CN105138983B - The pedestrian detection method divided based on weighting block model and selective search - Google Patents
The pedestrian detection method divided based on weighting block model and selective search Download PDFInfo
- Publication number
- CN105138983B CN105138983B CN201510518178.XA CN201510518178A CN105138983B CN 105138983 B CN105138983 B CN 105138983B CN 201510518178 A CN201510518178 A CN 201510518178A CN 105138983 B CN105138983 B CN 105138983B
- Authority
- CN
- China
- Prior art keywords
- filter
- pedestrian
- indicate
- picture
- sample
- 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
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
- G06V40/164—Detection; Localisation; Normalisation using holistic features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of pedestrian detection method divided based on weighting block model and selective search, is made of weighting block model training part and pedestrian detection part.Model construction process: the construction process based on deformable part model, it is different according to the circumstance of occlusion at each position of pedestrian under crowd scene, different weights is provided with to different components, and be trained weight using hidden support vector machines.Pedestrian detection process: extracting the HOG feature pyramid in picture, carries out selective search segmentation to picture, includes more objects with less valid window to reach, and herein on basis, carries out pedestrian detection using the cascade detection cut based on threshold value.The method of the present invention solves the problems, such as the missing inspection of pedestrian under crowd scene from two angles, has many advantages, such as to reduce interference, improves pedestrian detection accuracy rate.
Description
Technical field
The present invention relates to a kind of pedestrian detection field, deformable part model and choosing of especially a kind of use based on weight
The pedestrian detection method of selecting property search segmentation.
Background technique
Pedestrian detection based on still image is exactly that the human body occurred in image is split from background and is accurately determined
Position, is widely used in the fields such as human-computer interaction, intelligent video monitoring, intelligent transportation, is the basis of target classification and behavior understanding,
It is simultaneously also one of the important topic in computer vision research field.But when pedestrian is as the main body detected, often go out
Now in sufficiently complex background, and pedestrian has both rigidity and flexible double characteristic, appearance vulnerable to dress, scale, block,
Posture and visual angle etc. influence, this but also pedestrian detection not only become computer vision perception area research one of emphasis,
It is one of the difficult point of computer vision perception area research.
Pedestrian detection algorithm at this stage is mainly to be studied from two angles: first, based on global feature
Pedestrian detection algorithm.This kind of algorithm is people to be regarded as an entirety to extract feature, is then instructed using classifier to it
Practice, to achieve the effect that detect human body.Second, the concept of introducing member and posture carries out human testing.This kind of algorithm
If being that human body is resolved into stem portion, feature extraction is carried out respectively to each section, then using classifier to each section
Feature training, and these features are added up carry out human testing.Compared with first kind algorithm, such methods can be more convenient
Processing joint prob, and in the future can be carried out human posture estimation provide fundamental basis.
Nowadays, the pedestrian detection method based on deformable part model be in terms of pedestrian detection more successful algorithm it
One, which mainly includes two parts: overall model and partial model.Overall model captures the entirety of object in low resolution
Profile, partial model capture the detail section of object in high-resolution.The model combines the Global Information of object well
And local message, it is contemplated that the flexible characteristic of pedestrian.But under crowd scene, due to the presence of partial occlusion, the wheel of pedestrian
Wide visibility is limited, and the verification and measurement ratio of deformable part model is declined.
Summary of the invention
It is an object of that present invention to provide a kind of pedestrian detection method divided based on weighting block model and selective search,
Make model that can more meet the rigid structure of pedestrian under crowded environment on the basis of meeting human body non-rigid features, reduce detection process
In block and bring interference, in conjunction with selective search dividing method and cascade detection algorithm, avoid the slightly remote pedestrian of missing inspection,
While not reducing detection speed, the accuracy rate of pedestrian detection is improved.
To achieve the above object, use following technical scheme: detection method of the present invention is instructed by weighting block model
Practice part and pedestrian detection part composition, method and step is as follows:
(1) weighting block model training
According to the circumstance of occlusion at each position of pedestrian under crowd scene, different weights is set to different components, is used
Hidden support vector machines is trained weight, and optimizes weight using stochastic gradient descent method;
(2) pedestrian detection
Pedestrian's video pictures are obtained, selective search segmentation is carried out to picture, the HOG feature in picture is extracted and constructs spy
Pyramid is levied, pedestrian is detected using the cascade detection method cut based on threshold value.
