CN207529395U - A kind of body gait behavior active detecting identifying system folded based on semanteme - Google Patents
A kind of body gait behavior active detecting identifying system folded based on semanteme Download PDFInfo
- Publication number
- CN207529395U CN207529395U CN201721209457.9U CN201721209457U CN207529395U CN 207529395 U CN207529395 U CN 207529395U CN 201721209457 U CN201721209457 U CN 201721209457U CN 207529395 U CN207529395 U CN 207529395U
- Authority
- CN
- China
- Prior art keywords
- gait
- semanteme
- raspberry
- behavior
- parameter
- 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.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
The utility model discloses a kind of body gait behavior active detecting identifying systems folded based on semanteme.Utilize the embedded gait behavioral value identifying system hardware of the structure low-power consumption such as three-dimensional sound field alignment system, vola field of force alignment system, HDMI high-definition cameras, high definition video collecting system and microcomputer Raspberry；The utility model proposes the gait semanteme energy diagram with temporal aspect, comprising the gait time information under different situations, by largely there is the gait semanteme energy diagram of temporal aspect, study and the predictive ability of gait behavior cognitive system can be enhanced.Meanwhile the utility model proposes the body gait behavior active detecting identification technology folded based on semanteme, is all with a wide range of applications in every field, mainly includes：The fields such as remote identification, abnormal gait behavioral value, pedestrian behavior prediction and massive video retrieval, have good economic and social benefit.
Description
Technical field
The utility model is related to a kind of body gait behavior active detecting identifying systems folded based on semanteme.
Background technology
The identification of gait behavioral value is that the human body behavior carried out using body gait, movement posture and body as main feature is examined
It surveys, analysis, understand and predict.
Gait behavioral value identification technology all shows the prospect of being widely applied in the every field of life, for example (1) is far
Apart from identification：Identification and certification of the remote non-contact lower completion to personnel identity；(2) abnormal gait behavioral value：It is main
Abnormal behaviour screening and analysis are carried out, hazardous act is warned, to improve security protection safeguard level under public arena；(3)
Pedestrian behavior is predicted：The behavior of pedestrian is predicted in real time, decision-making foundation is provided for the systems such as unmanned；(4) magnanimity regards
Frequency is retrieved：The screening of video is realized by body gait behavior, finds suspected crime molecule；(5) the elderly guards：Pass through step
State behavior falls down, suffers a shock and the situations such as empty hair disease monitor that the elderly faces.
Currently to the detection of body gait behavior, according to data source, wearable body gait behavioral value can be divided into
The method of method and non-wearing.Wearable method needs human body to wear the detection device of feature, mainly including motion sensor,
The components such as controller and battery, it needs the cooperation of human body, is not suitable for the application under remote and non-contact situation.
Non- wearable method is mainly tracked, detects, analyzes and is identified the gait of human body by video and image data
Behavior, it does not need to cooperating on one's own initiative for detection target, is a kind of contactless detection method.According in human testing and processing
In the process, if to use manikin, and method based on model and the method for non-model (statistics) can be divided into.
Using wearable sports equipment, to detect and identify body gait behavior, there are many insufficient：Be not suitable for making in a wide range of
With particularly offender will not go to dress；Since the influences such as the size of wearable device and comfort level make people be unwilling to wear；
Human body movement data can only be got, it is impossible to the visual informations such as the image of human body are obtained, so as to lack effective visual analysis hand
Section.
Non- wearable mode, by being disposed in the camera of each position come detection and analysis human body.It is but traditional
Camera often all using fixed viewpoint and passive style of shooting, do not have active probe function, only when object appears in
Camera could find object, therefore moving object can actively avoid camera and reach the mesh for hiding monitoring within sweep of the eye
's.
Being divided into as unrelated and based on model the side of model of gait behavioral value is carried out by acquiring movement human image
Method.Wherein, the unrelated gait behavioral value method of model, gait behavior is analyzed and identifies by counting gait profile information.
Its shortcoming is to handle the influence of various covariant factors (block, wear clothes and visual angle etc.), detection and recognition effect well
It is showed under complex scene poor.Better performance is had when handling visual angle and covariant factor based on the method for model, but
It is to study the manikin (articulated model, hinge model, model of ellipse, joint skeleton model and surface shell model) used now
Lack body characteristics or low precision, and video and image are again too high as a kind of non-structured data dimension, greatly
Influence the effect to gait behavioral value and identification.
In short, although the research of gait and Activity recognition achieves many achievements, but due to the walking and movement of human body
Posture is influenced by various factors, for example deliberately hides shortage of data caused by camera shooting, and complex scene transformation increases people
Body divides difficulty, and clothing condition changes, blocks the superposition of the subjective and objective factors such as interference and different visual angles so that gait Activity recognition
The versatility and discrimination of algorithm be not still high, can not far be compared with human brain.
Utility model content
The technical issues of in order to solve the versatility and insufficient discrimination of current gait behavioral value identification, the utility model
A kind of body gait behavior active detecting identifying system folded based on semanteme that can realize accurate gait Activity recognition is provided
And method.
