CN108182382A - Based on the similar Activity recognition method and system of figure - Google Patents
Based on the similar Activity recognition method and system of figure Download PDFInfo
- Publication number
- CN108182382A CN108182382A CN201711275203.1A CN201711275203A CN108182382A CN 108182382 A CN108182382 A CN 108182382A CN 201711275203 A CN201711275203 A CN 201711275203A CN 108182382 A CN108182382 A CN 108182382A
- Authority
- CN
- China
- Prior art keywords
- data
- behavior
- sequence
- event
- activity recognition
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Abstract
The invention discloses a kind of Activity recognition method and system similar based on figure.This method includes the following steps:All kinds of initial data are acquired, and initial data is pre-processed using wireless sensor network;Based on treated, data carry out hierarchical diagram modeling, and data variation caused by occurring first with event builds Event Graph Model, then build behavior graph model to the user behavior sequence sets built by event.In occurrence diagram inquiry, by the feature structure of occurrence diagram according to significance level sequence construction feature structure sequence, the similarity query of figure is converted to the similarity query of feature structure sequence, and conventional sequence matching algorithm is improved;It in behavior figure inquiry, based on the similar feature selecting algorithm of structure, is inquired according to the figure querying method of feature based structure sequence, realizes Activity recognition.Hierarchical diagram proposed by the present invention models and schemes similar search algorithm so that Activity recognition accuracy and efficiency are significantly improved than method of the tradition based on statistics or sequential.
Description
Technical field
The present invention relates to the Activity recognition method and systems of wireless sensor network, and in particular to a kind of similar based on scheming
Activity recognition method and system.
Background technology
User behavior recognition refers to obtain the external manifestation of user behavior by technological means, on this basis by user's row
For the process for regarding as belonging to some classification.Behavior is defined as type of action or the behavior of detected object in the field
Pattern.Currently, aging of population is becoming a kind of global trend, and with this trend development, intelligence is carried out to the elderly
Healthy can close shield becomes one of urgent problem, by the common concern from various circles of society.Activity recognition technology passes through
Monitoring to the elderly's behavior can obtain its animation, it is ensured that life security in real time.Meanwhile current social is busy, fast section
The life style played has brought many psychological problems, and the problem is particularly prominent in the big crowd of teenager and operating pressure
Go out, in many cases since what abnormal behaviour was found causes great tragedy not in time.In addition, Activity recognition also smart home,
Intelligent office and movement, amusement etc. become the focus of attention of multiple industries with having extensive prospect and economic value.
The work in Activity recognition field is concentrated mainly on the Activity recognition based on computer vision for a long time, will image
Head perceptually means, are analyzed and processed by the image sequence to shooting, identify the behavior of user.But this method
Use in the presence of camera is limited by factors such as illumination, installation site and angles, there is monitoring blind area, hidden to user
It is private it is invasive it is relatively strong, lower deployment cost is higher and calculates the deficiencies of complicated, it is difficult to large-scale promotion application, use scope by compared with
Big limitation.
In addition to this, correlation of the data model selected in current behavior identification field between sensing data is examined
Consider not enough, search algorithm time complexity is higher, causes the performance that detection identifies bad.In terms of object is identified, active user
Activity recognition has focused largely on one simple Activity recognition, is mainly based upon static or sequential data.Also there are indivedual researchs
It is related to the identification of the single complex behavior in the case of including intersecting and being parallel, but computation complexity is higher.In addition, for more people
Complex behavior Study of recognition, although preliminary progress is had been achieved in computer vision field, in wireless sensor
Network field currently still lacks relevant achievement in research.And in practical applications, user's row based on wireless sensor network
For identifying system acquisition sensing data type it is more, quantity is big, relevance is close, how to organize these data, and therefrom take out
As the feature for going out user behavior finally realizes that the identification of user behavior is to be badly in need of the difficulties of solution.
Invention content
Goal of the invention:For above deficiency, the present invention propose it is a kind of in wireless sensor network based on scheming similar row
For recognition methods, effectively data can be modeled and accurately identify user behavior.
Another object of the present invention is to provide a kind of Activity recognition system similar based on figure.
Technical solution:A kind of Activity recognition method similar based on figure, includes the following steps:
1st, all kinds of initial data are acquired using wireless sensor network, and initial data is pre-processed.
Data scrubbing, data integration and screening are included to the pretreatment of initial data, wherein, data scrubbing includes filler
According to missing values, smooth noise data identify and delete isolated point;Data integration includes with screening will be from each sensor node
Data merge, and unified storage is in the database, then rejects redundant data therein.
