CN110263676A - A method of for generating high quality crowd density figure - Google Patents
A method of for generating high quality crowd density figure Download PDFInfo
- Publication number
- CN110263676A CN110263676A CN201910475281.9A CN201910475281A CN110263676A CN 110263676 A CN110263676 A CN 110263676A CN 201910475281 A CN201910475281 A CN 201910475281A CN 110263676 A CN110263676 A CN 110263676A
- Authority
- CN
- China
- Prior art keywords
- closeness
- characteristic pattern
- picture
- network
- probability
- 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
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of methods for generating high quality crowd density figure, comprising: collected picture is input to global closeness grade separation network, obtains corresponding global closeness Probability Characteristics figure；Collected picture is input to high dimensional feature figure estimation network, obtains high dimensional feature figure；Collected picture is input to local concentration grade separation network, obtains corresponding local concentration Probability Characteristics figure；Three above figure is merged, fusion feature figure is obtained；The fusion feature figure is passed through into convolution sum deconvolution processing respectively, obtains high-resolution characteristic pattern, the process of convolution for being then 1*1 using a convolution kernel size obtains high quality crowd density figure.The accuracy of crowd density figure and resolution ratio obtained using this programme is higher, is better able to meet the demand nowadays to accuracy rate in crowd's scene analysis.
Description
Technical field
The present invention relates to crowd's scene analysis technical fields under video monitoring, in particular to a kind of for generating high quality people
The method of group's density map.
Background technique
With the development of computer vision technique, monitoring camera is more and more used in people's life.Effective people
Group's scene analysis is most important to guarantee public safety.And crowd density figure is essential one in crowd's scene analysis
Ring, existing some people counting methods based on convolutional neural networks obtain very high accurate in the scene of high density
Rate, but picture middle-high density region and density regions are all over-evaluated or underestimated to these methods.
Therefore, how to obtain accurate, high quality density map is technical problem urgently to be solved.
Summary of the invention
The purpose of the present invention is: a kind of method for generating high quality crowd density figure is proposed, to meet nowadays to people
The demand of accuracy rate in group's scene analysis.
The technical solution adopted by the present invention to solve the technical problems is:
A method of for generating high quality crowd density figure, include the following steps:
S1, collected picture is input to global closeness grade separation network, it is general obtains corresponding global closeness
Rate distribution characteristics figure；
S2, collected picture is input to high dimensional feature figure estimation network, obtains high dimensional feature figure；
S3, collected picture is input to local concentration grade separation network, it is general obtains corresponding local concentration
Rate distribution characteristics figure；
S4, using image channel blending algorithm, merge the global closeness probability distribution graph achieved above, the height
Dimensional feature figure and the local concentration Probability Characteristics figure, obtain fusion feature figure；
S5, the fusion feature figure is passed through to convolution sum deconvolution processing respectively, obtains high-resolution characteristic pattern, then
The process of convolution for being 1*1 using a convolution kernel size, is mapped to final high quality for the high-resolution characteristic pattern
Crowd density figure.
Technical solution is advanced optimized, the obtaining step of the overall situation closeness Probability Characteristics figure is as follows:
S11, first according to image data concentrate number scale, by image data concentrate picture closeness be divided into
5 grades: very intensive, intensive, medium, sparse and very sparse；
S12, picture to be processed is inputted into the global closeness grade separation network, obtains picture to be processed and is belonging respectively to
The probability value of 5 grades；
One S13, creation empty 5* (W/8) * (H/8) characteristic pattern, wherein W and H is the width and height of picture to be processed, then
Sorter network is obtained to obtain global closeness Probability Characteristics figure in the characteristic pattern of 5 probability value filling creations.
As technical solution is advanced optimized, the overall situation closeness grade separation network is based on a VGG16 convolution
Neural network structure, and keep 13 layers of convolution in VGG16 convolutional neural networks structure constant, and modify its three layers of full articulamentums
Port number be respectively 512,256 and 5.
Technical solution is advanced optimized, the specific obtaining step of the high dimensional feature figure is as follows:
S21, high dimensional feature figure estimation network is built, high dimensional feature figure estimation network includes the first branching networks and the
Two branching networks；First branching networks remove three layers of full articulamentum by VGG16 and obtain；Second branching networks are one
Three-layer coil accumulates network structure, and its convolution kernel size is 5*5；
S22, collected picture is separately input into first branching networks and the second branching networks, to obtain phase
The characteristic pattern answered；
S23, the characteristic pattern that two branching networks obtain is merged, obtains high dimensional feature figure.