In order to realize above-mentioned technical purpose, the technical solution of the utility model is,
A kind of body gait behavior active detecting identifying system folded based on semanteme installs solid mechanical knot including system
Structure, three-dimensional sound field alignment system, plantar pressure field alignment system, the high definition video collecting and processing system with holder, large capacity magnetic
Disk array memory, Raspberry Pi microcomputer systems, high-speed mobile communications module and big data service hardware platform；
The system installation solid mechanical structure includes rotatable chassis module 11 and system mounting framework and bottom plate 5,
The system mounting framework and bottom plate 5 is fixed in rotatable chassis module 11；
The plantar pressure field alignment system includes the vola field of force alignment sensor array being made of multiple sensors
13, the vola field of force alignment sensor array 13 is set on ground；
The three-dimensional sound field alignment system includes multiple sound collection sensors, voice signal modulate circuit, embedded
Controller and wireless transport module, sound collection sensor ring system for winding installation solid mechanical structure setting are described
Embedded controller communicates to connect sound collection sensor, voice signal modulate circuit and wireless transport module respectively；
The Raspberry Pi microcomputer systems include Raspberry Pi microcomputers 4, bluetooth keyboard
7th, bluetooth mouse 10 and HDMI interface liquid crystal display 3, the bluetooth keyboard 7, bluetooth mouse 10 and HDMI interface liquid crystal
Show that device 3 is communicatively connected to Raspberry Pi microcomputers 4 respectively；
The high definition video collecting and processing system with holder includes HDMI high-definition cameras 1 and USB interfaces HDMI high
Clear video frequency collection card 2, the HDMI high-definition cameras 1 are communicatively connected to USB interface HDMI high definition video collectings card 2, described
USB interface HDMI high definition video collectings card 2 be communicatively connected to Raspberry Pi microcomputer systems；
The high-speed mobile communications module includes wireless WIFI module 8,4G high-speed communications module 9 and internet wireless road
By device 14, the wireless WIFI module 8,4G high-speed communications module 9 and internet wireless router 14 are communicatively connected to respectively
Raspberry Pi microcomputer systems；
The big data service hardware platform includes gait semanteme big data platform 15, the gait semanteme big data
Platform 15 is set to distal end and is connect with Raspberry Pi microcomputer system telecommunications；
The large capacity disc array memory is communicatively connected to Raspberry Pi microcomputer systems.
The system, the vola field of force sensor array are using camera as coordinate center, and monitoring area is drawn
It is divided into I × J small square area, vola field of force sensor, all I × J biography is placed in each grid spaces
Sensor forms sensor array.
A kind of body gait behavioral value recognition methods folded based on semanteme, using the system, including following step
Suddenly：
A, monitoring field is entered by three-dimensional sound field alignment system and plantar pressure field alignment system active probe and moves people
Body position information, and microcomputer Raspberry control systems are transferred data to, control system realizes video camera by holder
Rotation and track up；
B, structure unified standard not homomorphs and attitude data parametric human body gait behavior database, according to it is main into
Analysis and canonical correlation analysis determine the importance in gait behavioural analysis of Geometrical Parameter and skeletal joint parameter
Weights；
C, divide the movement human profile taken, choose body in three-dimensional gait behavior database and posture is close
Three-dimensional (3 D) manikin is parameterized, the naturally semantic parameter Estimation object function of three-dimensional (3 D) manikin is built therebetween, passes through this
Semantic objects function is realized under the conditions of missing or redundancy interference 2D body gait profiles to three-dimensional (3 D) manikin body and bone
Effective estimation of joint parameter value；
D, direct method is conjugated by optimizing the improvement of initial value, solves the naturally semantic parameter Estimation target letter of three-dimensional (3 D) manikin
Number, extracts gait pattern nature semantic description parameter；
E, the class brain semanteme folding encoder represented based on two-dimentional sparse distribution is carried out to gait behavior class brain semantic feature；
F, class brain and, class brain or the sampling of class brain and class brain polymerizing energy figure are carried out to gait behavior class brain grapheme
It calculates, and retrieval, filtering, classification and the identification of gait behavior is completed using public action identification database.
The method, the step A include step in detail below：
1) sound is acquired by three-dimensional sound field alignment system, denoising, divides, store and transmit function；
2) it acquires and demarcates location information, using camera as coordinate center, choose N number of different monitoring area coordinate Ln=
(Xn,Yn,Zn), n=1...N, artificial simulates human motion noise at selected coordinate, utilizes three dimensional sound sound field location hardware
Platform carries out the acquisition and storage of voice data, enables LnK sound collected by position and after the segmentation and normalization of progress
Data are
3) using Tensorflow deep learning frames, using LeNet-5 convolutional neural networks models, with all acoustic fields
DataFor sample, it is learning objective by its corresponding actual coordinate, carries out deep learning, and the model after learn is joined
It counts to establish model；
4) the model F obtained in step 3) is utilizedDeep, to estimate the sound position data in monitoring area under any position
LEstimation=FDeep(SL)=(Xe,Ye,Ze)；
5) by vola field of force sensor array, position of human body data are acquired by plantar pressure field alignment system, it is right
Signal where sensor on grid carries out Software Coding, is converted into its position data as output, exists when there is moving object station
When on pressure field sensor array, a position coordinates are exported, are enabled as LFeet=(Xf,Yf,Zf)；
6) processing is weighted to two position signals, to step 1) -4) acoustic field localization method and step 5) pressure field
The recognition accuracy of localization method is counted respectively, and it is respectively m% and n% to enable final discrimination, then respective weights k1=m/
(m+n) and k2=n/ (m+n), the location information that final active probe goes out are expressed as：Lp=k1LEstimation+k2LFeet=(Xp,Yp,
Zp)。
7) after Raspberry control systems receive and calculate final position data, PWM pulse-width signals are sent, are led to
Rotation of the motor Controlling model realization to camera is crossed, camera is finally made to be directed toward the position coordinates (X that active probe goes outp,Yp,
Zp), and start acquisition and pursuit movement somatic data.
