CN109064484A - Crowd movement's Activity recognition method with momentum Fusion Features is divided based on subgroup - Google Patents
Crowd movement's Activity recognition method with momentum Fusion Features is divided based on subgroup Download PDFInfo
- Publication number
- CN109064484A CN109064484A CN201810236397.2A CN201810236397A CN109064484A CN 109064484 A CN109064484 A CN 109064484A CN 201810236397 A CN201810236397 A CN 201810236397A CN 109064484 A CN109064484 A CN 109064484A
- Authority
- CN
- China
- Prior art keywords
- point
- group
- sub
- signature tracking
- indicate
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20112—Image segmentation details
- G06T2207/20164—Salient point detection; Corner detection
Abstract
The invention discloses a kind of crowd movement's Activity recognition methods divided based on subgroup with momentum Fusion Features, this method is tracked first with angle point and the method for background modeling, obtain the space time information of moving target in video image frame, the area of space information of population distribution in Utilization prospects, spatially adjoining crowd is divided into several sub-groups, sub-group is further divided by the motion relevance in a period of time, obtains the sub-group with Movement consistency;Secondly on the basis of sub-group segmentation, three momentum features of crowd movement is extracted and are merged;Finally it is trained the pixel characteristic of the feature of fusion and video frame as the input of differential cyclic convolution neural network, using the method for handmarking by training video fragment label at different description vocabulary, with the result of the data point reuse differential cyclic convolution neural network of tape label, good training result is obtained, the motor behavior that can effectively identify crowd, reaches preferable effect.
Description
Technical field
The present invention relates to a kind of crowd movement's Activity recognition methods divided based on subgroup with momentum Fusion Features, mainly
Crowd movement track is extracted using Harris Corner Detection Algorithm, mixed Gaussian background modeling extracts the foreground features of scene, into
Row subgroup divides.Momentum feature extraction is carried out on the basis of subgroup, will extract the input of three-momentum feature video data
To differential cyclic convolution neural metwork training, it is converted into crowd behaviour label, reaches the target of crowd movement's Activity recognition, belongs to
Image procossing, video detection and artificial intelligence interleaving techniques application field.
Background technique
The purpose of crowd movement's Activity recognition is to pass through motion profile and foreground extraction from sequence image for dense population
It is divided into subgroup, crowd movement's Activity recognition is carried out on the basis of subgroup.To the activity recognition of group level increasingly at
For a hot issue of computer vision field, in intelligent video monitoring, public safety, sports etc. have extensively
Application.Mainly there are Harris Corner Detection Algorithm, mixed Gaussian for crowd movement's Activity recognition algorithm in video image frame
Background modeling method, momentum Feature fusion.
(1) Harris's Corner Detection Algorithm: the algorithm is carried out on any direction on the image using a fixed window
Latter two situation, the pixel grey scale variation degree in window, when there are the cunnings on any direction are compared before sliding and slide in sliding
Grey scale change degree is larger when dynamic, then it is assumed that there are angle points in the window.Angle point while retaining image graphics important feature,
The data volume that information can be effectively reduced keeps the content of its information very high, effectively improves the speed of calculating, is conducive to figure
The reliable matching of picture, so that being treated as possibility in real time.
(2) mixed Gaussian background modeling method: its basic thought is to compare input picture with background model, according to
The information such as difference judge not meet the abnormal conditions of background model, to distinguish foreground pixel and background pixel.This method
The distribution situation that gray value in an image is indicated with grey level histogram, utilizes this statistical result, it is assumed that picture in image sequence
The distribution Normal Distribution function of plain gray value, is split image.
(3) momentum Feature fusion: this method is from crowd's subgroup level, using the group of crowd massing as grinding
Study carefully target, constructs the momentum feature as unit of subgroup.All sons in scene are detected by greatest hope optimization algorithm
Group extracts the momentum features such as the communality, stability and conflicting of group for the group detected, constitutes towards subgroup
The momentum feature of group, group's momentum feature can improve the dependence due to personal momentum feature to scene, with group's unit structure
The independent momentum feature of scene is built, the robustness and scalability of crowd movement's behavioural analysis are promoted.
Summary of the invention
Present invention aims at crowd movement's Activity recognition methods, propose a kind of based on subgroup division and momentum feature
Crowd movement's Activity recognition method of fusion is solved and is divided caused by crowd's overlapping in microcosmic segmentation under the crowd is dense scene
The problem of crowd behaviour details caused by segmentation granularity is excessive in fault and macroscopic view segmentation is ignored, and propose on this basis
Carry out group movement momentum Fusion Features training pattern, can effectively identify the motor behavior of crowd.