2nd, based on treated, data carry out hierarchical diagram modeling, include the following steps:
21) sensing data model is established based on data snapshot figure, data variation caused by event generation is recycled to establish
Event Graph Model builds sample event chart database;
22) it using the similar search algorithm query sample event chart database of the figure of feature based structure sequence, obtains similar
Event;
23) event graphic sequence is built based on similar case, identify basic act and builds behavior sequence, to being built by event
Behavior sequence collection establish behavior graph model, build behavior chart database;
24) using based on the similar feature selecting algorithm User behavior chart database of structure, Activity recognition result is obtained.
Wherein, structure Event Graph Model includes in step 21):Data in certain time are observed based on pretreated data
Corresponding changing value obtains the consecutive variations of event evolution data, data increment is applied to graph model, obtains data increasing
Spirogram;The data variation caused by graph model occurs using event using data increment nomography and represents different classes of event
Feature completes the structure of Event Graph Model.
The similar search algorithm of the figure of feature based structure sequence includes in step 22):By the feature structure of occurrence diagram according to
Significance level sorts, construction feature structure sequence, and the similarity query of figure is converted to the similarity query of feature structure sequence;
Then using the maximum common subsequence algorithm of weighting and the sequence editing distance algorithm of weighting come query structure sequence, in sequence
On the basis of editing distance algorithm, introduce index decreased function and define weight.
Structure behavior graph model includes in step 23):All types of basic acts are trained, obtain corresponding thing
Then the average duration of part sequence identifies event graphic sequence from real-time sensor data stream, and utilizes sliding window side
Method is split event graphic sequence, and basic act is identified from sequence of events;Daily sensor real time data is carried out
The identification of basic act, and daily behavior sequence is constructed according to sequential relationship;By the data accumulation of certain collection period,
A behavior sequence set is formed, for behavior sequence set, according to the appropriate merging period, the multisequencing connection based on weighting is matched
Algorithm merges behavior sequence, is built into behavior figure, and the structure for completing sample behavior chart database is trained by duration data
It builds.
Included in step 24) based on the similar feature selecting algorithm of structure:Utilize the side of the similar sub-clustering of Frequent tree mining structure
Method chooses the feature structure that discrimination is best and mutual structural redundancy is small from Frequent tree mining set, reaches raising and look into
Ask the purpose of accuracy.
A kind of Activity recognition system similar based on figure, including:Data acquisition module, data management module, event monitoring
With identification module, behavior pattern sort module, data acquisition module is mainly by several sensor nodes and several sensors
The base station composition of node connection, data management module, event monitoring and identification module, behavior pattern sort module are calculated in terminal
It is realized on machine, wherein, sensor node is used to acquire all kinds of initial data and encapsulate framing, then be sent to wireless transmission method
The base station, the base station transfer data to the terminal computer by serial ports, and the terminal computer is to collected
Data are pre-processed, analyzed and are excavated, and finally identify the behavior of user.
Wherein, data management module is to the pretreatment process of data:First, the number lost using Mean Method fitting
According to;Secondly, the noise in smooth sensing data;Then, the singular point in sensing data is carried out using median filtering method
It rejects;Finally, the repeated sample that removal sample data is concentrated.
Event monitoring is used to complete the training and inquiry of Event Graph Model with identification module, wherein, model training includes:Number
Data preprocess unit is stored in sensor database by data are corrected, noise abnormality detection and redundancy remove;Then, according to
Different classes of event data trains different Event Graph Models, builds occurrence diagram by data augmentat ion, final establish includes
All types of sample event chart databases;Inquiry identification includes:The sensing data acquired in real time is handled and is modeled,
The sample graph built in query event figure, with sample event chart database carries out similarity query, identifies the type of the event.
Behavior pattern sort module is modeled and is inquired for consummatory behavior figure, wherein, behavior figure modeling process includes:It is based on
The average duration of corresponding sequence of events is obtained to the training of basic act;Then it is identified from real-time sensor data stream
Outgoing event sequence, then basic act is identified from sequence of events, construct behavior sequence according to sequential relationship;Based on weighting
Multisequencing Algorithm merges behavior sequence, is built into behavior figure, finally obtains sample behavior chart database;Behavior figure
Inquiry is completed to carry out behavior figure similarity query in sample behavior chart database, finds the behavior similar to its behavior pattern
Figure, obtains the classification results of behavior pattern.
Advantageous effect:Compared with prior art, the present invention it has the following advantages:
1st, data variation caused by hierarchical diagram modeling method proposed by the present invention occurs first with event builds event artwork
Then type builds behavior graph model to the user behavior sequence sets that are built by event, with traditional building based on statistics and sequential
Mould method is compared, and can preferably reflect the association between the data of description user behavior, and precision ratio is more excellent.
2nd, the present invention occurrence diagram inquiry in using weighting maximum common subsequence algorithm and weighting sequence editor away from
From algorithm, because it is contemplated that the importance of feature structure, sorts according to importance size, the corresponding feature vector of occurrence diagram is turned
Characteristic sequence is turned to, and weighting function is introduced when carrying out similarity calculation, so inquiry accuracy is generally better than other
Algorithms most in use.
