CN110163122A - A kind of crowded crowd's method for detecting abnormality and system based on semi-supervised dictionary learning - Google Patents
A kind of crowded crowd's method for detecting abnormality and system based on semi-supervised dictionary learning Download PDFInfo
- Publication number
- CN110163122A CN110163122A CN201910362661.1A CN201910362661A CN110163122A CN 110163122 A CN110163122 A CN 110163122A CN 201910362661 A CN201910362661 A CN 201910362661A CN 110163122 A CN110163122 A CN 110163122A
- Authority
- CN
- China
- Prior art keywords
- event
- normal
- anomalous
- mode
- video
- 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
- 238000000034 method Methods 0.000 title claims abstract description 33
- 230000005856 abnormality Effects 0.000 title claims abstract description 16
- 230000002547 anomalous effect Effects 0.000 claims abstract description 74
- 238000001514 detection method Methods 0.000 claims abstract description 35
- 238000012360 testing method Methods 0.000 claims abstract description 29
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 17
- 239000000203 mixture Substances 0.000 claims abstract description 9
- 238000012544 monitoring process Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 13
- 230000002159 abnormal effect Effects 0.000 claims description 8
- 238000004590 computer program Methods 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 5
- 238000011840 criminal investigation Methods 0.000 abstract description 2
- 238000012986 modification Methods 0.000 abstract 3
- 230000004048 modification Effects 0.000 abstract 3
- 238000000605 extraction Methods 0.000 abstract 1
- 238000002474 experimental method Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 241001269238 Data Species 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 208000003464 asthenopia Diseases 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/44—Event detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of crowded crowd's method for detecting abnormality and system based on semi-supervised dictionary learning, mainly solve the problem of that traditional accident detection algorithm does not model anomalous event in test process and with progress model modification over time.Implementation step is: (1) Video Events feature extraction；(2) building of null hypothesis and alternative hypothesis；(3) normal video event schema learns；(4) anomalous event pattern learning；(5) accident detection；(6) model modification.The present invention explores the composition of anomalous video event in test process, and the event schema acquired is made to be more suitable for test video.Model modification strategy proposed by the present invention keeps learnt Video Events model more accurate, improves the precision of abnormality detection, can be used for the fields such as intelligent monitoring, traffic control and criminal investigation auxiliary.
Description
Technical field
The invention belongs to technical field of information processing, in particular to accident detection technology under a kind of crowd scene,
It can be used for the fields such as public safety intelligent monitoring, traffic control and criminal investigation auxiliary.
Background technique
With the promotion that people's public safety is realized, more and more monitoring devices are used in security system.Currently, prison
Control system still relies on the judgement of people, not only time-consuming and laborious, but also due to visual fatigue, it is easy to cause the leakage of suspicious event
Inspection.For this purpose, the intelligent monitor system of Computer Automatic Recognition anomalous event is allowed to become the emphasis studied at present.However, abnormal thing
The automatic detection of part is very challenging, because abnormal type is varied and has unpredictability, people are difficult sieve
List all Exception Types being likely to occur under special scenes.
According in monitoring scene moving target number, anomalous event can be divided into anomalous event under sparse scene and gather around
Squeeze the anomalous event under scene.Under sparse scene, moving target is very rare, can be easy to detect and track target, and mention
Effective feature representation is taken, therefore the precision of abnormality detection is higher.Under crowd scene, moving target is very more, each other
Inevitably occur frequently to block, therefore the detection and tracking of target become extremely difficult.Based on this, generally use at present
Video Events are described in feature based on video block.
In recent years, have the algorithm largely about accident detection under crowd scene to be suggested.They are mostly from one group
Only go out normal event schema comprising the training focusing study of normal video event, then find out during the test and these
The unmatched Video Events of normal event mode are determined as anomalous event.Lu et al. document " C.Lu, J.Shi, and J.Jia,
Abnormal Event Detection at 150 FPS in Matlab,in proceedings of IEEE
International Conference on Computer Vision, 2013, one group of study is normal in pp.2720-2727 "
The small test sample of reconstructed error is defined as exception by Video Events mode.But such methods have ignored it is different in test process
The event schema of ordinary affair part, and this has very important effect the accurate differentiation of anomalous event；Meanwhile people are difficult
It includes all normal event modes that training data, which is concentrated,.
