CN105913008A - Crowd exceptional event detection method based on hypothesis examination - Google Patents
Crowd exceptional event detection method based on hypothesis examination Download PDFInfo
- Publication number
- CN105913008A CN105913008A CN201610213811.9A CN201610213811A CN105913008A CN 105913008 A CN105913008 A CN 105913008A CN 201610213811 A CN201610213811 A CN 201610213811A CN 105913008 A CN105913008 A CN 105913008A
- Authority
- CN
- China
- Prior art keywords
- frame
- sample set
- video
- crowd
- value
- 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.)
- Granted
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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a <{EN0}>crowd exceptional event detection method based on hypothesis examination, particularly targeting the crowd of medium density. The crowd exceptional event detection method comprises steps of using an optical flow method to capture crowd motion information so as to obtain crowd motion characteristic descriptor, extracting a characteristic vector of each pixel point from each video frame, adopting a hypothesis examination statistics method to perform classification based on motion characteristic descriptor, and performing comparison on a statistical magnitude value and a threshold value which are obtained by calculation of the hypothesis examination model so as to detect whether the exceptional event happens. The crowd exceptional event detection method not only eliminates the tedious pretreatment step in the feature extraction phase, but also uses the hypothesis examination model in the detection step, which greatly reduces the time complexity and calculation complexity during the detection process on the basis that the detection result is not reduced. The crowd exceptional event detection method not only can be used for global exceptional event detection, but also can be applied to local exceptional event detection.
Description
Technical field
The present invention relates to a kind of method of crowd's accident detection under public place, particularly to one
Kind based on the assumption that the method for crowd's accident detection of inspection.
Background technology
In recent years, security issues become increasingly urgent in public place, and the requirement to intelligent monitoring the most more comes
The highest, therefore, Video Analysis Technology has become the focus of research in many countries, along with video divides
Going deep into and systematization of analysis technical research, increasing problem also shows especially out.One of them weight
The crowd to public domain carries out effective monitoring to want problem to be how.To this end, each intelligent monitoring
Association area develops various technology.Wherein, crowd's accident detection based on video
Technology development is particularly swift and violent, and this technology can find to monitor the anomalous event in region in time, improves
The response of relevant departments and rescue efficiency, thus effectively reduce public's person and the loss of property.
For crowd's accident detection problem, it is broadly divided into two basic problems, i.e. elementary events
Represent and the foundation of accident detection model.Wherein elementary event represents and is broadly divided into based on low layer
The representations of events of visual signature and representations of events based on high-level semantics features.Special based on Low Level Vision
The representations of events levied is mainly some features that can simply extract of manual extraction, and these features are not necessarily
Having strong physical significance, its method mainly has: light stream, shade of gray, objective contour etc..
Representations of events based on high-level semantics features mainly needs to carry out data the process of complexity, its thing
Reason meaning is obvious.Common high-level semantics features is target space-time track etc..Wherein anomalous event inspection
Survey the foundation of model, mainly have: based on the accident detection model classified and cluster, based on pushing away
Disconnected accident detection model, accident detection model based on energy and based on reconstruct different
Ordinary affair part detection model etc..
Although the method for crowd's accident detection is varied, but most crowd's anomalous event inspection
Method of determining and calculating can not process in real time, and main difficult point is that the calculating that model inspection is time-consuming and higher is multiple
Miscellaneous degree.
Summary of the invention
For the problems referred to above, the invention discloses a kind of based on the assumption that crowd's anomalous event of inspection is examined
The method surveyed, the method, in feature extraction phases, still uses light stream to describe son as target travel,
In the abnormality detection stage, use the method event detection as crowd's exception of hypothesis testing.With biography
System method is compared, and the method does not only has preferable testing result, its time complexity and calculating complexity
Degree also has substantial degradation.
The technical scheme steps that the present invention solves the employing of its technical problem is as follows:
By the present invention in that by optical flow method, capture the movable information of crowd, thus obtain the motion of crowd
Feature Descriptor;For each frame of video, extract the characteristic vector of each pixel, use and assume inspection
The statistical method tested carries out describing the classification of son based on motion feature;Then it will be assumed that testing model calculates
The statistics value obtained compares with threshold value, thus whether detects the generation of anomalous event, is suitable for
Need the crowd's scene processed in real time.
