CN104933441B - Object detection system and method - Google Patents
Object detection system and method Download PDFInfo
- Publication number
- CN104933441B CN104933441B CN201510323680.5A CN201510323680A CN104933441B CN 104933441 B CN104933441 B CN 104933441B CN 201510323680 A CN201510323680 A CN 201510323680A CN 104933441 B CN104933441 B CN 104933441B
- Authority
- CN
- China
- Prior art keywords
- target object
- filter
- svm
- deformable part
- characteristic pattern
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
-
- 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
- G06V2201/08—Detecting or categorising vehicles
Abstract
The invention discloses a kind of object detection method, the method includes：Obtain the characteristic pattern and twice of resolution characteristics figure of input picture；The root filter of deformable part model acts on the characteristic pattern, captures the global characteristics of target object, and all parts filter of deformable part model acts on the characteristic pattern of twice of the resolution ratio, captures the local feature of target object；Multiple support vector machines graders based on the training of greatest hope EM algorithms identify the target object in the input picture according to the global characteristics and local feature of target object；Export testing result.A kind of object detection system is also disclosed accordingly, and not only accuracy of detection and performance have been able to meet actual requirement, but also are expected to reach the application of heavy industrialization in a short time.
Description
Technical field
The present invention relates to target detection technique field more particularly to a kind of object detection system and methods.
Background technology
Currently, target detection is academicly, there are many models, wherein deformable member model (Deformable Part
Model, DPM) it is object detection model in the most popular recently figure, due to that can accurately detect target by joyous
It meets, is to generally acknowledge best algorithm of target detection.
But at present using DPM target detection technique be applied to vehicle, pedestrian detection when, detection result still needs to
It is promoted.Existing target detection technique detection efficiency, accuracy of detection not only fail to reach actual requirement, but also are difficult to large-scale industry
Change application.
Invention content
The technical issues of failing to reach actual requirement for the existing target detection technique detection efficiency of solution, accuracy of detection, this
A kind of object detection system of inventive embodiments offer and method.
In order to achieve the above objectives, the technical solution of the embodiment of the present invention is realized in：
A kind of object detection method, the method includes：
Obtain the characteristic pattern and twice of resolution characteristics figure of input picture；
The root filter of deformable part model acts on the characteristic pattern, captures the global characteristics of target object, can
The all parts filter of deformation component model acts on the characteristic pattern of twice of the resolution ratio, captures the part of target object
Feature；
Multiple support vector machines (SVM, Support Vector Machine) point based on the training of greatest hope EM algorithms
Class device identifies the target object in the input picture according to the global characteristics and local feature of target object；
Export testing result.
Wherein, described filter constitutes Star Model with the multiple component filter.
Wherein, the component filter is position-movable, and the local feature includes that the metamorphosis of target object is special
Sign.
Wherein, EM algorithms training mixing SVM models are based on, by mutually compensating for acting between multiple Linear SVM models,
The target object that need to be detected is divided into multiple subclasses according to the difference at visual angle or shape, corresponding SVM is built for each subclass
Model obtains multiple SVM classifiers, to detect the target object under various different visual angles.
Wherein, the deformable part model is weighting deformable part model.
A kind of object detection system, the system comprises：
Characteristic pattern unit obtains the characteristic pattern and twice of resolution characteristics figure of input picture；
Deformable part model unit, including a root filter and multiple component filters, described filter are used for
It acts on the characteristic pattern, captures the global characteristics of target object, each component filter is for acting on described two
On the characteristic pattern of times resolution ratio, the local feature of target object is captured；
Grader unit includes the multiple SVM classifiers trained based on EM algorithms, for according to the described complete of target object
Office's feature and local feature, identify the target object in the input picture, and export testing result.
Wherein, described filter constitutes hub-and-spoke configuration with the multiple component filter.
Wherein, the component filter is position-movable, and the local feature includes that the metamorphosis of target object is special
Sign.
Wherein, the grader unit be based on EM algorithms train mixing SVM models, by multiple Linear SVM models it
Between mutually compensate for act on, the target object that need to be detected is divided into multiple subclasses according to the difference at visual angle or shape, is each
Subclass builds corresponding SVM models, multiple SVM classifiers is obtained, to detect the target object under various different visual angles.
Wherein, the deformable part model is weighting deformable part model.
