CN105354576B - A kind of method of target's feature-extraction, target's feature-extraction module, object module creation module and intelligent image monitoring device - Google Patents
A kind of method of target's feature-extraction, target's feature-extraction module, object module creation module and intelligent image monitoring device Download PDFInfo
- Publication number
- CN105354576B CN105354576B CN201510639001.5A CN201510639001A CN105354576B CN 105354576 B CN105354576 B CN 105354576B CN 201510639001 A CN201510639001 A CN 201510639001A CN 105354576 B CN105354576 B CN 105354576B
- Authority
- CN
- China
- Prior art keywords
- target
- frame
- sub
- histograms
- oriented gradients
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
Abstract
A kind of method of target's feature-extraction, comprising: the first step, target detection frame lock onto target class；Second step, segmentation object find frame, target detection frame are made to be divided into several sub- frames of part superposition；Third step calculates the color histogram in each sub- frame.Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can include by more sub- frames, that is to say, that more sub- frames can include contour area.Since the textural characteristics of the contour area color of target are more abundant, more sub- frames have covered this color and vein region abundant, and the color histogram of sub- frame can more react clarification of objective, are conducive to the reliability for improving target following.
Description
Technical field
The present invention relates to intelligent image monitoring fields, are related specifically to method, the target signature of a kind of target's feature-extraction
Extraction module, object module creation module and intelligent image monitoring device.
Background technique
With the arrival of the internet of things era, intellectual product deepens continuously in our daily life, and produces to previous intelligence
More stringent requirements are proposed for the performance of product.
By taking intelligent image monitors as an example, it mainly includes target discovery module, object module creation module and target following
Module.
Wherein target discovery module passes through object module for positioning possible target class after target class positioning is correct
Creation module carries out feature extraction to interested target, after target's feature-extraction, using target tracking module to target
It is tracked.
Current monitoring device is mainly that three-dimensional space scene is mapped to two-dimensional surface, loses such as space and geographical
The information such as the three-dimensional depth of positional relationship, target itself.
The loss of these target informations is so that clarification of objective extraction becomes most important, in the prior art due to target spy
It levies extracting method and color histogram extraction is mainly carried out using the method for not overlapped partitioning target detection frame, so that target following
Reliability is not high, is easily lost or tracking error target.
Summary of the invention
Problems solved by the invention is in the prior art since target's feature-extraction method mainly uses not overlapped partitioning mesh
The method of mark discovery frame carries out color histogram extraction, so that the reliability of target following is not high, is easily lost or tracking error mesh
Mark.
To solve the above problems, the present invention provides a kind of method of target's feature-extraction, comprising:
The first step, target detection frame lock onto target class；
Second step, segmentation object find frame, target detection frame are made to be divided into several sub- frames of part superposition；
Third step calculates the color histogram in each sub- frame.
Further, the sub- frame is rectangle or square.
Further, between adjacent sub- frame between be divided into 8 pixels, part superposition between adjacent sub- frame.
Further, between second step and third step further include:
Calculate the histograms of oriented gradients in target detection frame；
Judge whether the target class of target detection frame locking is correct, if incorrect, returns by histograms of oriented gradients
The first step, if correctly, continuing to execute.
Further, between the histograms of oriented gradients in second step and calculating target detection frame further include: by every height
Frame is divided into four rectangular elements.
Further, the histograms of oriented gradients in calculating target detection frame includes:
Calculate the histograms of oriented gradients in each rectangular element；
The histograms of oriented gradients of all rectangular elements of connecting obtains the histograms of oriented gradients of target detection frame.
Further, the histograms of oriented gradients in each rectangular element is 9 dimension histograms of oriented gradients.
Compared with prior art, technical solution of the present invention has the advantage that
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
Further, it can effectively judge that the target class in target detection frame is not by the histograms of oriented gradients of target class
It is correct target class, if it is not, progress target class positioning again can be returned directly, avoids unnecessary target signature
It extracts, advantageously reduces calculation amount, and improve target following efficiency.
