CN105426811B - A kind of crowd's abnormal behaviour and crowd density recognition methods - Google Patents
A kind of crowd's abnormal behaviour and crowd density recognition methods Download PDFInfo
- Publication number
- CN105426811B CN105426811B CN201510640197.XA CN201510640197A CN105426811B CN 105426811 B CN105426811 B CN 105426811B CN 201510640197 A CN201510640197 A CN 201510640197A CN 105426811 B CN105426811 B CN 105426811B
- Authority
- CN
- China
- Prior art keywords
- crowd
- frame
- vector
- video
- pixel
- 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
- 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
Abstract
The present invention discloses a kind of crowd's abnormal behaviour and crowd density recognition methods, comprising: calculates the displacement vector of the i-th frame of video and each pixel of i+1 interframe to be detected;It extracts moving region prospect and carries out binary conversion treatment, the moving region foreground image after obtaining binaryzation;The shift length of each pixel of moving region foreground image after calculating binaryzation rejects and is displaced lesser pixel;For remaining pixel, the second-order differential in the direction x and the direction y is sought, obtains space characteristics point;It calculates using each space characteristics point as the average displacement vector along the x-axis direction and along the y-axis direction of the certain specification neighborhood of core;For i+1 frame and the i-th+2 frame video, the step of repeating front;Calculate the i-th frame to i+1 inter-frame video vector sum i+1 frame to the i-th+2 inter-frame video vector difference, and carry out statistics generate global characteristics vector;By the input support vector machines training of global characteristics vector, classification based training is carried out to crowd's exception and crowd density respectively.
Description
Technical field
The present invention relates to technical field of computer vision, and in particular to a kind of crowd's abnormal behaviour and crowd density identification side
Method.
Background technique
Computer vision technique is a kind of technology studied and how to simulate biological vision using computer and its relevant device.
Picture or video are acquired by imaging devices such as video cameras and is handled, and obtain the three-dimensional information of corresponding scene, then transfer to count
Calculation machine replaces brain to complete processing and understand.The technology is related to multiple subjects, including image procossing, pattern-recognition, image analysis
With image understanding etc..Currently, computer vision technique is widely used to every field, as medical image processing, video monitoring,
Electronics bayonet, virtual reality, intelligent transportation etc..
As the development and monitoring instrument of computer vision technique are in the general of the public places such as subway station, airport
And crowd's abnormal behaviour and crowd density identification technology are successfully realized.But relevant theory and product are still seldom,
Current technology mainly pass through the sampling index comprehensive evaluations such as particle moving direction and particle movement speed crowd it is abnormal and
Crowd density.
Crowd density is evaluated, concrete scheme includes:
1. detecting the characteristic point that change in time and space is violent in crowd's video using three-dimensional Hessian matrix;
2. extracting the dynamic texture of characteristic point with space-time local binarization mode method, then spectrum point is carried out to dynamic texture
Analysis is to obtain the local feature of image;
3. analyzing using local feature of the statistics with histogram to entire image, crowd's global characteristics vector is obtained, so
Obtained crowd's global characteristics vector input support vector machines is estimated to the global density grade of crowd afterwards;
4. local feature is carried out to be processed and translated into color space, the chromatogram of crowd local density distribution can be obtained
Display.
For crowd's anomaly evaluation, concrete scheme includes:
1. being sampled for various sizes of video using the grid of different size, each sampling particle is considered as one
Individual calculates the weighting interreaction force between particle in conjunction with local density's information;
2. the local feature (weighting interreaction force) at pair each sampled point carries out statistics with histogram, by force direction point
For N number of section, size is divided into M section, and statistics generates the global characteristics vector of N × M dimension, and input support vector machines training carries out
Tagsort;
3. the classification results of pair support vector machines carry out median filtering to remove the trip point in classification results.
Although above scheme is able to carry out the identification of crowd's abnormal behaviour and crowd density, but due to by extraneous complex environment
Interference, Crowds Distribute is uneven, the jump of video interframe is big etc. be unfavorable for identification factor limitation, there are more defects, packet
It includes:
1. prospect and background can not be extracted, can only exist very in the information caused to entire video frame uniform sampling
More invalid informations, so that algorithm whole efficiency is not high.
2. for crowd density being calculated using two kinds of incoherent algorithms extremely with crowd, need to carry out two thus
Wheel machine learning training: the first round is that learning training is carried out to crowd density, and the crowd density recognition training of this wheel must be complete
At and obtain preferable achievement, can just carry out the training of second wheel crowd's anomalous identification in this way, whole flow process is caused to compare
Cumbersome, practicability is poor in the actual environment, and due to more using algorithm, operand is larger, causes cpu load high, calculates
Time-consuming for method.