The present invention is directed to based on the crowd's accident detection under global scene specifically comprising the following steps that
Step 1: pretreatment;
First from video flowing, decoding obtains frame of video, then each frame of video is carried out gaussian filtering,
Its concrete operations are: by each pixel in a template scan video frame, determine neighborhood by template
The weighted average gray value of interior pixel, and by this weighted average gray value alternate template central pixel point
Value;
Step 2: feature extraction;
Using the frame of video after step 1 gaussian filtering as input, use dense optical flow method
Horn-Schunck calculates the light flow valuve of adjacent two interframe in video flowing pixel-by-pixel to every two field picture;
Step 3: model is set up;
The method using hypothesis testing sets up model;
Step 4: the detection of anomalous event;
By the detection model using step 3 to set up, the anomalous event under global scene is detected;
This detection model is mainly each frame of video as the sample set detected, with putting down of multiple normal frame
All light flow valuve is as training sample set, calculates the studentization pole between training sample set and test sample collection
Difference statistic q value, and q value is made comparisons with threshold value, such as larger than threshold value, i.e. assume to be false, then
Can determine whether out that this test video frame is abnormal frame, otherwise, it is assumed that set up, then can determine whether out that this test regards
Frequently frame is normal frame.
The method using hypothesis testing described in step 3 sets up model, specific as follows:
3-1. process of data preprocessing
According to the result of the every two field picture light flow valuve calculated in step 2, as the sample needing detection
This collection, wherein, a sample set is made up of smooth flow valuves all in single-frame images;But due to single frames
In image, the quantity of pixel is more, the most just introduces the concept of integrogram, thus when reducing calculating
Between;Its formula is as follows:
I'(x, y)=I (x, y)+I'(x-1, y)+I'(x, y-1)-I'(x-1, y-1)
I'(x in formula, is y) that (x, the result after y) place calculates integrogram, (x y) is single-frame images to I at point
In point (x, y) the light flow valuve at place;
3-2. sets up hypothesis testing model
The general principle of hypothesis testing method is: the average assuming first that between sample set is equal,
Then, according to the maxima and minima numerology biochemistry extreme difference Distribution Value of average in sample set, and with
This compares with known level of signifiance selected threshold, thus judges to assume whether set up, if setting up,
Then think between sample set be do not have differentiated, if being false, then it is assumed that be differentiated, go forward side by side one
Step judges which sample set and other sample sets have difference;Use the sample of known normal event
This collection is made comparisons with test sample collection, as long as drawing and being unsatisfactory for supposing between two sample sets, can sentence
Breaking, this test sample be abnormal;Implement process as follows:
3-2-1, assumed condition H0: μ1=μ2;
μ1For the average of training sample set, μ2Average for test sample;The present invention mainly chooses multiple
The average of the light flow valuve of normal video frame is as training sample set, by single-frame images in test video stream
Light flow valuve is as test sample collection;
3-2-2, the studentization extreme difference distribution statistics amount calculated between training sample set and test sample collection;
μ in formuladnewAnd μtnewFor the light flow valuve of test sample collection and training sample set after data prediction
Average, snewFor the variance of test sample collection after data prediction and training sample set variance and
Square root;Wherein, m is the number of sample set, and value is 2;N is sample point in sample set
Number, C, for repairing inclined constant, is taken as 15;
3-2-3, the choosing of threshold value: for known level of signifiance α and the known free degree
(m, m (n-1)), obtains q according to studentization extreme difference distribution tableαThe value of (m, m (n-1)) is as threshold value;
3-2-4, judge criterion: if calculated studentization extreme difference distribution statistics amount q is more than or equal to threshold
Value qα(m, m (n-1)), then it is assumed that assumed condition H0It is false, i.e. variant between sample set, with this
Just can determine whether that this test video frame is abnormal frame, otherwise, if q is less than threshold value qα(m, m (n-1)) then thinks
Assumed condition H0Set up, i.e. think that this test video frame is normal frame.