The method, the step B include step in detail below：
1) it determines to include gender, height, weight, Body proportion, muscle size, trunk fertilizer according to human anatomy data
It is thin, horizontal trunk size, vertical trunk size, trunk displacement, belly size, pushing, upperarm length, upper arm thickness, forearm length, preceding
The arm girth of a garment, bust, chest upright position, waist size, head sizes, neck length, collar, the buttocks girth of a garment, buttocks size, buttocks
Body characteristics parameter including displacement, thigh length, thigh thickness, lower-leg length, shank thickness, foot breadth and foot length：
Choose be suitble to the human skeleton model of gait Activity recognition determine to include root node, skull, neck, back bone,
Breastbone, aitch bone, left shoulder bone, right shoulder bone, left clavicle, right clavicle, left arm bone, right arm bone, left hand anklebone, right hand anklebone, the left hand palm
3 D human body including bone, right hand metacarpal bone, left kneecap, right knee dice, left foot anklebone, right crus of diaphragm anklebone, left foot phalanx and right crus of diaphragm phalanx
Bone parameters：
Standardized human body's model is defined asWherein, P represents that 3 D human body point converges conjunction, and T is represented
Human body triangular plate information；
2) existing public 3 D human body gait behavior database is used, synthesizes 3 D human body gait data, structure one
The 3 D human body gait behavior library of the imparametrization of a different visual angles, body and attitude data
3) it using the model in imparametrization 3 D human body gait behavior library as reference, is set by three-dimensional process software different
Visual angle, body and skeletal joint parameter { β, φ } carry out Set criteria parameter manikin, obtain and given non-parametric modelDepending on
Parameterized human body model that angle, shape and posture are consistent and corresponding body and bone parametersConstruct system
The body of one standard and the parametric human body gait behavior database of attitude data
4) according to method in 3), the corresponding semanteme of all models in the body gait behavior library of imparametrization is estimated
Parameter builds the body matrix β of different human body modelMatrixWith joint matrix φMatrix：
5) enantiomorph matrix βMatrixPrincipal component analysis is carried out, the covariance matrix of body matrix is sought first, then asks association side
Difference characteristic value and feature vector, then by body characteristics value according to from big to small sequence sort, according to sequence magnitude relationship from
In select most important I kinds Geometrical Parameter in gait behavior library, and use ωi∈ [0...1] i-th kind of Geometrical Parameter of expression
Significance level；To skeletal joint matrix φMatrixAlso principal component analysis is carried out, and uses εj∈ [0...1] represents jth kind joint
The significance level of parameter, I kinds Geometrical Parameter collectively forms most important body gait behavior semantic feature with J kind joint parameters joins
Number Bsemantic={ βI,φJ}；
6) enantiomorph matrix βMatrixWith joint matrix φMatrixCanonical correlation analysis is carried out, determines each manikin language
Correlation between adopted parameter, passes through correlation matrixWithDescription.
The method, the step C include step in detail below：
Frame RGB figures are extracted from gait video, S (x, y) is enabled to represent this frame image information, is built by Gaussian Background
Body gait figure is split, is expressed as by mould method
Two-dimentional gait profile is enabled to be expressed asWherein α is gait
Walking visual angle, s 'α,nRepresent gait profile in two-dimentional gait figureIn coordinate information, N is body gait profile discrete point
Sum；The constant Hu squares of geometry of two-dimentional gait profile are calculated, are expressed as H (S 'α)={ M1′,...,M7′}；In three-dimensional standard gait
In behavior database, by carrying out 2D projected outlines on α visual angles to all threedimensional models, and corresponding Hu squares are calculated, according to Hu squares
Similar features, the initialization that body and posture are close with current two-dimentional gait profile is selected in standard parameter gait library
Three-dimensional (3 D) manikin is defined as：Y3D={ P, T, βI,φJ, wherein, P represents that 3 D human body point converges conjunction, and T represents human body triangular plate
Information, βIFor the body characteristics parameter of I kind human bodies, φJFor J kind skeletal joint angle parameters；Human body three-dimensional point cloud data is by triangle
Piece information, Geometrical Parameter and skeletal joint angle information determine jointly, meet：
Projection two-dimensional silhouette of the manikin for including body characteristics and skeletal joint data on α visual angles is enabled to beThe calculating of its Hu square is expressed asCalculate tripleplane's profile and two
Tie up the difference between gait profileAnd between them Hu squares difference
The two superposition is expressed as：η=δ+λ；
The object function finally constructed is defined as：Two-dimentional gait is estimated by solving optimization problem
Three-dimensional (3 D) manikin body β corresponding to profileIWith skeletal joint parameter phiJActual numerical value.
The method, the step D include step in detail below：
1) according to step 4, object function to be solved is defined as：
2) above-mentioned object function is divided into two steps to solve, Geometrical Parameter fixed firstSolve skeletal joint most
Excellent dataChoose the body gait behavior skeletal joint semantic feature initial parameter values of " I " posture
3) initial valueOptimization：It enablesFor ten canonical reference postures in gait cycle, their institutes are calculated
Corresponding two-dimensional projection image Zernike not bending moments, enable and areTwo dimension step to be estimated is calculated again
State imageZernike not bending momentsTwo-dimentional gait image to be estimated is asked for by optimization problem's
Zernike not bending momentsWith two-dimensional projection image Zernike not bending momentsMost like posture label between set enables
For opt, i.e.,And with posture corresponding to the label as an optimization after initial attitude, enable and be
4) from initial valueStart, withStart to search for for direction, obtain the common n direction of search
Minimum valueOn this basis, the new direction of search is calculatedAnd this side up calculate
Go out minimum value
5) new search direction is defined, with front and rear two object functionDifference the maximum determine：
In kth wheel, the direction for contributing maximum is：
6) the search new direction using the maximum direction of contribution as next step scans for, and the termination condition of iteration is：
7) optimal skeletal joint parameter is determined according to above-mentioned stepsAfterwards, fix bone joint parameter, and using above-mentioned
Solve optimal skeletal joint parameterThe step of solve Geometrical ParameterOptimal value.