The specific method is as follows for model training in step (1):
Positive sample the P=[(I of (1-1) input training pattern1,B1),…,(It,Bt)], negative sample N={ J1,…,Jk, instruction
Practice total sample number | D=P ∪ N |, given the number of iterations K and number of vessels m;
Wherein, I1The 1st picture that corresponding positive sample is concentrated, B1Corresponding picture I1In pedestrian's bounding box, ItCorresponding positive sample
T-th of picture of this concentration, BtCorresponding picture ItIn pedestrian's bounding box, J1The 1st picture that corresponding negative sample is concentrated, JkIt is right
K-th of picture for answering negative sample to concentrate, K are the number of iterations, and m is number of vessels, and m value is 3, are shown using three containers, point
The front, side and the half body blocked of pedestrian are not illustrated, and D is training sample sum;
(1-2) is carried out | D=P ∪ N | secondary circulation, the process recycled every time are as follows:
1. the feature vector of each training sample is extracted,
In formula, p0The position of corresponding root filter, is indicated, H corresponds to histogram of gradients feature by standard of the upper left corner
Pyramid diagram,Indicate the p in feature pyramid H0The feature vector of position, pnCorresponding n-th of means filter
Position,Indicate the p in feature pyramid HnThe feature vector of position,Indicate the 1st component filtering
The relative deformation feature vector of device and root filter,Indicate that n-th of means filter is opposite with root filter
Deformation characteristics vector;
2. picture is divided into 3 groups according to the vertical-horizontal proportion of multiple pedestrian's bounding boxes, respectively from pedestrian's half body, side and just
Face constructs 3 mixed models, and picture is divided into 3 groups here, is half body, side and this positive 3 angle because of pedestrian,
It can include most of pedestrian;
(1-3) carries out m circulation, and the process recycled every time is as follows:
1. initialization vector β=(F0,ω1·F1,…,ωn·Fn,d1,…,dn, b), F0Indicate root filter, Fi(1≤i
≤ n) indicate i-th of means filter, ωi(1≤i≤n) indicates the weight of i-th of means filter, di(1≤i≤n) is indicated
The relative deformation coefficient of i-th means filter and root filter, b indicate deviation;
2. initialization includes positive sampleAnd negative sampleSample pool, cj(1≤j≤m) is that j-th of model holds
Picture group corresponding to device,It is cjThe positive sample of group,It is cjThe negative sample of group,It is cjAll negative samples of group;
3. carrying out K iteration, the process of each iteration is as follows:
A) stochastic gradient descent method, optimization objective function are usedUsing it is hidden support to
Amount machine is trained it, i.e.,
In formula, ψ (H, z) is the feature vector of training sample;β is initialization vector;Z is a series of potential value, i.e. root
The position collection of filter and means filter;X indicates each sample;Z (x) indicates samplexPossible potential value set;It is cj
The positive sample of group;It is cjThe negative sample of group;cjIt is picture group corresponding to j-th of model-container, and 1≤j≤m;
B) data mining removesIn simple sample, increaseIn negative sample, and then expandSample pool;
(1-4) obtains weighting block model.