The present invention is directed to based on the crowd's accident detection under the scene of local specifically comprising the following steps that
Step 1: pretreatment;
From video flowing, decoding obtains frame of video, then each frame of video is carried out gaussian filtering, its tool
Gymnastics is made: by each pixel in a template scan video frame, determines picture in neighborhood by template
The weighted average gray value of element, and by the value of this weighted average gray value alternate template central pixel point;
Step 2: feature extraction;
Using the frame of video after step 1 gaussian filtering as input, use dense optical flow method
Horn-Schunck calculates the light flow valuve of adjacent two interframe in video flowing pixel-by-pixel to every two field picture;
Step 3: model is set up;
First frame of video is divided into the window of m1 × n1 size, then, according to calculated knot
Really, the light flow valuve extracting one group of m1 × n1 area size carries out statistical analysis to draw its distribution curve,
Thus find the equal Normal Distribution of light flow valuve of each window, use hypothesis testing the most equally
Method set up model, detailed process is as follows:
(1) assumed condition H0: μ1=μ2;
μ1For the average of training sample set, μ2Average for test sample;The present invention mainly chooses multiple
The average of the light flow valuve of the window of multiple m1 × n1 sizes of normal video frame is as training sample
Collection, using the light flow valuve of the window of each m1 × n1 size of single-frame images in test video stream as
Test sample collection;
(2) the studentization extreme difference distribution statistics amount between sample set is calculated;
In this place owing to using the calculated feature interpretation of dense optical flow method in step 2 in the reason stage
Sub-noise jamming is relatively big, and again using less m1 × n1 region as the target area of detection, this is just
Make the studentization extreme difference Distribution Value calculated not accurate enough, therefore, to the light stream in m1 × n1 region
Value carries out repairing operation partially, and computing formula is as follows:
X in formulanewFor the light flow valuve after optimizing in the window of m1 × n1 size, xwindowFor just
The light flow valuve of the window of the m1 × n1 size begun,For initial m1 × n1 size
The average light flow valuve of window, swindowLight flow valuve for the window of initial m1 × n1 size
Standard variance;
Then the window of the m1 × n1 size after optimizing is calculated training window sample set and survey
The studentization extreme difference Distribution Value of examination window sample set, computing formula is as follows:
μ in formuladnewAnd μtnewTest sample collection after optimizing for light flow valuve and training sample set equal
Value, snewAfter optimizing for light flow valuve the variance of variance and the training sample set of test sample collection and flat
Root;Wherein, m is the number of sample set, and n is the number of sample point in sample set, and C is for repairing partially
Constant, is taken as 15;
(3) the choosing of threshold value: for known level of signifiance α and the known free degree (m, m (n-1)),
Q can be obtained according to studentization extreme difference distribution tableαThe value of (m, m (n-1)) is as threshold value, and wherein m is sample
This number, n is the number of sample point in each sample;
(4) criterion is passed judgment on: if calculated studentization extreme difference Distribution Value q is more than or equal to threshold value
qα(m, m (n-1)), then it is assumed that assumed condition H0It is false, i.e. variant between sample set, with this just
Can determine whether that this test video frame is abnormal frame, otherwise, if q is less than threshold value qα(m, m (n-1)) then thinks false
If condition H0Set up, i.e. think that this test video frame is normal frame;
Step 4: the detection of anomalous event;
By the detection model using step 3 to set up, the anomalous event under the scene of local is detected;
This detection model is mainly the window of certain m1 × n1 size of frame of video as the sample detected
Collection, by the average light flow valuve of the windowd of multiple m1 × n1 sizes of multiple normal frame as training
Sample set, calculates the studentization extreme difference statistic q value between training sample set and test sample collection, and
Q value is made comparisons with threshold value, such as larger than threshold value, i.e. assume to be false, then can determine whether out that this test regards
Frequently frame is abnormal frame, otherwise, it is assumed that sets up, then can determine whether out that this test video frame is normal frame.
Beneficial effects of the present invention:
1. the present invention is in feature extraction phases, simply original frame of video has been carried out simple denoising,
Then carrying out the calculating of light flow valuve, this well reduces the complexity of calculating, and retains well
The movable information of crowd, this is particularly suitable in middle-high density crowd.
2. the present invention sets up and detection-phase, the method employing hypothesis testing at model.The method pin
To feature extraction phases, the noise that the calculating of optical flow method causes has good robustness, and the party
Method is not under conditions of reducing testing result, and its time complexity has compared to traditional method very well
Optimization.
Accompanying drawing explanation
Fig. 1 is the flow chart of crowd's accident detection in the case of global abnormal.
Fig. 2 is the image after frame of video carries out light flow valuve calculating in the case of global abnormal.
Fig. 3 is the testing result of frame of video in the case of global abnormal.
Fig. 4 is the flow chart of crowd's accident detection in the case of local anomaly.
Fig. 5 is the image after frame of video carries out light flow valuve calculating in the case of local anomaly.
Fig. 6 is the testing result of frame of video in the case of local anomaly.
Detailed description of the invention
Below in conjunction with the accompanying drawings, specific embodiments of the present invention are described in further detail.
Be illustrated in figure 1 based on the assumption that the method for the crowd's accident detection checked, for based on
Crowd's accident detection under global scene, its concrete steps describe as shown in Figure 1:
Step 1: pretreatment.
First from video flowing, decoding obtains frame of video, then each frame of video is carried out gaussian filtering,
Its concrete operations are: by each pixel in a template scan video frame, determine neighbour by template
The weighted average gray value of pixel in territory, and with this weighted average gray value alternate template center pixel
The value of point.
Described template, or claim convolution, mask, be size be the matrix of 0 and the 1 of N*N;
Step 2: feature extraction.
Using the frame of video after step 1 gaussian filtering as input, use dense optical flow method
Horn-Schunck calculates the light flow valuve of adjacent two interframe in video flowing pixel-by-pixel to every two field picture.Its
Result is as shown in Figure 2.
Step 3: model is set up.