3rd, it is proposed by the present invention based on similar behavior figure search algorithm is schemed, use the Frequent tree mining of reflection structural dependence
It is significantly more efficient as feature structure Candidate Set, and for the different feature structure selection algorithm of the practical proposition of different levels
The characteristic set of reflection user behavior feature is extracted, improves the accuracy of basic algorithm inquiry, there is preferable performance table
It is existing.
Description of the drawings
Fig. 1 is system hardware deployment scheme;
Fig. 2 is system software architecture diagram;
Fig. 3 is based on the similar user behavior detection of figure and identifying system flow chart.
Specific embodiment
Technical scheme of the present invention is described further below in conjunction with the accompanying drawings.
With reference to Fig. 1, the Activity recognition system similar based on figure proposed by the present invention mainly includes several in hardware deployment
Terminal node, base station, data processing server etc..Terminal node uses wireless sensor, and data processing server is using calculating
Machine, operation principle are:First, the sensor node automatic network-building of distribution starts to acquire Various types of data;Then, sensor node
Data are sent to base station by wireless signal;Finally, base station transfers data to terminal computer by serial ports.Terminal calculates
Machine is handled, analyzed and is excavated to collected data, finally identifies the behavior of user, finds the behavior pattern of user.
Hardware system is the physical basis of Activity recognition system as related software and the carrier of algorithm.During design not only
It needs to fully consider calculating, communication and the storage capacity of sensor node and the processing capacity of base station and node stability etc.
Factor, it is also necessary to consider many factors such as communication protocol and network topology, ensure the stabilization of whole system with it is reliable.Such as Fig. 1
Shown, in the present embodiment, a set of wireless sensor network user behavior detection has selected three classes sensor with identifying system:First
Class is environmental parameter sensor, including Temperature Humidity Sensor DHT11, illuminance sensor BH1750, sound transducer FC-04;
Second class is alignment sensor, including infrared sensor of the human body (PIR) D203S, ultrasonic sensor HC-SR04;Third class is
Article movable sensor, including door status sensor, dry-reed tube switch sensor, soft-touch control sensor, displacement switch sensor
With infrared-reflectance sensors (printer goes out paper sensor) TCRT5000.The hardware platform of base station is CC2530, but function and
General sensor nodes backboard has larger difference:First, the basic function of base station is that the information for sending sensor node forwards
It is broadcast in network to computer, and by the order that computer issues;Secondly, base station will also be responsible for the sensor of filtering useless
Nodal information.Base station is directly communicated by RS232 serial ports with computer.A PC machine is used to be calculated as terminal in the present embodiment
Machine, major function are:Receive the sensing data forwarded from base station;It is sent and ordered to network by base station;Run various data
Processing, modeling and detection recognizer;And data and recognition result are subjected to output preservation.System is assisted using ZigBee communication
View has used Mesh topological, which includes a coordinator node and a series of routing node and terminal node, routing section
Point is responsible for collecting and forwarding information, and terminal node is responsible for gathered data, and coordinator node is then responsible for the stabilization work of control whole system
Make.
With reference to Fig. 2, system of the present invention mainly includes following several modules in software architecture level:Data acquire
Module, data management module, event monitoring and identification module, behavior pattern sort module, wherein, data acquisition module mainly by
Several sensor nodes, the base station being connect with several sensor nodes composition, data management module, event monitoring and identification
Module, behavior pattern sort module realize that the function of each main modular is as described below on terminal computer.
It is data acquisition module first, data acquisition module mainly using bottom layer driving, is obtained not by sensor hardware
The data of same type, and data are carried out with necessary correction, compensation, compression and coding, it is placed on the message of respective type
(Message) in, according to the form of regulation, in addition necessary start bit and CRC check, form a frame, with the side of wireless transmission
Formula is sent to other nodes or base station.Meanwhile data acquisition module be also responsible for preserve certain time width historical data exist
In sensor node (Mote), for itself data processing and inquiry.
Interactive message is mainly the following type in systems:Action message:The main activity for including people or object
Information is obtained and is sent by above-mentioned third class article movable sensor;Environment message:Main include is obtained from all kinds of environmental sensors
The status information taken is obtained and is sent by first kind environmental parameter sensor;Location message:The main position for including people or object
Information is obtained and is sent by the second class alignment sensor;Control message:Mainly comprising various control instruction information, by computer
It is sent to sensor network.Meanwhile the link signal strength various kinds of sensors between node can obtain, the way for obtaining and sending
The acquisition of the parameters such as diameter and other temperature, illumination and sending method are identical.
