CN106295565A - Monitor event identifications based on big data and in real time method of crime prediction - Google Patents

Abstract

Monitor event identifications based on big data and in real time method of crime prediction.Comprise the following steps that step one, monitor data collection；Step 2, defines high-level semantic；Step 3, finds middle level attribute；Step 4, machine learning based on sequence of attributes；Step 5, behavioral pattern is predicted；Step 6, sets up crime probability model；Step 7, real-time prediction of criminality.The present invention utilizes machine learning method based on sequence of attributes not have the complete monitor video material of mark to process to magnanimity, and combine the experience of handling a case of Public Security Organs, obtain the behavioral pattern prediction that offender commits a crime, criminal tendency factor data storehouse is utilized to obtain crime probability model through machine learning, the two combines just can be with the generation of real-time estimate crime, provide different grades of alarm, the real video monitoring realizing intelligence, municipal public safety management is significant.

Description

Monitor event identifications based on big data and in real time method of crime prediction

Technical field:

The present invention relates to video monitoring and public safety technical field, particularly relate to a kind of monitor event based on big data Identify and real-time method of crime prediction.

Background technology:

Intelligent video analysis is mainly by obtaining the substantial amounts of sample material as machine learning, thus obtains analyzing mould Type, sample size is the most, the most accurate, and the analysis model obtained is the most accurate, to this end, people spend substantial amounts of manpower and time Collect sample, set up Sample Storehouse, such as National Institute of Standards and Technology (NIST) The LabelMeVideo [3] of the ImageNet [2] and MIT of TRECVID video frequency searching contest database [1], Stanford, The crucial problem in one, intellectual analysis field is that sample size causes the accuracy analyzing model inadequate not, and sample size is inadequate Generally refer to data deficiency mark, i.e. what these data are actually, and in event recognition field, traditional way is from video The various low-level image feature of middle extraction, such as based on light stream feature, HOG feature and spatio-temporal interest point (STIP) feature, is then directly used for setting up various model, such as Hidden Markov Models, Dynamic Bayesian Networks, Conditional Random Fields and Random Forest, the method for this " settling at one go " is for certain A little specific events design a model, and also can obtain good effect, but it lacks extensibility, when the event kind needing identification When class increases, it may be necessary to redesign corresponding model；It addition, directly utilize low-level image feature to be highly prone to the shadow of video performance Ringing, due to the motion change of target, the change at visual angle and the impact factor such as blocking, the event of same type may have difference Performance, cause event recognition failure.

Research in prediction of criminality field at present is concentrated mainly on: 1. the analysis of aetiology of crime, has from Psychological Angle Analyzing, also have the crime correlate analysis utilizing data mining technology, which the result of these researchs can teach that Factor has important impact to crime, but it can not be directly used in real-time prediction of criminality, because these are analyzed It is all to analyze qualitatively, the quantitative relationship between not crime factor and crime generation；2. criminal trend in large scale scope Prediction, is to be predicted rather than the number of crime case in a period of time of a certain area for single people on the whole The probability of crime is predicted.

Summary of the invention:

In order to solve the problem existing for background technology, the present invention provide a kind of monitor event identification based on big data and Method of crime prediction, comprises the following steps that in real time

Step one, monitors data collection；

Step 2, defines high-level semantic；

Step 3, finds middle level attribute；

Step 4, machine learning based on sequence of attributes；

Step 5, behavioral pattern is predicted；

Step 6, sets up crime probability model；

Step 7, real-time prediction of criminality.

The inventive method proposes machine learning method based on sequence of attributes and completes monitor video high-level semantic Understanding, the method has extensibility, and the change to video performance is insensitive.The mark utilizing network Shanghai amount monitors Video, we obtain its classification, and automatically extract out middle level attribute from contextual information, the most just obtain enough confession machines The sample of device study.Machine learning method based on sequence of attributes is utilized not have the complete monitor video material of mark to magnanimity Process, and combine the experience of handling a case of Public Security Organs, obtain the behavioral pattern prediction that offender commits a crime, utilize criminal tendency Factor data storehouse obtains crime probability model through machine learning, and the two combination just can be given with the generation of real-time estimate crime Different grades of alarm, the real video monitoring realizing intelligence, municipal public safety management is significant.