The method, the step E include step in detail below：
1) to human body body joint parameter βIAnd φJ, with reference to body matrix βMatrixWith skeletal joint matrix φMatrixIt carries out
The weights omega for the significance level that principal component analysis obtainsi∈ [0...1], i ∈ [0...I] and εj∈ [0...1], j ∈ [0...J]
Descending sequence respectively；
2) it chooses in Geometrical Parameter, a body characteristics parameter a of weight order maximummax,
Pass through semantic dependence on parameter matrix againCorrelative relationship between middle body characteristics, is determined and amaxMaximally related body characteristics parameter AndIt is tieed up using m × n
Binary matrix represented come the two-dimentional sparse distribution both realized；
3) two-dimentional sparse distribution representation method is as follows to encode the numerical steps of semantic parameter, and order includes gender, height, body
Weight, Body proportion, muscle size, the trunk girth of a garment, horizontal trunk size, vertical trunk size, trunk displacement, belly size, hand
Length, upperarm length, upper arm thickness, forearm length, the forearm girth of a garment, bust, chest upright position, waist size, head sizes, neck are long
Degree, collar, the buttocks girth of a garment, buttocks size, buttocks displacement, thigh length, thigh thickness, lower-leg length, shank thickness, foot breadth and
The minimum value of a certain item semantic feature parameter value including foot length is VminMaximum value is Vmax, the numberical range V between themrange
=Vmin-Vmax, determine VrangeQuantization step for h, choose w bit 1 to encode semantic parameter, calculating required all two
System quantity m=h+w-1, wherein w is bit 1, remaining is bit 0, for arbitrary data V ∈ [Vmin Vmax], determine w
Index position Indexs of the bit 1 in all binary numbers1=h* (V-Vmin)/Vrange, i.e., in n binary data sequence
In, I ndex1A position starts, and the binary number of continuous w bit puts 1, remaining is 0, by the above method, determines to close
Save binary number n, numerical index position Index needed for skeltal semantic characteristic parameter2, in the binary matrix of m × n dimensions
In, with Index1For abscissa, Index2For ordinate, central point (Index is marked1, Index2), using central point as the center of circle,
Using r as radius, a region is determined, the binary point in this region puts 1 entirely, and the number in remaining binary matrix is set to 0 entirely, with this
Realize the submatrix that the two-dimentional sparse distribution of the related semantic characteristic to two represents, is referred to as under all semantic parameters；
4) submatrix of generation is mapped using hash function：Harsh (x, y)=ix, y, converts thereof into fixation
The output of size；
5) according to the importance weight size and sequencing of characteristic parameter, repeat step 2) -4) in coding method, it is real
Now to the coding of all gait semanteme parameter submatrixs, then, whole submatrixs are combined, form the gait language after folding
Adopted binary features matrix, size is (Im) × (Jn) or is gait grapheme Imgsemantic, the pixel on figure is only
It is represented with binary number 1 or 0.
The method, the step F include step in detail below：
1) to compare test gait behavior grapheme and training the similarity degree of the gait behavior grapheme in library, using direct
Class brain bit arithmetic method.DefinitionWithTo embody the grapheme of different gait behaviors, the two is carried out and is grasped
Make, and calculate its single order normAccording to Duplication δ=ρ/size (Imgsemantic)
To determine the similarity degree of the two；
2) when the gait behavior grapheme of input determine it is similar with a certain grapheme in library is trained, then using class brain or fortune
Calculation method, the grapheme for having missing to input using grapheme information complete in library are repaired, and the gait after reparation is semantic
Figure is：
3) Preference retrieval and classification：Realization sub-sampling is weighted with hash function to gait grapheme, there is selection
The great gait behavior class brain semantic feature of the right of retention of property；It is gait grapheme Img to enable (x, y)semanticIn coordinate picture
Element, the gait grapheme generated after being weighted to it areIn formula, w for 0.0 to
Weights between 1.0；Gait grapheme after sub-sampling is classified or compared, to retrieve the similar gait behavior of feature
Data；
4) the composite calulation method of gait behavior text：To realize the structured representation to gait behavior sequence, by single step
State graphemeIt condenses together to obtain gait semanteme energy diagram by statistical method
New energy diagramThe statistical nature of each subgraph is embodied, Duplication is high in subgraph, and the feature in energy diagram is brighter
It is aobvious, low then on the contrary of Duplication；The comparison and classification acted using the realization of gait semanteme energy diagram to gait behavior sequence.
The utility model has technical effect that, utilizes three-dimensional sound field alignment system, vola field of force alignment system, HDMI high
The embedded gait behavioral value of the structure low-power consumption such as clear camera, high definition video collecting system and microcomputer Raspberry
Identifying system hardware；Carry out active probe using three-dimensional sound field alignment system and plantar pressure sensor array to enter in monitoring range
Movement human；By rotatable high-definition camera active tracing and acquisition human body movement data, and to collected human body
Video data carries out semantic folding；By the method for thinking of analogy human brain, by the variation of visual sensor input frequently and
The detailed bottom gait behavior signal of feature is transformed into high-rise semantic description, so as to fulfill feature to mesh by being abstracted layer by layer
Target identification conversion, passes through gait behavior big data, the retrieval and matching of implementation pattern on this basis.Utilize semantic Foldable square
The text signal that picture signal is converted into having semantic feature is handled, can both realize having for gait behavioural characteristic by method
Effect represents, can also complete gait behavioral data dimensionality reduction well；Resulting gait behavior semantic text structural data
It may be directly applied in cognition computation model, excavated for reality and social computing provides a kind of completely new versatility good gait row
It is excavated for signal.The utility model proposes the gait semanteme energy diagram with temporal aspect, include the gait under different situations
Temporal information by largely there is the gait semanteme energy diagram of temporal aspect, can enhance the study of gait behavior cognitive system
And predictive ability.Meanwhile the utility model proposes the body gait behavior active detecting identification technology folded based on semanteme,
Every field is all with a wide range of applications, and mainly includes：Remote identification, abnormal gait behavioral value, Hang Renhang
For the fields such as prediction and massive video retrieval, there is good economic and social benefit.
The utility model is described in further detail below in conjunction with the accompanying drawings.
Description of the drawings
Fig. 1 is a kind of body gait behavior active detecting identifying system structure chart folded based on semanteme of the utility model；
Fig. 2 is a kind of three-dimensional sound field positioning system circuit flow chart based on convolutional neural networks algorithm of the utility model；
Fig. 3 is a kind of plantar pressure field positioning system circuit flow chart based on array of pressure sensors of the utility model；
Fig. 4 is a kind of body gait behavioral value identifying system implementing procedure figure folded based on semanteme of the utility model；
Fig. 5 is that the utility model recognizes configuration diagram based on the body gait behavior class brain that semanteme folds；
Fig. 6 is the body gait behavioral value recognition methods schematic diagram that the utility model is folded based on semanteme；
Fig. 7 folds flow for the utility model gait behavior semanteme；
Fig. 8 is the utility model gait behavior semanteme folding process parameter submatrix generation method schematic diagram；
Fig. 9 is that schematic diagram is retrieved in gait behavior of the utility model based on gait grapheme；
Figure 10 is that gait behavior energy diagram of the utility model based on gait grapheme synthesizes schematic diagram.
