CN106933861A - A kind of customized across camera lens target retrieval method of supported feature - Google Patents
A kind of customized across camera lens target retrieval method of supported feature Download PDFInfo
- Publication number
- CN106933861A CN106933861A CN201511021989.5A CN201511021989A CN106933861A CN 106933861 A CN106933861 A CN 106933861A CN 201511021989 A CN201511021989 A CN 201511021989A CN 106933861 A CN106933861 A CN 106933861A
- Authority
- CN
- China
- Prior art keywords
- feature
- target
- result
- picture
- cascade
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
Abstract
The invention discloses a kind of customized across the camera lens target retrieval method of supported feature, i.e.,:Monitor video under the multiple cameras of collection, extracts target image, generation destination image data storehouse;Feature extraction is carried out to the view data in destination image data storehouse, target characteristic database is constituted;Selected inquiry picture, feature extraction is carried out to the picture, constitutes inquiry picture characteristic;Formulate self-defined cascade and the screening strategy of multiple features;According to the screening strategy formulated, using inquiring about picture searched targets property data base, in each screening process, the similarity score of each feature is calculated respectively, if any multiple features, the fraction of multiple features is cascaded in self-defined ratio, the result sequence sorted by similarity score size is obtained, then according to the target number upper limit and score threshold the selection result;Operating personnel remodify cascade and screening strategy according to the result accuracy rate of target retrieval, query process are repeated, to reach optimal retrieval effectiveness.
Description
Technical field
The present invention relates to a kind of target retrieval method, a kind of customized across the camera lens target inspection of supported feature is related in particular to
Suo Fangfa.The invention belongs to pattern-recognition and field of intelligent monitoring, specific objective in across camera video surveillance network is applied to
Retrieval.
Background technology
With the reduction and the popularization of network of hardware cost, video monitoring system is often using multiple video frequency pick-up head monitoring
Environment, and the target of motion in (it was found that and identification) monitoring scene is observed by the perception of monitoring personnel.Artificial place
The data of these magnanimity are managed, while expending huge manpower and financial resources, the efficiency for the treatment of is very low and easily because of monitoring personnel
Fatigue is neglected and is malfunctioned.Therefore, by it is fast automatic positioned from the monitor video of magnanimity, tracked and recognized it is interested
Moving target, some are high to realize intellectual traffic control and navigation, the search of stream of people's real-time monitoring, danger early warning, suspicious object etc.
Level application, is increasingly becoming very important research topic in computer vision field.
How to allow monitoring personnel in the several minutes that can be concentrated one's energy, efficiently capture clue and suspect object, this is one
Highly application oriented technical task, thus, across camera lens target retrieval technology is arisen at the historic moment.Across the notable of camera lens target retrieval
Characteristic be can the retrieval and inquisition target from magnanimity monitor video network within a very short time, be substantially shorter and search original video
Time, improve the accuracy of efficiency and manual identified.
CBIR is the technical foundation across camera lens target retrieval, and picture material is retouched by characteristics of image
State.Characteristics of image is the low-level visual feature for reflecting picture material, such as color, texture, or by being obtained after proper treatment
Side, angle, line, chromatic zones, shape, the spatial relationship of profile and objects in images etc..In across camera lens target retrieval,
Because the environment of IMAQ is different, quality, illumination condition, shooting angle of image etc. can all have larger difference, same
Individual pedestrian or motor vehicle can be very different in appearance under different cameras or in different time, cause we for
With knowing, another characteristic is highly unstable, therefore, existing technical scheme is carried out using only single features metric form mostly at present
Across camera lens target retrieval, it is impossible to complete the similarity measurement function across camera lens.
The content of the invention
The need for for across camera lens target retrieval under the diversified video monitoring environment of adaptation, goal of the invention of the invention is to provide one kind
Customized across the camera lens target retrieval method of supported feature.