Secondly data management module, the main function of data management module is the Various types of data of receiving sensor network,
Wireless sensor network database is stored data in after data prediction, is used for other modules.Data management module master
Complete following functions:
1) data receiver:The various types of data that sensor node is sent in sensor network are received, it will
The message of encapsulation in the packet is unsealed, and is sent to pretreatment unit;
2) data prediction:Data packet is decomposed, removes some noises and interference, and supplement necessary information, is wrapped
Time label, sensor node number etc. are included, is stored in database.Wherein, the time marks:It is whole in order to ensure the accuracy of modeling
The time of a system needs to synchronize.The present invention is using the widely used Flooding Time-Synchronization agreement FTSP in wireless sensor network
(Flooding Time Synchronization Protocol) realizes the time synchronization of whole system, when ensureing accurate
Between mark.FTSP agreements are excellent in many-sided performance such as Energy-aware, autgmentability, robustness, stability for wireless sensor network
It is different.Basic principle is:To the temporal information of the whole network node broadcasts coordinator node first by the way of flooding, when realizing in net
Between synchronize after, each node includes an a length of a game GlobalTime and local zone time LocalTime, on this node
Operation uses LocalTime, and when communicating with other nodes, local LocalTime is converted to GlobalTime;From other sections
When point receives data, the GlobalTime values of reception are converted into the LocalTime of oneself.
Data with the real world are generally susceptible to the invasion of noise data, missing and inconsistency.Data prediction
To improving the quality of data, the precision of data mining process and performance important in inhibiting below is improved.The data of the present invention are pre-
Processing mainly includes:Data scrubbing, data integration and screening, specific method are as follows:
(1) data scrubbing:Main processing shortage of data, smooth noise data, identify and delete isolated point.It is lacked for data
It loses, the common approach is that filling vacancy value using most likely value, i.e., speculates missing values using existing data information so that
Missing values are as possible close to actual value.The present invention is in the method to sensing data missing polishing, using identical sensor distance
Nearest data are filled a vacancy before current time, for DSG (Data Snapshot Graph, data snapshot figure) figure side
Link RSSI (the Received Signal Strength of the missing of upper weights newest missing data after the modeling moment
Indication, received signal strength indicator) value complement is neat, and specific priority is:Assuming that the RSSI missings of T0 moment A-B links,
Using the RSSI value polishing of B-A links, (the two is that the routing of same chain two different terminal As and B acquire RSSI respectively first
R values), secondly wait for the RSSI polishings of the newest A-B B-A links at next moment.For common noise data,
Including mistake or deviate desired isolated value, use of the invention carries out branch mailbox according to certain class data, is carried out according to average value
Noise data it is smooth.Such as it is directed to the shake of ultrasound data, it is necessary to carry out noise smoothing processing;Finally, it is filtered using intermediate value
Wave method rejects singular value present in sensing data, and the simple principle of median filtering method is exactly for same sensing
Each data point of device in time series all centered on this point, intercepts the data of adjacent certain length, and to this
A little data are ranked up, and original value is replaced with the median of the data in intercepted length.If the length of data is even number, use
The average value of two intermediate values replaces.
(2) data integration and screening:Multiple data such as each node of sensor network and user PC will exactly be come from
The data in source merge, and unified storage is in the database, and redundancy therein and the data value repeated are detected
With processing.Data screening is exactly that the data in sensor database are arranged, and removes the sample data of repetition, improves sample
The availability of data.
Based on the method for above-mentioned data prediction, data prediction flow of the invention is:First, using Mean Method,
The data that the data fitting between former and later two points is taken to lose;Secondly, the noise in smooth sensing data;Then, it uses
Median filtering method rejects the singular point in sensing data;Finally, it for repeated sample present in sample set, uses
Data integration reduces the quantity of training sample with screening removal redundancy.
Followed by occurrence diagram modeling with enquiry module and behavior pattern sort module, reference and with reference to Fig. 3, the present invention
It is proposed hierarchical diagram modeling and the user behavior recognition method based on the similar inquiry of figure.According to two event, behavior levels, exist first
System event layer proposes the sensing data modeling algorithm based on data snapshot figure and the occurrence diagram modeling based on data increment figure
Algorithm.Sensing data modeling algorithm based on data snapshot figure is the flow for DSG modelings, and it is fast exactly to build an overall situation
According to the process of figure, the basic procedure of DSG algorithms is:Modeling opportunity is determined according to the time that is triggered of sensor node first, so
The data of all nodes and communication link in the wireless sensor network at the moment are acquired afterwards, are built into a complete graph.
It is noted that T_0 moment not all node all can horse back gathered data, for example environment nodes are periodical acquisition numbers
According to, this just needs the data on shortage of data vertex in the data snapshot figure to current time T_0 to carry out polishing, and the present invention uses
By the data of corresponding vertex collection period nearest distance T_0 come polishing.The missing of other data, such as the missing of the RSSI value on side
Problem carries out polishing using the method identical with during data prediction.