The embodiment of the present invention realizes target detection by deformable part model and multiple SVM classifiers, can not only
It is obviously improved detection efficiency, accuracy of detection, accuracy of detection and performance have been able to meet actual requirement, and are expected to reach in a short time
To the application of heavy industrialization.It can apply in vehicle flowrate, people flow rate statistical, vehicle security drive, rule-breaking vehicle inspection
Survey etc..
Description of the drawings
In attached drawing (it is not necessarily drawn to scale), similar reference numeral phase described in different views
As component.Similar reference numerals with different letter suffix can indicate the different examples of similar component.Attached drawing with example and
Unrestricted mode generally shows each embodiment discussed herein.
Fig. 1 is the flow chart of object detection method of the embodiment of the present invention；
Fig. 2 is the composed structure schematic diagram of object detection system of the embodiment of the present invention；
Fig. 3 is that the multiple SVM of the embodiment of the present invention divide schematic diagram to the son of primal problem；
Fig. 4 is for the embodiment of the present invention in the case where combining EM algorithm frames to the training process schematic diagram of mixing SVM；
Fig. 5 is the detects schematic diagram that the embodiment of the present invention is directed to upright pedestrian；
Fig. 6 is that the embodiment of the present invention mixes SVM models when being tested on PASCAL VOC2007 data sets in 20 classes
On detection mean accuracy schematic diagram.
Specific implementation mode
Embodiment one
The embodiment of the present invention provides a kind of object detection method, as shown in Figure 1, mainly may include steps of：
Step 101, the characteristic pattern and twice of resolution characteristics figure of input picture are obtained；
Step 102, the root filter of deformable part model acts on the characteristic pattern, captures the overall situation of target object
The all parts filter of feature, deformable part model acts on the characteristic pattern of twice of the resolution ratio, captures object
The local feature of body；
Step 103, multiple SVM classifiers based on the training of EM algorithms are special according to the global characteristics of target object and part
Sign, identifies the target object in the input picture；
Wherein, in machine learning field, support vector machines (SVM, Support Vector Machine), which is one, prison
The learning model superintended and directed, commonly used to carry out pattern-recognition, classification and regression analysis.
Step 104, testing result is exported.
The object detection method of the embodiment of the present invention can be applied in the statistical item of vehicle flowrate, accuracy of detection and performance
Actual requirement is had reached, and is expected to reach the application of heavy industrialization in a short time.It can also be applied in the stream of people simultaneously
Measure statistics, the safe driving of automobile, rule-breaking vehicle detection etc..
Embodiment two
The embodiment of the present invention provides a kind of object detection system, as shown in Fig. 2, the system comprises：
Characteristic pattern unit 21 obtains the characteristic pattern and twice of resolution characteristics figure of input picture；
Deformable part model unit 22, including a root filter and multiple component filters, described filter are used
In acting on the characteristic pattern, the global characteristics of target object are captured, each component filter is described for acting on
On the characteristic pattern of twice of resolution ratio, the local feature of target object is captured；
Grader unit 23 includes the multiple SVM classifiers trained based on EM algorithms, for according to described in target object
Global characteristics and local feature identify the target object in the input picture, and export testing result.
The object detection system of the embodiment of the present invention can be applied in the statistical item of vehicle flowrate, accuracy of detection and performance
Actual requirement is had reached, and is expected to reach the application of heavy industrialization in a short time.It can also be applied in the stream of people simultaneously
Measure statistics, the safe driving of automobile, rule-breaking vehicle detection etc..
Embodiment three
The object detection method of the embodiment of the present invention and the specific implementation of system is described in detail in the present embodiment.
The object detection method and system of the embodiment of the present invention are based primarily upon deformable part model and the EM training of weighting
The technologies such as SVM are mixed, good effect is achieved in the detection of the similar objects such as pedestrian and vehicle.
Wherein, if deformable part model is a star mould being made of a root filter and dry part filter
Type, the deformable part model have considered the Global Information of target and the presentation information and its spatial relationship of each section, can
To extract than based on the abundanter information of whole method.
The main feature of the deformable part model has following two points：
(1) the feature description ability that overall situation and partial situation is combined, root filter are a kind of global template for target, can
To catch the global characteristics of target, and component filter is a kind of local template, and acts on the characteristic pattern of two times of resolution ratio
On, the local feature of target can be described effectively；
(2) for the descriptive power of target morphology variation, since root filter is rigid, do not have the energy of processing deformation
Power, and the position of component filter is transportable in deformable part model, which can effectively handle human body
Etc. targets metamorphosis.