Summary of the invention
In order to accurately differentiate that anomalous event, the present invention propose that a kind of crowded crowd based on semi-supervised dictionary learning examines extremely
Survey method constructs the accident detection frame based on semi-supervised learning, and anomalous event is defined as comprising anomalous event mode,
The higher Video Events of anomaly detector score simultaneously, during the test, picking out in training process automatically does not have just
Normal Video Events are updated normal event mode, improve model tormulation precision.
The technical solution of the invention is as follows provides a kind of crowded crowd abnormality detection side based on semi-supervised dictionary learning
Method, comprising the following steps:
Step 1, building data set；
Video data in collection monitoring system constructs data set, and data set is divided into training dataset and test number
According to collection, normal video data is only included in the training set；
Affair character extracts in step 2, data set；
Input video frame in data set is divided into spatially partly overlapping video block by (2a)；
(2b) calculates the gradient and light stream figure of every frame image, and to each video block extract respectively gradient and Optic flow information into
Column hisgram statistics, the local feature as Video Events are expressed；
Step 3, building are it is assumed that described assume to include null hypothesis H0With alternative hypothesis H1；
(3a) null hypothesis H0: normal video event only includes normal event mode, i.e. event y only use normal event mode into
The combination of row sparse linear, y=D α0+ ε, wherein D is normal event mode, α0It is sparse expression coefficient, ε is plant noise；
(3b) alternative hypothesis H1: anomalous video event is made of simultaneously normal and anomalous event mode, i.e. event y is wrapped simultaneously
Containing normal and anomalous event mode, y=D α1+ S β+ε, wherein S is anomalous event mode, α1It is sparse expression coefficient with β；
Step 4, the normal video event concentrated based on training data, learn normal event mode；
(4a) models plant noise using Gaussian mixtures, objective function are as follows:
Wherein, K is the number of gauss component, π0kWithBe respectively in null hypothesis the prior probability of k-th of gauss component and
Variance, I are unit matrixs,It represents with D α0iFor mean value,For the Gaussian Profile of variance；
(4b) optimizes the parameter in formula (1) objective function using expectation-maximization algorithm；Obtain normal event
The objective function of mode and its sparse coefficient:
Wherein,Representative sample yiIt is raw by k-th of gauss component
At probability；
(4c) formula (2) is optimized using iteration more new strategy；
Step 5 learns on the basis of the existing normal event mode that the test sample and step 4 of test data set obtain
Anomalous event mode；
(5a) models plant noise using Gaussian mixtures, objective function are as follows:
Wherein, K is the number of gauss component, π1kWithIt is the prior probability of k-th of gauss component in alternative hypothesis respectively
And variance, I are unit matrixs,It represents with D α1iFor mean value,For the Gaussian Profile of variance；
(5b) optimizes the objective function parameters of formula (5) using expectation-maximization algorithm；
(5c) is carried out down-sampled or is carried out a liter sampling to suspicious sample to normal sample, obtains study anomalous event mode
Objective function:
Wherein,ri=1/p (yi|H0)；
(5d) anomalous event Mode S and expression factor alpha1, β optimized using alternative and iterative algorithm；Specific optimization method
Same step (4c).
Step 6, accident detection；
(6a) constructs anomalous event detector Det, calculates compared to normal event mode is only utilized, introduces anomalous event mould
Promotion of the formula to expression precision, i.e.,Wherein p (y | H0) expression event belongs to the probability of null hypothesis, p (y | H1)
Expression event belongs to the probability of alternative hypothesis；
In (6b) detection process, if in event comprising anomalous event mode and anomalous event detector Det score be greater than it is pre-
If threshold xi, then determine the event for anomalous event；
Step 7, event schema updates；
In (7a) detection process, using comprising anomalous event mode, but anomalous event detector Det score is less than default threshold
The test sample of value ξ is updated normal event mode；
(7b) brings updated normal event mode D into formula (6), is updated to anomalous event mode, and with more
Normal and anomalous event mode detection anomalous event after new.