It is marked in figure：1 is HDMI high-definition cameras, and 2 be USB interface HDMI high definition video collecting cards, and 3 be HDMI interface liquid
Crystal display, 4 be Raspberry Pi microcomputers, and 5 be system mounting framework and bottom plate, and 6 store battle array for large capacity disc
Row, 7 be bluetooth keyboard, and 8 be wireless WIFI module, and 9 be 4G high-speed communication modules, and 10 be bluetooth mouse, and 11 be rotatable chassis
Module, 12 be three-dimensional sound field alignment system, 13 be vola field of force alignment sensor array, 14 be internet wireless router, 15
For gait semanteme big data platform.
Specific embodiment
1. the utility model one kind be based on three-dimensional sound field alignment system, plantar pressure field alignment system, big data platform and
The gait behavioral value identifying system hardware configuration of semantic folding is as shown in Figure 1.
2. a kind of circuit flow chart of the three-dimensional sound field alignment system based on convolutional neural networks algorithm of the utility model is shown in
Fig. 2.
3. a kind of circuit flow chart of the plantar pressure field alignment system based on array of pressure sensors of the utility model is shown in
Fig. 3.
4. the utility model one kind be based on three-dimensional sound field alignment system, plantar pressure field alignment system, big data platform and
The gait behavioral value identifying system implementing procedure figure of semantic folding is shown in Fig. 4.
5. the utility model is shown in Fig. 5 based on the body gait behavior class brain cognition framework that semanteme folds.
The utility model simulates the hierarchical structure of human brain based on the class brain cognition framework that semanteme folds, and bottom is former for gait
Beginning data input, and by image preprocessing layer, can obtain corresponding gait profile.Using intermediate parameterized human body model
Method of estimation, estimates the corresponding body of model and skeletal joint semantic description parameter, these semantic features include embodying shape
The static parameter (height, weight, fat or thin etc.) of body and the attitude parameter (skeletal joint angle-data) for embodying motion feature, they
All it is the high abstraction of underlying pixel data, meets the relation between Thinking, Language mode of the mankind, be text signal.Semantic foldable structure is carried out to it
Change and represent, by the gait grapheme of generation, it is carried out class brain with or sub-sampling and polymerizing energy figure calculating, can be further
Realize study, analysis and the understanding of gait behavior.
6. the utility model is shown in Fig. 6 based on the body gait behavioral value recognition methods implementing procedure that semanteme folds.
7. the utility model gait behavior semanteme, which folds implementing procedure, sees Fig. 7.
8th, gait behavioral value identifying system hardware platform of the structure based on big data platform and semantic folding, specifically
Process is as follows：
The gait behavioral value identifying system hardware platform of the structure based on big data platform and semantic folding by
System installation solid mechanical structure, high definition video collecting and processing system, large capacity disc array memory, Raspberry Pi are micro-
Type computer, high-speed mobile communications module and big data service hardware platform are formed.
1) system mounting framework and bottom plate 5 and rotatable chassis module 11 form the machinery of gait behavioral value identifying system
Structure；
2) HDMI high-definition cameras 1, USB interface HDMI high definition video collectings card 2, HDMI interface liquid crystal display 3,
Raspberry Pi microcomputers 4, large capacity disc storage array 6, bluetooth keyboard 7, wireless WIFI module 8,4G are logical at a high speed
Believe module 9, bluetooth mouse 10, three-dimensional sound field alignment system 12, vola field of force alignment sensor array 13, internet wireless routing
Device 14 and gait semanteme big data platform 15 are formed is put down based on three-dimensional sound field alignment system, vola field of force alignment system, big data
The gait behavioral value identifying system hardware platform of platform and semantic folding；
3) three-dimensional sound field alignment system 12, vola field of force alignment sensor array 13 and the miniature calculating of Raspberry Pi
Machine 4, which is realized, detects the active position for entering monitoring field movement human；
4) HDMI high-definition cameras 1, USB interface HDMI high definition video collectings card 2, Raspberry Pi microcomputers
4th, large capacity disc storage array 6 and gait semanteme big data platform 15, which are realized, adopts the high definition of body gait behavior vision data
Collection, storage and semantic folding；
5) bluetooth keyboard 7, bluetooth mouse 10 and HDMI interface liquid crystal display 3 realize the human-computer interaction of system；
6) wireless WIFI module 8,4G high-speed communications module 9 and internet wireless router 14, realize gait semantic data
Wireless transmission and telecommunications functions；
7) rotatable chassis module 11 realizes the rotation process of high-definition camera, is detected with reference to sound field and field of force active position
Information, better active tracing acquire body gait behavior video data；
8) gait semanteme big data platform 15 realizes the structure in the body gait behavior big data pond folded based on semanteme, and
It completes to the retrieval of gait behavior grapheme, filtering, classification and identification function.
9th, the movement human active position detection based on sound field positioning and pressure field positioning.
1) three-dimensional sound field positioning system hardware platform is built.In four different location placement of sounds acquisition sensings of monitoring scene
Device (A, B, C and D), voice signal modulate circuit, embedded controller and wireless transport module, realize the acquisition to sound, go
It makes an uproar, divides, storage and wireless transmission.
2) acquire and demarcate location information.Using camera as coordinate center, choose N=100 different monitoring areas and sit
Mark Ln=(Xn,Yn,Zn), n=1...N, artificial simulation human motion noise 10 times at selected coordinate utilize three dimensional sound
Field location hardware platform carries out the acquisition and storage of voice data, enables LnThe K=after segmentation and normalization collected by position
10 voice datas are
3) using Tensorflow deep learning frames, using LeNet-5 convolutional neural networks models, with all 1000
Sound field dataFor sample, it is learning objective by its corresponding actual coordinate, carries out deep learning, and after is learnt
Model parameter.
4) sound and coordinate data are trained using LeNet-5 convolutional neural networks, after the completion of training, is monitored using its estimation
Sound position data L in region under any positionEstimation=FDeep(SL)=(Xe,Ye,Ze), FDeepGod is obtained for study
Through network model.