Further, every sub- frame is divided into 4 rectangular elements, the 9 dimension direction histograms calculated in each rectangular element can
To improve the extraction quality of target direction histogram of gradients, target class can be increased and judge reliability.
The present invention also provides a kind of target's feature-extraction modules formed by the method that the above method is formed.
Compared with prior art, technical solution of the present invention has the advantage that
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
The present invention also provides a kind of object module creation modules formed by the method that the above method is formed.
Compared with prior art, technical solution of the present invention has the advantage that
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
The present invention also provides a kind of intelligent image monitoring device, the intelligent image monitoring device includes above-mentioned target signature
Extraction module, at least one of object module creation module.
Compared with prior art, technical solution of the present invention has the advantage that
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
Detailed description of the invention
Fig. 1 is the implementation method of target discovery module；
Fig. 2 is the method flow diagram of first embodiment of the invention；
Fig. 3 is the vector schematic diagram of 9 dimension histograms of oriented gradients in first embodiment of the invention.
Specific embodiment
In the prior art since target's feature-extraction method is mainly carried out using the method for not overlapped partitioning target detection frame
Color histogram extracts, so that the reliability of target following is not high, is easily lost or tracking error target.
To make the above purposes, features and advantages of the invention more obvious and understandable, with reference to the accompanying drawing to the present invention
Specific embodiment be described in detail.
The implementation method of target discovery module includes:
A large amount of target class profile information, i.e. histograms of oriented gradients are stored, in intelligent image monitoring system with convenient
Target class is identified, for example, a large amount of body configuration's data can be inputted to identify people；
After the target class that target discovery module has found that it is likely that, the histograms of oriented gradients of target class will be collected, and
It is matched with a large amount of histograms of oriented gradients is stored in intelligent image monitoring system, if matching degree is more than threshold value, is recognized
To there is target class.
Target discovery module can only identify target class, and cannot identify specific target.For example, if intelligent image monitors
A large amount of body configuration's data are stored in system, then target discovery module can only recognize in the visual field whether someone, but not
It can identify the specific features of people, that is, it can not navigate to some specific people.
The implementation method of object module creation module includes:
After target discovery module determines that target class occurs, object module creation module can collect single or multiple targets
Target signature information, i.e. color histogram.For example, someone in the target discovery module discovery visual field, will to a certain personal or
Some establish corresponding color histogram for certain.
After storing the color histogram of target, start to track it.
With reference to Fig. 1, target tracking module is mainly comprised the following modules:
Behavior prediction module, the frame recording shot by screen, uses gaussian random gait pattern or Monte Carlo mould
Type establishes probability recurrence model to the track of mobile target, and predicts the probability of target next frame position.This prediction can
To effectively improve the stability of computational efficiency and target following.
Characteristic extracting module and Characteristic Contrast module after having found that it is likely that target, need to carry out feature extraction to the target,
The color histogram of possible target is extracted, and the color histogram of the color histogram of extraction and object reference model is carried out
Whether comparison, the target to determine discovery are correct.
The result of Characteristic Contrast is inputted decision-making module by decision-making module, and whether the target to judge discovery is correct, if really
Surely the target found is wrong, and the spatial position of behavior prediction module change search target is notified by feedback mechanism.
First embodiment
With reference to Fig. 2, the present embodiment provides a kind of methods of target's feature-extraction, comprising:
The first step, target detection frame lock onto target class.
Before monitoring system starts shooting, a large amount of target class profile information, i.e. direction are stored in target discovery module
Histogram of gradients, these target class profile informations carry out automatic mesh by matching algorithm for monitoring system in shooting process later
Class search is marked, if matching degree is greater than threshold value, and is considered the target class for needing to track.
After monitoring system finds target class, the target in target class is locked by target detection frame, makes target
It is surrounded by target detection frame.The generally rectangular cross-section frame of target detection frame, target class can be other objects of people or movement.Such as
There are multiple targets in fruit target class, then can create multiple target detection frames so that comprising a mesh in each target detection frame
Mark.