Further, the step (7a) specifically:
It will include anomalous event mode, but anomalous event detector Det score is less than the collection of the test sample of preset threshold ξ
Conjunction is denoted asNormal mode is updated using following target formula
Wherein, W is diagonal matrix, and i-th of element is w on diagonal linei, D0For current normal event mode.
Further, normal sample is carried out in step (5c) down-sampled: selection p (y | H0) < p0Sample learning it is abnormal
Event schema, p0For parameter preset；
Sampling is carried out liter to suspicious sample: distributing a weight r for each sample, wherein r=1/p (y | H0)。
Further, the local feature using HOG and HOF feature as Video Events is expressed.
Further, step (4c) specifically:
Since each sample only includes limited event schema, coefficient vector α0iBe it is sparse, i.e.,
||α0i||0Calculate α0iThe number of middle nonzero element, t represent the upper limit number of nonzero element.Above formula is using projection ladder
Degree descent method optimizes.About dictionary D, objective function can be write
Wherein, W is diagonal matrix, and i-th of element is w on diagonal linei.Constraint conditionPurpose is word in order to prevent
Allusion quotation element value is excessive.Above formula is square constrained least square problem an of standard, can be carried out by Lagrange duality algorithm
Optimization.
The present invention also provides a kind of crowded crowd's abnormality detection system based on semi-supervised dictionary learning, including processor and
Memory is characterized in that in the memory and stores computer program, when computer program is run in the processor,
Execute above-mentioned crowded crowd's method for detecting abnormality based on semi-supervised dictionary learning.
It the present invention also provides a kind of computer readable storage medium, is characterized in that and stores computer program, count
Calculation machine program, which is performed, realizes above-mentioned crowded crowd's method for detecting abnormality based on semi-supervised dictionary learning.
The beneficial effects of the present invention are:
Composition of the present invention due to considering anomalous video event in test process proposes a kind of new accident detection
Frame.Plant noise is modeled using mixed Gauss model, improves the ability to express of model.During the test,
It adaptively detects the normal video event of less generation, and Video Events model is updated, improve the table of model
Up to precision.
Detailed description of the invention
Fig. 1 is that the present invention is based on the accident detection method flow diagrams of semi-supervised dictionary learning；
Specific embodiment
The step of realizing below in conjunction with drawings and the specific embodiments to the present invention is described in further detail:
Referring to Fig.1, the step of present invention realizes is as follows:
Step 1, collection monitoring system data, construct data set, and data set includes training dataset and test data set, instruction
Practice to concentrate and only includes normal video data.
Affair character extracts in step 2, data set:
Input video frame in data set is divided into spatially partly overlapping video block by (2a).
(2b) calculates the gradient and light stream figure of every frame image, and in each video block respectively to gradient and Optic flow information into
Column hisgram counts to obtain the local feature expression of Video Events；HOG (histogram of is chosen in the present embodiment
) and HOF (histogram of optical flow) feature gradient.
Step 3, building are assumed:
(3a) null hypothesis H0: normal video event only includes normal event mode, i.e. event y can only use normal event mould
Formula carries out sparse linear combination, y=D α0+ε.Wherein, D is normal event mode, α0It is sparse expression coefficient, ε is making an uproar for model
Sound.
(3b) alternative hypothesis H1: anomalous video event is made of simultaneously normal and anomalous event mode, i.e. event y is wrapped simultaneously
Containing normal and anomalous event mode, y=D α1+Sβ+ε.Wherein, S is anomalous event mode, α1It is sparse expression coefficient with β.
Step 4, the normal video event concentrated based on training data learns normal thing using maximum- likelihood estimation
Part mode:
(4a) is in general, plant noise ε is considered sampling from independent identically distributed Gaussian Profile.However, in a practical situation,
The expression and modeling of Video Events are extremely complex, and single Gaussian Profile is difficult to noise accurate modeling.For this purpose, the present invention uses
Gaussian mixtures model noise.Theoretically, when gauss component is enough, Gaussian mixtures can be approached arbitrarily
Continuous density function.Based on this, objective function can be write
Wherein, K is the number of gauss component, π0kWithBe respectively in null hypothesis the prior probability of k-th of gauss component and
Variance, I are unit matrixs.