The method of step (2) pedestrian detection is as follows:
(2-1) inputs picture to be detected;
(2-2) carries out selective search segmentation to picture, and removal can not have the region of pedestrian, obtain R region;
(2-3) extracts the HOG feature of picture, construction feature pyramid;
(2-4) carries out PCA dimensionality reduction to weighting block model, is converted into cascade model;
(2-5) carries out K circulation, carries out cascade detection to each region respectively, finds out hypothesis score
In formula, p0The position of corresponding root filter, pnThe position of corresponding n-th of means filter, using the upper left corner as standard into
Row indicates that H corresponds to histogram of gradients feature pyramid,Indicate the p in feature pyramid H0The feature vector of position,
F0Indicate root filter, piThe position of corresponding i-th of means filter,Indicate the p in feature pyramid HiPosition
Feature vector, FiIndicate i-th of means filter, ωiIndicate the weight of i-th of means filter, dxiIndicate i-th of component mistake
Range difference of the filter relative to the X-direction between the ideal position and physical location of root filter, dyiIndicate i-th
Range difference of the part filter relative to the Y direction between the ideal position and physical location of root filter,
Indicate the relative deformation feature vector of i-th of means filter Yu root filter, diIndicate that i-th of means filter and root filter
The relative deformation coefficient of device, b indicate deviation;
score(p0,…,pn) it is made of four parts:For the score of root filter,
For the score of the means filter with weight,Ideal for means filter relative to root filter
The deformation score of position and physical location;
If score (p0,…,pn) it is greater than cascaded thresholds, then it is assumed that this assumes region overlay pedestrian area, referred to as candidate
It is assumed that otherwise think this assume region be not covered with pedestrian area, extract include pedestrian bounding box;
(2-6) carries out duplicate removal to the bounding box for including pedestrian, removes duplicate region;
(2-7) is got the bid out pedestrian with rectangle frame in picture, obtains final pedestrian's detection block.
Compared with prior art, the present invention has the advantage that
1, different from existing deformable part model, the weighting block model in the present invention more meets crowd scene downlink
The rigidity of people and nonrigid characteristic reduce the leakage that interference brought by blocking in detection process improves pedestrian under crowd scene
Inspection rate.
2, this method introduces selective search segmentation from reducing search space and accelerating the two angles of matching speed
It is detected with cascade, avoids the slightly remote pedestrian of missing inspection, while not reducing detection speed, improve the accurate of pedestrian detection
Rate, comprehensive speed, precision and the robustness for improving pedestrian detection algorithm.
3, this method real-time with higher and processing speed, method is succinct, practical, there is higher efficiency.
Detailed description of the invention
Fig. 1 is the procedure chart of the method for the present invention.
Fig. 2 is the flow chart of weighting block model training in the method for the present invention.
Fig. 3 is the flow chart of the pedestrian detection based on selective search segmentation in the method for the present invention.
Specific embodiment
Method proposed in this paper is described further below with reference to the HOG feature pyramid detection process of pedestrian:
In the flow chart of the method for the present invention as shown in Figure 1, it is trained model first, detailed process is followed successively by trained sample
This, initialization root filter, using standard SVM training, merge container, using LSVM training, initialisation unit filter, again
Using LSVM training, update component weight, obtain weighting block model.Then image detection is carried out, detailed process is followed successively by defeated
Enter picture, be split (cutting procedure includes selective search, obtains useful bounding box) to picture, weighting block model
Cascade model is converted to, the feature pyramid for obtaining picture, cascade detection, pedestrian area is obtained, the duplicate region of removal, obtains
Final pedestrian detection result.
The specific method is as follows:
Detection method of the present invention is made of weighting block model training part and pedestrian detection part, wherein
(1) weighting block model training
According to the circumstance of occlusion at each position of pedestrian under crowd scene, different weights is set to different components, is used
Hidden support vector machines is trained weight, and optimizes weight using stochastic gradient descent method;It is as shown in Figure 2:
Positive sample the P=[(I of (1-1) input training pattern1,B1),…,(It,Bt)], negative sample N={ J1,…,Jk, instruction
Practice total sample number | D=P ∪ N, given the number of iterations K and number of vessels m;
Wherein, I1The 1st picture that corresponding positive sample is concentrated, B1Corresponding picture I1In pedestrian's bounding box, ItCorresponding positive sample
T-th of picture of this concentration, BtCorresponding picture ItIn pedestrian's bounding box, J1The 1st picture that corresponding negative sample is concentrated, JkIt is right
K-th of picture for answering negative sample to concentrate, K are the number of iterations, and m is number of vessels, and m value is 3, show to use three containers,