It is of the present invention a kind of to be divided based on subgroup and crowd movement's Activity recognition methods of momentum Fusion Features includes
Following steps:
Step 1): user inputs continuous video, video is divided into continuous video frame, by each above-mentioned video frame
Motion information per single pedestrian as signature tracking a point P, point P is with a four dimensional vector P=(Px,Py,Pv,Pd) carry out table
Show, the Px、PyIndicate the space coordinate of character trace point, PvIndicate the displacement of the point, PdIndicate the movement of the point
Direction, PdValue isThe point set of all signature tracking points of picture frame is denoted as OI={ P1,P2,P3,P4};
Step 2): the moving characteristic of sub-group is determined by momentum feature, with the intracorporal signature tracking point of sub-group and subgroup
Based on, define three kinds of different momentum features: direction of motion consistency, spatial stability, crowd's abrasion interference;Every height
It include H signature tracking point, i.e. C in groupk=(P1,P2,...,PH);
Step 3): calculating the average value of the description factor in continuous 5 frame, constructs a vector with three average values The image of a triple channel is collectively constituted, 224 × 224 × 3 dimension datas is formed and is input to differential recurrence
Convolutional neural networks DRCNN is trained, and is converted into 4096 dimensional feature vectors, the differential cyclic convolution neural network be by
The shot and long term that (Group -16 the Visual Geometry) model of VGG -16 and 3 layer heaps are folded remembers recurrent neural network LSTM
It is connected in end-to-end model, crowd behaviour label is finally converted for feature vector using output function, using artificial mark
According to behavior main body, behavior scene, the difference of behavior itself, labeled as difference are occurred for training video segment by the method for note
Description vocabulary, with the data point reuse differential cyclic convolution neural network of tape label as a result, realizing crowd movement's Activity recognition.
Wherein,
The step 1) specifically:
Step 1.1): obtaining the location information of signature tracking point in successive video frames by Harris's Corner Detection Algorithm,
The foreground features of target group are obtained, Harris's angle point track algorithm is to carry out any side on the image using a fixed window
Latter two situation, the pixel grey scale variation degree in window, when there are any directions are compared before sliding and slide in upward sliding
On sliding when grey scale change degree it is larger, then it is assumed that there are angle points in the window, by signature tracking each in successive video frames
The position of point is together in series, and obtains the motion profile T of each signature tracking point, all signature tracking point motion profile collection are combined into
TI={ T1,T2,T3,T4};
Step 1.2): foreground extraction is carried out using mixed Gaussian background modeling method, for pixel in current video frame
Gray value, when the difference of the mean value of the S Gaussian Profile in Gaussian Mixture background meets formula:
Be considered as successful match, i.e., the pixel is background, wherein I (x, y, t) indicate the pixel (x, y) in the pixel value of t moment,
Indicate that the average gray value of the S Gaussian Profile t moment, λ indicate the multiple coefficient of standard deviation,Indicate the S Gaussian Profile t
The variance of moment gray value, the space size and distance relation with surrounding group that target group are obtained by foreground extraction,
Referred to as prospect patch, patch set are denoted as BI={ B1,B2,B3,...,Bk, it is marked off spatially by relationship change spatially
The individual closed on;
Step 1.3): using the two kinds of space time informations marked off, pod is divided;
The step 1.3) specifically:
Step 1.3.1): by the set O comprising signature tracking pointIIf being divided into the subset comprising doing, it is expressed as OI=
{CI,FI, the CI={ C1,C2,...,CKIt is the point set that picture frame has Movement consistency after dividing, it constitutes crowd and draws
The sub-group divided, point set FIIt is then the point being removed, the spatial information of patch is indicated with rectangular area, with patch boundary pixel position
It sets coordinate value and marks off a rectangular area, the coordinate of the rectangle lower right corner and upper left angle point is obtained, if the coordinate of signature tracking point P
Position is in the profile of patch, then the point is divided into the point of the patch, otherwise rejects the point;
Step 1.3.2): the attribute P of signature tracking point PdIt indicates the movement relation of signature tracking point between frame and frame, calculates
The displacement vector values of signature tracking point P and the cosine value of X-axis angle, obtain angular separation θ, 0~2 π of angular separation are divided into
12 equal portions, each section are respectively labeled as Di(i=1,2,3..., 12) is PdAssignment, specific division methods formula are as follows:
By the division of patch range and the constraint of the direction of motion, the point of signature tracking similar in movement tendency is divided into one
A sub-group;
Step 1.3.3): abnormal point is corrected, in K point of proximity for calculating signature tracking point P, the attribute value P of each point of proximityd
The frequency of occurrences, the maximum mark value of the frequency of occurrences are Di, signature tracking point PdIt is denoted as Dj, for all signature tracking points, work as i+
1=j or i-1=j, by the P of PdValue is modified to Di;Rejecting abnormalities point: it calculates and is transported in the L point of proximity of signature tracking point P with it
The number I of the dynamic identical point of trend, sets critical value M, a M≤L, as I < M, P point is considered abnormal point, from sub-group Ck
Middle rejecting.