Then when carrying out similarity query to occurrence diagram, by the feature for excavating sample occurrence diagram and query event figure
Structure by the feature structure of occurrence diagram by the descending sequence of importance, forms the feature structure sequence for representing occurrence diagram.Figure
Similarity query problem be converted to the similarity query problem of sequence, for two kinds of common sequence similarity computational methods
LCS (Longest Common Subsequence, maximum common subsequence), SED (Sequence Edit Distance, sequence
Row editing distance) propose improved weighting algorithm WLCS (Weight Longest Common Sequence), WSED (Weight
Sequence Edit Distance) characteristic sequence is inquired.Secondly in system action layer, with reference to the identification of basic act
And user behavior sequence construct method, it is proposed for the characteristics of behavior sequence and the weighting multisequencing connection of continuous coupling is encouraged to match
The modeling method of algorithm WMSA (Weighted Multiple Sequence Alignment).In behavior figure inquiry phase,
On occurrence diagram search algorithm basis, it can be caused to inquire because there are too many redundancies for original feature selecting algorithm is used
The problem of accuracy rate is relatively low proposes a kind of frequent feature structure selection algorithm similar based on structure according to the characteristics of behavior figure
SSFS (Structural Similarity on Feature Selection), utilizes the side of the similar sub-clustering of Frequent tree mining structure
Method reduces redundancy, improves feature structure discrimination, increases inquiry accuracy rate.It is described in detail below.
The detection of event is in the bottom of hierarchical logic with identification in entire software architecture, and main realize includes occurrence diagram
The training of model and inquiry two parts.
1) model training:It collects data sample and trains model.First, data pre-processing unit by data by school
Just, noise abnormality detection and removal are stored in sensor database;Then, it is trained according to the event data of different classifications different
Event Graph Model, be finally built into comprising all types of event chart databases.The structure of occurrence diagram is mainly increased by data
Amount method is completed, and main Constructed wetlands are comprehensively to reflect the behavior of user to wireless sensor network number as far as possible in occurrence diagram
The variation caused by, final form is undirected label figure.The figure modeling method based on data increment, formalization are fixed
Justice is as follows:
R=<V,E,ID,DV,DE,fv,gv,he,tv>
Wherein, R represents data increment figure;V represents the nonempty set on vertex in figure;E={ vivj| node viWith vjBetween
Nonoriented edge };The number set of ID representative sensor nodes;DV={ DVi| the monitor value at sensor node the latter moment is with before
The difference of one moment monitor value };DE={ DEi| side EiData value differences of the corresponding communication link RSSI between two moment };
fv:V → ID is the labeling function on vertex, reflects the vertex in figure and the correspondence of sensor node;he:E → DE is the number on side
According to function, reflect the degree of association between two endvertexes on side;gv:V → DV is the data function on vertex, can obtain this vertex
Data information;tvIt represents the time, refers to the v moment.By defining it is found that DIG (Data Increment Graph, data increment figure)
It is to be calculated by the data of continuous acquisition, is a non-directed graph.Each sensor node (vertex in DIG) data are by latter
A moment monitoring data subtract previous moment monitoring data and obtain, it describes certain continuous time inner vertex data and side
Value increase situation of change.Here the temporal information of DIG is stored in KS vertex when relevant DSG is modeled.
2) inquiry identification:Real-time sensing data is handled and modeled, query event figure is built, with occurrence diagram number
Similarity query is carried out according to the sample graph in library, identifies the type of the event.The present invention proposes feature based structure sequence
Scheme similar search algorithm, employ the maximum common subsequence algorithm WLCS of weighting and the sequence editing distance algorithm of weighting
WSED.On the basis of sequence editing distance algorithm SED, introduce index decreased function and define weight.The big subtotal of weight of element i
Calculating formula is:
W (i)=α-(i+1)
Wherein, α usually takes the integer more than 1, and 2 are uniformly taken in system.I represents the process in the editing distance of the sequence of calculation
In, the corresponding position in the characteristic sequence of query graph (is inserted and revises) to the operation of element, i is smaller it can be seen from formula, says
Bright element is in FSQThe position of (characteristic sequence for referring to query graph) is more forward, and the importance of the element (feature structure) is also bigger.Cause
This, more forward element operation, " editor " cost is higher.
After completing the work of the two stages, it is also necessary to recognition result to the corresponding types in sample event chart database
Sample graph is corrected and optimizes.