The front, side and the half body blocked of pedestrian are shown respectively, D is training sample sum;
(1-2) is carried out | D=P ∪ N | secondary circulation, the process recycled every time are as follows:
1. the feature vector of each training sample is extracted,
In formula, p0The position of corresponding root filter, is indicated, H corresponds to histogram of gradients feature by standard of the upper left corner
Pyramid,Indicate the p in feature pyramid H0The feature vector of position, pnThe position of corresponding n-th of means filter
It sets,Indicate the p in feature pyramid HnThe feature vector of position, dx1Indicate the 1st means filter relative to root
The range difference of X-direction between the ideal position and physical location of filter, dy1Indicate that the 1st means filter is opposite
The range difference of Y direction between the ideal position and physical location of root filter, dxnIndicate n-th of means filter
Relative to the range difference of the X-direction between the ideal position and physical location of root filter, dynIndicate n-th of component mistake
Range difference of the filter relative to the Y direction between the ideal position and physical location of root filter,It indicates
The relative deformation feature vector of 1st means filter and root filter,Indicate n-th of means filter with
The relative deformation feature vector of root filter;
2. picture is divided into 3 groups according to the vertical-horizontal proportion of multiple pedestrian's bounding boxes, respectively from pedestrian's half body, side and just
Face constructs 3 mixed models, and picture is divided into 3 groups here, is half body, side and this positive 3 angle because of pedestrian,
It can include most of pedestrian;
(1-3) carries out m circulation, and the process recycled every time is as follows:
1. initialization vector β=(F0,ω1·F1,…,ωn·Fn,d1,…,dn, b), F0Indicate root filter, Fi(1≤i
≤ n) indicate i-th of means filter, ωi(1≤i≤n) indicates the weight of i-th of means filter, di(1≤i≤n) is indicated
The relative deformation coefficient of i-th means filter and root filter, b indicate deviation;
2. initialization includes positive sampleAnd negative sampleSample pool, cj(1≤j≤m) is that j-th of model holds
Picture group corresponding to device,It is cjThe positive sample of group,It is cjThe negative sample of group,It is cjAll negative samples of group;
3. carrying out K iteration, the process of each iteration is as follows:
A) stochastic gradient descent method, optimization objective function are usedUsing it is hidden support to
Amount machine is trained it, i.e.,
In formula, ψ (H, z) is the feature vector of training sample;β is initialization vector;Z is a series of potential value, i.e. root
The position collection of filter and means filter;It is cjThe positive sample of group;It is cjThe negative sample of group;cj(1≤j≤m) is jth
Picture group corresponding to a model-container;
B) data mining removesIn simple sample, increaseIn negative sample, and then expandSample pool;
(1-4) obtains weighting block model.
(2) pedestrian detection
Pedestrian's video pictures are obtained, selective search segmentation is carried out to picture, the HOG feature in picture is extracted and constructs spy
Pyramid is levied, pedestrian is detected using the cascade detection method cut based on threshold value.It is as shown in Figure 3:
(2-1) inputs picture to be detected;
(2-2) carries out selective search segmentation to picture, and removal can not have the region of pedestrian, obtain R region;
(2-3) extracts the HOG feature of picture, construction feature pyramid;
(2-4) carries out PCA dimensionality reduction to weighting block model, is converted into cascade model;
(2-5) carries out K circulation, carries out cascade detection to each region respectively, finds out hypothesis score
In formula, p0The position of corresponding root filter, pnThe position of corresponding n-th of means filter, using the upper left corner as standard into
Row indicates that H corresponds to histogram of gradients feature pyramid,Indicate the p in feature pyramid H0The feature vector of position,
F0Indicate root filter, piThe position of corresponding i-th of means filter,Indicate the p in feature pyramid HiPosition
Feature vector, FiIndicate i-th of means filter, ωiIndicate the weight of i-th of means filter, dxiIndicate i-th of component mistake
Range difference of the filter relative to the X-direction between the ideal position and physical location of root filter, dyiIndicate i-th
Range difference of the part filter relative to the Y direction between the ideal position and physical location of root filter,
Indicate the relative deformation feature vector of i-th of means filter Yu root filter, diIndicate that i-th of means filter and root filter
The relative deformation coefficient of device, b indicate deviation;
score(p0,…,pn) it is made of four parts:For the score of root filter,For the score of the means filter with weight,For means filter relative to
The ideal position of root filter and the deformation score of physical location;
If score (p0,…,pn) it is greater than cascaded thresholds, then it is assumed that this assumes region overlay pedestrian area, referred to as candidate
It is assumed that otherwise think this assume region be not covered with pedestrian area, extract include pedestrian bounding box;
(2-6) carries out duplicate removal to the bounding box for including pedestrian, removes duplicate region;
(2-7) is got the bid out pedestrian with rectangle frame in picture, obtains final pedestrian's detection block.