Accompanying drawing illustrates:

Fig. 1 is the schematic diagram that present invention discover that middle level attribute；

Fig. 2 is present invention machine learning based on sequence of attributes schematic diagram.

Detailed description of the invention:

For the defect of event recognition field traditional method, we have proposed based on sequence of attributes (Attribute Sequence) machine learning method identifies monitor event, its objective is to be reached to high level by the study of middle level attribute Semantic mapping, attribute is the more weak semantic description in a kind of middle level between low-level image feature and high-level semantic, from bottom Features training obtain, additionally, attribute is not by the constraint of semantic category (Semantic category), same attribute may Being present in multiple classification, the combination of many attribute can express certain emerging semantic category, so a lot of high-level semantics Identify and just can be realized by the study of middle level attribute.

Monitoring data collection:

At big data age, the generation of data can be uploaded more than 500,000,000 figures with geometric growth, every day, the whole world Sheet, the video just having 20 hours durations per minute shared, and more crucially these multi-medium datas shared are general the most Through being marked, i.e. they have clear and definite title, and some also carries descriptive word, specific to monitor video, its Title typically reflects certain security incident, and these monitor videos have been completed the extraction of scene (Scene), eliminate Unrelated and invalid content, utilizes the monitor video that these have marked, sets up some typical security incidents automatically Sample Storehouse, the Sample Storehouse so set up is magnanimity, and is sustainable growth.

Definition high-level semantic:

The monitor event identified as required, defines the set of a set of keyword, such as, plunders, thrustes into, and these are crucial Word is exactly high-level semantic.Therefore, on the one hand, on network, search for title contain the video of these keywords, and title or The contextual information of video should occur " monitoring " printed words, under being downloaded the most in the lump by the contextual information relevant to this video Come, for video is described the audio-frequency information of character, utilizes voice to turn text technology after extracting and be converted into word, On the other hand, the monitor video that public security department is the most archived, archive information has the explanation to video content, it is also possible to make Use for assistance data.

Discovery middle level attribute:

The monitor event identified for needs, re-defines a set of ratio more complete middle level property set, in complete Manual definition Layer property set has careless omission unavoidably, and incomplete middle level property set can affect the effectiveness of machine learning, it is therefore desirable to a kind of The method automatically extracting middle level attribute, the most just obtains enough samples for machine learning.We utilize with video together The contextual information downloaded, and the descriptive information having been converted in the video of word, use natural language processing technique, carry Take out descriptive words therein (generally noun and verb), the descriptive word that same category of monitor video is extracted Language carries out cluster analysis, finds out the word that the frequency of occurrences is higher, and carries out extensive to word, the knapsack of such as Mr. Zhang, Wang Nv The wallet of scholar is all " people X and object X ", as it is shown in figure 1, so obtained middle level attribute has static state, such as people 1,2,3 ..., Vehicle 1,2,3 ..., people X and object X etc., also have dynamic, such as, run, fall down to the ground, the middle level of all categories monitor video is belonged to Property collection merge and just obtained a set of ratio more complete middle level property set, for new monitor event, as long as adding corresponding Middle level attribute, has extendible learning capacity.

Machine learning based on sequence of attributes:

As in figure 2 it is shown, middle level attribute can remove the coupling between monitor video and its high-level semantic classification, attribute is not subject to The constraint of semantic category, same attribute is likely to be present in multiple classification, therefore each attribute can be carried out across classification Practise, train corresponding grader, after the training completing grader, to training data application class device, obtain corresponding attribute, The temporal information of record attribute simultaneously, i.e. it at which frame (or which frame) is extracted, so between each attribute Just having had temporal precedence relationship, constituted sequence of attributes, only federation properties and temporal information just can accurately identify monitoring Event, such as: consideration " robbery " event, 3 attributes therein: people 1 and object 1, people 2 and object 1, people 2 run, certain on the time Be the former front, the posterior order of the latter, so, just obtained sequence of attributes by the training of Massive Sample and arrived high-level semantic Map, when carrying out the test of unknown sample, based on middle level attributive classification device, obtain the sequence of attributes of input monitoring video, and root According to the mapping relations of sequence of attributes to high-level semantic, obtain the high-level semantic of test monitoring video.