The locking of target can be monitoring system and choose automatically, can also be with human intervention, to find interested target.
Target can be individually, or multiple.
Second step, segmentation object find frame, and target detection frame is made to be divided into several sub- frame compositions of part superposition.
In a particular embodiment, the sub- frame is rectangle or square, part superposition between adjacent sub- frame.Sub- frame number
Depending on the spacing between the size and adjacent sub- frame of target detection frame.For the target detection frame of 64x128 pixel, if adjacent
8 pixels are divided between sub- frame, then target detection frame can be divided into 105 sub- frames.
Sub- frame quantity be easy to cause the characteristic of subsequent extracted too many too much and makes to handle time extension and deteriorate monitoring
Real-time；And the quantity such as fruit frame is very little, and be easy to cause the clarification of objective of extraction very little and cause target following can
It is reduced by property.
In a particular embodiment, every sub- frame can also be divided into 4 rectangular elements, to further increase target direction
The extraction quality of histogram of gradients can increase the reliability of target class judgement.
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
Third step calculates the histograms of oriented gradients in target detection frame.
After the segmentation of target detection frame, the histograms of oriented gradients in target detection frame is by calculating in each rectangular element
Histograms of oriented gradients react.
Histograms of oriented gradients calculates gradient magnitude and the direction of gray level image first, in each rectangular element according to
Amplitude size to reform the histograms of oriented gradients of a rectangular element to the counting weighting on different directions.It connects all squares
The histograms of oriented gradients of shape unit has just obtained the histograms of oriented gradients of entire target detection frame.
Histograms of oriented gradients is mainly used for defining for objective contour, that is to say, that can be anti-by histograms of oriented gradients
Mirror the appearance profile of target.
With reference to Fig. 3, in a particular embodiment, the direction gradient that can be calculated on 9 dimension directions in each rectangular element is straight
Fang Tu, wherein 8 are histograms of oriented gradients related with direction, 1 be independent of direction histograms of oriented gradients.
Since target detection frame is divided into 105 sub- frames, every sub- frame is divided into 4 rectangular elements, each rectangular element
9 dimension histograms of oriented gradients are calculated, therefore each target detection frame has 3780 dimension histograms of oriented gradients.
4th step, object judgement.
Judge whether the target class of target detection frame locking is correct, if incorrect, returns by histograms of oriented gradients
The first step continues to execute if correct.Each histograms of oriented gradients can be indicated by a gradient vector H.
It is compared with the inner product of gradient vector H with threshold value by device C before adjudicating, then judges target if it is greater than threshold value
It was found that in frame being correct target class, it is wrong to be otherwise considered as target class positioning.Following formula is to determine that target correctly calculates public affairs
Formula:
HC > Cthresh。
Can effectively judge whether correct the target class in target detection frame is by the histograms of oriented gradients of target
Target class avoids unnecessary target's feature-extraction, has if it is not, progress target class positioning again can be returned directly
Conducive to reduction calculation amount, and improve target following efficiency.
5th step extracts color histogram.
Color histogram is the ratio that the different colours in each sub- frame of statistics account for, and can reflect mesh by color histogram
Whether correct target feature compares the target that can be determined in target frame by the color histogram with reference model.
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
The method of above-described target's feature-extraction is used for target tracking module, in other embodiments, target signature
The method of extraction can be used for object module creation module, for extracting the clarification of objective that will be tracked.
Second embodiment
Second embodiment of the invention provides a kind of target's feature-extraction module formed by the method for first embodiment.
Compared with prior art, technical solution of the present invention has the advantage that
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
3rd embodiment
Third embodiment of the invention provides a kind of object module creation module formed by the method for first embodiment.
Compared with prior art, technical solution of the present invention has the advantage that
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
Fourth embodiment
Fourth embodiment of the invention provides a kind of intelligent image monitoring device, and the intelligent image monitoring device includes second
Target's feature-extraction module in embodiment, at least one of the object module creation module in 3rd embodiment.