It is finally behavior pattern sort module, the modeling of consummatory behavior figure and behavior patterns mining.Behavior figure is modeled
And inquiry, classify according to behavior pattern to user, be the key problem of Activity recognition research.Processing procedure can also be divided into
Two parts of modeling and inquiry, wherein modeled segments mainly comprise the following steps:
1) all types of basic acts are trained, obtain the average duration of corresponding sequence of events, then
Event graphic sequence is identified from real-time sensor data stream, and event graphic sequence is split using sliding window method, from
Basic act is identified in sequence of events.
2) identification of basic act is carried out to daily sensor real time data, and is constructed according to sequential relationship daily
Behavior sequence.
3) by the data accumulation of certain collection period, a behavior sequence set is formed, for behavior sequence set, is pressed
According to the appropriate merging period, the multisequencing Algorithm WMSA based on weighting merges behavior sequence, is built into behavior figure.
The structure for completing sample behavior chart database is trained by sufficient data.
Wherein, the weighting multisequencing Algorithm WMSA that the behavior figure of the structure user is proposed is in modeling period
The behavior sequence of user merges, and completes the figure modeling to user behavior.When being merged to two behavior sequences, preferentially
Consider that the connection of continuous behavior segment is matched, and realize improved multisequencing connection on this basis and match.For given comparison sequence
P and Q is arranged, constructs similarity matrix.It is character match score to enable x, and y is to mismatch score, and z is gap penalty.Match in tradition connection
In algorithm, when there is continuous k vacancy, then gap penalty is kz.It is p to the gap penalty occurred first in innovatory algorithm,
When there is continuous k vacancy, gap penalty is p+ (k-1) × z (p>>Z), while continuous coupling character is encouraged, for multiple
The character of continuous coupling, which is introduced, encourages function e (S) to increase matching score, and wherein S is the similarity score of continuous coupling character.
It encourages function e (S) that there is increase weight for the character of multiple continuous couplings, stresses continuously to match character, it can be with
Sum function is interpreted as, the character of continuous coupling is more, and bonus point is more.The formula of WMSA represents as follows:
The formula uses the thought of Dynamic Programming, takes the maximum value of four kinds of situations.Wherein, PiFor i-th of character of P, QjFor
J-th of character of Q;F (i, j) represents the optimal similarity score between P and Q;T (i, j) represents current similarity score;s(i,
J) matching score for two characters is represented;E (S) is continuous coupling bonus point function;Max () expressions are maximized.
The groundwork of inquiry phase is exactly to carry out similarity query in sample behavior chart database to User behavior figure,
The behavior figure similar to its behavior pattern is found, completes the inquiry to behavior figure, obtains the classification results of behavior pattern.Described
Based in the similar feature selecting algorithm SSFS of structure, choose discrimination most preferably from Frequent tree mining set and tie between each other
The small feature structure of structure redundancy achievees the purpose that improve query accuracy.The structural similarity of Frequent tree mining is calculated using maximum
Public subgraph MCS (Maximal Common Subgraph) is measured, and the calculation formula of MCS is as follows in the present invention:
Wherein, q and g represents two figures, and Sim (q, g) represents figure q and schemes the structural similarity of the Frequent tree mining of g, MCS (q,
G) it represents figure q and schemes the node number of the public subgraph of maximum of g, ωijAnd ωi’j’The corresponding side e of q and g are represented respectivelyijAnd ei′j′
On weights, β represents to weigh the parameter of side and node significance level, β ∈ (0,1), and β takes 0.6 in the present invention.The first half of formula
Point it is the measurement to the node similarity of graph structure q and g, latter half is the measurement of opposite side and weights.Linear group of the two
Close the similarity represented between cum rights graph structure.
Finally, with the correct behavior figure classified, the sample behavior figure that classification is corresponded in behavior chart database is carried out excellent
Change, improve the query accuracy of follow-up same type behavior figure.
Claims (10)
- A kind of 1. Activity recognition method similar based on figure, which is characterized in that include the following steps:All kinds of initial data are acquired, and initial data is pre-processed using wireless sensor network;Based on treated, data carry out hierarchical diagram modeling, include the following steps:Sensing data model is established based on data snapshot figure, data variation caused by recycling event occurs establishes event artwork Type builds sample event chart database;Using the similar search algorithm query sample event chart database of the figure of feature based structure sequence, similar case is obtained;Event graphic sequence is built based on similar case, identify basic act and builds behavior sequence, the behavior to being built by event Sequence sets establish behavior graph model, build behavior chart database;Using based on the similar feature selecting algorithm User behavior chart database of structure, Activity recognition result is obtained.
- 2. the Activity recognition method similar based on figure according to claim 1, which is characterized in that the initial data is located in advance Reason includes data scrubbing, data integration and screening, wherein,The data scrubbing includes filling data missing values, and smooth noise data identify and delete isolated point;The data integration includes merging the data from each sensor node, and be uniformly stored in database with screening In, then reject redundant data therein.