The step 2) specifically:
Step 2.1): the coordinate of each signature tracking point and position in sub-group direction of motion consistency feature extraction: are calculated
Average value is pipetted, the center-of-mass coordinate and average displacement of each sub-group in continuous several frames is obtained, acquires the totality of each sub-group
Movement tendency vectorAccording to formula:The overall movement for calculating each sub-group becomes
The velocity-dependent of gesture vector and the vector, whereinIndicate the movement tendency vector of each sub-group,Indicate k-th of neighbour
The movement tendency vector of contact, N represents the sub-group number marked off, describedIt is higher to be worth bigger expression velocity-dependent;_
Step 2.2): spatial stability momentum feature extraction: spatial stability refers to each signature tracking o'clock in a timing
Between stable neighbours are kept in range, keep specific topological structure within a certain period of time;Define sub-group in each feature with
Track point PiT moment stability by formula:It acquires, wherein N indicates to draw
The sub-group number separated, PiIndicate ith feature trace point, | N1(Pi)\N1(Pi)T| it indicates 1 into T moment sub-group
A signature tracking point abutment points in maintain to stablize the mean number of the constant point of syntople with it, K indicates most adjacent
The number of point;Stability of each signature tracking point at a distance from the abutment points in its neighborhood can use formula in sub-group:It indicates, wherein N indicates the number of the sub-group marked off, PiIndicate ith feature trace point,Generation
The average distance of table signature tracking point and its k point of proximity, above two stability is added, and constitutes subgroup monolithic stability
Property, with ω (Ck) indicate;
Step 2.3): crowd's abrasion interference momentum feature extraction: conflicting calculation formula are as follows:Wherein N indicates the number of the sub-group marked off,
PiIndicate ith feature trace point,Indicate the movement tendency vector of each sub-group,Indicate the movement of k-th of abutment points
Trend vector, avrg (Nother(Pi)) represent in the adjoining of the signature tracking point in a sub-group comprising in other sub-groups
The average value of signature tracking point, α and β are weight coefficient.
Step 1.2) Plays difference multiple coefficient lambda takes 2.5.
The utility model has the advantages that a kind of crowd movement's behavior knowledge divided based on subgroup with momentum Fusion Features proposed by the present invention
Other method, specifically has the beneficial effect that:
1, the present invention is to carrying out subgroup with crowd under the crowd is dense scene and divide from the angle of sub-group to be analyzed not
The difficulty that only can solve the disengaging movement individual from pod can also obtain group being considered as an entirety as research
The internal feature ignored when object.
2, the present invention proposes a kind of algorithm based on space-time restriction, by the temporal movement in athletic group between individual
The foundation that relationship and propinquity spatially are divided as group, two kinds of conditions mutually constrain, and group, which is divided into, has movement
The sub-group of consistency.This algorithm can be suitable for the monitoring video flow of different groups density and observation visual angle, and realize letter
Single, the speed of service is fast.
3, the present invention is trained the three dimensional video data taken out using differential cyclic convolution neural network, by feature
The step of extraction and parameter learning, separates, group's video suitable for various resolution ratio and mobile context.
Detailed description of the invention
Fig. 1 is subgroup cluster partition algorithm flow chart.
Fig. 2 is momentum Feature Fusion Algorithm flow chart.
Specific embodiment
The some embodiments of attached drawing of the present invention are described in more detail below.
User inputs 15 seconds continuous videos, and a picture frame, i.e. D were divided into every 0.5 secondI={ D1,D2,D3,...,
D30, DtFor the picture frame of t moment, the crowd with Movement consistency is subjected to subgroup division under the crowd is dense scene,
Algorithm flow chart is as shown in Figure 1.