5) vola field of force sensor array is built.Using camera as coordinate center, monitoring area is divided into I × J=20
× 20 small square areas place vola field of force sensor, all I × J=20 × 20 in each grid spaces
A sensor forms sensor array.Pressure sensor when no human body passes through, fixed output relevant voltage (be usually 0V,
0) digital signal corresponds to, when there is human body to stand when on sensor grid, output voltage generates variation (being more than 0V), voltage letter
1 is expressed as after number word.Software Coding is carried out to the signal on all grids, is converted into its position data as exporting, therefore
When having moving object station when on pressure field sensor array, a position coordinates are exported, are enabled as LFeet=(Xf,Yf,Zf)。
6) processing is weighted to two position signals, the position weight that acoustic field estimates is generally less than the position of pressure field
Weight (when particularly sound sound is smaller) is put, the location information that final active probe goes out is expressed as：Lp=0.3LEstimation+
0.7LFeet=(Xp,Yp,Zp)。
7) after Raspberry control systems receive and calculate final position data, pwm control signal is sent, passes through electricity
Machine Controlling model realizes the rotation to camera, and camera is finally made to be directed toward the position coordinates (X that active probe goes outp,Yp,Zp), and
Start acquisition and pursuit movement somatic data.
10th, high-definition gait behavior video data acquiring and storage.
1) it in Raspberry PI systems, is programmed using Python, the moving object in high-definition camera head region is used
Gauss models and frame differential method is detected；
2) it when being tested with movement human appearance in the video sequence, stores it in large capacity disc array, with
For processing such as next step analysis and calculating.
11st, definition suitable for gait behavioural analysis standard parameter manikin, build unified standard body and
The parametric human body gait behavior database of attitude data.Each language is determined by principal component analysis and CCA canonical correlation analysis
The effect size of adopted parameter and the correlation matrix between them
1) standardized human body's model is defined：Wherein, P represents that 3 D human body point converges conjunction, refers to
Manikin body data in Makehuman softwares, each model are made of more than 15000 a vertex, and T represents human body triangular plate letter
Breath, one of triangular plate are made of three clouds.More than 100 kinds of Geometrical Parameter quantity, is expressed as
(height, weight and waistline etc.) chooses CMU Mocap human skeleton models to determine that the skeletal joint of 3 D human body is joined(20 remainders such as left and right elbow joint, left and right foot joint and left and right ankle-joint).
2) with existing 3 D human body gait behavior database KY4D (Kyushu University 4D Gait
Database), the 3 D human body gait behavior library of an imparametrization is built by the method for point cloud compressing
Data instruction shares K=42 sample.
3) using the model in 3 D human body gait behavior library as reference, pass through Makehuman and Blender three-dimensional process systems
Different bodies and skeletal joint parameter { β, φ } is set to carry out Set criteria parameter manikin, is obtained and 42 models in databaseParameterized human body model that shape is consistent with posture and corresponding body and bone parametersStructure
Go out the not homomorphs of unified standard and the parametric human body gait behavior database of attitude data
4) using all model semantics parameters estimated, the body matrix β of different human body model is builtMatrixAnd pass
Save matrix φMatrix：
5) enantiomorph matrix βMatrixPCA principal component analysis is carried out, so as to select most important I=in gait behavior library
30 kinds of Geometrical Parameters, preceding 30 parameter attributes of selected characteristic value maximum, and use ω according to contribution degreei∈ [0...1] is represented
The significance level of Geometrical Parameter.To skeletal joint matrix φMatrixAlso PCA principal component analysis is carried out, and according to the tribute of characteristic value
Degree of offering use εj∈ [0...1] determines the significance level of all joint parameters.I=30 kinds Geometrical Parameter and J=24 kinds joint
Parameter collectively forms most important body gait behavior semantic feature parameter Bsemantic={ βI,φJ}。
6) enantiomorph matrix βMatrixWith joint matrix φMatrixCanonical correlation analysis is carried out, determines each manikin language
Correlation between adopted parameter, passes through correlation matrixWithDescription.
12nd, divide the movement human profile taken, choose body in three-dimensional gait behavior database and posture approaches
Parametrization three-dimensional (3 D) manikin as initial model, build three-dimensional (3 D) manikin semanteme parameter Estimation mesh naturally therebetween
Scalar functions by this semantic objects function, are realized under the conditions of missing or redundancy interference 2D body gait profiles to 3 D human body
Effective estimation of model body and skeletal joint parameter value.
1) frame RGB figures are extracted from gait video, human body clothing color is close with background color in the figure, passes through height
Body gait figure is split, is expressed as by this background modeling method or background subtraction methodDue to foreground and background
Similar, the outline data split exists to be lacked to a certain degree.
2) by outline segmentation, two-dimentional gait profile during gait walking visual angle α=90 ° is extracted：
Profile discrete point is unified for N=128；Calculate two-dimentional gait
The constant Hu squares of geometry of profile, are expressed as H (S 'α)={ M '1,...,M′7}。
3) three-dimensional (3 D) manikin is defined：Y3D={ P, T, βI,φJ, three-dimensional (3 D) manikin is rotated into α=90 °, then projected
To two-dimensional space, correspondence profile isThe calculating of its Hu square is expressed as
In three-dimensional standard gait behavior database, by carrying out 2D projected outlines on the visual angle of α=90 ° to all threedimensional models, and
Calculate corresponding Hu squares, according to the similar features of Hu squares, selected in standard parameter gait library body and posture phase with it is current
Two-dimentional gait profile S 'αThe initialization three-dimensional (3 D) manikin being close.
4) difference selected by calculating between three-dimensional parameter model projection profile and two-dimentional gait profile
And between them Hu squares differenceThe two superposition is expressed as：η=δ+λ.
5) object function finally constructed is defined as：
13rd, direct method is conjugated by optimizing the improvement of initial value, solves the naturally semantic parameter Estimation target of three-dimensional (3 D) manikin
Function extracts gait pattern nature semantic description parameter.
