to a super-pixel figure;2) using the back of the body
Scape subtraction sums up in the point that each pixel in foreground class or background classes;3) method is cut using figure to carry out carefully the foreground target detected
Change.
The present invention uses following technical scheme to solve above-mentioned technical problem:
The present invention provides a kind of target extraction method of accurate robust, including step in detail below:
Step 1, input video stream generates the corresponding super-pixel figure of each video frame;
Step 2, foreground target is detected based on super-pixel;
Step 3, the foreground target that step 2 detects is refined.
As the further technical solution of the present invention, in step 1 using the method for cluster by the pixel cluster of video frame at
Super-pixel, so that video frame is converted to the representation of super-pixel figure, specially:
1) it initializes:For video frame t, randomly chooses N number of point and be used as seed, be expressed asK=0,1 ..., N-1;
2) it clusters:WithIt is N number of super-pixel the pixel cluster of video frame t as seed;
3) seed generates:Calculate k-th of super-pixel ktGeometric center, and using it as new seed
As the further technical solution of the present invention, super-pixel is indicated using mixed Gauss model in step 2, and utilize the back of the body
Scape subtraction is detected foreground target, specially:
1) it initializes:If video frame t is not initial frame, go to step 2);Otherwise, for k-th of super-pixel kt, build
Vertical mixed Gauss modelWherein, L is the number of Gaussian function, and x is i-th of Gaussian function
Weight coefficient, gi(x;μi,σi) it is i-th of variance μiAnd meansquaredeviationσiGaussian function;
2) classify:For k-th of super-pixel kt, calculate each gi(x;μi,σi) mahalanobis distance, if from ktTo gi(x;
μi,σi) distanceThen ktBelong to foreground, gos to step 4);Otherwise belong to background, go to step 3);
3) it updates:Mixed Gauss model is updated with moving average;
4) it post-processes:Execute foreground target Refinement operation.
As the further technical solution of the present invention, step 3 includes two steps:1) foreground target restores;2) boundary refines.
As the further technical solution of the present invention, foreground target restores to cut method using figure, i.e.,:Before being super-pixel segmentation
Two class of scape and background, specially:
1) bounding box based on current super-pixel figure testing result is generated;
2) it is all super-pixel structure figures in bounding boxSo that:Each super-pixel is a node, and any two exists
It is connected with a line between two adjacent nodes of space;
3) schemingOn the segmentation result that is improved using the figure method of cutting of standard.
As the present invention further technical solution, boundary refine the step of it is as follows:
1) for each foreground target being resumed in super-pixel figure, a narrowband is drawn along its boundary;
2) on original video frame the fine boundary of method extraction Pixel-level is cut using figure;
3) output boundary.
The present invention has the following technical effects using above technical scheme is compared with the prior art:The present invention is that one kind exists
The method of extraction foreground target in video frame realizes the external Objective extraction in static or close static background.The present invention's
Advantage is:Target can accurately be extracted and identify boundary, efficiency high robust is good.
Description of the drawings
Fig. 1 is flow chart of the method for the present invention.
Specific implementation mode
Technical scheme of the present invention is described in further detail below in conjunction with the accompanying drawings:
The present invention proposes a kind of method cut based on super-pixel and figure, and the basic thought of the method is:First, each
Video frame is expressed as sparse form, i.e., each frame is converted to a super-pixel figure;2) each pixel is returned using background subtraction
It ties in foreground class or background classes;3) method is cut using figure to refine the foreground target detected.
Technical scheme of the present invention is as shown in Figure 1:
A method of detecting external object under video camera fixed scene, this method is suitable for executing in computing device,
Include the following steps:
Step 1, video frame is converted to the representation of super-pixel figure.