Preceding 15 frame video image frame is taken, per single pedestrian as a tracking characteristics trace point, with a four dimensional vector P
=(Px,Py,Pv,Pd) indicate the motion information of signature tracking point P, wherein Px,PyIndicate the space coordinate of signature tracking point, PvTable
Show the displacement of the point, PdIndicate that the direction of motion of the point, value are15 are obtained by Harris Corner Detection Algorithm
The co-ordinate position information of signature tracking point in video frame connects the coordinate position of signature tracking point each in successive video frames
The motion profile T of each signature tracking point is obtained, all signature tracking point motion profile collection are combined into TI={ T1,T2,T3,T4,
Every track includes the set T of several signature tracking pointsi={ P1,P2,P3,...,Pk, continuous motion profile embodies one section
The movement relation of time in-group.
Foreground extraction is carried out using mixed Gaussian background modeling method, for the gray value of pixel in current video frame, such as
The difference of the mean value of k-th Gaussian Profile meets formula in fruit and Gaussian Mixture background:It is considered as
Successful match, the i.e. pixel are background, and wherein λ is the multiple coefficient of standard deviation, and λ value is 2.5.It is obtained by foreground extraction
The space size of target group and distance relation with surrounding group, referred to as prospect patch, patch set are denoted as BI={ B1,
B2,B3,...,Bk, the individual spatially closed on is marked off by relationship change spatially.
Using the two kinds of space time informations marked off, pod is divided.By the set O comprising signature tracking pointI
If being divided into the subset comprising doing, it is expressed as OI={ CI,FI, wherein CI={ C1,C2,...,CKIt is picture frame by dividing
Afterwards with the point set of Movement consistency, the sub-group that crowd divides, point set F are constitutedIIt is then the point being removed.It is indicated with rectangle frame
The spatial information of patch marks off a rectangular area with patch boundary location of pixels coordinate value, obtains the rectangle lower right corner and a left side
The point is divided into the patch if the coordinate position of signature tracking point P is in the profile of patch by the coordinate of upper angle point
Point, otherwise reject the point.The attribute P of signature tracking point PdIndicate the movement relation of signature tracking point between frame and frame.First
Calculate the motion vector of signature tracking point PWith the cosine value of X-axis angle, to obtain angular separation θ.It will be square
12 equal portions are divided into 0~2 π of angle, each section is respectively labeled as Di(i=1,2,3 ..., 12) is PdAssignment is specific to divide
Method are as follows:
By the division of patch range and the constraint of the direction of motion, the point of signature tracking similar in these movement tendencies is divided
At a sub-group.
Abnormal point is modified and is rejected.Firstly, carrying out the amendment of abnormal point, K for calculating signature tracking point P face
In near point, the attribute value P of each point of proximitydThe maximum mark value of the frequency of occurrences is Di, signature tracking point PdIt is denoted as Dj, for all
Signature tracking point, if i+1=j or i-1=j, then just by the P of PdValue is modified to Di;Secondly, abnormity point elimination is calculated
The number I of the identical point of trend is moved in the L point of proximity of signature tracking point P.A critical value M (M≤L) is set, M
Value is 5, if I < M, P point is considered abnormal point, from sub-group CkMiddle rejecting.
If the moving characteristic momentum feature of sub-group determines, based on the intracorporal signature tracking point of sub-group and subgroup,
Define three kinds of different momentum features: direction of motion consistency, spatial stability, crowd's abrasion interference.Assuming that each subgroup
It include H signature tracking point, i.e. C in bodyk=(P1,P2,...,PH)。
Direction of motion consistency feature extraction: the coordinate of each signature tracking point and displacement in sub-group are calculated and is averaged
Value obtains the center-of-mass coordinate and average displacement of each sub-group in continuous several frames, acquires the overall movement trend of each sub-group
VectorAccording to formula:Calculate the overall movement trend vector of each sub-group with
The velocity-dependent of the vector, whereinIndicate the movement tendency vector of each sub-group,Indicate the fortune of k-th of abutment points
Dynamic trend vector, N represents the sub-group number marked off, describedIt is higher to be worth bigger expression velocity-dependent.