In order to realize real-time prediction of criminality, we combine behavioral pattern prediction and crime probability model.

Behavioral pattern is predicted:

Identify that monitor event can only produce alarm when event occurs, for some premeditated criminal behavior, tend to Discovery clues and traces in monitor video before crime occurs, the and (network that these monitor videos usually do not have on network On general scene time only event occurs of monitor video), and the most do not mark, utilize machine based on sequence of attributes Device learns, and magnanimity can not had the complete monitor video material of mark to process by we, identifies the criminal behavior needing prediction, After positioning this criminal behavior and subject of crime, with subject of crime as target, in the monitor video before criminal behavior occurs, search The appearance of rope subject of crime, obtains time, place, the information such as hover/stay, trail, and aggregated analysis also combines Public Security Organs The experience of handling a case obtain corresponding behavioral pattern, with this, criminal behavior is predicted.

Set up crime probability model:

But, rely solely on the generation of behavioral pattern prediction crime, higher rate of false alarm can be caused, it practice, crime divides The educational level of son, family background, economic situation, inter personal contact, situation of serving a sentence etc. often have bigger difference with ordinary people, As the factor of prediction crime, and according to this citizen can be set up a set of criminal tendency factor data storehouse.

Utilizing the data base of the most a set of magnanimity scale, Macro or mass analysis also combines the experience of handling a case of Public Security Organs and obtains accordingly Behavioral pattern, can set up crime probability model:

G=w₁α₁+w₂α₂+w₃α₃+… (1)

The input of this model is criminal tendency factor-alpha_i, it is output as crime advanced warning grade G, w_iFor weight.In order to obtain a_iNeed Will be by criminal tendency factor quantification, such as to educational level, can be by primary school/junior middle school, special secondary school/senior middle school, junior college/university, research Life is quantified as 1, and 0.75,0.5,0.25, maximum is 1, is sequentially reduced backwardWherein N is the possible value of the criminal tendency factor Number (being exactly 4 for educational level), should follow a principle, the i.e. the biggest quantized value of the probability of crime quantifying when The biggest, in order to obtain weight w_i, need to train first with the criminal tendency factor of existing offender and non-offender One SVM classifier

θ₁α₁+θ₂α₂+θ₃α₃+…b (2)

a_iFor the criminal tendency factor after quantifying, θ_iFor coefficient, b is constant.Make a=[a₁, a₂, a₃...], θ=[θ₁, θ₂, θ₃...], C=[1 ,-1] represents crime and non-crime respectively, then the training process of this SVM classifier is

min||θ||²

s.t.C(θ^TA+b) >=1 (to all samples) (3)

Solve this optimization problem, obtain parameter θ and the b of grader.The meaning of θ is each criminal tendency factor pair crime Whether contribution, θ_iShow that the most greatly this criminal tendency factor is the most important, therefore w_iCan be calculated as below

$w_i=\frac{{\mathrm{\θ}}_i}{{\Σ}_i{\mathrm{\θ}}_i}---\left(4\right)$

So w_iJust can reflect the importance of the corresponding criminal tendency factor, and meet ∑_iw_i=1.

Prediction of criminality in real time:

Bonding behavior model prediction and crime probability model can be carried out real-time prediction of criminality.Monitoring camera pair Main body in picture carries out real-time behavior prediction, starts face recognition module when meeting specific behavioral pattern, confirms it Transfer the respective record in criminal tendency factor data storehouse after identity and input crime probability model, obtaining crime early warning etc. Level, further, after being quantified by crime advanced warning grade G (M quantized interval), can provide the crime generation alarm of M grade, example Such as M=4, just there is the alarm of 4 grades (normal, should be noted that, dangerous, abnormally dangerous), coordinate color to represent, monitor in public security Display whole city dangerous hot spots scattergram real-time on the giant-screen at center, quickly, reasonably distributing police strength to public security department provides The foundation of science.

Claims (1)

1. monitor event identifications based on big data and in real time method of crime prediction, it is characterised in that comprise the following steps that

Step one, monitors data collection；

Step 2, defines high-level semantic；

Step 3, finds middle level attribute；

Step 4, machine learning based on sequence of attributes；

Step 5, behavioral pattern is predicted；

Step 6, sets up crime probability model；

Step 7, real-time prediction of criminality.