3. proposing " pressure model " based on " social force model " theory when calculating crowd's exception, essence is that crowd is close
Degree is merged with crowd's abnormal behaviour, and the judgement of the higher crowd's abnormal behaviour of crowd density is more sensitive, and opposite crowd density is got over
The judgement more sensitivity of low crowd behaviour exception is lower, broadly combines crowd density and crowd extremely increase erroneous detection in this way
Rate (FN) and omission factor (FP).
Summary of the invention
Present invention aim to address the defects of the prior art, provide crowd's abnormal behaviour of a kind of high efficiency and accuracy
With crowd density recognition methods, the technical solution adopted is as follows:
A kind of crowd's abnormal behaviour and crowd density recognition methods, comprising:
S1. the displacement vector of the i-th frame of video and each pixel of i+1 interframe to be detected is calculated;
S2. it extracts moving region prospect and carries out binary conversion treatment, the moving region foreground image after obtaining binaryzation;
S3. the shift length for calculating each pixel of the moving region foreground image after binaryzation, it is minimum to reject displacement
Pixel;
S4. for remaining pixel, the second-order differential in the direction x and the direction y is sought, obtains space characteristics point;
S5. along the x-axis direction and along the y-axis direction flat using each space characteristics point as the certain specification neighborhood of core is calculated
Equal displacement vector;
S6. for i+1 frame and the i-th+2 frame video, step S1 to S5 is repeated;
S7. calculate the i-th frame to i+1 inter-frame video displacement vector and i+1 frame to the i-th+2 inter-frame video displacement swear
The difference of amount, and carry out statistics and generate global characteristics vector;
S8. by the input support vector machines training of global characteristics vector, classify respectively to crowd's exception and crowd density
Training, to be identified while realization to crowd density and crowd's exception.
Preferably, calculating the i-th frame of video to be detected in the step S1 using optical flow method and i+1 interframe being each
The displacement vector of pixel.
Preferably, extracting moving region prospect using interframe difference method in the step S2.
Preferably, in the step S5, calculate using each space characteristics point as 3 × 3 neighborhoods of core along the x-axis direction
Average displacement vector along the y-axis direction.
Preferably, in the step S7, using histogram to the vector sum i+1 frame of the i-th frame to i+1 inter-frame video
Difference to the vector of the i-th+2 inter-frame video is counted.
Preferably, being anomaly divided into normal and abnormal two classes as a result, being divided into crowd density to crowd in the step S8
Extremely low, basic, normal, high and high five classes result.
Compared with prior art, beneficial effects of the present invention: the present invention has carried out moving region foreground extraction, eliminates big
Measure garbage and interference;It carries out the abnormal calculating with crowd density of crowd simultaneously by using a kind of algorithm, greatly reduces
Operand improves efficiency, and does not interfere between each other;By by the direction of displacement of each sampled point and displacement two
Parameter is converted into is compared continuous two inter-frame video picture displacements phasor difference again, i.e., 2-D data is converted to a dimension
According to so that the robustness of algorithm is more preferable.Traditional technology counts two features of direction of displacement and displacement of sampled point,
9 different directions are divided into, the 7 different velocity magnitude shift values in each direction point cause the separator bar in each direction attached
Closely the error as caused by tiny angle change is larger, and traditional technology is the displacement method for normalizing to each direction
It is divided into 7 grades, and displacement is divided into 50 sections by this method, and is the absolute value as unit of pixel, so that system
Count it is more accurate, and eliminate by normalization cause similitude.
Detailed description of the invention
Fig. 1 is flow chart of the present invention;
Fig. 2 is the present embodiment experiment figure;
Fig. 3 is to carry out the image that binary conversion treatment obtains to the moving region prospect of extraction;
Fig. 4 is the second-order differential for asking pixel the direction x and the direction y, the schematic diagram of obtained space characteristics point;
Fig. 5 be calculate the i-th frame to i+1 inter-frame video vector sum i+1 frame to the i-th+2 inter-frame video vector
Difference, and the image counted using feature histogram.
Specific embodiment
Present invention is further described in detail with reference to the accompanying drawings and examples.