Embodiment described above only describe the preferred embodiments of the invention, not to model of the invention
It encloses and is defined, without departing from the spirit of the design of the present invention, those of ordinary skill in the art are to technical side of the invention
The various changes and improvements that case is made should all be fallen into the protection scope that claims of the present invention determines.
Claims (1)
1. a kind of pedestrian detection method divided based on weighting block model and selective search, which is characterized in that the detection
Method is made of weighting block model training part and pedestrian detection part, and method and step is as follows:
(1) weighting block model training
According to the circumstance of occlusion at each position of pedestrian under crowd scene, different weights is set to different components, uses hidden branch
It holds vector machine and is trained weight, and weight is optimized using stochastic gradient descent method;
(2) pedestrian detection
Pedestrian's video pictures are obtained, selective search segmentation is carried out to picture, extract the HOG feature in picture and construction feature gold
Word tower detects pedestrian using the cascade detection method cut based on threshold value;
The specific method is as follows for model training in the step (1):
Positive sample the P=[(I of (1-1) input training pattern1,B1),…,(It,Bt)], negative sample N={ J1,…,Jk, training sample
This sum | D=P ∪ N |, given the number of iterations K and number of vessels m;
Wherein, I1The 1st picture that corresponding positive sample is concentrated, B1Corresponding picture I1In pedestrian's bounding box, ItCorresponding positive sample collection
In t-th of picture, BtCorresponding picture ItIn pedestrian's bounding box, J1The 1st picture that corresponding negative sample is concentrated, JkIt is corresponding negative
K-th of picture in sample set, m value are 3, show to use three containers, show respectively the front of pedestrian, side and
The half body blocked;
(1-2) is carried out | D=P ∪ N | secondary circulation, the process recycled every time are as follows:
1. the feature vector of each training sample is extracted,
In formula, p0The position of corresponding root filter, is indicated, H corresponds to histogram of gradients feature gold word by standard of the upper left corner
Tower,Indicate the p in feature pyramid H0The feature vector of position, pnThe position of corresponding n-th of means filter,Indicate the p in feature pyramid HnThe feature vector of position, dx1Indicate the 1st means filter relative to root mistake
The range difference of X-direction between the ideal position and physical location of filter, dy1Indicate the 1st means filter relative to
The range difference of Y direction between the ideal position and physical location of root filter, dxnIndicate n-th of means filter phase
For the range difference of the X-direction between the ideal position and physical location of root filter, dynIndicate n-th of component filtering
Range difference of the device relative to the Y direction between the ideal position and physical location of root filter,Indicate the 1st
The relative deformation feature vector of a means filter and root filter,Indicate n-th of means filter and root mistake
The relative deformation feature vector of filter;
2. picture is divided into 3 groups according to the vertical-horizontal proportion of multiple pedestrian's bounding boxes, come respectively from pedestrian's half body, side and front
Construct 3 mixed models;
(1-3) carries out m circulation, and the process recycled every time is as follows:
1. initialization vector β=(F0,ω1·F1,…,ωn·Fn,d1,…,dn, b), F0Indicate root filter, FiIt indicates i-th
Means filter, ωiIndicate the weight of i-th of means filter, diIndicate that i-th of means filter is opposite with root filter
Deformation coefficient, b indicate deviation;
2. initialization includes positive sampleAnd negative sampleSample pool, cjIt is figure corresponding to j-th of model-container
Piece group,It is cjThe positive sample of group,It is cjThe negative sample of group,It is cjAll negative samples of group;
3. carrying out K iteration, the process of each iteration is as follows:
A) stochastic gradient descent method, optimization objective function are usedUsing hidden support vector machines
It is trained, i.e.,
In formula, ψ (H, z) is the feature vector of training sample;β is initialization vector;Z is a series of potential value, i.e., root filters