Technical solution is advanced optimized, the specific obtaining step of the local concentration Probability Characteristics figure is as follows:
S31, first according to image data concentrate number scale, by image data concentrate picture closeness be divided into
Five grades: very intensive, intensive, medium, sparse and very sparse；
S32, using the sliding window of a 64x64 on collected picture interception image, and the image that will be truncated to
It is input in a local concentration grade separation network, to obtain the probability that the video in window closeness belongs to five grades
Value；The local concentration grade separation network is made of 5 layers of convolutional layer and 3 layers of full articulamentum；
S33, the characteristic pattern for creating an empty 5*W*H, wherein W and H is the width and height for inputting picture, then in characteristic pattern
On find position corresponding with the sliding window width, and insert the window closeness class probability value；
S34, the characteristic pattern size that will be filled with closeness class probability value are adjusted to 5* (W/8) * (H/8), and then obtain institute
State local concentration Probability Characteristics figure.
The beneficial effects of the present invention are: the accuracy of crowd density figure and resolution ratio using this programme acquisition are higher, more
The demand nowadays to accuracy rate in crowd's scene analysis can be met.
Detailed description of the invention
Fig. 1 is general flow chart of the invention.
Fig. 2 is the acquisition flow chart of global closeness grade separation network.
Fig. 3 is the acquisition flow chart of high dimensional feature figure estimation network.
Fig. 4 is the acquisition flow chart of local concentration grade separation network.
Specific embodiment
Below in conjunction with attached drawing.The present invention will be further described.
Implementation process of the invention is based primarily upon three network structures, is respectively as follows: global density collection grade separation network, height
Dimensional feature figure estimates network and local concentration grade separation network.
Wherein, global closeness grade separation network is based on a VGG16 convolutional neural networks structure, and holding
13 layers of convolution in VGG16 convolutional neural networks structure are constant, and the port number for modifying its three layers full articulamentum be respectively 512,
256 and 5.
Wherein, high dimensional feature figure estimation network includes the first branching networks and the second branching networks；First branched network
Network removes three layers of full articulamentum by VGG16 and obtains；Second branching networks are that a three-layer coil accumulates network structure, and its convolution
Core size is 5*5.
Wherein, local concentration grade separation network is made of 5 layers of convolutional layer and 3 layers of full articulamentum.
Main-process stream of the invention is as shown in Figure 1, a kind of method for generating high quality crowd density figure, including walks as follows
It is rapid:
S1, collected picture is input to global closeness grade separation network, it is general obtains corresponding global closeness
Rate distribution characteristics figure；
S2, collected picture is input to high dimensional feature figure estimation network, obtains high dimensional feature figure；
S3, collected picture is input to local concentration grade separation network, it is general obtains corresponding local concentration
Rate distribution characteristics figure；
S4, using image channel blending algorithm, merge the global closeness probability distribution graph achieved above, the height
Dimensional feature figure and the local concentration Probability Characteristics figure, obtain fusion feature figure；
S5, the fusion feature figure is passed through to convolution sum deconvolution processing respectively, obtains high-resolution characteristic pattern, then
The process of convolution for being 1*1 using a convolution kernel size, is mapped to final high quality for the high-resolution characteristic pattern
Crowd density figure.
Wherein, the specific obtaining step of global closeness Probability Characteristics figure is as shown in Fig. 2, include the following:
S11, first according to image data concentrate number scale, by image data concentrate picture closeness be divided into
5 grades: very intensive, intensive, medium, sparse and very sparse；
S12, picture to be processed is inputted into global closeness grade separation network, obtains picture to be processed and is belonging respectively to 5
The probability value of grade；
One S13, creation empty 5* (W/8) * (H/8) characteristic pattern, wherein W and H is the width and height of picture to be processed, then
Sorter network is obtained to obtain global closeness Probability Characteristics figure in the characteristic pattern of 5 probability value filling creations.
Wherein, the specific obtaining step of high dimensional feature figure is as shown in figure 3, include the following:
S21, high dimensional feature figure estimation network is built, high dimensional feature figure estimates that network includes the first branching networks and second point
Branch network；First branching networks remove three layers of full articulamentum by VGG16 and obtain；Second branching networks are one three layers
Convolutional network structure, and its convolution kernel size is 5*5；
S22, collected picture is separately input into first branching networks and the second branching networks, to obtain phase
The characteristic pattern answered；
S23, the characteristic pattern that two branching networks obtain is merged, obtains high dimensional feature figure.