Parameter in (4b) formula (1) uses expectation maximization (Expectationmaximization, abbreviation EM) algorithm
It optimizes.In general, the maximization of formula (1) can be realized by optimizing its lower bound.About normal event mode and its sparse
The objective function of coefficient can be write
Wherein,Representative sample yi is the probability generated by k-th of gauss component.
Formula (2) is optimized using iteration more new strategy.Wherein, since each sample only includes limited event mould
Formula, therefore coefficient vector α0iBe it is sparse, i.e.,
||α0i||0Calculate α0iThe number of middle nonzero element, t represent the upper limit number of nonzero element.Above formula is using projection ladder
Degree descent method optimizes.About dictionary D, objective function can be write
Wherein, W is diagonal matrix, and i-th of element is w on diagonal linei.Constraint conditionPurpose is word in order to prevent
Allusion quotation element value is excessive.Above formula is square constrained least square problem an of standard, can be carried out by Lagrange duality algorithm
Optimization.
Step 5, learn anomalous event mode.
(5a) learns anomalous event on the basis of the test sample of input test data set and existing normal event mode
Mode models plant noise using Gaussian mixtures, objective function are as follows:
Wherein, K is the number of gauss component, π1kWithIt is the prior probability of k-th of gauss component in alternative hypothesis respectively
And variance, I are unit matrixs,It represents with D α1iFor mean value,For the Gaussian Profile of variance；
Parameter in (5b) formula (5) equally uses EM algorithm to optimize.
(5c) due to the rare characteristic of anomalous video event, be significantly larger than can for the quantity of normal sample in input test sample
The quantity of the exceptional sample of energy.
In order to solve the problems, such as imbalanced training sets, the present invention is handled using two strategies: 1) normal sample drop and is adopted
Sample, and selection p (y | H0) < p0Sample learning anomalous event mode, i.e. the sample that cannot express very well of selection normal event mode
(p0For parameter preset)；2) a liter sampling is carried out to suspicious sample, distributes a weight, the weight and normal event for each sample
Mode expression precision be in inverse ratio, i.e. r=1/p (y | H0), so that more suspicious sample liter sampling number is more.Based on this, learn
The objective function for practising anomalous event mode can be write
Wherein,Anomalous event Mode S and expression factor alpha1, β optimized using alternative and iterative algorithm, with
Step (4b) is consistent.
Step 6, accident detection.
(6a) constructs anomalous event detector, calculates compared to normal event mode is only utilized, introduces anomalous event mode
Promotion to expression precision, i.e.,
In (6b) test process, anomalous event is determined as comprising anomalous event mode by the present invention, while anomalous event is visited
Survey the Video Events that device score is greater than preset threshold ξ (i.e. β ≠ 0&Det > ξ).
Step 7, event schema updates.
There are four types of situations altogether for the result of (7a) accident detection: 1) β=0&Det > ξ；2) β=0&Det≤ξ；3)β≠
0&Det > ξ；4)β≠0&Det≤ξ.The first situation can not occur, because β=0 will make alternative hypothesis degenerate for null hypothesis,
The score of detector will level off to 1 at this time；Second situation indicates normal event；The third is the case where being abnormal；4th
In the case of kind, test sample had both included anomalous event mode, while these promotions of anomalous event mode to event representation precision
It is again not high, it is consequently belonging to the normal event of less generation, the present invention is updated normal event mode using these events.
(7b) remembers that the normal event sample set of these less generations isNormal mode is carried out using following target formula
It updates
Wherein, W is diagonal matrix, and i-th of element is w on diagonal linei。D0For current normal event mode.Constraint condition
Purpose is to prevent dpIt is excessive so that A the value is too small.
(7c) brings updated normal event mode D into formula (6), is updated to anomalous event mode, and with more
Normal and anomalous event mode detection anomalous event after new.
Effect of the invention can be described further by following experiment.