1) according to step 6, object function to be solved is：
2) Geometrical Parameter fixed firstSolve skeletal joint optimal dataChoose the human body of " I " posture
Gait behavior skeletal joint semantic feature initial parameter values
3) to first attitude valueIt optimizes：It enablesFor ten canonical reference postures in gait cycle,
The Zernike of their corresponding two-dimensional projection images not bending moment is calculated, enables and is
Two-dimentional gait image to be estimated is calculated againZernike not bending moments, enable and beBy building optimization problemIt asks forWithMost like posture label between collection is enabled as opt, with the label
Corresponding posture as an optimization after initial attitude, enable and be
4) from initial valueStart, withStart to search for for direction, obtain common n=24 search
The minimum value in directionOn this basis, the new direction of search is calculatedAnd this side up
Calculate minimum value
5) new search direction is defined, with front and rear two object functionDifference the maximum determine：
In kth wheel, the direction for contributing maximum is：
6) the search new direction using the maximum direction of contribution as next step scans for, and the termination condition of iteration is：
7) optimal skeletal joint parameter is determinedAfterwards, then fix bone joint parameter, it is solved with reference to above-mentioned alternative manner
Geometrical ParameterOptimal value.
14th, class brain and, class brain or the sampling of class brain and class brain polymerizing energy figure are carried out to gait behavior class brain grapheme
It calculates, retrieval, filtering, classification and the knowledge method for distinguishing of gait behavior is completed using gait behavior big data.
1) to human body body joint parameter βIAnd φJ, with reference to the significance level obtained in step 1 by principal component analysis
Weights omegai∈ [0...1] (i ∈ [0...I]) and εj∈ [0...1] (j ∈ [0...J]) descending sequences respectively.
2) it chooses in Geometrical Parameter, a body characteristics parameter a of weight order maximummax,
Pass through semantic dependence on parameter again
MatrixCorrelative relationship between middle body characteristics, is determined and amaxMaximally related body characteristics parameter AndUtilize m × n
The binary matrix of dimension represents come the two-dimentional sparse distribution both realized.
3) two-dimentional sparse distribution representation method is to encode the number of semantic parameter by the binary number 1 and 0 of fixed quantity
Value.The minimum value for enabling " height " semantic feature parameter value is Vmin=0cm maximum values are Vmax=200cm, the numerical value between them
Range Vrange=Vmin-Vmax=200cm.Determine VrangeQuantization step for h=100, choose w=5 bit 1 to encode language
Adopted parameter calculates required all binary number m=h+w-1=104 (w be bit 1 remaining be bit 0).For appointing
Anticipate data V ∈ [Vmin Vmax]=170cm determines index position Indexs of the w bit 1 in all binary numbers1=h* (V-
Vmin)/Vrange=170, i.e., in n binary data sequence, I ndex1Centered on=170, the two of continuous w=5 bit
System number puts 1, remaining is 0.By the above method, determine needed for an other semantic feature parameter " weight=80Kg "
Binary number n, numerical index position Index2=80.In the binary matrix of m × n dimensions, with Index1For abscissa,
Index2For ordinate, central point ((Index is marked1,Index2)=(70,80), using central point as the center of circle, using r=2 as half
Diameter, determines a region, and the binary point in this region puts 1 entirely, and the number in remaining binary matrix is set to 0 entirely, is realized pair with this
The two-dimentional sparse distribution of two related Chinese language characteristics represents, is referred to as the submatrix under all semantic parameters (see Fig. 8-a
It is shown).
4) it is wider for the value range of human body semanteme parameter, and the size of gait semanteme parameter submatrix is certain asks
Topic maps the submatrix of above method generation using hash function：Harsh (x, y)=ix, y is converted thereof into solid
Determine the output ((see shown in Fig. 8-b) of size.).
5) according to the weights of importance value of characteristic parameter, above-mentioned coding method is repeated, is realized to all gait semanteme parameters
Then the coding of submatrix, whole submatrixs is combined, form the gait semantic binary eigenmatrix (step after folding
State grapheme Imgsemantic, as shown in Figure 7), pixel is only represented with binary one or 0 on figure.
15th, class brain and, class brain or the sampling of class brain and class brain polymerizing energy figure are carried out to gait behavior class brain grapheme
It calculates, and retrieval, filtering, classification and the identification of gait behavior is completed using public action identification database.Public training data
Library can be used such as The UCF Dataset, The Hollywood Dataset etc..
1) similarity system design for two gait behavior graphemes of realization, using direct class brain bit arithmetic method.It takes same
The image of the same gait sequential of people by the utility model method, generates gait grapheme, respectivelyWithThe two is carried out and is operated, and calculates its single order normCalculate overlapping
Rate δ=ρ/size (Imgsemantic), the two Duplication is more than 90%；To the gait grapheme of different human body difference gait sequential
It is compared, Duplication is generally less than 50%.By comparing Duplication, it can detect the walking posture of different sequential, compare
The similitude of two gait graphemes.
2) a frame gait figure is taken, artificial horizontal stripe is carried out to it and blocks dividing processing (profile generation missing data), to original
The gait profile schemed and have missing data calculates its gait grapheme respectively.There is the semanteme of missing to input using original gait figure
Figure is repaired, the gait grapheme after reparation：By comparing, there is missing
Gait data show and repaired that likelihood is generally higher than 95% after reparation.
3) Preference retrieval and classification are (see Fig. 9)：Gait grapheme with hash function is weighted, son can be realized
It adopts, selectively retains the gait behavior class brain semantic feature being concerned about.It is gait grapheme Img to enable (x, y)semanticIn seat
Pixel is marked, it is weighted, Geometrical Parameter position weighted value is 1, and attitude parameter position weight is 0, and the gait of generation is semantic
FigureOnly embody the body characteristics of each gait behavior model.To the step after sub-sampling
State grapheme is classified or is compared, and can retrieve the similar body gait behavioral data of body；
4) the composite calulation method of gait behavior text：To realize the structured representation to gait behavior sequence, by I=10
The single gait grapheme of frameIt condenses together to obtain gait semanteme energy diagram by statistical method(see Figure 10)；New energy diagramEmbody the statistical nature of each subgraph, subgraph
Middle Duplication is high, and the feature in energy diagram is more apparent, low then on the contrary of Duplication.To MoBo (CMU Motion of Body)
Body gait in database generates gait semanteme energy diagram according to a little methods, then carries out classification comparison, average recognition rate reaches
To 95%.Preferably gait sequence figure can be described by gait energy diagram.