To achieve the above object, the present invention uses following technical scheme:A kind of customized across camera lens target retrieval of supported feature
Method, it comprises the following steps:
Monitor video under A, the multiple cameras of collection, by computer vision techniques such as foreground detection, target followings,
Extract target image, generation destination image data storehouse;
B, feature extraction is carried out to the view data in destination image data storehouse, and combining target picture and space time information,
Constitute target characteristic database;
C, selected inquiry picture, feature extraction is carried out to the picture;
D, the self-defined cascade for formulating multiple features and screening strategy;
In the self-defined cascade and screening strategy, n times screening process, wherein N > 0, first time screening process can be carried out
Input include inquiring about the characteristic and target characteristic database of picture, the input of remaining each screening process is inquiry picture
Characteristic and last time the selection result, this selection result again as input next time, last time the selection result
As the retrieval result that this method is exported;
Cascade each time and screening process can use M feature, and wherein M > 0 need to such as be carried out using multiple features to feature
Cascade, when such as being cascaded using two features, the fractional result that characteristic matching is set respectively is x and y, carries out fraction normalizing
After change, as a result respectively x ' and y ', after weighted sum, fractional result is z=ax '+by ' (wherein a > 0 and b > 0, a and b
Value can by machine learning algorithm be fitted optimum value), matching result is ranked up according to fraction z then;
If there is provided similarity score threshold value, filters out qualified result, if there is provided object filtering number, according to
Fraction height takes the number upper limit fractional result;
E, the screening strategy according to formulation, using picture searched targets property data base is inquired about, in each screening process,
The similarity score of each feature is calculated respectively, if any multiple features, the fraction of multiple features is cascaded in self-defined ratio,
The result sequence sorted by similarity score size is obtained, then according to the target number upper limit and score threshold the selection result;
F, operating personnel remodify and cascade and screening strategy according to the result accuracy rate of target retrieval, repeat step E,
To reach optimal retrieval effectiveness.
Global characteristics are divided into the extraction that destination image data, inquiry picture carry out feature in the step B and step C
Extract the extraction with local feature;
The global characteristics include RGB, HSV, color characteristic, textural characteristics and shape facility,
The local feature includes SIFT feature, LBP features and Harr features.
Brief description of the drawings
Fig. 1 is self-defined across the camera lens target retrieval Method And Principle block diagram of supported feature of the present invention;
Fig. 2 is the theory diagram of the multiple features cascade that the present invention is used and screening strategy.
Specific embodiment
The present invention is illustrated below in conjunction with the accompanying drawings.
As shown in figure 1, customized across the camera lens target retrieval method of supported feature that the present invention is provided is:
Monitor video under A, the multiple cameras of collection, by computer vision techniques such as foreground detection, target followings,
Extract target image, generation destination image data storehouse.
The present invention carries out foreground detection using the background technology of wiping out, and obtaining prospect using present image and background image difference transports
Dynamic region.Background image is in general to be updated by the method for background modeling and obtained, and multimode state property is also dynamic background modeling
A factor that must take into consideration, such as leaf swing, the ripple of water etc..Mixed Gauss model is this quick change for the treatment of
The practical method of situation, its advantage can be that current most stable of pattern is selected in the mode of fixed number as the back of the body
Scape, and real-time online updates the Gaussian Distribution Parameters of each mode;Meanwhile, by using target tracking technology, identification is continuous
The same target occurred in time, is retrieved with reducing to the target for repeating.
B, feature extraction is carried out to the view data in destination image data storehouse, and combining target picture and space time information,
Constitute target characteristic database.
C, selected inquiry picture, feature extraction is carried out to the picture.
Extraction and local feature that feature extraction is divided into global characteristics are carried out to destination image data, inquiry picture in the present invention
Extraction.Wherein, global characteristics are local mainly including RGB, HSV, color characteristic, textural characteristics and shape facility etc.
Feature is mainly including SIFT feature (feature of local invariable rotary), LBP features and Harr features etc..