Spatial stability momentum feature extraction: spatial stability refers to that each signature tracking point is protected within certain time
Keep steady fixed neighbours, keeps specific topological structure within a certain period of time;Define each signature tracking point P in sub-groupiIn t
It carves, the field of K most abutment points compositions is Nt(Pi), stability can be by such as formula:It acquires, wherein N indicates the sub-group number marked off, PiIt indicates i-th
Signature tracking point, | N1(Pi)\N1(Pi)T| expression is tieed up in 1 to a characteristic point in T moment sub-group abutment points with it
It keeps steady and determines the mean number of the constant point of syntople, K indicates the number of most abutment points;In sub-group each signature tracking point with
The stability of the distance of abutment points in its neighborhood can use formula:It indicates, wherein N expression marks off
Sub-group number, PiIndicate ith feature trace point,The average distance of signature tracking point and its k point of proximity is represented,
Above two stability is added, subgroup overall stability is constituted, with ω (Ck) indicate.
Group's abrasion interference momentum feature extraction: conflicting calculation formula are as follows: conflicting calculation formula are as follows:Wherein N indicates the number of the sub-group marked off,
PiIndicate ith feature trace point, whereinIndicate the movement tendency vector of each sub-group,Indicate k-th of abutment points
Movement tendency vector, avrg (Nother(Pi)) represent in the adjoining of the signature tracking point in a sub-group comprising other subgroups
The average value of signature tracking point in body, α and β are weight coefficient.
The mean value calculation of the description factor in preceding 5 frame is come out, constructs a vector with three average valuesSimilar to the rgb value of pixel, the image of a triple channel is collectively constituted, form 224 × 224 ×
3 dimension datas are input to differential cyclic convolution neural network (DRCNN) and are trained, and are converted into 4 096 as feature vector.It is wherein micro-
The shot and long term memory for dividing cyclic convolution neural network VGG -16 (Group -16 Visual Geometry) model and 3 layer heaps to fold
Recurrent neural network (Long Short-Term Memory, LSTM) is connected in end-to-end model, and it is trained accurate to improve
Rate finally uses output function to convert crowd behaviour label for feature vector, using the method for handmarking by training video
According to behavior main body, behavior scene occur for segment, and the difference of behavior itself is marked labeled as different description vocabulary with band
The data point reuse differential cyclic convolution neural network of note as a result, realize crowd movement's Activity recognition.
Claims (5)
1. a kind of crowd movement's Activity recognition method divided based on subgroup with momentum Fusion Features, which is characterized in that the party
Method the following steps are included:
Step 1): user inputs continuous video, video is divided into continuous video frame, by the above-mentioned every list of each video frame
A pedestrian as signature tracking a point P, point P motion information with a four dimensional vector P=(Px, Py, Pv, Pd) indicate, institute
State Px、PyIndicate the space coordinate of character trace point, PvIndicate the displacement of the point, PdIndicate the direction of motion of the point,
PdValue isThe point set of all signature tracking points of picture frame is denoted as OI={ P1, P2, P3, P4};
Step 2): the moving characteristic of sub-group is determined by momentum feature, using the intracorporal signature tracking point of sub-group and subgroup as base
Plinth defines three kinds of different momentum features: direction of motion consistency, spatial stability, crowd's abrasion interference;Each sub-group
In include H signature tracking point, i.e. Ck=(P1, P2..., PH);
Step 3): calculating the average value of the description factor in continuous 5 frame, constructs a vector with three average valuesω
(C), ρ (C)), the image of a triple channel is collectively constituted, 224 × 224 × 3 dimension datas is formed and is input to differential recursive convolution mind
It is trained through network DRCNN, is converted into 4096 dimensional feature vectors, the differential cyclic convolution neural network is by VGG-16 mould
The shot and long term memory recurrent neural network LSTM that type and 3 layer heaps are folded is connected in end-to-end model, finally uses output function
Crowd behaviour label is converted by feature vector, is led training video segment according to behavior using the method for handmarking
Body, behavior scene, the difference of behavior itself are followed labeled as different description vocabulary with the data point reuse differential of tape label
Ring convolutional neural networks as a result, realize crowd movement's Activity recognition.