Wherein, the specific obtaining step of local concentration Probability Characteristics figure is as shown in figure 4, include the following:
S31, first according to image data concentrate number scale, by image data concentrate picture closeness be divided into
Five grades: very intensive, intensive, medium, sparse and very sparse；
S32, using the sliding window of a 64x64 on collected picture interception image, and the image that will be truncated to
It is input in a local concentration grade separation network, to obtain the probability that the video in window closeness belongs to five grades
Value；Local concentration grade separation network is made of 5 layers of convolutional layer and 3 layers of full articulamentum；
S33, the characteristic pattern for creating an empty 5*W*H, wherein W and H is the width and height for inputting picture, then in characteristic pattern
On find position corresponding with sliding window width, and insert the window closeness class probability value；
S34, the characteristic pattern size that will be filled with closeness class probability value are adjusted to 5* (W/8) * (H/8), and then acquisition office
Portion's closeness Probability Characteristics figure.
The advantages of basic principles and main features and this programme of this programme have been shown and described above.The technology of the industry
Personnel are it should be appreciated that this programme is not restricted to the described embodiments, and the above embodiments and description only describe this
The principle of scheme, under the premise of not departing from this programme spirit and scope, this programme be will also have various changes and improvements, these changes
Change and improvement is both fallen within the scope of claimed this programme.This programme be claimed range by appended claims and its
Equivalent thereof.
Claims (5)
1. a kind of method for generating high quality crowd density figure, which comprises the steps of:
S1, collected picture is input to global closeness grade separation network, obtains corresponding global closeness probability point
Cloth characteristic pattern；
S2, collected picture is input to high dimensional feature figure estimation network, obtains high dimensional feature figure；
S3, collected picture is input to local concentration grade separation network, obtains corresponding local concentration probability point
Cloth characteristic pattern；
S4, using image channel blending algorithm, it is special to merge the global closeness probability distribution graph achieved above, the higher-dimension
Sign figure and the local concentration Probability Characteristics figure, obtain fusion feature figure；
S5, the fusion feature figure is passed through to convolution sum deconvolution processing respectively, high-resolution characteristic pattern is obtained, then passes through again
The process of convolution that a convolution kernel size is 1*1 is crossed, the high-resolution characteristic pattern is mapped to final high quality crowd
Density map.
2. a kind of method for generating high quality crowd density figure as shown in claim 1, which is characterized in that the overall situation
The obtaining step of closeness Probability Characteristics figure is as follows:
S11, first according to image data concentrate number scale, by image data concentrate picture closeness be divided into 5
Grade: very intensive, intensive, medium, sparse and very sparse；
S12, picture to be processed is inputted into the global closeness grade separation network, obtains picture to be processed and is belonging respectively to 5
The probability value of grade；
One S13, creation empty 5* (W/8) * (H/8) characteristic pattern, wherein W and H is the width and height of picture to be processed, then will be divided
Class network obtains obtaining global closeness Probability Characteristics figure in the characteristic pattern of 5 probability value filling creations.
3. a kind of method for generating high quality crowd density figure as shown in claim 2, which is characterized in that the overall situation
Closeness grade separation network is based on a VGG16 convolutional neural networks structure, and holding VGG16 convolutional neural networks structure
In 13 layers of convolution it is constant, and the port number for modifying its three layers full articulamentum is respectively 512,256 and 5.
4. a kind of method for generating high quality crowd density figure as shown in claim 1, which is characterized in that the higher-dimension
The specific obtaining step of characteristic pattern is as follows:
S21, high dimensional feature figure estimation network is built, the high dimensional feature figure estimation network includes the first branching networks and second point
Branch network；First branching networks remove three layers of full articulamentum by VGG16 and obtain；Second branching networks are one three layers
Convolutional network structure, and its convolution kernel size is 5*5；
S22, collected picture is separately input into first branching networks and the second branching networks, to obtain corresponding
Characteristic pattern；
S23, the characteristic pattern that two branching networks obtain is merged, obtains high dimensional feature figure.