1) global characteristics extraction, is carried out to the view data in destination image data storehouse and inquiry picture.
Conventional global characteristics extracting method includes sub-space learning method, and this kind of method uses PCA technologies, increment R-SVD
Some basic images are extracted as feature etc. method, then carry out backprojection reconstruction on these basic images to original image, follow-up place
Reason, the method can effectively antagonize illumination and the change of posture during tracking.
2) local shape factor, is carried out to the view data in destination image data storehouse and inquiry picture.
SIFT feature extracting method:Difference of Gaussian filter under image and different scale is carried out into convolution algorithm, is being obtained
Image on the method that is positioned using multiresolution local extremum extract the point of safes in graphical rule space and (be referred to as key
Point), then the key point for obtaining is described using Local gradient direction histogram vectors, for follow-up pattern match.
LBP features are a kind of effective local texture description methods, are had to rotation and illumination variation a certain degree of non-quick
Perception, is widely used in the fields such as texture modeling and recognition of face.Harr features are a kind of based on the calculation of rectangular image block difference
The character description method of son, can define various Operator structures and yardstick in practical application.With reference to multi-level sampling structure and
The method of integrogram, Harr features have the quick advantage of calculating, are widely used in the detection of object, tracking and identification.
The extracting method of LBP features and Harr features is prior art, be will not be repeated here.
D, the self-defined cascade for formulating multiple features and screening strategy.
As shown in Fig. 2 in the self-defined cascade and screening strategy, n times screening process, wherein N > 0 can be carried out, the
The input of primary screening process include inquire about picture characteristic and target characteristic database, remaining each screening process it is defeated
Enter the characteristic and the selection result of last time to inquire about picture, this selection result again as input next time, finally
The retrieval result that primary screening result is exported as this method;
Cascade each time and screening process can use M feature, and wherein M > 0 need to such as be carried out using multiple features to feature
Cascade, when such as being cascaded using two features, the fractional result that characteristic matching is set respectively is x and y, carries out fraction normalizing
After change, as a result respectively x ' and y ', after weighted sum, fractional result be z=ax '+by ' (wherein a > 0 and b > 0, a and
The value of b can be fitted optimum value by machine learning algorithm), matching result is ranked up according to fraction z then, if setting
Similarity score threshold value is put, qualified result has been filtered out, if there is provided object filtering number, takes according to fraction height
The number upper limit fractional result.
E, the screening strategy according to formulation, using picture searched targets property data base is inquired about, in each screening process,
The similarity score of each feature is calculated respectively, if any multiple features, the fraction of multiple features is cascaded in self-defined ratio,
The result sequence sorted by similarity score size is obtained, then according to the target number upper limit and score threshold the selection result.
F, operating personnel remodify and cascade and screening strategy according to the result accuracy rate of target retrieval, repeat step E,
To reach optimal retrieval effectiveness.
When carrying out feature extraction to destination image data, inquiry picture in the present invention, to improve operational efficiency, the present invention is first
Dimensionality reduction is first carried out to the feature extracted using PCA technologies, then by the cascade of self-defined multiple features and screening strategy and is counted
The similarity of each feature is calculated, Stepwise Screening goes out similar target, finally according to the matching result under particular video frequency image-context
Feedback information, makes feature cascade and the screening strategy of optimization, to reach the accuracy rate of optimal target retrieval.
Beneficial effects of the present invention:
1st, using PCA (principal component analysis) dimensionality reduction technology, the accuracy rate and efficiency of algorithm of target retrieval can be improved.
2nd, because the environment of IMAQ is different, quality, illumination condition, shooting angle of image etc. can all have larger area
Not, under different image-contexts, different feature cascade and screening strategies may be adapted to, if in laboratory environments,
Fixed feature cascade and screening strategy is determined using limited sample, is then released product, it is difficult to suitable for various videos
Under image-context the need for across camera lens target retrieval.Therefore, the strategy of the cascade of flexible configuration feature and screening, in varying environment
Different sample trainings can be used down, the feature for being optimized is cascaded and screening strategy, so as to obtain optimal target inspection
Rope accuracy rate, makes have good adaptability and autgmentability with the product of the technology.