2. a kind of crowd movement Activity recognition side divided based on subgroup with momentum Fusion Features according to claim 1
Method, which is characterized in that the step 1) specifically:
Step 1.1): obtaining the location information of signature tracking point in successive video frames by Harris's Corner Detection Algorithm, obtains
The foreground features of target group, Harris's angle point track algorithm are carried out on any direction on the image using a fixed window
Sliding, compare with sliding latter two situation before sliding, pixel grey scale variation degree in window, when there are on any direction
Grey scale change degree is larger when sliding, then it is assumed that there are angle points in the window, by signature tracking point each in successive video frames
Position is together in series, and obtains the motion profile T of each signature tracking point, and all signature tracking point motion profile collection are combined into TI=
{T1, T2, T3, T4};
Step 1.2): foreground extraction is carried out using mixed Gaussian background modeling method, for the gray scale of pixel in current video frame
Value, when the difference of the mean value of the S Gaussian Profile in Gaussian Mixture background meets formula:Just recognize
For successful match, i.e., the pixel is background, wherein I (x, y, t) indicate the pixel (x, y) in the pixel value of t moment,It indicates
The average gray value of the S Gaussian Profile t moment, λ indicate the multiple coefficient of standard deviation,Indicate the S Gaussian Profile t moment
The variance of gray value, by the space size of foreground extraction acquisition target group and the distance relation with surrounding group, referred to as
Prospect patch, patch set are denoted as BI={ B1, B2, B3..., Bk, it is marked off by relationship change spatially and is spatially closed on
Individual;
Step 1.3): using the two kinds of space time informations marked off, pod is divided;
3. a kind of crowd movement Activity recognition side divided based on subgroup with momentum Fusion Features according to claim 1
Method, which is characterized in that the step 1.3) specifically:
Step 1.3.1): by the set O comprising signature tracking pointIIf being divided into the subset comprising doing, it is expressed as OI={ CI,
FI, the CI={ C1, C2..., CKIt is that picture frame divides after dividing with the point set of Movement consistency, composition crowd
Sub-group, point set FIIt is then the point being removed, the spatial information of patch is indicated with rectangular area, with patch boundary location of pixels seat
Scale value marks off a rectangular area, the coordinate of the rectangle lower right corner and upper left angle point is obtained, if the coordinate position of signature tracking point P
In profile in patch, then the point is divided into the point of the patch, otherwise rejects the point;
Step 1.3.2): the attribute P of signature tracking point PdIt indicates the movement relation of signature tracking point between frame and frame, calculates feature
The displacement vector values of trace point P and the cosine value of X-axis angle, obtain angular separation θ, 0~2 π of angular separation are divided into 12 etc.
Part, each section is respectively labeled as Di(i=1,2,3..., 12) is PdAssignment, specific division methods formula are as follows:
By the division of patch range and the constraint of the direction of motion, the point of signature tracking similar in movement tendency is divided into a son
Group;
Step 1.3.3): abnormal point is corrected, in K point of proximity for calculating signature tracking point P, the attribute value P of each point of proximitydOccur
Frequency, the maximum mark value of the frequency of occurrences are Di, signature tracking point PdIt is denoted as Dj, for all signature tracking points, work as i+1=j
Or i-1=j, by the P of PdValue is modified to Di;Rejecting abnormalities point: it calculates and moves in the L point of proximity of signature tracking point P
The number I of the identical point of gesture sets critical value M, a M≤L, as I < M, P point is considered abnormal point, from sub-group CkIn
It rejects.
4. a kind of crowd movement Activity recognition side divided based on subgroup with momentum Fusion Features according to claim 1
Method, which is characterized in that the step 2) specifically:
Step 2.1): direction of motion consistency feature extraction: the coordinate of each signature tracking point and displacement in sub-group are calculated and is taken
Average value obtains the center-of-mass coordinate and average displacement of each sub-group in continuous several frames, acquires the overall movement of each sub-group
Trend vectorAccording to formula:Calculate the overall movement trend of each sub-group to
The velocity-dependent of amount and the vector, whereinIndicate the movement tendency vector of each sub-group,Indicate k-th of abutment points
Movement tendency vector, N represents the sub-group number that marks off, describedIt is higher to be worth bigger expression velocity-dependent;_
Step 2.2): spatial stability momentum feature extraction: spatial stability refers to each signature tracking point in certain time model
The stable neighbours of interior holding are enclosed, keep specific topological structure within a certain period of time;Define each signature tracking point in sub-group
PiT moment stability by formula:It acquires, wherein N expression marks off
Sub-group number, PiIndicate ith feature trace point, | N1(Pi)\N1(Pi)T| it indicates 1 to one in T moment sub-group
Maintain to stablize the mean number of the constant point of syntople in the abutment points of a signature tracking point with it, K indicates most abutment points
Number;Stability of each signature tracking point at a distance from the abutment points in its neighborhood can use formula in sub-group:It indicates, wherein N indicates the number of the sub-group marked off, PiIndicate ith feature trace point,Generation
The average distance of table signature tracking point and its k point of proximity, above two stability is added, and constitutes subgroup monolithic stability
Property, with ω (Ck) indicate;
Step 2.3): crowd's abrasion interference momentum feature extraction: conflicting calculation formula are as follows:Wherein N indicates the number of the sub-group marked off,
PiIndicate ith feature trace point,Indicate the movement tendency vector of each sub-group,Indicate the movement of k-th of abutment points
Trend vector, avrg (Nother(Pi)) represent in the adjoining of the signature tracking point in a sub-group comprising special in other sub-groups
The average value of trace point is levied, α and β are weight coefficient.