1. simulated conditions
It is Intel (R) Core i3-2130 3.4GHZ, memory 16G, WINDOWS 7 behaviour that the present invention, which is in central processing unit,
Make in system, the emulation carried out with MATLAB software.
Video database used in experiment is by Hong Kong Chinese University in the Avenue data set of publication in 2013, test
It altogether include 14 class anomalous events in video, comprising: run, stay, shed.Training dataset and test data set separately include
16 sections and 21 sections of video datas, the length of every section of video it is not consistent.It is concentrated in test data, every frame video has destination layer
True mark, i.e., mark abnormal target area with box.The data set includes 30,652 frame video datas, spatial resolution altogether
It is 360 × 640.
2. emulation content
Firstly, completing inventive algorithm, (abnormality detection based on semi-supervised dictionary learning is calculated on Avenue data set
Method) experiment with the algorithm of Lu.Then, the precision of destination layer accident detection, i.e. detection zone and true exceptions area are counted
The ratio of domain overlapping area and the gross area is greater than Detection accuracy when preset value θ.Quantitative detection accuracy is as shown in table 1.
1 destination layer detection accuracy of table
The result for wherein comparing algorithm Lu comes from document:
C.Lu,J.Shi,and J.Jia,“Abnormal Event Detection at 150 FPS in Matlab,”
in proceedings of IEEE International Conference on Computer Vision,2013,
pp.2720-2727.
As it can be seen from table 1 at all detection threshold value θ, the present invention can the method than Lu preferably complete destination layer
The detection of anomalous event.This is because the mode that the present invention explores anomalous event in test process forms, and tectonic model
More new strategy realizes the promotion in model tormulation ability and precision.
Claims (7)
1. a kind of crowded crowd's method for detecting abnormality based on semi-supervised dictionary learning, which comprises the following steps:
Step 1, building data set；
Video data in collection monitoring system constructs data set, data set is divided into training dataset and test data set,
Normal video data is only included in the training set；
Affair character extracts in step 2, data set；
Input video frame in data set is divided into spatially partly overlapping video block by (2a)；
(2b) calculates the gradient and light stream figure of every frame image, and extracts gradient and Optic flow information progress respectively directly to each video block
Side's figure statistics is expressed as the local feature of Video Events；
Step 3, building are it is assumed that described assume to include null hypothesis H0With alternative hypothesis H1；
(3a) null hypothesis H0: normal video event only includes normal event mode, i.e. event y is only carried out with normal event mode dilute
Dredge linear combination, y=D α0+ ε, wherein D is normal event mode, α0It is sparse expression coefficient, ε is plant noise；
(3b) alternative hypothesis H1: anomalous video event is made of simultaneously normal and anomalous event mode, i.e. event y includes just simultaneously
Normal and anomalous event mode, y=D α1+ S β+ε, wherein S is anomalous event mode, α1It is sparse expression coefficient with β；
Step 4, the normal video event concentrated based on training data, learn normal event mode；
(4a) models plant noise using Gaussian mixtures, objective function are as follows:
Wherein, K is the number of gauss component, π0kWithIt is the prior probability of k-th gauss component and side in null hypothesis respectively
Difference, I are unit matrixs,It represents with D α0iFor mean value,For the Gaussian Profile of variance；
(4b) optimizes the parameter in formula (1) objective function using expectation-maximization algorithm；Obtain normal event mode
And its objective function of sparse coefficient:
Wherein,Representative sample yiIt is to be generated by k-th of gauss component
Probability；
(4c) optimizes formula (2) using iteration more new strategy；
Step 5, study is abnormal on the basis of the existing normal event mode that the test sample and step 4 of test data set obtain
Event schema；
(5a) models plant noise using Gaussian mixtures, objective function are as follows:
Wherein, K is the number of gauss component, π1kWithIt is the prior probability of k-th gauss component and side in alternative hypothesis respectively
Difference, I are unit matrixs,It represents with D α1iFor mean value,For the Gaussian Profile of variance；
(5b) optimizes the objective function parameters of formula (5) using expectation-maximization algorithm；
(5c) is carried out down-sampled or is carried out a liter sampling to suspicious sample to normal sample, obtains the target of study anomalous event mode
Function:
Wherein,ri=1/p (yi|H0)；
(5d) anomalous event Mode S and expression factor alpha1, β optimized using alternative and iterative algorithm；
Step 6, accident detection；
(6a) constructs anomalous event detector Det, calculates compared to normal event mode is only utilized, introduces anomalous event mode pair
The promotion of precision is expressed, i.e.,Wherein p (y | H0) expression event belongs to the probability of null hypothesis, p (y | H1) table
Show that event belongs to the probability of alternative hypothesis；
In (6b) detection process, if comprising anomalous event mode and anomalous event detector Det score is greater than default threshold in event
Value ξ then determines the event for anomalous event；
Step 7, event schema updates；
In (7a) detection process, using comprising anomalous event mode, but anomalous event detector Det score is less than preset threshold ξ
Test sample normal event mode is updated；
(7b) brings updated normal event mode D into formula (6), is updated to anomalous event mode, and with after update
Normal and anomalous event mode detection anomalous event.