In step 1, similarity of the pixel of video frame based on some scale is spatially clustered into many groups, these
Group's i.e. so-called " super-pixel ".A large amount of pixel cluster at super-pixel, magnitude substantially reduces, therefore for the processing of super-pixel
It is much higher more than the treatment effeciency for pixel.In addition, super-pixel has stronger iamge description ability compared to pixel.For this purpose,
Efficient super-pixel clustering algorithm, such as average drifting (mean shift) method, watershed (water shed) can be used in we
Method etc..
The algorithm for generating super-pixel figure is as follows:
1) it initializes:For video frame t, randomly chooses N number of point and be used as seed, be expressed asK=0,1 ..., N-1;
2) it clusters:WithIt is N number of super-pixel the pixel cluster of video frame t as seed;
3) seed generates:Calculate k-th of super-pixel ktGeometric center, and using it as new seed
For k-th of super-pixel k in video frame ttThere are corresponding super-pixel k' in video frame t+1t+1.In order to look for
To corresponding to ktSuper-pixel k't+1, we only need to track by seed in video frame t+1The super-pixel of generation, is used in combination this
As a result the model in foreground detection is updated.
Step 2, foreground target is detected based on super-pixel figure.
In foreground target detection based on super-pixel, we can subtract method using existing background pixel-based,
We use mixed Gauss model to indicate super-pixel for this, wherein mixed Gauss model is several Gaussian functions with some weight
The model that ratio is summed to form.
K-th of super-pixel k in video frame t backgroundstIt is expressed as with mixed Gauss model (GMM)It is high based on mixing
Under this model, super-pixel k is corresponded in video frame t+1tSuper-pixel k't+1If k't+1It arrivesAny component distance
More than some threshold value (variance of such as 2 times this components), then k't+1It can be classified as foreground, be otherwise classified as background and for updating
Step 2 is described as follows:
1) it initializes:If video frame t is not initial frame, go to step 2);Otherwise, for k-th of super-pixel kt, build
Vertical mixed Gauss modelWherein, L is the number of Gaussian function, and x is i-th of Gaussian function
Weight coefficient, gi(x;μi,σi) it is i-th of variance μiAnd meansquaredeviationσiGaussian function.
2) classify:For k-th of super-pixel kt, calculate each gi(x;μi,σi) geneva (Mahalanobis) distance, from
ktTo gi(x;μi,σi) distance useIt indicates.IfThen ktBelong to foreground, gos to step 4);Otherwise belong to
In background, go to step 3).
3) it updates:Mixed Gauss model is updated with moving average
4) it post-processes:Execute foreground target Refinement operation.
Step 3, the foreground target detected is refined.
Compared to traditional target detection pixel-based, had in terms of efficiency and robustness based on the method for super-pixel bright
Aobvious advantage, however disadvantage there are two it:1) object boundary detected is not very accurate, an and often part for target;
2) object boundary detected is very coarse.Therefore micronization processes must be carried out to boundary to obtain more accurate and fine side
Boundary.
Boundary refines and processing includes two steps:1) foreground target restores;2) boundary refines.Foreground target restore purpose be
Certain error detections be background foreground restored, boundary refinement purpose be reach the fine target side of Pixel-level
Boundary.
Foreground target restores to cut method using figure, i.e.,:It is two class of foreground and background super-pixel segmentation.It is each in step 2
Super-pixel is all to be indicated with mixed Gauss model, and all foreground super-pixel is indicated with a mixed Gauss model here,
Same had powerful connections super-pixel is indicated with another mixed Gauss model.This representation can describe video frame well, from
And the super-pixel being accidentally classified as in background is restored in foreground.
Shown in foreground target recovery is as follows:
1) bounding box based on current super-pixel figure testing result is generated;
2) it is all super-pixel structure figures in bounding boxSo that:Each super-pixel is a node, and any two exists
It is connected with a line between two adjacent nodes of space;
3) schemingOn the segmentation result that is improved using the figure method of cutting of standard.