Therefore, the hub-and-spoke configuration constituted using root filter and component filter can effectively promote detection performance.
The detecting system proposes following two points on the basis of deformable part model and improves to promote deformable part
The detection performance of part model：
1) weighting block model.The power of comprehensive analysis all parts classifier performance, can be reinforced by weighting block
The classification performance of marking area strengthens utilization of some components in detection result.
Here, weighting block model is the deformable part model weighted, and different weights is subject to different components,
Be equivalent to the corresponding detection score of each component of deformable part model be multiplied by a weight make more effective component by
Pay attention to more, the maximized effect for playing all parts, to obtain better detection result.
2) since single Linear SVM model can not effectively solve the classification of two classification problems, and it is non-thread
Property SVM models computation complexity again it is very high, it is difficult to achieve the purpose that detect in real time.The embodiment of the present invention is used to be calculated by EM
Method training mixing SVM models.
EM algorithms, i.e. EM algorithm (Expectation Maximization Algorithm, but it is maximum to translate expectation
Change algorithm), it is a kind of iterative algorithm, the maximum for the probability parameter model containing hidden variable (hidden variable) is seemingly
So estimation or maximum a posteriori estimate.In statistics calculates, EM algorithms are found in probability (probabilistic) model
The algorithm of parameter maximal possibility estimation or MAP estimation, wherein probabilistic model depend on the hidden variable that can not be observed
(Latent Variable).Greatest hope is frequently used in the data clusters (Data of machine learning and computer vision
Clustering) field.
EM algorithms are alternately calculated by two steps：The first step is to calculate it is expected (E), is showed using to hidden variable
There is estimated value, calculates its maximum likelihood estimator；Second step is to maximize (M), maximizes the maximum likelihood value acquired in E steps
Carry out the value of calculating parameter.The estimates of parameters found in M steps is used for during next E step calculates, this process constantly alternately into
Row, iteration use EM steps, until convergence.
In the embodiment of the present invention, is trained based on EM algorithms and mutually compensate for acting between multiple Linear SVM models, held always
Continue until convergence.And the critical issue of the model is interaction and the distribution method between each submodel, is instructed using EM algorithms
Practice mixing SVM models, convert the output of SVM to probability, then utilizes the frame joint training of EM mixing SVM models,
It can make to mutually compensate between the detection performance of each subclass model in this way, by its convergence of experimental verification and effectively
Property.
In the embodiment of the present invention, by mutually compensating for acting between multiple Linear SVM models, the target that need to be detected is pressed
It is divided into several subclasses according to the difference at visual angle or shape so that the similarity between each subclass is very high, is then each subclass
Corresponding detection model is built, difference can be overcome in the class of target category to detection performance by the combination of multiple SVM models
Influence reach the target for promoting detection result to accurately detect target under various different visual angles.
Fig. 3 and Fig. 4 show respectively by multiple SVM to primal problem son divide and under based on EM algorithm frames it is right
Mix the training process of SVM.
As shown in figure 4, mixing SVM training process include the following steps, wherein input include positive sample P, negative sample N and
Initialization model ω, β, output include new detection mathematical model βnew, ωnew：
Step 401：All samples are first divided into M classes before training, follow-up training is done with grader ω, β per class.For institute
Some positive sample collection (xi,yi) execute following iterative step：
Step 402：The sample of input is calculated with SVM classifier, the mould that the result of output is described according to following formula (1)
Type carries out probability conversion.
Step 403：Execute M operations；
For each sample, according to pk(xi) value be ranked up, maximum pk(xi) positive sample where sample set will
It is sent to step 4 and is iterated calculating.
Step 404：For by the calculated sample set of EM methods, weighted value is calculated by following formula (2), (3)
α；
Step 405：Parameter ω, β of training weighting block SVM on sample set is obtained in EM methods, obtains new testing number
Learn model βnew, ωnew。
As shown in figure 5, the object detection system of the embodiment of the present invention and method to be applied to the detection of upright pedestrian, tool
Body realizes that process may include：
Step a1, input picture；
Step a2 obtains the characteristic pattern on the characteristic pattern and twice of resolution ratio of input picture；
Step a3, root filter act on the characteristic pattern, obtain the response image of root filter；All parts filter
The characteristic pattern in twice of resolution ratio is acted on, the response image of component filter is obtained, and after further obtaining range conversion
Response；
Step a4 merges the response image of the response image of root filter and all parts filter, by multiple
SVM classifier identifies the target in input picture；
Step a5 by testing result and is exported.