Which frame of video in crowd's accident detection under global scene, mainly detection video set
For there being anomalous event to occur.Therefore, if using the method for hypothesis testing to be modeled, then need to sentence
The disconnected sample set being made up of the light flow valuve of single-frame images whether Normal Distribution.Count for step 2
The result obtained, draws its distribution curve to the light flow valuve size of every two field picture and carries out statistical analysis,
It appeared that the equal Normal Distribution of light flow valuve of every two field picture, hypothesis testing therefore can be used
Method sets up model.The present invention uses the method for hypothesis testing that detection target is set up model, its tool
Body process is as follows:
3-1. process of data preprocessing
According to the result of the every two field picture light flow valuve calculated in step 2, as needing detection
Sample set, wherein, a sample set is made up of smooth flow valuves all in single-frame images.But due to
In single-frame images, the quantity of pixel is more, the most just introduces the concept of integrogram, thus reduces
The calculating time.Its formula is as follows:
I'(x, y)=I (x, y)+I'(x-1, y)+I'(x, y-1)-I'(x-1, y-1)
I'(x in formula, is y) that (x, the result after y) place calculates integrogram, (x y) is single-frame images to I at point
In point (x, y) the light flow valuve at place.
3-2. sets up hypothesis testing model
According to the statistic analysis result of step 3, it meets the precondition of hypothesis testing method,
But owing to data are pre-processed, therefore pretreated data are further analyzed,
After same drafting pretreatment, the distribution curve of data carries out statistical analysis, it appeared that pretreated
Data set still Normal Distribution, therefore can use the method for hypothesis testing as anomalous event
Detection model.
The general principle of hypothesis testing method is: the average assuming first that between sample set is equal,
Then, according to the maxima and minima numerology biochemistry extreme difference Distribution Value of average in sample set, and
Compare with known level of signifiance selected threshold with this, thus judge to assume whether set up, if becoming
Vertical, then it is assumed that between sample set be do not have differentiated, if being false, then it is assumed that be differentiated,
And judge which sample set and other sample sets have difference further.But in the present invention,
We use the sample set of known normal event to make comparisons with test sample collection, therefore, as long as
Go out and be unsatisfactory for supposing between two sample sets, may determine that this test sample is abnormal.Specifically
Realize process as follows:
(1) assumed condition H0: μ1=μ2;
μ1For the average of training sample set, μ2Average for test sample.The present invention mainly chooses many
The average of the light flow valuve of individual normal video frame is as training sample set, by single frames figure in test video stream
The light flow valuve of picture is as test sample collection.
(2) the studentization extreme difference distribution statistics amount between training sample set and test sample collection is calculated;
μ in formuladnewAnd μtnewFor test sample collection after data prediction and the light stream of training sample set
The average of value, snewVariance for the variance of test sample collection after data prediction Yu training sample set
Root sum square.Wherein, m is the number of sample set, and value is 2;N is sample in sample set
The number of point, C, for repairing inclined constant, is taken as 15.
(3) the choosing of threshold value: for known level of signifiance α and the known free degree
(m, m (n-1)), obtains q according to studentization extreme difference distribution tableαThe value of (m, m (n-1)) is as threshold value;
(4) criterion is passed judgment on: if calculated studentization extreme difference distribution statistics amount q is more than or equal to threshold
Value qα(m, m (n-1)), then it is assumed that assumed condition H0It is false, i.e. variant between sample set, with this
Just can determine whether that this test video frame is abnormal frame, otherwise, if q is less than threshold value qα(m, m (n-1)) then recognizes
For assumed condition H0Set up, i.e. think that this test video frame is normal frame.
Step 4: the detection of anomalous event.
By the detection model using step 3 to set up, the anomalous event under global scene is detected.
This detection model is mainly each frame of video as the sample set detected, by multiple normal frame
Average light flow valuve, as training sample set, calculates the student between training sample set and test sample collection
Change extreme difference statistic q value, and q value is made comparisons with threshold value, such as larger than threshold value, i.e. can not assume
Vertical, then can determine whether out that this test video frame is abnormal frame, otherwise, it is assumed that set up, then can determine whether out
This test video frame is normal frame.About the testing result under global scene as shown in Figure 3.
Be illustrated in figure 4 based on the assumption that the method for the crowd's accident detection checked, for based on
The locally crowd's accident detection under scene, its concrete steps describe as shown in Figure 4:
Step 1: pretreatment.
From video flowing, decoding obtains frame of video, then each frame of video is carried out gaussian filtering, its
Concrete operations are: by each pixel in a template scan video frame, determine neighborhood by template
The weighted average gray value of interior pixel, and by this weighted average gray value alternate template central pixel point
Value.
Described template, or claim convolution, mask, be size be the matrix of 0 and the 1 of N*N;
Step 2: feature extraction.