Claims (2)
1. a kind of body gait behavior active detecting identifying system folded based on semanteme, which is characterized in that installed including system
Solid mechanical structure, three-dimensional sound field alignment system, plantar pressure field alignment system, the high definition video collecting processing system with holder
System, large capacity disc array memory, Raspberry Pi microcomputer systems, high-speed mobile communications module and big data clothes
Business hardware platform；
The system installation solid mechanical structure includes rotatable chassis module 11 and system mounting framework and bottom plate 5, described
System mounting framework and bottom plate 5 be fixed in rotatable chassis module 11；
The plantar pressure field alignment system includes the vola field of force alignment sensor array 13 being made of multiple sensors, institute
The vola field of force alignment sensor array 13 stated is set on ground；
The three-dimensional sound field alignment system includes multiple sound collection sensors, voice signal modulate circuit, embedded Control
Device and wireless transport module, sound collection sensor ring system for winding installation solid mechanical structure setting, the insertion
Formula controller communicates to connect sound collection sensor, voice signal modulate circuit and wireless transport module respectively；
The Raspberry Pi microcomputer systems include Raspberry Pi microcomputers 4, bluetooth keyboard 7, indigo plant
Tooth mouse 10 and HDMI interface liquid crystal display 3, the bluetooth keyboard 7, bluetooth mouse 10 and HDMI interface liquid crystal display 3
It is communicatively connected to Raspberry Pi microcomputers 4 respectively；
The high definition video collecting and processing system with holder includes HDMI high-definition cameras 1 and USB interface HDMI high definitions regard
Frequency capture card 2, the HDMI high-definition cameras 1 are communicatively connected to USB interface HDMI high definition video collectings card 2, the USB
Interface HDMI high definition video collectings card 2 is communicatively connected to Raspberry Pi microcomputer systems；
The high-speed mobile communications module includes wireless WIFI module 8,4G high-speed communications module 9 and internet wireless router
14, the wireless WIFI module 8,4G high-speed communications module 9 and internet wireless router 14 are communicatively connected to respectively
Raspberry Pi microcomputer systems；
The big data service hardware platform includes gait semanteme big data platform 15, the gait semanteme big data platform
15 are set to distal end and are connect with Raspberry Pi microcomputer system telecommunications；
The large capacity disc array memory is communicatively connected to Raspberry Pi microcomputer systems.
2. system according to claim 1, which is characterized in that the vola field of force alignment sensor array is to image
Head is coordinate center, and monitoring area is divided into I × J small square area, and a vola is placed in each grid spaces
Field of force sensor, I × J all sensors form sensor array.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201721209457.9U CN207529395U (en) | 2017-09-20 | 2017-09-20 | A kind of body gait behavior active detecting identifying system folded based on semanteme |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201721209457.9U CN207529395U (en) | 2017-09-20 | 2017-09-20 | A kind of body gait behavior active detecting identifying system folded based on semanteme |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN207529395U true CN207529395U (en) | 2018-06-22 |
Family
ID=62569965
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201721209457.9U Active CN207529395U (en) | 2017-09-20 | 2017-09-20 | A kind of body gait behavior active detecting identifying system folded based on semanteme |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN207529395U (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111045726A (en) * | 2018-10-12 | 2020-04-21 | 上海寒武纪信息科技有限公司 | Deep learning processing device and method supporting coding and decoding |
-
2017
- 2017-09-20 CN CN201721209457.9U patent/CN207529395U/en active Active
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111045726A (en) * | 2018-10-12 | 2020-04-21 | 上海寒武纪信息科技有限公司 | Deep learning processing device and method supporting coding and decoding |
| CN111045726B (en) * | 2018-10-12 | 2022-04-15 | 上海寒武纪信息科技有限公司 | Deep learning processing device and method supporting coding and decoding |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107423730A (en) | A kind of body gait behavior active detecting identifying system and method folded based on semanteme | |
| CN108830150B (en) | One kind being based on 3 D human body Attitude estimation method and device | |
| CN104715493B (en) | A kind of method of movement human Attitude estimation | |
| CN103268495B (en) | Human body behavior modeling recognition methods based on priori knowledge cluster in computer system | |
| CN112101176B (en) | User identity recognition method and system combining user gait information | |
| CN107423729A (en) | A kind of remote class brain three-dimensional gait identifying system and implementation method towards under complicated visual scene | |
| CN110222665A (en) | Human motion recognition method in a kind of monitoring based on deep learning and Attitude estimation | |
| CN106022213A (en) | Human body motion recognition method based on three-dimensional bone information | |
| Arif et al. | Automated body parts estimation and detection using salient maps and Gaussian matrix model | |
| CN111160294B (en) | Gait recognition method based on graph convolution network | |
| CN102074034A (en) | Multi-model human motion tracking method | |
| CN104063721B (en) | A kind of human behavior recognition methods learnt automatically based on semantic feature with screening | |
| CN105354555B (en) | A kind of three-dimensional face identification method based on probability graph model | |
| CN108846343B (en) | Multi-task collaborative analysis method based on three-dimensional video | |
| CN111062340A (en) | Abnormal gait behavior identification method based on virtual posture sample synthesis | |
| Ma et al. | Human motion gesture recognition based on computer vision | |
| Arif et al. | Human pose estimation and object interaction for sports behaviour | |
| Liu et al. | The study on human action recognition with depth video for intelligent monitoring | |
| CN207529395U (en) | A kind of body gait behavior active detecting identifying system folded based on semanteme | |
| Li et al. | Human motion representation and motion pattern recognition based on complex fuzzy theory | |
| Sharma et al. | Various approaches of human activity recognition: a review | |
| CN107692984A (en) | A kind of fatigue detection method and device | |
| CN207529394U (en) | A kind of remote class brain three-dimensional gait identifying system towards under complicated visual scene | |
| Batool et al. | Fundamental Recognition of ADL Assessments Using Machine Learning Engineering | |
| Li et al. | Human sports action and ideological and PoliticalEvaluation by lightweight deep learning model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| GR01 | Patent grant | ||
| GR01 | Patent grant |