2. crowded crowd's method for detecting abnormality according to claim 1 based on semi-supervised dictionary learning, which is characterized in that
(7a) specifically:
It will include anomalous event mode, but the set of test sample of the anomalous event detector Det score less than preset threshold ξ is remembered
ForNormal mode is updated using following target formula
Wherein, W is diagonal matrix, and i-th of element is w on diagonal linei, D0For current normal event mode.
3. crowded crowd's method for detecting abnormality according to claim 2 based on semi-supervised dictionary learning, which is characterized in that
In step (5c):
Normal sample is carried out down-sampled: selection p (y | H0) < p0Sample learning anomalous event mode, p0For parameter preset；
Sampling is carried out liter to suspicious sample: distributing a weight r for each sample, wherein r=1/p (y | H0)。
4. crowded crowd's method for detecting abnormality according to claim 2 based on semi-supervised dictionary learning, it is characterised in that:
Local feature using HOG and HOF feature as Video Events is expressed.
5. crowded crowd's method for detecting abnormality according to claim 2 based on semi-supervised dictionary learning, which is characterized in that
Step (4c) specifically:
Coefficient vector α0i:
||α0i||0Calculate α0iThe number of middle nonzero element, t represents the upper limit number of nonzero element, using Projected descent method
It optimizes；
About dictionary D, objective function is write:
Wherein, W is diagonal matrix, and i-th of element is w on diagonal linei, constraint conditionPass through Lagrange duality algorithm
It optimizes.
6. a kind of crowded crowd's abnormality detection system based on semi-supervised dictionary learning, including processor and memory, feature
It is: stores computer program in the memory, when computer program is run in the processor, perform claim requires 1 to 4
Method described in one.
7. a kind of computer readable storage medium, it is characterised in that: store computer program, computer program is performed reality
The existing any method of Claims 1-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910362661.1A CN110163122A (en) | 2019-04-30 | 2019-04-30 | A kind of crowded crowd's method for detecting abnormality and system based on semi-supervised dictionary learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910362661.1A CN110163122A (en) | 2019-04-30 | 2019-04-30 | A kind of crowded crowd's method for detecting abnormality and system based on semi-supervised dictionary learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110163122A true CN110163122A (en) | 2019-08-23 |
Family
ID=67633141
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910362661.1A Pending CN110163122A (en) | 2019-04-30 | 2019-04-30 | A kind of crowded crowd's method for detecting abnormality and system based on semi-supervised dictionary learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110163122A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113011399A (en) * | 2021-04-28 | 2021-06-22 | 南通大学 | Video abnormal event detection method and system based on generation cooperative judgment network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106033548A (en) * | 2015-03-13 | 2016-10-19 | 中国科学院西安光学精密机械研究所 | Crowd abnormity detection method based on improved dictionary learning |
| CN107633331A (en) * | 2017-09-26 | 2018-01-26 | 北京福布罗科技有限公司 | Time series models method for building up and device |
| CN108256296A (en) * | 2017-12-29 | 2018-07-06 | 北京科迅生物技术有限公司 | Data processing method and device |
| CN108846852A (en) * | 2018-04-11 | 2018-11-20 | 杭州电子科技大学 | Monitor video accident detection method based on more examples and time series |
-
2019
- 2019-04-30 CN CN201910362661.1A patent/CN110163122A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106033548A (en) * | 2015-03-13 | 2016-10-19 | 中国科学院西安光学精密机械研究所 | Crowd abnormity detection method based on improved dictionary learning |