Using the frame of video after step 1 gaussian filtering as input, use dense optical flow method
Horn-Schunck calculates the light flow valuve of adjacent two interframe in video flowing pixel-by-pixel to every two field picture, its
Result is as shown in Figure 5.
Step 3: model is set up.
Which region in detection based on local anomaly event, mainly detection frame of video is different for having
Ordinary affair part occurs.Therefore, if using the method for hypothesis testing to be modeled, then need to judge by list
The sample set whether Normal Distribution of the light flow valuve composition in certain region of two field picture.First will regard
Frequently frame is divided into the window of m1 × n1 size, then, according to the calculated result of step 2, takes out
The light flow valuve taking one group of m1 × n1 area size carries out statistical analysis to draw its distribution curve, permissible
Find the equal Normal Distribution of light flow valuve of each window, therefore can use the side of hypothesis testing
Method sets up model.
The present invention uses the method for hypothesis testing that detection target is set up model, and its detailed process is as follows:
(1) assumed condition H0: μ1=μ2;
μ1For the average of training sample set, μ2Average for test sample.The present invention mainly chooses many
The average of the light flow valuve of the window of multiple m1 × n1 sizes of individual normal video frame is as training sample
This collection, by the light flow valuve of the window of each m1 × n1 size of single-frame images in test video stream
As test sample collection.
(2) the studentization extreme difference distribution statistics amount between sample set is calculated;
In this place owing to using the calculated feature of dense optical flow method in step 2 to retouch in the reason stage
State sub-noise jamming relatively big, and again using less m1 × n1 region as the target area of detection, this
The studentization extreme difference Distribution Value allowing for calculating is not accurate enough, therefore, to m1 × n1 district in the present invention
Light flow valuve in territory carries out repairing operation partially, and computing formula is as follows:
X in formulanewFor the light flow valuve after optimizing in the window of m1 × n1 size, xwindowFor just
The light flow valuve of the window of the m1 × n1 size begun,For initial m1 × n1 size
The average light flow valuve of window, swindowLight flow valuve for the window of initial m1 × n1 size
Standard variance.
Then the window of the m1 × n1 size after optimizing is calculated training window sample set and survey
The studentization extreme difference Distribution Value of examination window sample set, computing formula is as follows:
μ in formuladnewAnd μtnewTest sample collection after optimizing for light flow valuve and training sample set equal
Value, snewAfter optimizing for light flow valuve the variance of variance and the training sample set of test sample collection and
Square root.Wherein, m is the number of sample set, and n is the number of sample point in sample set, and C is for repairing
Constant, is taken as 15 partially.
(3) the choosing of threshold value: for known level of signifiance α and the known free degree
(m, m (n-1)), can obtain q according to studentization extreme difference distribution tableαThe value of (m, m (n-1)) is as threshold
Value, wherein m is the number of sample, and n is the number of sample point in each sample.
(4) criterion is passed judgment on: if calculated studentization extreme difference Distribution Value q is more than or equal to threshold value
qα(m, m (n-1)), then it is assumed that assumed condition H0It is false, i.e. variant between sample set, with this just
Can determine whether that this test video frame is abnormal frame, otherwise, if q is less than threshold value qα(m, m (n-1)) then thinks
Assumed condition H0Set up, i.e. think that this test video frame is normal frame.
Step 4: the detection of anomalous event.
By the detection model using step 3 to set up, the anomalous event under the scene of local is detected.
This detection model is mainly the window of certain m1 × n1 size of frame of video as the sample detected
This collection, by the average light flow valuve of the windowd of multiple m1 × n1 sizes of multiple normal frame as instruction
Practice sample set, calculate the studentization extreme difference statistic q value between training sample set and test sample collection,
And q value is made comparisons with threshold value, such as larger than threshold value, i.e. assume to be false, then can determine whether out this survey
Examination frame of video is abnormal frame, otherwise, it is assumed that sets up, then can determine whether out that this test video frame is normal
Frame.About the testing result under the scene of local as shown in Figure 6.
Claims (5)
1. based on the assumption that inspection crowd's accident detection method, it is characterised in that by use light stream
Method, the movable information of capture crowd, thus the motion feature obtaining crowd describes son;Regard for each
Frequently frame, extracts the characteristic vector of each pixel, uses the statistical method of hypothesis testing to carry out based on fortune
The classification of dynamic Feature Descriptor;Then it will be assumed that testing model calculated statistics value does with threshold value
Relatively, thus whether detect the generation of anomalous event, be suitable for the crowd's scene needing to process in real time.