5. a kind of crowd movement Activity recognition side divided based on subgroup with momentum Fusion Features according to claim 2
Method, which is characterized in that step 1.2) Plays difference multiple coefficient lambda takes 2.5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810236397.2A CN109064484B (en) | 2018-03-21 | 2018-03-21 | Crowd movement behavior identification method based on fusion of subgroup component division and momentum characteristics |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810236397.2A CN109064484B (en) | 2018-03-21 | 2018-03-21 | Crowd movement behavior identification method based on fusion of subgroup component division and momentum characteristics |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109064484A true CN109064484A (en) | 2018-12-21 |
CN109064484B CN109064484B (en) | 2022-02-08 |
Family
ID=64819919
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810236397.2A Active CN109064484B (en) | 2018-03-21 | 2018-03-21 | Crowd movement behavior identification method based on fusion of subgroup component division and momentum characteristics |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109064484B (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109829382A (en) * | 2018-12-30 | 2019-05-31 | 北京宇琪云联科技发展有限公司 | The abnormal object early warning tracing system and method for Behavior-based control feature intelligent analysis |
CN109918996A (en) * | 2019-01-17 | 2019-06-21 | 平安科技(深圳)有限公司 | The illegal action identification method of personnel, system, computer equipment and storage medium |
CN110263690A (en) * | 2019-06-12 | 2019-09-20 | 成都信息工程大学 | A kind of group behavior feature extraction based on small group and description method and system |
CN110443789A (en) * | 2019-08-01 | 2019-11-12 | 四川大学华西医院 | A kind of foundation and application method of immunofixation electrophoresis figure automatic identification model |
CN110472604A (en) * | 2019-08-20 | 2019-11-19 | 中国计量大学 | A kind of pedestrian based on video and crowd behaviour recognition methods |
CN111160150A (en) * | 2019-12-16 | 2020-05-15 | 盐城吉大智能终端产业研究院有限公司 | Video monitoring crowd behavior identification method based on depth residual error neural network convolution |
CN111429185A (en) * | 2020-03-27 | 2020-07-17 | 京东城市(北京)数字科技有限公司 | Crowd portrait prediction method, device, equipment and storage medium |
CN111933298A (en) * | 2020-08-14 | 2020-11-13 | 医渡云(北京)技术有限公司 | Crowd relation determination method, device, electronic equipment and medium |
CN112850436A (en) * | 2019-11-28 | 2021-05-28 | 宁波微科光电股份有限公司 | Pedestrian trend detection method and system of elevator intelligent light curtain |
CN113283387A (en) * | 2021-06-23 | 2021-08-20 | 华东交通大学 | Group abnormal behavior detection method and device |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106529465A (en) * | 2016-11-07 | 2017-03-22 | 燕山大学 | Pedestrian cause and effect relation identification method based on momentum kinetic model |
-
2018
- 2018-03-21 CN CN201810236397.2A patent/CN109064484B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106529465A (en) * | 2016-11-07 | 2017-03-22 | 燕山大学 | Pedestrian cause and effect relation identification method based on momentum kinetic model |
Non-Patent Citations (2)
Title |
---|
HAKAN BILEN等: "Dynamic Image Networks for Action Recognition", 《CVPR》 * |
成金庚等: "结合群组动量特征与卷积神经网络的人群行为分析", 《科学技术与工程》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109829382A (en) * | 2018-12-30 | 2019-05-31 | 北京宇琪云联科技发展有限公司 | The abnormal object early warning tracing system and method for Behavior-based control feature intelligent analysis |
CN109829382B (en) * | 2018-12-30 | 2020-04-24 | 北京宇琪云联科技发展有限公司 | Abnormal target early warning tracking system and method based on intelligent behavior characteristic analysis |
CN109918996A (en) * | 2019-01-17 | 2019-06-21 | 平安科技(深圳)有限公司 | The illegal action identification method of personnel, system, computer equipment and storage medium |
CN110263690A (en) * | 2019-06-12 | 2019-09-20 | 成都信息工程大学 | A kind of group behavior feature extraction based on small group and description method and system |