Compared with prior art, technical solution of the present invention has the advantage that
Target detection frame is divided into several sub- frame compositions of part superposition, objective contour region can be by more sub- frames
Included, that is to say, that more sub- frames can include contour area.Due to target contour area color textural characteristics more
Add abundant, therefore more sub- frame has covered this color and vein region abundant, and the color histogram of sub- frame more can be anti-
Clarification of objective is answered, the reliability for improving target following is conducive to.
Although present disclosure is as above, present invention is not limited to this.Anyone skilled in the art are not departing from this
It in the spirit and scope of invention, can make various changes or modifications, therefore protection scope of the present invention should be with claim institute
Subject to the range of restriction.
Claims (4)
1. a kind of method of target's feature-extraction characterized by comprising
The first step, target detection frame lock onto target class；
Second step, segmentation object find frame, target detection frame are made to be divided into several sub- frames of part superposition；The sub- frame be rectangle or
Person's square, every sub- frame are divided into 4 rectangular elements；
Third step calculates the histograms of oriented gradients in target detection frame；Histograms of oriented gradients in the target detection frame
It is reacted by calculating the histograms of oriented gradients in each rectangular element；Specifically, the gradient width of gray level image is calculated first
Angle value and direction reform a rectangle to the counting weighting on different directions according to amplitude size in each rectangular element
The histograms of oriented gradients of unit, the histograms of oriented gradients for all rectangular elements of connecting just have obtained entire target detection frame
Histograms of oriented gradients；
4th step, object judgement；Judge whether the target class of target detection frame locking is correct by histograms of oriented gradients, often
A histograms of oriented gradients indicates by a gradient vector H, by the inner product of device C and gradient vector H before adjudicating and threshold value into
Row compares, and then judges that in target detection frame be correct target class if it is greater than threshold value, is otherwise considered as target class and is located
Accidentally；
5th step calculates the color histogram in each sub- frame.
2. the method for target's feature-extraction as described in claim 1, which is characterized in that be divided into 8 pictures between adjacent sub- frame
Element, part superposition between adjacent sub- frame.
3. the method for target's feature-extraction as described in claim 1, which is characterized in that the 4th step further include:
Judge whether the target class of target detection frame locking is correct, if incorrect, returns to first by histograms of oriented gradients
Step, if correctly, continuing to execute.
4. the method for target's feature-extraction as described in claim 1, which is characterized in that the direction gradient in each rectangular element is straight
Side's figure is 9 dimension histograms of oriented gradients.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510639001.5A CN105354576B (en) | 2015-09-30 | 2015-09-30 | A kind of method of target's feature-extraction, target's feature-extraction module, object module creation module and intelligent image monitoring device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510639001.5A CN105354576B (en) | 2015-09-30 | 2015-09-30 | A kind of method of target's feature-extraction, target's feature-extraction module, object module creation module and intelligent image monitoring device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105354576A CN105354576A (en) | 2016-02-24 |
| CN105354576B true CN105354576B (en) | 2019-06-11 |
Family
ID=55330543
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510639001.5A Active CN105354576B (en) | 2015-09-30 | 2015-09-30 | A kind of method of target's feature-extraction, target's feature-extraction module, object module creation module and intelligent image monitoring device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105354576B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108460899A (en) * | 2017-02-20 | 2018-08-28 | 深圳怡化电脑股份有限公司 | A kind of characteristics of image recognition methods and device |
| CN108874269B (en) * | 2017-05-12 | 2020-12-29 | 北京臻迪科技股份有限公司 | Target tracking method, device and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103714181A (en) * | 2014-01-08 | 2014-04-09 | 天津大学 | Stratification specific figure search method |