The most according to claim 1 based on the assumption that inspection crowd's accident detection method, its
It is characterised by specifically comprising the following steps that for based on the crowd's accident detection under global scene
Step 1: pretreatment
First from video flowing, decoding obtains frame of video, then each frame of video is carried out gaussian filtering,
Its concrete operations are: by each pixel in a template scan video frame, determine neighborhood by template
The weighted average gray value of interior pixel, and by this weighted average gray value alternate template central pixel point
Value;
Step 2: feature extraction;
Using the frame of video after step 1 gaussian filtering as input, use dense optical flow method Horn-Schunck
Every two field picture is calculated pixel-by-pixel the light flow valuve of adjacent two interframe in video flowing;
Step 3: model is set up;
The method using hypothesis testing sets up model;
Step 4: the detection of anomalous event;
By the detection model using step 3 to set up, the anomalous event under global scene is detected;
This detection model is mainly each frame of video as the sample set detected, with putting down of multiple normal frame
All light flow valuve is as training sample set, calculates the studentization pole between training sample set and test sample collection
Difference statistic q value, and q value is made comparisons with threshold value, such as larger than threshold value, i.e. assume to be false, then
Can determine whether out that this test video frame is abnormal frame, otherwise, it is assumed that set up, then can determine whether out that this test regards
Frequently frame is normal frame.
The most according to claim 2 based on the assumption that inspection crowd's accident detection method, its
Be characterised by the template described in step 1, or claim convolution, mask, be size be the 0 and 1 of N*N
Matrix.
The most according to claim 2 based on the assumption that inspection crowd's accident detection method, its
It is characterised by that the method using hypothesis testing described in step 3 sets up model, specific as follows:
3-1. process of data preprocessing
According to the result of the every two field picture light flow valuve calculated in step 2, as the sample needing detection
This collection, wherein, a sample set is made up of smooth flow valuves all in single-frame images;But due to single frames
In image, the quantity of pixel is more, the most just introduces the concept of integrogram, thus when reducing calculating
Between;Its formula is as follows:
I'(x, y)=I (x, y)+I'(x-1, y)+I'(x, y-1)-I'(x-1, y-1)
I'(x in formula, is y) that (x, the result after y) place calculates integrogram, (x is y) that single-frame images exists to I at point
Point (x, y) the light flow valuve at place;
3-2. sets up hypothesis testing model
The general principle of hypothesis testing method is: the average assuming first that between sample set is equal,
Then, according to the maxima and minima numerology biochemistry extreme difference Distribution Value of average in sample set, and with
This compares with known level of signifiance selected threshold, thus judges to assume whether set up, if setting up,
Then think between sample set be do not have differentiated, if being false, then it is assumed that be differentiated, go forward side by side one
Step judges which sample set and other sample sets have difference;Use the sample of known normal event
This collection is made comparisons with test sample collection, as long as drawing and being unsatisfactory for supposing between two sample sets, can sentence
Breaking, this test sample be abnormal;Implement process as follows:
3-2-1, assumed condition H0: μ1=μ2;
μ1For the average of training sample set, μ2Average for test sample;The present invention mainly chooses multiple
The average of the light flow valuve of normal video frame is as training sample set, by single-frame images in test video stream
Light flow valuve is as test sample collection;
3-2-2, the studentization extreme difference distribution statistics amount calculated between training sample set and test sample collection;
μ in formuladnewAnd μtnewFor the light flow valuve of test sample collection and training sample set after data prediction
Average, snewFor the variance of test sample collection after data prediction and training sample set variance and
Square root;Wherein, m is the number of sample set, and value is 2;N is the individual of sample point in sample set
Number, C, for repairing inclined constant, is taken as 15;
3-2-3, the choosing of threshold value: for known level of signifiance α and the known free degree (m, m (n-1)),
Q is obtained according to studentization extreme difference distribution tableαThe value of (m, m (n-1)) is as threshold value;
3-2-4, judge criterion: if calculated studentization extreme difference distribution statistics amount q is more than or equal to threshold value
qα(m, m (n-1)), then it is assumed that assumed condition H0It is false, i.e. variant between sample set, just may be used with this
Judge that this test video frame is abnormal frame, otherwise, if q is less than threshold value qα(m, m (n-1)) then thinks hypothesis
Condition H0Set up, i.e. think that this test video frame is normal frame.