Here, in testing result, each target detected is irised out in a manner of picture frame in the input image.
The object detection system and method for the embodiment of the present invention have been successfully applied in the statistics of vehicle flowrate, accuracy of detection
Actual requirement is had reached with performance, and is expected to reach the application of heavy industrialization in a short time.It can also be applied to simultaneously
People flow rate statistical, the safe driving of automobile, rule-breaking vehicle detection etc..
Improved mixing SVM models are tested on PASCAL VOC2007 data sets, in identical test-strips
The performance of the object detection system of the embodiment of the present invention improves 2.1%mAP on the whole under part, specifically the detection in 20 classes
Mean accuracy table as shown in FIG. 6.
The foregoing is only a preferred embodiment of the present invention, is not intended to limit the scope of the present invention.
Claims (8)
1. a kind of object detection method, which is characterized in that the method includes：
The target object that need to be detected is divided into multiple subclasses according to the difference at visual angle or shape, is built for each subclass corresponding
SVM models obtain multiple SVM classifiers；
Multiple SVM classifiers are trained using EM algorithms, including：E is operated：Convert the output of SVM classifier to probability；M is operated：
Sample set where choosing the positive sample of maximum probability calculates the value of the SVM classifier parameter；Described in iteration E operation and it is described
M is operated, until convergence；
Obtain the characteristic pattern and twice of resolution characteristics figure of input picture；
The root filter of deformable part model acts on the characteristic pattern of the input picture, and the overall situation for capturing target object is special
Sign, all parts filter of deformable part model act on twice of resolution characteristics figure, capture target object
Local feature；
Multiple support vector machines graders based on the training of greatest hope EM algorithms according to the global characteristics of target object and
Local feature identifies the target object in the input picture；
Export testing result.
2. according to the method described in claim 1, it is characterized in that：Described filter is constituted with all parts filter
Star Model.
3. method according to claim 1 or 2, which is characterized in that
The component filter it is position-movable, the local feature includes the morphological change characteristics of target object.
4. according to the method described in claim 1, it is characterized in that：
The deformable part model is weighting deformable part model.
5. a kind of object detection system, it is characterised in that the system comprises：
Grader unit is every for the target object of detection will to be needed to be divided into multiple subclasses according to the difference at visual angle or shape
A subclass builds corresponding SVM models, obtains multiple SVM classifiers；Multiple SVM classifiers are trained using EM algorithms, including：E
Operation：Convert the output of SVM classifier to probability；M is operated：Sample set where choosing the positive sample of maximum probability calculates
The value of the SVM classifier parameter；E operations and M operations described in iteration, until convergence；
Characteristic pattern unit, the characteristic pattern for obtaining input picture and twice of resolution characteristics figure；
Deformable part model unit, including a root filter and multiple component filters, described filter is for acting on
On the characteristic pattern of the input picture, the global characteristics of target object are captured, the multiple component filter is for acting on
On twice of resolution characteristics figure, the local feature of target object is captured；
The grader unit includes the multiple SVM classifiers trained based on EM algorithms, for according to the described complete of target object
Office's feature and local feature, identify the target object in the input picture, and export testing result.
6. system according to claim 5, which is characterized in that described filter is constituted with the multiple component filter
Hub-and-spoke configuration.
7. system according to claim 5 or 6, which is characterized in that
The component filter it is position-movable, the local feature includes the morphological change characteristics of target object.