The result that foreground is restored similarly is to obtain a super-pixel figure, and wherein foreground super-pixel is marked as foreground.Before
Scape recovery usually can restore to obtain almost entire foreground target, but due to super-pixel, obtained boundary is not very smart
Carefully, it must continue to refine into row bound thus.The step of boundary refines is as follows:
1) for each foreground target being resumed in super-pixel figure, a narrowband is drawn along its boundary;
2) on original video frame the fine boundary of method extraction Pixel-level is cut using figure;
3) output boundary.
The above, the only specific implementation mode in the present invention, but scope of protection of the present invention is not limited thereto, appoints
What is familiar with the people of the technology within the technical scope disclosed by the invention, it will be appreciated that expects transforms or replaces, and should all cover
Within the scope of the present invention, therefore, the scope of protection of the invention shall be subject to the scope of protection specified in the patent claim.
Claims (6)
1. a kind of target extraction method of accurate robust, which is characterized in that including step in detail below:
Step 1, input video stream generates the corresponding super-pixel figure of each video frame;
Step 2, foreground target is detected based on super-pixel;
Step 3, the foreground target that step 2 detects is refined.
2. a kind of target extraction method of accurate robust according to claim 1, which is characterized in that using poly- in step 1
The method of class by the pixel cluster of video frame at super-pixel, so that video frame is converted to the representation of super-pixel figure, specifically
For:
1) it initializes:For video frame t, randomly chooses N number of point and be used as seed, be expressed asK=0,1 ..., N-1;
2) it clusters:WithIt is N number of super-pixel the pixel cluster of video frame t as seed;
3) seed generates:Calculate k-th of super-pixel ktGeometric center, and using it as new seed
3. a kind of target extraction method of accurate robust according to claim 2, which is characterized in that using mixed in step 2
It closes Gauss model and indicates super-pixel, and foreground target is detected using background subtraction, specially:
1) it initializes:If video frame t is not initial frame, go to step 2);Otherwise, for k-th of super-pixel kt, establish mixed
Close Gauss modelWherein, L is the number of Gaussian function, and x is the power of i-th of Gaussian function
Weight coefficient, gi(x;μi,σi) it is i-th of variance μiAnd meansquaredeviationσiGaussian function;
2) classify:For k-th of super-pixel kt, calculate each gi(x;μi,σi) mahalanobis distance, if from ktTo gi(x;μi,σi)
DistanceThen ktBelong to foreground, gos to step 4);Otherwise belong to background, go to step 3);
3) it updates:Mixed Gauss model is updated with moving average;
4) it post-processes:Execute foreground target Refinement operation.
4. a kind of target extraction method of accurate robust according to claim 1, which is characterized in that step 3 includes two steps:
1) foreground target restores;2) boundary refines.
5. a kind of target extraction method of accurate robust according to claim 4, which is characterized in that foreground target recovery is adopted
Method is cut with figure, i.e.,:It is two class of foreground and background super-pixel segmentation, specially:
1) bounding box based on current super-pixel figure testing result is generated;
2) it is all super-pixel structure figures in bounding boxSo that:Each super-pixel is a node, and any two is in space
It is connected with a line between two adjacent nodes;
3) schemingOn the segmentation result that is improved using the figure method of cutting of standard.