| CN104077605A (en) * | 2014-07-18 | 2014-10-01 | 北京航空航天大学 | Pedestrian search and recognition method based on color topological structure |
| CN104636754A (en) * | 2015-01-31 | 2015-05-20 | 华南理工大学 | Intelligent image classifying method based on tongue body partition color feature |
-
2015
- 2015-09-30 CN CN201510639001.5A patent/CN105354576B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103714181A (en) * | 2014-01-08 | 2014-04-09 | 天津大学 | Stratification specific figure search method |
| CN104077605A (en) * | 2014-07-18 | 2014-10-01 | 北京航空航天大学 | Pedestrian search and recognition method based on color topological structure |
| CN104636754A (en) * | 2015-01-31 | 2015-05-20 | 华南理工大学 | Intelligent image classifying method based on tongue body partition color feature |
Non-Patent Citations (1)
| Title |
|---|
| 视频监控中运动目标检测与跟踪算法研究;肖和平;《中国优秀硕士学位论文全文数据库 信息科技辑》;20131215;第2013年卷(第S2期);第I138-1503页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105354576A (en) | 2016-02-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105405154B (en) | Target object tracking based on color-structure feature | |
| Kratz et al. | Tracking with local spatio-temporal motion patterns in extremely crowded scenes | |
| Wojek et al. | Monocular 3d scene understanding with explicit occlusion reasoning | |
| Chan et al. | Counting people with low-level features and Bayesian regression | |
| CN103745203B (en) | View-based access control model notes the object detecting and tracking method with average drifting | |
| Chu et al. | Tracking human under occlusion based on adaptive multiple kernels with projected gradients | |
| CN101789124B (en) | Segmentation method for space-time consistency of video sequence of parameter and depth information of known video camera | |
| CN108537212B (en) | Student behavior detection method based on motion estimation | |
| CN110008867A (en) | A kind of method for early warning based on personage's abnormal behaviour, device and storage medium | |
| CN111652136B (en) | Pedestrian detection method and device based on depth image | |
| CN104200485A (en) | Video-monitoring-oriented human body tracking method | |
| CN107833239B (en) | Optimization matching target tracking method based on weighting model constraint | |
| CN104598883A (en) | Method for re-recognizing target in multi-camera monitoring network | |
| KR101130963B1 (en) | Apparatus and method for tracking non-rigid object based on shape and feature information | |
| CN109636828A (en) | Object tracking methods and device based on video image | |
| CN101324958A (en) | Method and apparatus for tracking object | |
| CN110111362A (en) | A kind of local feature block Similarity matching method for tracking target | |
| CN103714556A (en) | Moving target tracking method based on pyramid appearance model | |
| Joshi et al. | A low cost and computationally efficient approach for occlusion handling in video surveillance systems | |
| CN114187665A (en) | Multi-person gait recognition method based on human body skeleton heat map | |
| CN110348332A (en) | The inhuman multiple target real-time track extracting method of machine under a kind of traffic video scene | |
| CN109255802A (en) | Pedestrian tracting method, device, computer equipment and storage medium | |
| CN104156982B (en) | Motion target tracking method and device | |
| CN105354576B (en) | A kind of method of target's feature-extraction, target's feature-extraction module, object module creation module and intelligent image monitoring device | |
| CN102750522A (en) | Method for tracking targets |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP02 | Change in the address of a patent holder | ||
| CP02 | Change in the address of a patent holder |
Address after: Room 503, Building No. 588 Zixing Road, Minhang District, Shanghai, 200241 Patentee after: SHANGHAI SHENGYAO INTELLIGENT TECHNOLOGY CO., LTD. Address before: 200000 Free Trade Zone, Shanghai, Pudong New Area new Jinqiao Road, building 27, building 2, floor 13 Patentee before: SHANGHAI SHENGYAO INTELLIGENT TECHNOLOGY CO., LTD. |
|
| EE01 | Entry into force of recordation of patent licensing contract | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20160224 Assignee: Guangxi Sheng Yao science and technology of aviations Co.,Ltd. Assignor: SHANGHAI SHENGYAO INTELLIGENT SCIENCE & TECHNOLOGY Co.,Ltd. Contract record no.: X2021310000018 Denomination of invention: A method of target feature extraction, a target feature extraction module, a target model creation module and an intelligent image monitoring device are provided Granted publication date: 20190611 License type: Common License Record date: 20210408 |