- 3. the Activity recognition method similar based on figure according to claims 1, which is characterized in that the structure occurrence diagram Model includes the following steps:The corresponding changing value of data in certain time is observed based on pretreated data, obtains the company of event evolution data Continuous variation, applies to graph model by data increment, obtains data increment figure;The data variation caused by graph model occurs using event using data increment nomography and represents different classes of event Feature completes the structure of Event Graph Model.
- 4. the Activity recognition method similar based on figure according to claim 1, which is characterized in that the feature based structure The similar search algorithm of figure of sequence includes:The feature structure of occurrence diagram is sorted according to significance level, construction feature structure sequence, The similarity query of figure is converted to the similarity query of feature structure sequence;Then it is calculated using the maximum common subsequence of weighting Method and weighting sequence editing distance algorithm carry out query structure sequence, wherein, the maximum common subsequence algorithm of the weighting and The sequence editing distance algorithm of weighting is on the basis of sequence editing distance algorithm, introduces index decreased function and defines weight.
- 5. the Activity recognition method similar based on figure according to claim 1, which is characterized in that the structure behavior artwork Type includes the following steps:All types of basic acts are trained, obtain the average duration of corresponding sequence of events, then from real-time Sensor data stream in identify event graphic sequence, and event graphic sequence is split using sliding window method, from event sequence Basic act is identified in row;The identification of basic act is carried out to daily sensor real time data, and daily behavior sequence is constructed according to sequential relationship Row;By the data accumulation of certain collection period, a behavior sequence set is formed, for behavior sequence set, according to appropriate The merging period, the multisequencing Algorithm based on weighting merges behavior sequence, behavior figure is built into, by lasting number The structure of sample behavior chart database is completed according to training.
- 6. the Activity recognition method similar based on figure according to claim 1, which is characterized in that described similar based on structure Feature selecting algorithm include:Using the method for the similar sub-clustering of Frequent tree mining structure, discrimination is chosen from Frequent tree mining set The small feature structure of best and mutual structural redundancy achievees the purpose that improve query accuracy.
- 7. a kind of Activity recognition system similar based on figure, which is characterized in that including:Data acquisition module, data management module, Event monitoring and identification module, behavior pattern sort module, the data acquisition module is mainly by several sensor nodes and institute State the base station composition of several sensor node connections, the data management module, the event monitoring and identification module, the row It is realized on terminal computer for mode classification module, wherein,The sensor node is used to acquire all kinds of initial data and encapsulate framing, then be sent to the base with wireless transmission method Stand, the base station transfers data to the terminal computer by serial ports, the terminal computer to collected data into Row pretreatment, analysis and excavation, finally identify the behavior of user.
- 8. the Activity recognition system similar based on figure according to claim 7, which is characterized in that the data management module Pretreatment process to data is:First, the data lost using Mean Method fitting;Secondly, in smooth sensing data Noise;Then, the singular point in sensing data is rejected using median filtering method;Finally, removal sample data is concentrated Repeated sample.
- 9. the Activity recognition system similar based on figure according to claim 7, which is characterized in that the event monitoring is with knowing Other module is used to complete the training and inquiry of Event Graph Model, wherein,The model training includes:Data pre-processing unit is by data are corrected, noise abnormality detection and redundancy remove, deposit Sensor database;Then, different Event Graph Models is trained according to different classes of event data, passes through data augmentat ion structure Occurrence diagram is built, it is final to establish comprising all types of sample event chart databases;The inquiry identification includes:The sensing data acquired in real time is handled and modeled, query event figure is built, with sample Sample graph in present event chart database carries out similarity query, identifies the type of the event.