Embodiment: as shown in Figure 1, being handled using method of the invention Fig. 2, comprising:
S1. the displacement vector of the i-th frame of video and each pixel of i+1 interframe to be detected is calculated;
S2. it extracts moving region prospect and carries out binary conversion treatment, the moving region foreground image after obtaining binaryzation;
S3. the shift length for calculating each pixel of the moving region foreground image after binaryzation, it is minimum to reject displacement
Pixel;
S4. for remaining pixel, the second-order differential in the direction x and the direction y is sought, obtains space characteristics point;
S5. along the x-axis direction and along the y-axis direction flat using each space characteristics point as the certain specification neighborhood of core is calculated
Equal displacement vector;
S6. for i+1 frame and the i-th+2 frame video, step S1 to S5 is repeated;
S7. calculate the i-th frame to i+1 inter-frame video displacement vector and i+1 frame to the i-th+2 inter-frame video displacement swear
The difference of amount, and carry out statistics and generate global characteristics vector;
S8. by the input support vector machines training of global characteristics vector, classify respectively to crowd's exception and crowd density
Training, to be identified while realization to crowd density and crowd's exception.
In the step S1, the position of the i-th frame of video and each pixel of i+1 interframe to be detected is calculated using optical flow method
Move vector.
In the step S2, moving region prospect is extracted using interframe difference method.
In the step S5, calculating using each space characteristics point is 3 × 3 neighborhoods of core along the x-axis direction and along y-axis side
To average displacement vector.
In the step S7, using histogram to the vector sum i+1 frame of the i-th frame to i+1 inter-frame video to the i-th+2 frame
Between the difference of vector of video counted, as shown in Figure 5.
In the step S8, normal and abnormal two classes are anomaly divided into crowd as a result, to crowd density be divided into it is extremely low, low,
Middle and high and high five classes result.
The present embodiment has carried out moving region foreground extraction, eliminates a large amount of garbages and interference;By using one kind
Algorithm carries out the abnormal calculating with crowd density of crowd simultaneously, greatly reduces operand, improves efficiency, and do not have between each other
There is interference;By converting two parameters of the direction of displacement of each sampled point and displacement to continuous two inter-frame video figures
Image displacement phasor difference is compared again, i.e., 2-D data is converted to one-dimensional data, so that the robustness of algorithm is more preferable.
Claims (6)
1. a kind of crowd's abnormal behaviour and crowd density recognition methods characterized by comprising
S1. the displacement vector of the i-th frame of video and each pixel of i+1 interframe to be detected is calculated;
S2. it extracts moving region prospect and carries out binary conversion treatment, the moving region foreground image after obtaining binaryzation;
S3. the shift length for calculating each pixel of the moving region foreground image after binaryzation, rejects and is displaced the smallest picture
Vegetarian refreshments;
S4. for remaining pixel, the second-order differential in the direction x and the direction y is sought, obtains space characteristics point;
S5. it calculates using each space characteristics point as the average bit along the x-axis direction and along the y-axis direction of the certain specification neighborhood of core
Move vector;
S6. for i+1 frame and the i-th+2 frame video, step S1 to S5 is repeated;
S7. calculate the i-th frame to i+1 inter-frame video displacement vector and i+1 frame to i-th+2 inter-frame video displacement vector
Difference, and carry out statistics and generate global characteristics vector;
S8. by the input support vector machines training of global characteristics vector, classification based training is carried out to crowd's exception and crowd density respectively,
To be identified while realization to crowd density and crowd's exception.
2. a kind of crowd's abnormal behaviour according to claim 1 and crowd density recognition methods, which is characterized in that the step
In rapid S1, the displacement vector of the i-th frame of video and each pixel of i+1 interframe to be detected is calculated using optical flow method.
3. a kind of crowd's abnormal behaviour according to claim 1 and crowd density recognition methods, which is characterized in that the step
In rapid S2, moving region prospect is extracted using interframe difference method.
4. a kind of crowd's abnormal behaviour according to claim 1 and crowd density recognition methods, which is characterized in that the step
In rapid S5, calculates and sweared by the average displacement along the x-axis direction and along the y-axis direction of 3 × 3 neighborhoods of core of each space characteristics point
Amount.
5. a kind of crowd's abnormal behaviour according to claim 1 and crowd density recognition methods, which is characterized in that the step
In rapid S7, using histogram to the vector of the vector sum i+1 frame of the i-th frame to i+1 inter-frame video to the i-th+2 inter-frame video
Difference is counted.