8. system according to claim 5, it is characterised in that：
The deformable part model is weighting deformable part model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510323680.5A CN104933441B (en) | 2015-06-12 | 2015-06-12 | Object detection system and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510323680.5A CN104933441B (en) | 2015-06-12 | 2015-06-12 | Object detection system and method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104933441A CN104933441A (en) | 2015-09-23 |
| CN104933441B true CN104933441B (en) | 2018-08-10 |
Family
ID=54120601
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510323680.5A Active CN104933441B (en) | 2015-06-12 | 2015-06-12 | Object detection system and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104933441B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105225523B (en) * | 2015-10-15 | 2018-01-02 | 浙江宇视科技有限公司 | A kind of parking space state detection method and device |
| CN106295523A (en) * | 2016-08-01 | 2017-01-04 | 马平 | A kind of public arena based on SVM Pedestrian flow detection method |
| CN106407943A (en) * | 2016-09-28 | 2017-02-15 | 天津工业大学 | Pyramid layer positioning based quick DPM pedestrian detection method |
| CN106548145A (en) * | 2016-10-31 | 2017-03-29 | 北京小米移动软件有限公司 | Image-recognizing method and device |
| CN106909936B (en) * | 2017-01-21 | 2020-02-21 | 江苏大学 | Vehicle detection method based on double-vehicle deformable component model |
| CN106874873A (en) * | 2017-02-16 | 2017-06-20 | 上海建炜信息技术有限公司 | Rapid pedestrian detection method based on DPM |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101794515A (en) * | 2010-03-29 | 2010-08-04 | 河海大学 | Target detection system and method based on covariance and binary-tree support vector machine |
| CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
| CN102819753A (en) * | 2012-07-17 | 2012-12-12 | 华中科技大学 | Object detection method based on local model employing maximal sub-diagram |
| CN104268539A (en) * | 2014-10-17 | 2015-01-07 | 中国科学技术大学 | High-performance human face recognition method and system |
-
2015
- 2015-06-12 CN CN201510323680.5A patent/CN104933441B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101794515A (en) * | 2010-03-29 | 2010-08-04 | 河海大学 | Target detection system and method based on covariance and binary-tree support vector machine |
| CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
| CN102819753A (en) * | 2012-07-17 | 2012-12-12 | 华中科技大学 | Object detection method based on local model employing maximal sub-diagram |
| CN104268539A (en) * | 2014-10-17 | 2015-01-07 | 中国科学技术大学 | High-performance human face recognition method and system |
Non-Patent Citations (3)
| Title |
|---|
| "Human Detection in Images via Piecewise Linear Support Vector Machines";Qixiang Ye ET AL;《IEEE TRANSACTIONS ON IMAGE PROCESSING》;20130228;第778-780、783页、图4、图7 * |
| "基于可变形部件模型融合上下文的物体类检测";肖志勇;《中国优秀硕士学位论文全文数据库 信息科技辑》;20140915;第15-23页、图2.8 * |
| "基于支持向量机的目标检测算法综述";郭明玮 等;《控制与决策》;20140228;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104933441A (en) | 2015-09-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104933441B (en) | Object detection system and method | |
| Zhang et al. | Histogram of 3d facets: A characteristic descriptor for hand gesture recognition | |
| Ali et al. | A real-time deformable detector | |
| CN104166841A (en) | Rapid detection identification method for specified pedestrian or vehicle in video monitoring network | |
| CN104915673B (en) | A kind of objective classification method and system of view-based access control model bag of words | |
| CN103902978B (en) | Face datection and recognition methods | |
| CN103778436B (en) | A kind of pedestrian's attitude detecting method based on image procossing | |
| CN104063719A (en) | Method and device for pedestrian detection based on depth convolutional network | |
| CN103942577A (en) | Identity identification method based on self-established sample library and composite characters in video monitoring | |
| CN102467657A (en) | Gesture recognizing system and method | |
| Lin et al. | Live Face Verification with Multiple Instantialized Local Homographic Parameterization. | |
| Weiyao et al. | Human action recognition using multilevel depth motion maps | |
| CN103745233B (en) | The hyperspectral image classification method migrated based on spatial information | |
| CN103793926A (en) | Target tracking method based on sample reselecting | |
| WO2017116879A1 (en) | Recognition of hand poses by classification using discrete values | |
| CN103839066A (en) | Feature extraction method based on biological vision | |
| Putro et al. | Lightweight convolutional neural network for real-time face detector on cpu supporting interaction of service robot | |
| Ghorbel et al. | An extension of kernel learning methods using a modified log-euclidean distance for fast and accurate skeleton-based human action recognition | |
| CN103824062A (en) | Motion identification method for human body by parts based on non-negative matrix factorization | |
| CN103714340A (en) | Self-adaptation feature extracting method based on image partitioning | |
| JP5704909B2 (en) | Attention area detection method, attention area detection apparatus, and program | |
| Wang et al. | A new hand gesture recognition algorithm based on joint color-depth superpixel earth mover's distance | |
| Chiu et al. | See the difference: Direct pre-image reconstruction and pose estimation by differentiating hog | |
| CN103020631B (en) | Human movement identification method based on star model | |
| CN108564595A (en) | Image tracking method and device, electronic equipment, storage medium, program |
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 |