The most according to claim 1 based on the assumption that inspection crowd's accident detection method, its
It is characterised by specifically comprising the following steps that for based on the crowd's accident detection under the scene of local
Step 1, from video flowing decoding obtain frame of video, then each frame of video is carried out gaussian filtering,
Its concrete operations are: by each pixel in a template scan video frame, determine neighborhood by template
The weighted average gray value of interior pixel, and by this weighted average gray value alternate template central pixel point
Value;
Step 2: feature extraction;
Using the frame of video after step 1 gaussian filtering as input, use dense optical flow method Horn-Schunck
Every two field picture is calculated pixel-by-pixel the light flow valuve of adjacent two interframe in video flowing;
Step 3: model is set up;
First frame of video is divided into the window of m1 × n1 size, then, is calculated according to step 2
Result, the light flow valuve extracting one group of m1 × n1 area size carries out statistical to draw its distribution curve
Analysis, thus find the equal Normal Distribution of light flow valuve of each window, therefore same use assumes inspection
The method tested sets up model, and detailed process is as follows:
(1) assumed condition H0: μ1=μ2;
μ1For the average of training sample set, μ2Average for test sample;The present invention mainly chooses multiple
The average of the light flow valuve of the window of multiple m1 × n1 sizes of normal video frame as training sample set,
Using the light flow valuve of the window of each m1 × n1 size of single-frame images in test video stream as test
Sample set;
(2) the studentization extreme difference distribution statistics amount between sample set is calculated;
In this place owing to using the calculated feature interpretation of dense optical flow method in step 2 in the reason stage
Sub-noise jamming is relatively big, and again using less m1 × n1 region as the target area of detection, this just makes
The studentization extreme difference Distribution Value that must calculate is not accurate enough, therefore, enters the light flow valuve in m1 × n1 region
Row repaiies operation partially, and computing formula is as follows:
X in formulanewFor the light flow valuve after optimizing in the window of m1 × n1 size, xwindowFor initially
The light flow valuve of window of m1 × n1 size,Window for initial m1 × n1 size
Average light flow valuve, swindowThe standard side of light flow valuve for the window of initial m1 × n1 size
Difference;
Then the window of the m1 × n1 size after optimizing is calculated training window sample set and test
The studentization extreme difference Distribution Value of window sample set, computing formula is as follows:
μ in formuladnewAnd μtnewTest sample collection after optimizing for light flow valuve and the average of training sample set,
snewAfter optimizing for light flow valuve the variance of variance and the training sample set of test sample collection and square
Root;Wherein, m is the number of sample set, and n is the number of sample point in sample set, C for repairing inclined constant,
It is taken as 15;
(3) the choosing of threshold value: for known level of signifiance α and the known free degree (m, m (n-1)),
Q can be obtained according to studentization extreme difference distribution tableαThe value of (m, m (n-1)) is as threshold value, and wherein m is sample
This number, n is the number of sample point in each sample;
(4) criterion is passed judgment on: if calculated studentization extreme difference Distribution Value q is more than or equal to threshold value
qα(m, m (n-1)), then it is assumed that assumed condition H0It is false, i.e. variant between sample set, just may be used with this
Judge that this test video frame is abnormal frame, otherwise, if q is less than threshold value qα(m, m (n-1)) then thinks hypothesis
Condition H0Set up, i.e. think that this test video frame is normal frame;
Step 4: the detection of anomalous event;
By the detection model using step 3 to set up, the anomalous event under the scene of local is detected;
This detection model is mainly the window of certain m1 × n1 size of frame of video as the sample detected
Collection, by the average light flow valuve of the windowd of multiple m1 × n1 sizes of multiple normal frame as training sample
This collection, calculates the studentization extreme difference statistic q value between training sample set and test sample collection, and by q
Value is made comparisons with threshold value, such as larger than threshold value, i.e. assumes to be false, then can determine whether out this test video frame
For abnormal frame, otherwise, it is assumed that establishment, then can determine whether out that this test video frame is normal frame.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610213811.9A CN105913008B (en) | 2016-04-07 | 2016-04-07 | Based on the assumption that the crowd's accident detection method examined |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610213811.9A CN105913008B (en) | 2016-04-07 | 2016-04-07 | Based on the assumption that the crowd's accident detection method examined |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105913008A true CN105913008A (en) | 2016-08-31 |
CN105913008B CN105913008B (en) | 2019-03-01 |
Family
ID=56744703
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610213811.9A Active CN105913008B (en) | 2016-04-07 | 2016-04-07 | Based on the assumption that the crowd's accident detection method examined |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105913008B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106951574A (en) * | 2017-05-03 | 2017-07-14 | 牡丹江医学院 | A kind of information processing system and method based on computer network |