The position collection of device and means filter;X indicates each sample;Z (x) indicates the possible potential value set of sample x;It is cjGroup
Positive sample;It is cjThe negative sample of group;cjIt is picture group corresponding to j-th of model-container, and 1≤j≤m;
B) data mining removesIn simple sample, increaseIn negative sample, and then expandSample pool;
(1-4) obtains weighting block model;
The method of step (2) pedestrian detection is as follows:
(2-1) inputs picture to be detected;
(2-2) carries out selective search segmentation to picture, and removal can not have the region of pedestrian, obtain R region;
(2-3) extracts the HOG feature of picture, construction feature pyramid;
(2-4) carries out PCA dimensionality reduction to weighting block model, is converted into cascade model;
(2-5) carries out K circulation, carries out cascade detection to each region respectively, finds out hypothesis score
In formula, p0The position of corresponding root filter, pnThe position of corresponding n-th of means filter carries out table by standard of the upper left corner
Showing, H corresponds to histogram of gradients feature pyramid,Indicate the p in feature pyramid H0The feature vector of position, F0Table
Show root filter, piThe position of corresponding i-th of means filter,Indicate the p in feature pyramid HiThe feature of position
Vector, FiIndicate i-th of means filter, ωiIndicate the weight of i-th of means filter, dxiIndicate i-th of means filter
Relative to the range difference of the X-direction between the ideal position and physical location of root filter, dyiIndicate i-th of component mistake
Range difference of the filter relative to the Y direction between the ideal position and physical location of root filter,It indicates
The relative deformation feature vector of i-th means filter and root filter, diIndicate i-th of means filter and root filter
Relative deformation coefficient, b indicate deviation;
score(p0,…,pn) it is made of four parts:For the score of root filter,For band
There is the score of the means filter of weight,Ideal position for means filter relative to root filter
With the deformation score of physical location;
If score (p0,…,pn) be greater than cascaded thresholds, then it is assumed that this assumes region overlay pedestrian area, it is referred to as candidate it is assumed that
Otherwise think this assume region be not covered with pedestrian area, extract include pedestrian bounding box;
(2-6) carries out duplicate removal to the bounding box for including pedestrian, removes duplicate region;
(2-7) is got the bid out pedestrian with rectangle frame in picture, obtains final pedestrian's detection block.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510518178.XA CN105138983B (en) | 2015-08-21 | 2015-08-21 | The pedestrian detection method divided based on weighting block model and selective search |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510518178.XA CN105138983B (en) | 2015-08-21 | 2015-08-21 | The pedestrian detection method divided based on weighting block model and selective search |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105138983A CN105138983A (en) | 2015-12-09 |
CN105138983B true CN105138983B (en) | 2019-06-28 |
Family
ID=54724327
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510518178.XA Active CN105138983B (en) | 2015-08-21 | 2015-08-21 | The pedestrian detection method divided based on weighting block model and selective search |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105138983B (en) |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105975925B (en) * | 2016-05-03 | 2019-05-03 | 电子科技大学 | Partial occlusion pedestrian detection method based on joint-detection model |
CN106295523A (en) * | 2016-08-01 | 2017-01-04 | 马平 | A kind of public arena based on SVM Pedestrian flow detection method |
CN108009466B (en) * | 2016-10-28 | 2022-03-15 | 北京旷视科技有限公司 | Pedestrian detection method and device |
CN106778665B (en) * | 2016-12-29 | 2019-09-17 | 浙江大华技术股份有限公司 | A kind of vehicle window detection method and device |
CN107133607B (en) * | 2017-05-27 | 2019-10-11 | 上海应用技术大学 | Demographics' method and system based on video monitoring |
CN107463954B (en) * | 2017-07-21 | 2019-10-25 | 华中科技大学 | A kind of template matching recognition methods obscuring different spectrogram picture |
CN109727188A (en) * | 2017-10-31 | 2019-05-07 | 比亚迪股份有限公司 | Image processing method and its device, safe driving method and its device |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102810161A (en) * | 2012-06-07 | 2012-12-05 | 江苏物联网研究发展中心 | Method for detecting pedestrians in crowding scene |