3rd, in the present invention feature cascade and screening strategy is highly configurable, and being mainly reflected in screening process number of times can match somebody with somebody
The feature put, screened use every time is configurable, the normalization of multiple features cascades and weight proportion is configurable, each similarity
The threshold value of fraction is configurable, object filtering number is configurable etc..Meanwhile, feature cascade and screening strategy can be preserved, very just
Just reuse.
4th, the present invention has good autgmentability, as long as the certain algorithm interface of agreement, can very easily by new feature
Matching algorithm is added in feature cascade and screening strategy.
The above is presently preferred embodiments of the present invention and its know-why used, and is come for those skilled in the art
Say, without departing from the spirit and scope of the present invention, any equivalent transformation based on the basis of technical solution of the present invention,
Simple replacement etc. is obvious to be changed, and is belonged within the scope of the present invention.
Claims (2)
1. a kind of customized across the camera lens target retrieval method of supported feature, it is characterised in that:It comprises the following steps:
Monitor video under A, the multiple cameras of collection, by computer vision techniques such as foreground detection, target followings,
Extract target image, generation destination image data storehouse;
B, feature extraction is carried out to the view data in destination image data storehouse, and combining target picture and space time information,
Constitute target characteristic database;
C, selected inquiry picture, feature extraction is carried out to the picture;
D, the self-defined cascade for formulating multiple features and screening strategy;
In the self-defined cascade and screening strategy, n times screening process, wherein N > 0, first time screening process can be carried out
Input include inquiring about the characteristic and target characteristic database of picture, the input of remaining each screening process is inquiry picture
Characteristic and last time the selection result, this selection result again as input next time, last time the selection result
As the retrieval result that this method is exported;
Cascade each time and screening process can use M feature, and wherein M > 0 need to such as be carried out using multiple features to feature
Cascade, when such as being cascaded using two features, the fractional result that characteristic matching is set respectively is x and y, carries out fraction normalizing
After change, as a result respectively x ' and y ', after weighted sum, fractional result is z=ax '+by ' (wherein a > 0 and b > 0, a and b
Value can by machine learning algorithm be fitted optimum value), matching result is ranked up according to fraction z then;
If there is provided similarity score threshold value, filters out qualified result, if there is provided object filtering number, according to
Fraction height takes the number upper limit fractional result;
E, the screening strategy according to formulation, using picture searched targets property data base is inquired about, in each screening process,
The similarity score of each feature is calculated respectively, if any multiple features, the fraction of multiple features is cascaded in self-defined ratio,
The result sequence sorted by similarity score size is obtained, then according to the target number upper limit and score threshold the selection result;
F, operating personnel remodify and cascade and screening strategy according to the result accuracy rate of target retrieval, repeat step E,
To reach optimal retrieval effectiveness.