6. a kind of target extraction method of accurate robust according to claim 4, which is characterized in that the step of boundary refines
It is as follows:
1) for each foreground target being resumed in super-pixel figure, a narrowband is drawn along its boundary;
2) on original video frame the fine boundary of method extraction Pixel-level is cut using figure;
3) output boundary.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810184857.1A CN108549833A (en) | 2018-03-07 | 2018-03-07 | A kind of target extraction method of accurate robust |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810184857.1A CN108549833A (en) | 2018-03-07 | 2018-03-07 | A kind of target extraction method of accurate robust |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108549833A true CN108549833A (en) | 2018-09-18 |
Family
ID=63516326
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810184857.1A Withdrawn CN108549833A (en) | 2018-03-07 | 2018-03-07 | A kind of target extraction method of accurate robust |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108549833A (en) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103164858A (en) * | 2013-03-20 | 2013-06-19 | 浙江大学 | Adhered crowd segmenting and tracking methods based on superpixel and graph model |
CN103871076A (en) * | 2014-02-27 | 2014-06-18 | 西安电子科技大学 | Moving object extraction method based on optical flow method and superpixel division |
KR101532320B1 (en) * | 2014-04-18 | 2015-07-22 | 국방과학연구소 | Method for detecting a moving object using stereo camera installed in autonomous vehicle |
CN106981068A (en) * | 2017-04-05 | 2017-07-25 | 重庆理工大学 | A kind of interactive image segmentation method of joint pixel pait and super-pixel |
CN107016691A (en) * | 2017-04-14 | 2017-08-04 | 南京信息工程大学 | Moving target detecting method based on super-pixel feature |
-
2018
- 2018-03-07 CN CN201810184857.1A patent/CN108549833A/en not_active Withdrawn
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103164858A (en) * | 2013-03-20 | 2013-06-19 | 浙江大学 | Adhered crowd segmenting and tracking methods based on superpixel and graph model |
CN103871076A (en) * | 2014-02-27 | 2014-06-18 | 西安电子科技大学 | Moving object extraction method based on optical flow method and superpixel division |
KR101532320B1 (en) * | 2014-04-18 | 2015-07-22 | 국방과학연구소 | Method for detecting a moving object using stereo camera installed in autonomous vehicle |
CN106981068A (en) * | 2017-04-05 | 2017-07-25 | 重庆理工大学 | A kind of interactive image segmentation method of joint pixel pait and super-pixel |
CN107016691A (en) * | 2017-04-14 | 2017-08-04 | 南京信息工程大学 | Moving target detecting method based on super-pixel feature |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Yao et al. | Multi-layer background subtraction based on color and texture | |
Chen et al. | Efficient hierarchical method for background subtraction | |
Desa et al. | Image subtraction for real time moving object extraction | |
CN110929593A (en) | Real-time significance pedestrian detection method based on detail distinguishing and distinguishing | |
Roy et al. | Real-time adaptive histogram min-max bucket (HMMB) model for background subtraction | |
Fu et al. | Context-aware and depthwise-based detection on orbit for remote sensing image | |
Angelo | A novel approach on object detection and tracking using adaptive background subtraction method | |
Devi et al. | A survey on different background subtraction method for moving object detection | |
CN112883827B (en) | Method and device for identifying specified target in image, electronic equipment and storage medium | |
CN108549833A (en) | A kind of target extraction method of accurate robust | |
Wei et al. | An improved method of motion detection based on temporal difference | |
Tang et al. | Fast background subtraction using improved GMM and graph cut | |
He et al. | A Pedestrian Detection Method Using SVM and CNN Multistage Classification. | |
Chen et al. | Objects detecting based on adaptive background models and multiple cues | |
Lee et al. | Probabilistic background subtraction in a video-based recognition system. | |
Yu et al. | Background modeling with extracted dynamic pixels for pumping unit surveillance | |
Vaca-Castano et al. | Improving egocentric vision of daily activities | |
Davarpanah et al. | Using multi-scale filtering to initialize a background extraction model | |
Subetha et al. | Fusion-based Gaussian mixture model for background subtraction from videos | |
Fei et al. | A Review of Abnormal Personnel Behavior Detection for Safety Protection in Public Places | |
Zhong et al. | Robust background modeling via standard variance feature | |
Zhang et al. | Target Detection Algorithm Based on Region Segmentation in Moving Background | |
Lu et al. | Hierarchical foreground detection in dynamic background | |
Kiliroor et al. | Shadow Detection and Removal Using Dynamic Thresholding Approach for Driverless Car | |
Bulla et al. | Fake shadow detection using local HOG features |
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 | ||
WW01 | Invention patent application withdrawn after publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180918 |