| CN107633331A (en) * | 2017-09-26 | 2018-01-26 | 北京福布罗科技有限公司 | Time series models method for building up and device |
| CN108256296A (en) * | 2017-12-29 | 2018-07-06 | 北京科迅生物技术有限公司 | Data processing method and device |
| CN108846852A (en) * | 2018-04-11 | 2018-11-20 | 杭州电子科技大学 | Monitor video accident detection method based on more examples and time series |
Non-Patent Citations (4)
| Title |
|---|
| YACHUANG FENG等: "Learning deep event models for crowd anomaly detection", 《NEUROCOMPUTING》 * |
| YUAN YUAN等: "Statistical Hypothesis Detector for Abnormal Event Detection in Crowded Scenes", 《IEEE TRANSACTIONS ON CYBERNETICS》 * |
| YUAN YUAN等: "Structured dictionary learning for abnormal event detection in crowded scenes", 《PATTERN RECOGNITION》 * |
| 冯亚闯: "视频中的异常事件检测算法研究", 《中国博士学位论文全文数据库_信息科技辑》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113011399A (en) * | 2021-04-28 | 2021-06-22 | 南通大学 | Video abnormal event detection method and system based on generation cooperative judgment network |
| CN113011399B (en) * | 2021-04-28 | 2023-10-03 | 南通大学 | Video abnormal event detection method and system based on generation cooperative discrimination network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110929578B (en) | Anti-shielding pedestrian detection method based on attention mechanism | |
| CN105787458B (en) | The infrared behavior recognition methods adaptively merged based on artificial design features and deep learning feature | |
| CN106203331B (en) | A kind of crowd density evaluation method based on convolutional neural networks | |
| CN111860160B (en) | Method for detecting wearing of mask indoors | |
| CN110213244A (en) | A kind of network inbreak detection method based on space-time characteristic fusion | |
| CN107016357A (en) | A kind of video pedestrian detection method based on time-domain convolutional neural networks | |
| CN107622258A (en) | A kind of rapid pedestrian detection method of combination static state low-level image feature and movable information | |
| CN109086672A (en) | A kind of recognition methods again of the pedestrian based on reinforcement learning adaptive piecemeal | |
| CN104992223A (en) | Intensive population estimation method based on deep learning | |
| CN107563349A (en) | A kind of Population size estimation method based on VGGNet | |
| CN109615014A (en) | A kind of data sorting system and method based on the optimization of KL divergence | |
| Yang et al. | Fruit Target Detection Based on BCo‐YOLOv5 Model | |
| CN112434599B (en) | Pedestrian re-identification method based on random occlusion recovery of noise channel | |
| CN107491749A (en) | Global and local anomaly detection method in a kind of crowd's scene | |
| CN110008853A (en) | Pedestrian detection network and model training method, detection method, medium, equipment | |
| CN110084812A (en) | A kind of terahertz image defect inspection method, device, system and storage medium | |
| CN113158983A (en) | Airport scene activity behavior recognition method based on infrared video sequence image | |
| CN112347930A (en) | High-resolution image scene classification method based on self-learning semi-supervised deep neural network | |
| CN106599834A (en) | Information pushing method and system | |
| Sun et al. | Prediction model for the number of crucian carp hypoxia based on the fusion of fish behavior and water environment factors | |
| CN110163122A (en) | A kind of crowded crowd's method for detecting abnormality and system based on semi-supervised dictionary learning | |
| CN111860097B (en) | Abnormal behavior detection method based on fuzzy theory | |
| Du | An anomaly detection method using deep convolution neural network for vision image of robot | |
| CN116884192A (en) | Power production operation risk early warning method, system and equipment | |
| Pang et al. | Salient object detection via effective background prior and novel graph |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190823 |
|
| RJ01 | Rejection of invention patent application after publication |