6. a kind of crowd's abnormal behaviour according to claim 1 and crowd density recognition methods, which is characterized in that the step
In rapid S8, normal and abnormal two classes are anomaly divided into as a result, being divided into extremely low, basic, normal, high and high five class to crowd density to crowd
As a result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510640197.XA CN105426811B (en) | 2015-09-28 | 2015-09-28 | A kind of crowd's abnormal behaviour and crowd density recognition methods |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510640197.XA CN105426811B (en) | 2015-09-28 | 2015-09-28 | A kind of crowd's abnormal behaviour and crowd density recognition methods |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105426811A CN105426811A (en) | 2016-03-23 |
CN105426811B true CN105426811B (en) | 2019-03-15 |
Family
ID=55505011
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510640197.XA Active CN105426811B (en) | 2015-09-28 | 2015-09-28 | A kind of crowd's abnormal behaviour and crowd density recognition methods |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105426811B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110796035A (en) * | 2019-10-14 | 2020-02-14 | 上海复瞰科技有限公司 | People entering and exiting counting method based on human shape detection and speed calculation |
CN111079536B (en) * | 2019-11-18 | 2023-08-29 | 高新兴科技集团股份有限公司 | Behavior analysis method, storage medium and device based on human body key point time sequence |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101159859A (en) * | 2007-11-29 | 2008-04-09 | 北京中星微电子有限公司 | Motion detection method, device and an intelligent monitoring system |
CN102324016A (en) * | 2011-05-27 | 2012-01-18 | 郝红卫 | Statistical method for high-density crowd flow |
CN103793920A (en) * | 2012-10-26 | 2014-05-14 | 杭州海康威视数字技术股份有限公司 | Retro-gradation detection method based on video and system thereof |
CN104123544A (en) * | 2014-07-23 | 2014-10-29 | 通号通信信息集团有限公司 | Video analysis based abnormal behavior detection method and system |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP6112801B2 (en) * | 2012-08-22 | 2017-04-12 | キヤノン株式会社 | Image recognition apparatus and image recognition method |
-
2015
- 2015-09-28 CN CN201510640197.XA patent/CN105426811B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101159859A (en) * | 2007-11-29 | 2008-04-09 | 北京中星微电子有限公司 | Motion detection method, device and an intelligent monitoring system |
CN102324016A (en) * | 2011-05-27 | 2012-01-18 | 郝红卫 | Statistical method for high-density crowd flow |
CN103793920A (en) * | 2012-10-26 | 2014-05-14 | 杭州海康威视数字技术股份有限公司 | Retro-gradation detection method based on video and system thereof |
CN104123544A (en) * | 2014-07-23 | 2014-10-29 | 通号通信信息集团有限公司 | Video analysis based abnormal behavior detection method and system |
Non-Patent Citations (1)
Title |
---|
"基于视频的人群数量统计及异常检测方法研究";陈禹;《中国优秀硕士学位论文全文数据库 信息科技辑》;20150715(第7期);I138-1352页正文第7页第3-5段,第21页第2段,第25页第1段,第37页第2-6段,第38页第1段,39页第1-3段,6段,图4.6 |
Also Published As
Publication number | Publication date |
---|---|
CN105426811A (en) | 2016-03-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109670441B (en) | Method, system, terminal and computer readable storage medium for realizing wearing recognition of safety helmet | |
CN104217419B (en) | Human body detection device and method and human body counting device and method | |
WO2020253308A1 (en) | Human-machine interaction behavior security monitoring and forewarning method for underground belt transportation-related personnel | |
US9025875B2 (en) | People counting device, people counting method and people counting program | |
CN104361611B (en) | Group sparsity robust PCA-based moving object detecting method | |
CN104240264B (en) | The height detection method and device of a kind of moving object | |
CN104835175B (en) | Object detection method in a kind of nuclear environment of view-based access control model attention mechanism | |
CN104504377B (en) | A kind of passenger on public transport degree of crowding identifying system and method | |
CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
CN104166841A (en) | Rapid detection identification method for specified pedestrian or vehicle in video monitoring network | |
CN107967440A (en) | A kind of monitor video method for detecting abnormality based on multizone mutative scale 3D-HOF | |
CN105404857A (en) | Infrared-based night intelligent vehicle front pedestrian detection method | |
CN101431664A (en) | Automatic detection method and system for intensity of passenger flow based on video image | |
CN107657244A (en) | A kind of human body tumble behavioral value system and its detection method based on multiple-camera | |
CN104781848A (en) | Image monitoring apparatus for estimating gradient of singleton, and method therefor | |
CN106815578A (en) | A kind of gesture identification method based on Depth Motion figure Scale invariant features transform | |
CN109145696B (en) | Old people falling detection method and system based on deep learning | |
CN112926522B (en) | Behavior recognition method based on skeleton gesture and space-time diagram convolution network | |
CN105740751A (en) | Object detection and identification method and system | |
CN103413154A (en) | Human motion identification method based on normalized class Google measurement matrix | |
CN112801037A (en) | Face tampering detection method based on continuous inter-frame difference | |
CN105426811B (en) | A kind of crowd's abnormal behaviour and crowd density recognition methods | |
CN105809092A (en) | Population target detection method and device thereof | |
CN104123569B (en) | Video person number information statistics method based on supervised learning | |
CN104718560B (en) | Image monitoring apparatus for estimating size of singleton, and method therefor |
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 |