5. a kind of method for generating high quality crowd density figure as shown in claim 1, which is characterized in that the part
The specific obtaining step of closeness Probability Characteristics figure is as follows:
S31, first according to image data concentrate number scale, by image data concentrate picture closeness be divided into five
Grade: very intensive, intensive, medium, sparse and very sparse；
S32, using the sliding window of a 64x64 on collected picture interception image, and by the image being truncated to input
Into a local concentration grade separation network, to obtain the probability value that the video in window closeness belongs to five grades；
The local concentration grade separation network is made of 5 layers of convolutional layer and 3 layers of full articulamentum；
S33, the characteristic pattern for creating an empty 5*W*H, wherein W and H is the width and height for inputting picture, is then looked on characteristic pattern
To position corresponding with the sliding window width, and insert the window closeness class probability value；
S34, the characteristic pattern size that will be filled with closeness class probability value are adjusted to 5* (W/8) * (H/8), and then obtain the office
Portion's closeness Probability Characteristics figure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910475281.9A CN110263676A (en) | 2019-06-03 | 2019-06-03 | A method of for generating high quality crowd density figure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910475281.9A CN110263676A (en) | 2019-06-03 | 2019-06-03 | A method of for generating high quality crowd density figure |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110263676A true CN110263676A (en) | 2019-09-20 |
Family
ID=67916592
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910475281.9A Pending CN110263676A (en) | 2019-06-03 | 2019-06-03 | A method of for generating high quality crowd density figure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110263676A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110807409A (en) * | 2019-10-30 | 2020-02-18 | 上海眼控科技股份有限公司 | Crowd density detection model training method and crowd density detection method |
| CN111310805A (en) * | 2020-01-22 | 2020-06-19 | 中能国际建筑投资集团有限公司 | Method, device and medium for predicting density of target in image |
-
2019
- 2019-06-03 CN CN201910475281.9A patent/CN110263676A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110807409A (en) * | 2019-10-30 | 2020-02-18 | 上海眼控科技股份有限公司 | Crowd density detection model training method and crowd density detection method |
| CN111310805A (en) * | 2020-01-22 | 2020-06-19 | 中能国际建筑投资集团有限公司 | Method, device and medium for predicting density of target in image |
| CN111310805B (en) * | 2020-01-22 | 2023-05-30 | 中能国际高新科技研究院有限公司 | Method, device and medium for predicting density of target in image |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107480772B (en) | License plate super-resolution processing method and system based on deep learning | |
| US10958854B2 (en) | Computer-implemented method for generating an output video from multiple video sources | |
| CN110008832A (en) | Based on deep learning character image automatic division method, information data processing terminal | |
| Zhao et al. | Pyramid global context network for image dehazing | |
| CN110324626A (en) | A kind of video coding-decoding method of the dual code stream face resolution ratio fidelity of internet of things oriented monitoring | |
| CN102360498A (en) | Reconstruction method for image super-resolution | |
| CN106952286A (en) | Dynamic background Target Segmentation method based on motion notable figure and light stream vector analysis | |
| CN102034267A (en) | Three-dimensional reconstruction method of target based on attention | |
| CN108376247A (en) | Strategic coarse-fine combined sea-land separation method applied to optical remote sensing ship detection | |
| CN112528913A (en) | Grit particulate matter particle size detection analytic system based on image | |
| CN106548494A (en) | A kind of video image depth extraction method based on scene Sample Storehouse | |
| CN110263676A (en) | A method of for generating high quality crowd density figure | |
| CN106447718B (en) | A kind of 2D turns 3D depth estimation method | |
| CN114820579A (en) | Semantic segmentation based image composite defect detection method and system | |
| CN103871089B (en) | Image superpixel meshing method based on fusion | |
| CN115984747A (en) | Video saliency target detection method based on dynamic filter | |
| Qiao et al. | A novel segmentation based depth map up-sampling | |
| CN111105354A (en) | Depth image super-resolution method and device based on multi-source depth residual error network | |
| CN111104532B (en) | RGBD image joint recovery method based on double-flow network | |
| CN106157251A (en) | A kind of face super-resolution method based on Cauchy's regularization | |
| CN114913095B (en) | Depth deblurring method based on domain adaptation | |
| CN111626944A (en) | Video deblurring method based on space-time pyramid network and natural prior resistance | |
| CN103578121A (en) | Motion detection method based on shared Gaussian model in disturbed motion environment | |
| CN101276476A (en) | Process for the separating prospect background of 2D cartoon animation | |
| CN103530879B (en) | Pedestrian's color extraction method under special scenes |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190920 |
|
| WD01 | Invention patent application deemed withdrawn after publication |