CN108881952A (en) * | 2018-07-02 | 2018-11-23 | 上海商汤智能科技有限公司 | Video generation method and device, electronic equipment and storage medium |
CN109255660A (en) * | 2018-09-25 | 2019-01-22 | 科达集团股份有限公司技术分公司 | A kind of advertising accounts optimization method using the unbalanced data of height |
CN113051971A (en) * | 2019-12-27 | 2021-06-29 | 沈阳新松机器人自动化股份有限公司 | Method and system for estimating number of people in abnormal emergency of group |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103413321A (en) * | 2013-07-16 | 2013-11-27 | 南京师范大学 | Crowd behavior model analysis and abnormal behavior detection method under geographical environment |
CN103854027A (en) * | 2013-10-23 | 2014-06-11 | 北京邮电大学 | Crowd behavior identification method |
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 |
CN105389567A (en) * | 2015-11-16 | 2016-03-09 | 上海交通大学 | Group anomaly detection method based on a dense optical flow histogram |
-
2016
- 2016-04-07 CN CN201610213811.9A patent/CN105913008B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103413321A (en) * | 2013-07-16 | 2013-11-27 | 南京师范大学 | Crowd behavior model analysis and abnormal behavior detection method under geographical environment |
CN103854027A (en) * | 2013-10-23 | 2014-06-11 | 北京邮电大学 | Crowd behavior identification method |
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 |
CN105389567A (en) * | 2015-11-16 | 2016-03-09 | 上海交通大学 | Group anomaly detection method based on a dense optical flow histogram |
Non-Patent Citations (2)
Title |
---|
吴新宇等: "基于视频的人群异常事件监测综述", 《电子测量与仪器学报》 * |
赖作镁等: "背景运动补偿和假设检验的目标检测算法", 《光学精密工程》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106951574A (en) * | 2017-05-03 | 2017-07-14 | 牡丹江医学院 | A kind of information processing system and method based on computer network |
CN106951574B (en) * | 2017-05-03 | 2019-06-14 | 牡丹江医学院 | A kind of information processing system and method based on computer network |
CN108881952A (en) * | 2018-07-02 | 2018-11-23 | 上海商汤智能科技有限公司 | Video generation method and device, electronic equipment and storage medium |
CN109255660A (en) * | 2018-09-25 | 2019-01-22 | 科达集团股份有限公司技术分公司 | A kind of advertising accounts optimization method using the unbalanced data of height |
CN109255660B (en) * | 2018-09-25 | 2021-09-21 | 浙文互联集团股份有限公司技术分公司 | Advertisement account optimization method using highly unbalanced data |
CN113051971A (en) * | 2019-12-27 | 2021-06-29 | 沈阳新松机器人自动化股份有限公司 | Method and system for estimating number of people in abnormal emergency of group |
Also Published As
Publication number | Publication date |
---|---|
CN105913008B (en) | 2019-03-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105744232B (en) | A kind of method of the transmission line of electricity video external force damage prevention of Behavior-based control analytical technology | |
CN103839065B (en) | Extraction method for dynamic crowd gathering characteristics | |
CN104268528B (en) | A kind of crowd massing method for detecting area and device | |
CN110956094A (en) | RGB-D multi-mode fusion personnel detection method based on asymmetric double-current network | |
CN110059694A (en) | The intelligent identification Method of lteral data under power industry complex scene | |
CN106250936A (en) | Multiple features multithreading safety check contraband automatic identifying method based on machine learning | |
CN105260749B (en) | Real-time target detection method based on direction gradient binary pattern and soft cascade SVM | |
CN103871077B (en) | A kind of extraction method of key frame in road vehicles monitoring video | |
CN105913008A (en) | Crowd exceptional event detection method based on hypothesis examination | |
CN110633678B (en) | Quick and efficient vehicle flow calculation method based on video image | |
CN111062967B (en) | Electric power business hall passenger flow statistical method and system based on target dynamic tracking | |
CN108804987B (en) | Door opening and closing state detection method and device and people flow detection system | |
CN114333070A (en) | Examinee abnormal behavior detection method based on deep learning | |
CN106709438A (en) | Method for collecting statistics of number of people based on video conference | |
CN104504377A (en) | Bus passenger crowding degree identification system and method | |
CN102332094A (en) | Semi-supervised online study face detection method | |
CN103945089A (en) | Dynamic target detection method based on brightness flicker correction and IP camera | |
CN112084928A (en) | Road traffic accident detection method based on visual attention mechanism and ConvLSTM network | |
CN106127124A (en) | The automatic testing method of the abnormal image signal in region, taxi front row | |
CN111027440B (en) | Crowd abnormal behavior detection device and detection method based on neural network | |
CN109711313B (en) | Real-time video monitoring method for identifying sewage poured into river channel | |
CN103366163B (en) | Face detection system and method based on incremental learning | |
CN102625028B (en) | The method and apparatus that static logos present in video is detected | |
CN116977917A (en) | Infrared image pedestrian detection method | |
CN116596895A (en) | Substation equipment image defect identification method and system |
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 | ||
EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20160831 Assignee: HANGZHOU ZHISHU TECHNOLOGY Co.,Ltd. Assignor: HANGZHOU DIANZI University Contract record no.: X2020330000124 Denomination of invention: Detection method of crowd abnormal events based on hypothesis testing Granted publication date: 20190301 License type: Common License Record date: 20201226 |
|
EE01 | Entry into force of recordation of patent licensing contract |