2. customized across the camera lens target retrieval method of a kind of supported feature according to claim 1, it is characterised in that:
It is divided into the extraction of global characteristics in the step B and step C to the extraction that destination image data, inquiry picture carry out feature
With the extraction of local feature;
The global characteristics include RGB, HSV, color characteristic, textural characteristics and shape facility,
The local feature includes SIFT feature, LBP features and Harr features.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201511021989.5A CN106933861A (en) | 2015-12-30 | 2015-12-30 | A kind of customized across camera lens target retrieval method of supported feature |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201511021989.5A CN106933861A (en) | 2015-12-30 | 2015-12-30 | A kind of customized across camera lens target retrieval method of supported feature |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106933861A true CN106933861A (en) | 2017-07-07 |
Family
ID=59442471
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201511021989.5A Pending CN106933861A (en) | 2015-12-30 | 2015-12-30 | A kind of customized across camera lens target retrieval method of supported feature |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106933861A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107449371A (en) * | 2017-07-21 | 2017-12-08 | 上海理工大学 | The method of optimum choice standard spherical mirror head |
CN108777779A (en) * | 2018-06-12 | 2018-11-09 | 北京京东金融科技控股有限公司 | A kind of intelligent device, method, medium and the electronic equipment of video capture equipment |
CN108932509A (en) * | 2018-08-16 | 2018-12-04 | 新智数字科技有限公司 | A kind of across scene objects search methods and device based on video tracking |
CN109063656A (en) * | 2018-08-08 | 2018-12-21 | 厦门市美亚柏科信息股份有限公司 | A kind of method and device carrying out face inquiry using multiple face engines |
CN109255072A (en) * | 2018-08-15 | 2019-01-22 | 腾讯科技(深圳)有限公司 | Information recalls method and device, computer storage medium, electronic equipment |
CN110188217A (en) * | 2019-05-29 | 2019-08-30 | 京东方科技集团股份有限公司 | Image duplicate checking method, apparatus, equipment and computer-readable storage media |
CN110399528A (en) * | 2019-08-05 | 2019-11-01 | 北京深醒科技有限公司 | A kind of automatically across feature inference formula target retrieval method |
CN111814690A (en) * | 2020-07-09 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method and device and computer readable storage medium |
CN114419525A (en) * | 2022-03-30 | 2022-04-29 | 成都考拉悠然科技有限公司 | Harmful video detection method and system |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101510257A (en) * | 2009-03-31 | 2009-08-19 | 华为技术有限公司 | Human face similarity degree matching method and device |
CN102184186A (en) * | 2011-04-12 | 2011-09-14 | 宋金龙 | Multi-feature adaptive fusion-based image retrieval method |
CN102201061A (en) * | 2011-06-24 | 2011-09-28 | 常州锐驰电子科技有限公司 | Intelligent safety monitoring system and method based on multilevel filtering face recognition |
CN103617233A (en) * | 2013-11-26 | 2014-03-05 | 烟台中科网络技术研究所 | Method and device for detecting repeated video based on semantic content multilayer expression |
CN104778675A (en) * | 2015-04-28 | 2015-07-15 | 中国矿业大学 | Coal mining fully mechanized excavation face dynamic video image fusion method |
-
2015
- 2015-12-30 CN CN201511021989.5A patent/CN106933861A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101510257A (en) * | 2009-03-31 | 2009-08-19 | 华为技术有限公司 | Human face similarity degree matching method and device |
CN102184186A (en) * | 2011-04-12 | 2011-09-14 | 宋金龙 | Multi-feature adaptive fusion-based image retrieval method |
CN102201061A (en) * | 2011-06-24 | 2011-09-28 | 常州锐驰电子科技有限公司 | Intelligent safety monitoring system and method based on multilevel filtering face recognition |
CN103617233A (en) * | 2013-11-26 | 2014-03-05 | 烟台中科网络技术研究所 | Method and device for detecting repeated video based on semantic content multilayer expression |
CN104778675A (en) * | 2015-04-28 | 2015-07-15 | 中国矿业大学 | Coal mining fully mechanized excavation face dynamic video image fusion method |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107449371A (en) * | 2017-07-21 | 2017-12-08 | 上海理工大学 | The method of optimum choice standard spherical mirror head |