- 10. the Activity recognition system similar based on figure according to claim 7, which is characterized in that the behavior pattern point Generic module is modeled and is inquired for consummatory behavior figure, wherein,The behavior figure modeling process includes:Based on obtaining when averagely continuing of corresponding sequence of events to the training of basic act Between;Then sequence of events is identified from real-time sensor data stream, then basic act is identified from sequence of events, according to Sequential relationship constructs behavior sequence;Multisequencing Algorithm based on weighting merges behavior sequence, is built into behavior Figure, finally obtains sample behavior chart database;The behavior figure inquiry is completed to carry out behavior figure similarity query in sample behavior chart database, finds and its behavior mould The similar behavior figure of formula, obtains the classification results of behavior pattern.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711275203.1A CN108182382A (en) | 2017-12-06 | 2017-12-06 | Based on the similar Activity recognition method and system of figure |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711275203.1A CN108182382A (en) | 2017-12-06 | 2017-12-06 | Based on the similar Activity recognition method and system of figure |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108182382A true CN108182382A (en) | 2018-06-19 |
Family
ID=62545661
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711275203.1A Pending CN108182382A (en) | 2017-12-06 | 2017-12-06 | Based on the similar Activity recognition method and system of figure |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108182382A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111524318A (en) * | 2020-04-26 | 2020-08-11 | 中控华运(厦门)集成电路有限公司 | Intelligent health condition monitoring method and system based on behavior recognition |
CN111695426A (en) * | 2020-05-08 | 2020-09-22 | 北京邮电大学 | Behavior pattern analysis method and system based on Internet of things |
CN111860661A (en) * | 2020-07-24 | 2020-10-30 | 中国平安财产保险股份有限公司 | Data analysis method and device based on user behavior, electronic equipment and medium |
CN116705180A (en) * | 2023-08-08 | 2023-09-05 | 山东北国发展集团有限公司 | N2O catalytic decomposition monitoring method and system based on multidimensional data analysis |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104102910A (en) * | 2014-08-07 | 2014-10-15 | 吉林农业大学 | Sports video tactical behavior recognition method based on space-time local mode |
-
2017
- 2017-12-06 CN CN201711275203.1A patent/CN108182382A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104102910A (en) * | 2014-08-07 | 2014-10-15 | 吉林农业大学 | Sports video tactical behavior recognition method based on space-time local mode |
Non-Patent Citations (1)
Title |
---|
陈虎: "基于图相似的传感器复杂事件检测技术研究", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111524318A (en) * | 2020-04-26 | 2020-08-11 | 中控华运(厦门)集成电路有限公司 | Intelligent health condition monitoring method and system based on behavior recognition |
CN111695426A (en) * | 2020-05-08 | 2020-09-22 | 北京邮电大学 | Behavior pattern analysis method and system based on Internet of things |
CN111695426B (en) * | 2020-05-08 | 2024-01-05 | 北京邮电大学 | Behavior pattern analysis method and system based on Internet of things |
CN111860661A (en) * | 2020-07-24 | 2020-10-30 | 中国平安财产保险股份有限公司 | Data analysis method and device based on user behavior, electronic equipment and medium |
CN111860661B (en) * | 2020-07-24 | 2024-04-30 | 中国平安财产保险股份有限公司 | Data analysis method and device based on user behaviors, electronic equipment and medium |
CN116705180A (en) * | 2023-08-08 | 2023-09-05 | 山东北国发展集团有限公司 | N2O catalytic decomposition monitoring method and system based on multidimensional data analysis |
CN116705180B (en) * | 2023-08-08 | 2023-10-31 | 山东北国发展集团有限公司 | N2O catalytic decomposition monitoring method and system based on multidimensional data analysis |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108182382A (en) | Based on the similar Activity recognition method and system of figure | |
CN102098175B (en) | Alarm association rule obtaining method of mobile internet | |
CN110147743A (en) | Real-time online pedestrian analysis and number system and method under a kind of complex scene | |
CN106844585A (en) | A kind of time-space relationship analysis system based on multi-source Internet of Things location aware | |
CN106599230A (en) | Method and system for evaluating distributed data mining model | |
CN100595782C (en) | Classification method for syncretizing optical spectrum information and multi-point simulation space information | |
CN106709035A (en) | Preprocessing system for electric power multi-dimensional panoramic data | |
CN111126658A (en) | Coal mine gas prediction method based on deep learning | |
CN1561503A (en) | Classification of objects through model ensembles | |
CN110135354A (en) | A kind of change detecting method based on outdoor scene threedimensional model | |
CN105139035A (en) | Mixed attribute data flow clustering method for automatically determining clustering center based on density | |
CN112291807A (en) | Wireless cellular network traffic prediction method based on deep migration learning and cross-domain data fusion | |
CN106649809A (en) | Internet of Things analysis method | |
CN109630095A (en) | A kind of rod-pumped well operating mode's switch method and system based on multi-angle of view study | |
CN112381171B (en) | Multi-sensor node missing data filling method based on combined model | |
CN111008337A (en) | Deep attention rumor identification method and device based on ternary characteristics | |
CN110188085A (en) | Quality data model method for building up between a kind of tobacco volume hired car | |
CN115376318B (en) | Traffic data compensation method based on multi-attribute fusion neural network | |
CN109582714A (en) | A kind of government affairs item data processing method based on time fading correlation | |
CN106777279A (en) | A kind of time-space relationship analysis system | |
CN110009416A (en) | A kind of system based on big data cleaning and AI precision marketing | |
CN105323024A (en) | Network signal intensity detecting and fusing method | |
Ramkumar et al. | A survey on mining multiple data sources | |
CN117172509A (en) | Construction project distribution system based on decoration construction progress analysis | |
CN108846043A (en) | Network trace mining analysis method and system based on internet big data |
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: 20180619 |
|
RJ01 | Rejection of invention patent application after publication |