CN103617426A (en) * | 2013-12-04 | 2014-03-05 | 东北大学 | Pedestrian target detection method under interference by natural environment and shelter |
CN104217225A (en) * | 2014-09-02 | 2014-12-17 | 中国科学院自动化研究所 | A visual target detection and labeling method |
-
2015
- 2015-08-21 CN CN201510518178.XA patent/CN105138983B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102810161A (en) * | 2012-06-07 | 2012-12-05 | 江苏物联网研究发展中心 | Method for detecting pedestrians in crowding scene |
CN103617426A (en) * | 2013-12-04 | 2014-03-05 | 东北大学 | Pedestrian target detection method under interference by natural environment and shelter |
CN104217225A (en) * | 2014-09-02 | 2014-12-17 | 中国科学院自动化研究所 | A visual target detection and labeling method |
Non-Patent Citations (4)
Title |
---|
Cascade Object Detection with Deformable Part Models;Pedro F. Felzenszwalb et al;《2010 IEEE conference on Computer vision and pattern recognition》;20100618;第2241-2248页 |
Weighted Deformable Part Model for Robust Human Detection;Tianshuo Li et al;《ICIC 2014 LNCS 8588》;20141231;第764-775页 |
基于DPM模型的行人检测技术的研究;熊聪;《电子设计工程》;20141231;第22卷(第23期);第172-173页 |
基于可变形部件模型的人体级联检测;张闯;《中国优秀硕士学位论文全文数据库信息科技辑》;20140915(第09期);第I138-1035页 |
Also Published As
Publication number | Publication date |
---|---|
CN105138983A (en) | 2015-12-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105138983B (en) | The pedestrian detection method divided based on weighting block model and selective search | |
CN110084292B (en) | Target detection method based on DenseNet and multi-scale feature fusion | |
CN107239740B (en) | A kind of SAR image automatic target recognition method of multi-source Fusion Features | |
CN103310194B (en) | Pedestrian based on crown pixel gradient direction in a video shoulder detection method | |
CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
CN105046195B (en) | Human bodys' response method based on asymmetric generalized gaussian model | |
CN106874894A (en) | A kind of human body target detection method based on the full convolutional neural networks in region | |
CN106709568A (en) | RGB-D image object detection and semantic segmentation method based on deep convolution network | |
CN107464252A (en) | A kind of visible ray based on composite character and infrared heterologous image-recognizing method | |
CN102214291A (en) | Method for quickly and accurately detecting and tracking human face based on video sequence | |
CN107798691B (en) | A kind of unmanned plane independent landing terrestrial reference real-time detection tracking of view-based access control model | |
CN104408482A (en) | Detecting method for high-resolution SAR (Synthetic Aperture Radar) image object | |
CN103942577A (en) | Identity identification method based on self-established sample library and composite characters in video monitoring | |
CN101464948A (en) | Object identification method for affine constant moment based on key point | |
CN103903013A (en) | Optimization algorithm of unmarked flat object recognition | |
CN105574527A (en) | Quick object detection method based on local feature learning | |
CN105893946A (en) | Front face image detection method | |
CN102902967A (en) | Method for positioning iris and pupil based on eye structure classification | |
CN103886589A (en) | Goal-oriented automatic high-precision edge extraction method | |
CN104821010A (en) | Binocular-vision-based real-time extraction method and system for three-dimensional hand information | |
CN110599463B (en) | Tongue image detection and positioning algorithm based on lightweight cascade neural network | |
CN105938551A (en) | Video data-based face specific region extraction method | |
CN106980825A (en) | A kind of face posture sorting technique based on normalization pixel difference feature | |
CN101957993A (en) | Adaptive infrared small object detection method | |
CN113850221A (en) | Attitude tracking method based on key point screening |
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 | ||
TR01 | Transfer of patent right | ||
TR01 | Transfer of patent right |
Effective date of registration: 20210623 Address after: 066006 1st floor, software center, No.2 Weishui Road, Qinhuangdao Economic and Technological Development Zone, Hebei Province Patentee after: Hebei yandayan Soft Information System Co.,Ltd. Address before: 066004 No. 438 west section of Hebei Avenue, seaport District, Hebei, Qinhuangdao Patentee before: Yanshan University |