CN107449371B (en) * | 2017-07-21 | 2019-11-01 | 上海理工大学 | The method of optimum choice standard spherical mirror head |
CN108777779A (en) * | 2018-06-12 | 2018-11-09 | 北京京东金融科技控股有限公司 | A kind of intelligent device, method, medium and the electronic equipment of video capture equipment |
CN109063656A (en) * | 2018-08-08 | 2018-12-21 | 厦门市美亚柏科信息股份有限公司 | A kind of method and device carrying out face inquiry using multiple face engines |
CN109255072A (en) * | 2018-08-15 | 2019-01-22 | 腾讯科技(深圳)有限公司 | Information recalls method and device, computer storage medium, electronic equipment |
CN109255072B (en) * | 2018-08-15 | 2023-04-14 | 腾讯科技(深圳)有限公司 | Information recall method and device, computer storage medium and electronic equipment |
CN108932509A (en) * | 2018-08-16 | 2018-12-04 | 新智数字科技有限公司 | A kind of across scene objects search methods and device based on video tracking |
CN110188217A (en) * | 2019-05-29 | 2019-08-30 | 京东方科技集团股份有限公司 | Image duplicate checking method, apparatus, equipment and computer-readable storage media |
US11886492B2 (en) | 2019-05-29 | 2024-01-30 | Boe Technology Group Co., Ltd. | Method of matching image and apparatus thereof, device, medium and program product |
CN110399528A (en) * | 2019-08-05 | 2019-11-01 | 北京深醒科技有限公司 | A kind of automatically across feature inference formula target retrieval method |
CN110399528B (en) * | 2019-08-05 | 2023-11-07 | 北京深醒科技有限公司 | Automatic cross-feature reasoning type target retrieval method |
CN111814690A (en) * | 2020-07-09 | 2020-10-23 | 浙江大华技术股份有限公司 | Target re-identification method and device and computer readable storage medium |
CN111814690B (en) * | 2020-07-09 | 2023-09-01 | 浙江大华技术股份有限公司 | Target re-identification method, device and computer readable storage medium |
CN114419525A (en) * | 2022-03-30 | 2022-04-29 | 成都考拉悠然科技有限公司 | Harmful video detection method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106933861A (en) | A kind of customized across camera lens target retrieval method of supported feature | |
Ranjan et al. | Learning to count everything | |
Yang et al. | Visual perception enabled industry intelligence: state of the art, challenges and prospects | |
Asha et al. | Vehicle counting for traffic management system using YOLO and correlation filter | |
Junos et al. | Automatic detection of oil palm fruits from UAV images using an improved YOLO model | |
CN109766873B (en) | Pedestrian re-identification method based on hybrid deformable convolution | |
DE112009000485T5 (en) | Object comparison for tracking, indexing and searching | |
Singh et al. | Animal detection in man-made environments | |
CN109344842A (en) | A kind of pedestrian's recognition methods again based on semantic region expression | |
CN102346854A (en) | Method and device for carrying out detection on foreground objects | |
CN111199556A (en) | Indoor pedestrian detection and tracking method based on camera | |
Tao et al. | Smoke vehicle detection based on multi-feature fusion and hidden Markov model | |
Bhagat et al. | WheatNet-lite: A novel light weight network for wheat head detection | |
Dou et al. | Background subtraction based on deep convolutional neural networks features | |
Hussain et al. | A simple and efficient deep learning-based framework for automatic fruit recognition | |
Sun et al. | Exploiting deeply supervised inception networks for automatically detecting traffic congestion on freeway in China using ultra-low frame rate videos | |
Dodia et al. | A comparison of YOLO based vehicle detection algorithms | |
CN116912763A (en) | Multi-pedestrian re-recognition method integrating gait face modes | |
Bansal et al. | Dynamic texture using deep learning | |
Bousnina et al. | Learning system for mobile robot detection and tracking | |
Praveena et al. | Hybrid clustering algorithm and feed-forward neural network for satellite image classification | |
CN113569912A (en) | Vehicle identification method and device, electronic equipment and storage medium | |
Qin et al. | Scene segmentation based on seeded region growing for foreground detection | |
Liu et al. | A Probabilistic Framework Based on KDE-GMM hybrid model for moving object segmentation in dynamic scenes | |
CN112101433B (en) | Automatic lane-dividing vehicle counting method based on YOLO V4 and DeepSORT |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170707 |