CN110443789A (en) * | 2019-08-01 | 2019-11-12 | 四川大学华西医院 | A kind of foundation and application method of immunofixation electrophoresis figure automatic identification model |
CN110443789B (en) * | 2019-08-01 | 2021-11-26 | 四川大学华西医院 | Method for establishing and using immune fixed electrophoretogram automatic identification model |
CN110472604B (en) * | 2019-08-20 | 2021-05-14 | 中国计量大学 | Pedestrian and crowd behavior identification method based on video |
CN110472604A (en) * | 2019-08-20 | 2019-11-19 | 中国计量大学 | A kind of pedestrian based on video and crowd behaviour recognition methods |
CN112850436A (en) * | 2019-11-28 | 2021-05-28 | 宁波微科光电股份有限公司 | Pedestrian trend detection method and system of elevator intelligent light curtain |
CN111160150A (en) * | 2019-12-16 | 2020-05-15 | 盐城吉大智能终端产业研究院有限公司 | Video monitoring crowd behavior identification method based on depth residual error neural network convolution |
CN111429185A (en) * | 2020-03-27 | 2020-07-17 | 京东城市(北京)数字科技有限公司 | Crowd portrait prediction method, device, equipment and storage medium |
CN111429185B (en) * | 2020-03-27 | 2023-06-02 | 京东城市(北京)数字科技有限公司 | Crowd figure prediction method, device, equipment and storage medium |
CN111933298A (en) * | 2020-08-14 | 2020-11-13 | 医渡云(北京)技术有限公司 | Crowd relation determination method, device, electronic equipment and medium |
CN111933298B (en) * | 2020-08-14 | 2024-02-13 | 医渡云(北京)技术有限公司 | Crowd relation determining method and device, electronic equipment and medium |
CN113283387A (en) * | 2021-06-23 | 2021-08-20 | 华东交通大学 | Group abnormal behavior detection method and device |
Also Published As
Publication number | Publication date |
---|---|
CN109064484B (en) | 2022-02-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109064484A (en) | Crowd movement's Activity recognition method with momentum Fusion Features is divided based on subgroup | |
CN104077605B (en) | A kind of pedestrian's search recognition methods based on color topological structure | |
CN104867161B (en) | A kind of method for processing video frequency and device | |
CN105528794B (en) | Moving target detecting method based on mixed Gauss model and super-pixel segmentation | |
CN109035293B (en) | Method suitable for segmenting remarkable human body example in video image | |
CN104268583B (en) | Pedestrian re-recognition method and system based on color area features | |
CN105893936B (en) | A kind of Activity recognition method based on HOIRM and Local Feature Fusion | |
CN105320917B (en) | A kind of pedestrian detection and tracking based on head-shoulder contour and BP neural network | |
CN107103326A (en) | The collaboration conspicuousness detection method clustered based on super-pixel | |
CN101599179A (en) | Method for automatically generating field motion wonderful scene highlights | |
CN103929685A (en) | Video abstract generating and indexing method | |
CN107153824A (en) | Across video pedestrian recognition methods again based on figure cluster | |
CN107273905A (en) | A kind of target active contour tracing method of combination movable information | |
CN105046720B (en) | The behavior dividing method represented based on human body motion capture data character string | |
CN107341445A (en) | The panorama of pedestrian target describes method and system under monitoring scene | |
CN107230188A (en) | A kind of method of video motion shadow removing | |
CN109766822A (en) | Gesture identification method neural network based and system | |
CN103853794B (en) | Pedestrian retrieval method based on part association | |
Komorowski et al. | Deepball: Deep neural-network ball detector | |
CN108830882A (en) | Video abnormal behaviour real-time detection method | |
CN110363197A (en) | Based on the video area-of-interest exacting method for improving visual background extraction model | |
CN105956604A (en) | Action identification method based on two layers of space-time neighborhood characteristics | |
Fang et al. | Pedestrian attributes recognition in surveillance scenarios with hierarchical multi-task CNN models | |
CN110322479B (en) | Dual-core KCF target tracking method based on space-time significance | |
Zhang et al. | An Improved Computational Approach for Salient Region Detection. |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |