CN109492581A - A kind of human motion recognition method based on TP-STG frame - Google Patents
A kind of human motion recognition method based on TP-STG frame Download PDFInfo
- Publication number
- CN109492581A CN109492581A CN201811328308.3A CN201811328308A CN109492581A CN 109492581 A CN109492581 A CN 109492581A CN 201811328308 A CN201811328308 A CN 201811328308A CN 109492581 A CN109492581 A CN 109492581A
- Authority
- CN
- China
- Prior art keywords
- frame
- target
- prediction
- bounding box
- space
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The invention discloses a kind of human motion recognition methods based on TP-STG frame, and this method includes: using video information as input, SVM classifier is added in priori knowledge, proposes posteriority criterion to remove non-personnel targets;Personnel targets are partitioned into detection algorithm by target positioning, and are exported in a manner of target frame and coordinate information, provide input data for human body critical point detection;Physical feeling positioning and correlation degree analysis are carried out to extract whole human body key point informations using improved gesture recognition algorithms, form crucial point sequence;Space-time diagram is constructed on crucial point sequence by action recognition algorithm, is applied in multilayer space-time diagram convolution operation, and the classification of motion is carried out by Softmax classifier, realizes the human action identification under complex scene.Method of the invention combines the actual scene of ocean platform for the first time, and the TP-STG frame of proposition attempts to identify for the first time the activity of the worker on offshore drilling platform using the method for target detection, gesture recognition and space-time diagram convolution.
Description
Technical field
The invention belongs to computer visions and field of image processing, are related to a kind of human action knowledge based on TP-STG frame
Other method.
Background technique
It with the universal of monitoring camera and is widely applied, the video data of magnanimity brings huge pressure to manual identified
Power carries out analysis to video data using artificial mode and studies and judges, manpower, experience and the analysis ability 'bottleneck' restrictions of monitoring personnel
Intelligent behavior differentiates the overall efficiency of application.In recent years, with the continuous propulsion of research, the research of human action identification is achieved
Certain progress.Traditional method is most commonly seen in these three methods of template matching, three dimensional analysis and time series, but calculates
It measures bigger, vulnerable to the interference of noise, lacks robustness, and the considerations of to action behavior mode globality and analysis of overall importance
Deficiency, the feature extracted is few and simple, causes the accuracy rate of identification lower.
Previous human action recognizer effect under single special scenes is more prominent, but most of human body action recognitions
Algorithm be only applicable to simple scenario, influenced by environmental factor very big.When traditional algorithm be applied to complex scene in when by
It is influenced in by factors such as mixed and disorderly backgrounds, is difficult to be correctly detecting human action feature, recognition effect sharply declines, complicated field
Human action under scape is identified as a problem urgently to be resolved.
Summary of the invention
The object of the present invention is to provide a kind of human motion recognition methods based on TP-STG frame, and this method solve existing
There is technology to the problem that human action recognition effect in complex scene is poor and error is big, the human body that can be used under complex scene is dynamic
It identifies, target positioning and gesture recognition and estimation is carried out to personnel in dynamic scene, realize personnel targets movement in image
Precisely identification.
In order to achieve the above object, the present invention provides a kind of human motion recognition methods based on TP-STG frame, should
Method includes:
(S100) using video information as input, SVM classifier is added in priori knowledge, proposes posteriority criterion to go
Unless personnel targets;
(S200) personnel targets are partitioned into detection algorithm by target positioning, and in a manner of target frame and coordinate information
Output, provides input data for human body critical point detection;
(S300) physical feeling positioning and correlation degree analysis are carried out to extract entirely using improved gesture recognition algorithms
Portion's human body key point information forms crucial point sequence;
(S400) space-time diagram is constructed on crucial point sequence by action recognition algorithm, be applied in multilayer space-time picture scroll
Product operation, and the classification of motion is carried out by Softmax classifier, realize the human action identification under complex scene.
Wherein, in the step S100, for the environmental quality under concrete scene, the elder generation being suitble under the scene is matched
Knowledge is tested, such as this complex scene of offshore oil platform, due to the safety clothes color and certain cylindrical tube colors of personnel targets
Quite similar with form, color, texture and shape feature are difficult to differentiate between, and often obscure two using conventional model under simple scenario
Person leads to higher rate of false alarm;For such problem, the present invention proposes priori knowledge SVM classifier is added, to detection target
SVM pre-training is carried out with target is obscured, the non-personnel targets that will identify that are considered as negative sample removal, reduce the calculating of negative sample
Amount, improves the accuracy rate of next stage target detection.
Preferably, the target, which is positioned with detection algorithm, includes:
(S210) image data is converted video data to by data prediction, and carries out sample label operation, made
For the input data of algorithm;
(S211) input picture is divided into N*N grid, r bounding box is predicted to each grid by feature extraction,
As soon as if the center of an object is fallen in a bounding box, then this grid is responsible for detecting this object;
(S212) single convolutional network is run on the image obtains the confidence score calculating of bounding box, these confidence levels point
Number reflects the accuracy of credibility and target in prediction block in bounding box comprising target;
(S213) it by increasing the costing bio disturbance of bounding box coordinates prediction, and reduces to the confidence for not including object boundary frame
The prediction of degree is lost, and model early stage diverging and unstable is prevented;
(S214) bounding box width and height are normalized according to a certain percentage, so that they fall between zero and one, is obtained
The target category probability and bounding box coordinates information finally predicted.
Preferably, in the step S212, in the calculating process of confidence score, need to define predicted boundary frame
With the intersection degree of actual boundary frame, in this, as the calculation basis of confidence score, if do not deposited in predicted boundary frame unit
In target, then confidence score should be zero;Otherwise confidence score is equal to the intersection PIA between prediction block and true object boundary frame
With the product of real goal frame ground truth, thus the definition of confidence level is indicated are as follows:
In formula (1), Cr indicates confidence level, Gr(Object)Indicate real goal frame,Indicate prediction block and real goal
Intersection between bounding box.
Preferably, costing bio disturbance is completed by overall goal loss function in the step S213, wherein coordinate is pre-
The loss function L (x, y, w, h) of survey are as follows:
In formula (2), x, y indicate that the coordinate of the central point of the bounding box relative to grid cell, w and h indicate entire input
The width and height of image, andRespectively indicate i xi,yi,wi,hiMean value,Expression judges i-th of network
In j-th of grid whether be responsible for this target object, λ indicates weight.
The loss function L (obj) of bounding box confidence level prediction containing real goal is indicated are as follows:
In formula (3), CiIndicate the confidence value of i-th of bounding box, and for the bounding box confidence level without real goal
The loss function L (noobj) of prediction is indicated are as follows:
The loss function L (class) of the prediction of target category is indicated are as follows:
In formula (5), pi(c) it is expressed as the probability of i-th of target category,The probability for being expressed as i target category is equal
Value, after acquiring each loss function, each loss function Weighted Fusion obtains final goal loss function L (f):
L (f)=L (x, y, w, h)+L (obj)+L (noobj)+L (class) (6)
In formula (6), final goal loss function L (f) is that loss function, the bounding box containing real goal that coordinate is predicted are set
The prediction of the loss function of reliability prediction, the bounding box confidence level loss function predicted without real goal and target category
Loss function weighted sum.
Preferably, the improved gesture recognition algorithms include:
(S310) color image of the w*h size obtained using target detection on last stage as input;
(S311) multiple dimensioned mode is taken, it is wild to expand perception according to 1.0 to 1.2 times of ratios;
(S312) a Feature Mapping F is obtained by the feature extraction of preceding 8 layer network of VGG;
(S313) network is divided into two loop branches, and a branch is used to predict the two-dimentional confidence level figure of body part position
S carries out physical feeling and positions to obtain all visible key points of human body, another branch is for prediction pixel point in skeleton
Two-dimensional vector field L is associated degree analyzing and obtains the sightless key point of human body;
(S314) loop branches obtain one group of S using characteristic pattern F as input1,L1;
(S315) the output S of the above branch of branch's difference aftert-1,Lt-1With characteristic pattern F as input, constantly into
Row iteration;
(S316) by p stage final output S (p) and L (p);
(S317) S, the predicted value and ground truth (S of L are calculated*,L*) between L2 norm, the ground of S and L
Truth is calculated according to the 2D point of mark, if some key point mark missing, does not calculate the value of the point, final output is all
The information of key point.
Preferably, in the step S313, two loop branches return S and L respectively, each stage calculates primary
Loss S and L and is originally inputted later and continues to input next stage and be trained;With the increase of the number of iterations,
On the basis of the key point position known, closed using the relative position that the shift length between vector is established between partes corporis humani position
System, to realize the prediction and estimation of the invisible key point of human body.
Preferably, in the step S317, pass through the 2D point X marked in imagej,kCalculate the ground truth of S
(S*), wherein Xj,kIndicate the calculation method of the jth kind key point of k-th of people in image,Expression meets normal distribution;Work as picture
The close annotation point X of vegetarian refreshments Pj,kWhen, reach the peak value of normal curve, then the S of jth kind key point is k in image in every image
The normal distribution peak value of people
In formula (7), σ is the parameter of normal distribution, and for S, every one kind key point has a channel channel, is generated
It is the peak value of response for retaining each point according to the method that multiple distributions are maximized when groundtruth, works as peak valueThe value of the jth kind key point of k-th of people is obtained when obtaining maximum value
Pass through two key point X of k-th of peoplej1,k,Xj2,kBetween the unit vector of any pixel P calculate L's
groundtruth(Lb *), wherein L indicates the affinity of human body, and k indicates k-th of people, j1And j2Expression two being capable of phase
Joint even, such as head are directly connected by neck with trunk, and b indicates the position of b kind human body.
For the ground truth (L of Lb *) calculate it is relative complex, as the key point X of k-th of people in imagej1,kIt is directed toward
Xj2,kWhen, unit vectorPoint situation discussion is needed to calculate, as shown in formula (9):
Wherein, v size and Orientation is fixed, xj,kIndicate the position of j-th of key point of k-th of people, pixel P is
It is no can fall in human body limb it is dry on need to meet two conditions, as threshold range:
In formula (10), lb,kAnd σlThe dry length and width of human body limb is respectively indicated, by mutually similar to all personnel's target
Other limb is dry to carry out equalization, so that the number of active lanes of the output of L is equal with dry kind of number of limb, acquires mean value
In formula (11), the two-dimensional vector field mean value at the position b in every imageIndicate k people in the position of pixel P
The average value for setting the affine force vector in position at place, as key point dj1And dj2And known to the affine force vector in position before them
Later, by calculating two key point dj1And dj2Between line vector sum line on each pixel the affine force vector in position it
Between dot product integral as the position correlation R between two key points:
In formula (12), w indicates the weight between two key points, and for value between [0,1], P (w) indicates weighting function,
Lb(P (w)) indicates the vector of the position affinity of each pixel on line.
Preferably, the action recognition algorithm includes:
(S410) continuous sequence of the human body key point obtained using gesture recognition on last stage as input;
(S411) using graph structure rule, space and timing structure information is made full use of to construct space-time diagram;
(S412) on space-time diagram using multilayer graph convolution operation to extract high-level characteristic, corresponding space division rule;
(S413) the neighborhood subset number for judging space-time node of graph, designs corresponding space division rule and determines and use
Rule;
(S414) classification of motion is carried out using the Softmax classification of motion device of standard;
(S415) output action class label and corresponding movement scoring.
Preferably, in the step S411, when constructing space-time diagram, inside each video frame, according to human body from
Right skeleton connection relationship constructs space diagram, while the identical key point of adjacent two frame is connected and composed timing side, all input frames
In key point configuration node collection V, all directed edges constitute side collection E, obtain space-time diagram G=(V, E) according to above-mentioned rule, from
The spatial information of skeleton key point so is remained, and the motion profile of key point is showed in the form of timing side.
Specifically, node set V={ v in figureti| t=1,2...T, i=1,2...N } comprising all on crucial point sequence
Artis, wherein T indicates video frame number, and N indicates the number of all key points of human body, is set as 18.When constructing space-time diagram,
Feature vector F (the v of t frame in key point, i-th artisti) be made of the coordinate information and confidence level of key point;
There are two subsets to form by the set E on side, is the link E of each video frame frame intrinsic articulation point respectivelys={ vtivtj| (i, j) ∈ P } with
And the link E of different video frame interframet={ vtiv(t+1)i, wherein P indicates the set of all key points of human body, and i, j are respectively
Two any joints in set of keypoints.
Preferably, in the step S413, the adjacent pixel set of center pixel, i.e. neighborhood collection is suitable according to space
Sequence is divided into a series of set T, and each set just includes one pixel of image, these set constitute one stroke of neighborhood collection
Point, if 1 neighborhood of node is divided into a subset, divided labeled as unique;If 1 neighborhood of node is divided into two subsets,
That is the subset of node itself and neighbors subset, labeled as the division based on distance;If 1 neighborhood of node is divided into three sons
Collection, including on node itself, spatial position compare this node closer to the neighbors set of entire skeleton center of gravity and further from weigh
The neighbors set of the heart is defined centripetal movement with centrifugal movement according to motion analysis, divides labeled as steric configuration.
Human motion recognition method based on TP-STG frame of the invention solves the prior art to complex scene servant
The problem that body action recognition error is big and real-time is poor, has the advantage that
(1) method of the invention combines the actual scene of ocean platform for the first time, and the TP-STG frame of proposition is attempted to make for the first time
Worker's activity on offshore drilling platform is identified with the method for target detection, gesture recognition and space-time diagram convolution;
(2) method of the invention is applied in this complex scene of ocean platform, is examined to reduce cylindrical tube to target
The influence of survey proposes a kind of data prediction scheme based on SVM, improves the accuracy rate of target identification;
(3) method of the invention proposes improved gesture recognition algorithms in the case where there is veil, utilizes target detection
Result realize the invisible key point of human body detection and estimation, remove the miscellaneous work of artificial label target from;
(4) method of the invention is sufficiently used spatial structural form by constructing space-time diagram on crucial point sequence
And timing structure information, it is applied in multilayer space-time diagram convolution operation, and by the progress classification of motion of Softmax classifier and in advance
It surveys, realizes the human action identification under complex scene.
Detailed description of the invention
Fig. 1 is the structure flow chart of the human motion recognition method based on TP-STG frame in the present invention.
Fig. 2 is the flow chart of target positioning and detection algorithm in the present invention.
Fig. 3 is the flow chart of improved gesture recognition algorithms in the present invention.
Fig. 4 is the flow chart of action recognition algorithm in the present invention.
Specific embodiment
Below in conjunction with drawings and examples, the following further describes the technical solution of the present invention.
A kind of human motion recognition method based on TP-STG frame, as shown in Figure 1, being of the invention based on TP-STG frame
The structure flow chart of the human motion recognition method of frame, this method include:
(S100) using video information as input, feature extraction is carried out, then SVM classifier is added in priori knowledge, is proposed
Posteriority criterion is to remove non-personnel targets;
(S200) personnel targets are partitioned into detection algorithm by target positioning, and in a manner of target frame and coordinate information
Output, provides information using feature selecting for human body critical point detection;
(S300) physical feeling positioning and correlation degree analysis are carried out to detect entirely using improved gesture recognition algorithms
Portion's human body key point information forms crucial point sequence;
(S400) space-time diagram is constructed on crucial point sequence by action recognition algorithm, be applied in multilayer space-time picture scroll
Product operation, and the classification of motion is carried out by Softmax classifier, realize that the human action under complex scene is classified and estimated.
Wherein, in the step S100, using video information as input, pass through data prediction and sample label
It converts video data to the image data that can input depth network, is examined using the image data set trainer's target marked
Survey model.For the environmental quality under concrete scene, match the priori knowledge being suitble under the scene, as offshore oil platform this
Complex scene, since the safety clothes color and certain cylindrical tube colors and form of personnel targets are quite similar, color, texture and
Both shape feature is difficult to differentiate between, often obscured using conventional model under simple scenario, lead to higher rate of false alarm;For such
Problem, the present invention propose priori knowledge SVM classifier is added, to detection target and obscure target progress SVM pre-training, will know
Not Chu non-personnel targets be considered as negative sample removal, reduce the calculation amount of negative sample, improve next stage target detection
Accuracy rate.
As shown in Fig. 2, for the flow chart of target positioning and detection algorithm of the invention, personnel targets positioning and detection algorithm
Process includes:
(S210) image data is converted video data to by data prediction, and carries out sample label operation, made
For the input data of algorithm;
(S211) input picture is divided into N*N grid, r bounding box is predicted to each grid by feature extraction,
Using object judgement mechanism, if as soon as the center of an object is fallen in a bounding box, then this grid is responsible for detecting this
A object, each grid cell predict that the r bounding box and confidence score of these frames carry out next if target is not present
The detection of a grid;
(S212) single convolutional network is run on the image obtains the confidence score calculating of bounding box, these confidence levels point
Number reflects the accuracy of credibility and target in prediction block in bounding box comprising target, is selected by Confidence
Select the calculation of loss function;
(S213) it by increasing the costing bio disturbance of bounding box coordinates prediction, and reduces to the confidence for not including object boundary frame
The prediction of degree is lost, and model early stage diverging and unstable is prevented;
(S214) bounding box width and height are normalized according to a certain percentage, so that they fall between zero and one, is obtained
The target category probability and bounding box coordinates information finally predicted.
Target positioning and detection algorithm solve target detection as single regression problem, directly input from image pixel
It is exported to bounding box coordinates and class probability, uses the bounding box for disposably predicting target contained by all grids on the image, determines
Position reliability and all categories probability vector disposably solve problem, and only running single convolutional network can then predict
Target category and its position.
Design for network, initial convolutional layer is used to extract feature from image, and full articulamentum is used to predict to export
Destination probability and position coordinates.Entire detection network has 24 3 × 3 convolutional layers and 2 full articulamentums, alternate 1 × 1 reduction zone
For reducing the feature space from previous layer.In order to enable the network to receive the input picture of sizes, the present invention is reduced
Full articulamentum in network structure, because full articulamentum necessarily requires to output and input regular length feature vector, if will
As soon as whole network becomes a full convolutional network, detected then can be inputted to sizes.Meanwhile full convolutional network phase
The spatial positional information of target can preferably be retained for full articulamentum.In addition, in order to promote wisp detection effect, this hair
The bright pond number of layers eliminated in network structure, so that final characteristic pattern size is bigger, characteristic pattern size depends on original image ruler
It is very little, but the size of characteristic pattern is necessary for odd number, guarantees that centre has a position that can save the target at original image center with this.
For Confidence, algorithm sets 0.6 for initial threshold, and the object type greater than the threshold value is considered as target class
Not, otherwise it is considered as target to be not present.Target is judged with the presence or absence of the calculation for determining loss function, containing true according to confidence level
The loss function L (obj) of the bounding box confidence level prediction of target is indicated are as follows:
In formula (3), CiIndicate the confidence value of i-th of bounding box, and for the bounding box confidence level without real goal
The loss function L (noobj) of prediction is indicated are as follows:
In formula (4),It indicates to judge j-th of grid whether not responsible this target object, λ in i-th of network
Indicate weight.
As shown in figure 3, for the flow chart of improved gesture recognition algorithms in the present invention, which includes:
(S310) color image of the w*h size obtained using target detection on last stage as input;
(S311) multiple dimensioned mode is taken, it is wild to expand perception according to 1.0 to 1.2 times of ratios;
(S312) a Feature Mapping F is obtained by the feature extraction of preceding 8 layer network of VGG;
(S313) by key-point analysis as a result, algorithm is divided into two loop branches, a branch is for predicting body
The two-dimentional confidence level figure S of position position, carries out physical feeling positioning and critical point detection obtains all visible key points of human body, separately
One branch is used for two-dimensional vector field L of the prediction pixel point in skeleton, is associated degree analyzing and position affinity calculates
Obtain the information of the invisible key point of human body;
(S314) when current generation t≤p, loop branches obtain one group of S using characteristic pattern F as input1,L1;
(S315) the output S of the above branch of branch's difference aftert-1,Lt-1With characteristic pattern F as input, constantly into
Row iteration;
(S316) by p stage final output S (p) and L (p);
(S317) S, the predicted value and ground truth (S of L are calculated*,L*) between L2 norm, the ground of S and L
Truth is calculated according to the 2D point of mark, if some key point mark missing, does not calculate the value of the point, final output is all
The information of key point.
In step S313, image is analyzed by the convolutional network in preceding 8 layer network, two loop branches return S respectively
And L, the confidence level figure of physical feeling is iteratively predicted in branch one and human body affinity is predicted in branch two,
It is each drop applications different loss function at the end of each stage, S and L and being originally inputted is continued under input later
One stage is trained.In the predicted value and true confidence map of estimation, the connection between key point is lost using L2,
Spatially loss function is weighted, to solve some practical problems.With the increase of the number of iterations, in known key point
On the basis of position, the relative positional relationship between partes corporis humani position is established using the shift length between vector, to realize
The prediction and estimation of the invisible key point of human body.
Expression for position affinity, it is a 2D vector set, each 2D vector set can encode one
The position and direction of limbs, while retaining the position and direction information between the region for indicating limbs.For belonging to specific limbs
Each pixel at position, 2D vector coding are directed toward from a part of limbs the direction of another part, each type
The limbs two-part affinity that has its corresponding polymerization connected.The confidence map one of these position affinity and physical feeling
It rises and carries out combination learning and prediction by CNN, effectively can carry out Attitude estimation for more people, while in the feelings for guaranteeing precision
Under condition, it can accomplish real-time effect.
As shown in figure 4, for the flow chart of action recognition algorithm in the present invention, which includes:
(S410) continuous sequence of the human body key point obtained using gesture recognition on last stage as input;
(S411) using graph structure rule, space and timing structure information is made full use of to construct space-time diagram;
(S412) on space-time diagram using multilayer graph convolution operation to extract high-level characteristic, corresponding space division rule;
(S413) the neighborhood subset number for judging space-time node of graph, designs corresponding space division rule and determines and use
Rule;
(S414) classification of motion is carried out using the Softmax classification of motion device of standard;
(S415) output action class label and corresponding movement scoring.
In step S413, convolution operation on the image pixel set that center pixel is adjacent, i.e. neighborhood collection is according to sky
Between sequence be such as from left to right divided into a series of set T from top to bottom, each set just includes a pixel, these set
Just constitute a division of neighborhood collection.The parameter of convolution kernel and the subset number and feature vector length in this division
It is related, as long as defining certain division rule and defining referring to image convolution the parameter of convolution kernel.
For different figure convolutional networks, targeted convolution operation is defined, is just reduced to design corresponding division rule
Then.To one there are the division rule of T subset, the parameter of convolution kernel includes T part, each partial parameters quantity and feature
Vector is the same.In the convolution operation that a window size is 3 × 3, the neighborhood of a pixel is divided into 9 according to spatial order
A subset, including upper left, on, upper right, left, middle and right, lower-left, lower and bottom right, each subset includes a pixel.Convolution kernel
Parameter includes 9 parts, and each position is consistent with the feature vector length of characteristic pattern, and image convolution can be considered as picture scroll product and exist
A kind of application in rule mesh trrellis diagram.
For skeleton division rule spatially, if 1 neighborhood of node is divided into a subset, drawn labeled as unique
Point;If 1 neighborhood of node is divided into two subsets, the i.e. subset of node itself and neighbors subset, labeled as based on distance
It divides;If by 1 neighborhood of node be divided into three subsets, including node itself, on spatial position than this node closer to entire
The neighbors set of skeleton center of gravity and neighbors set further from center of gravity transport centripetal movement with centrifugation according to motion analysis
It is dynamic to be defined, it is divided labeled as steric configuration.By using different skeleton division rules, realize classification of motion device to movement
The prediction and scoring of classification.
Method of the invention is using advanced technologies such as deep learning, intelligent video behavioural analysis and big data analysis to video
In data personnel targets carry out intellectual analysis, quickly study and judge and track and identify, realize abnormal operation video assistant analysis and
The early warning and alarming of security incident.Specifically, in oil field, by analyzing the movement of drilling production operation personnel, Neng Gouji
Shi Faxian personnel targets are potentially dangerous, enhance the quick-reaction capability to abnormal conditions.In addition, being analyzed by computer video
Technology positions personnel targets in dynamic scene, identifies and critical point detection, carries out point of target action on this basis
Analyse and study and judge, reduce the time of manual intervention, avoid caused by being produced because of personal accident and violation operation to oil field interference and
Economic loss saves manpower and material resources to ensure safety production, improves production management level.In addition, being based on TP-STG
The human motion recognition method of frame realizes personnel's action recognition and analysis under complex scene, is applicable not only to oilfield,
Also there is important practical application value in the other fields such as medical treatment and security.
In conclusion the human motion recognition method of the invention based on TP-STG frame is dynamic to human body under complex scene
It is fast and accurately identified, can be applied to multiple fields and carry out target positioning and detection, gesture recognition, critical point detection
And the differentiation and analysis of behavior and movement.
It is discussed in detail although the contents of the present invention have passed through above preferred embodiment, but it should be appreciated that above-mentioned
Description is not considered as limitation of the present invention.After those skilled in the art have read above content, for of the invention
A variety of modifications and substitutions all will be apparent.Therefore, protection scope of the present invention should be limited to the appended claims.
Claims (7)
1. a kind of human motion recognition method based on TP-STG frame, feature and specific step is as follows:
(S100) using video information as input, priori knowledge is added SVM classifier, propose posteriority criterion to go unless
Personnel targets;
(S200) personnel targets are partitioned into detection algorithm by target positioning, and defeated in a manner of target frame and coordinate information
Out, input data is provided for human body critical point detection;
(S300) physical feeling positioning and correlation degree analysis are carried out to extract whole people using improved gesture recognition algorithms
Body key point information forms crucial point sequence;
(S400) space-time diagram is constructed on crucial point sequence by action recognition algorithm, is applied to grasp in multilayer space-time diagram convolution
Make, and the classification of motion is carried out by Softmax classifier, realizes the human action identification under complex scene;
In the step S100, for the environmental quality under concrete scene, the priori knowledge being suitble under the scene is matched, such as
This complex scene of offshore oil platform, due to personnel targets safety clothes color and certain cylindrical tube colors and form very
Both similar, color, texture and shape feature are difficult to differentiate between, often obscured using conventional model under simple scenario, cause higher
Rate of false alarm;For such problem, the present invention proposes priori knowledge SVM classifier is added, and to detection target and obscures target
SVM pre-training is carried out, the non-personnel targets that will identify that are considered as negative sample removal, reduce the calculation amount of negative sample, improve
The accuracy rate of next stage target detection.
2. the human motion recognition method according to claim 1 based on TP-STG frame, which is characterized in that the mesh
Position, which is demarcated, with detection algorithm includes:
(S210) image data is converted video data to by data prediction, and carries out sample label operation, as calculation
The input data of method;
(S211) input picture is divided into N*N grid, r bounding box is predicted to each grid by feature extraction, if
As soon as the center of an object is fallen in a bounding box, then this grid is responsible for detecting this object;
(S212) single convolutional network is run on the image and obtains the confidence score calculating of bounding box, these confidences are anti-
Accuracy of the credibility and target in bounding box comprising target in prediction block is reflected;
(S213) it by increasing the costing bio disturbance of bounding box coordinates prediction, and reduces to the confidence level for not including object boundary frame
Prediction loss, prevent model early stage diverging and it is unstable;
(S214) bounding box width and height are normalized according to a certain percentage, so that they fall between zero and one, is obtained final
The target category probability and bounding box coordinates information of prediction.
3. the human motion recognition method according to claim 2 based on TP-STG frame, which is characterized in that described
In step S212, in the calculating of confidence score, need to define the intersection degree of predicted boundary frame and actual boundary frame, with this
As the calculation basis of confidence score, if target is not present in predicted boundary frame unit, confidence score should be zero;
Otherwise confidence score is equal to intersection PIA and real goal frame ground between prediction block and true object boundary frame
The product of truth, thus the definition of confidence level indicates are as follows:
In formula (1), Cr indicates confidence level, Gr(Object)Indicate real goal frame,Indicate prediction block and true object boundary
Intersection between frame.
4. the human motion recognition method according to claim 2 based on TP-STG frame, which is characterized in that described
In step S213, costing bio disturbance is completed by overall goal loss function, wherein the loss function L (x, y, w, h) of coordinate prediction
Are as follows:
In formula (2), x, y indicate that the coordinate of the central point of the bounding box relative to grid cell, w and h indicate entire input picture
Width and height, andRespectively indicate i xi,yi,wi,hiMean value,Expression judges in i-th of network
Whether j-th of grid is responsible for this target object, and λ indicates weight;
The loss function L (obj) of bounding box confidence level prediction containing real goal is indicated are as follows:
In formula (3), CiIndicate the confidence value of i-th of bounding box, and the bounding box confidence level without real goal is predicted
Loss function L (noobj) is indicated are as follows:
The loss function L (class) of the prediction of target category is indicated are as follows:
In formula (5), pi(c) it is expressed as the probability of i-th of target category,It is expressed as the mathematical expectation of probability of i target category, is asked
After obtaining each loss function, each loss function Weighted Fusion obtains final goal loss function L (f):
L (f)=L (x, y, w, h)+L (obj)+L (noobj)+L (class) (6)
In formula (6), final goal loss function L (f) is the loss function of coordinate prediction, the bounding box confidence level containing real goal
The loss of the prediction of the loss function of prediction, the bounding box confidence level loss function predicted without real goal and target category
Function weighted sum.
5. the human motion recognition method according to claim 1 based on TP-STG frame, which is characterized in that described changes
Into gesture recognition algorithms include:
(S310) color image of the w*h size obtained using target detection on last stage as input;
(S311) multiple dimensioned mode is taken, it is wild to expand perception according to 1.0 to 1.2 times of ratios;
(S312) a Feature Mapping F is obtained by the feature extraction of preceding 8 layer network of VGG;
(S313) network is divided into two loop branches, and a branch is used to predict the two-dimentional confidence level figure S of body part position, into
Row physical feeling positions to obtain all visible key points of human body, another branch is used for two dimension of the prediction pixel point in skeleton
Vector field L is associated degree analyzing and obtains the sightless key point of human body;
(S314) loop branches obtain one group of S using characteristic pattern F as input1,L1;
(S315) the output S of the above branch of branch's difference aftert-1,Lt-1With characteristic pattern F as inputting, constantly change
Generation;
(S316) by p stage final output S (p) and L (p);
(S317) S, the predicted value and ground truth (S of L are calculated*,L*) between L2 norm, the ground truth of S and L
It is calculated according to the 2D of mark point, if some key point mark missing, does not calculate the value of the point, all key points of final output
Information;
In the step S313, two loop branches return S and L respectively, each stage calculates primary loss, later handle
It S and L and is originally inputted and continues to input next stage and be trained;With the increase of the number of iterations, in known key point
On the basis of position, the relative positional relationship between partes corporis humani position is established using the shift length between vector, to realize
The prediction and estimation of the invisible key point of human body.
6. the human motion recognition method according to claim 1 based on TP-STG frame, which is characterized in that described is dynamic
Include: as recognizer
(S410) continuous sequence of the human body key point obtained using gesture recognition on last stage as input;
(S411) using graph structure rule, space and timing structure information is made full use of to construct space-time diagram;
(S412) on space-time diagram using multilayer graph convolution operation to extract high-level characteristic, corresponding space division rule;
(S413) the neighborhood subset number for judging space-time node of graph designs corresponding space division rule and determines the rule used
Then;
(S414) classification of motion is carried out using the Softmax classification of motion device of standard;
(S415) output action class label and corresponding movement scoring.
7. the human motion recognition method according to claim 6 based on TP-STG frame, which is characterized in that described
In step S411, when constructing space-time diagram, inside each video frame, space is constructed according to the natural skeleton connection relationship of human body
Figure, while the identical key point of adjacent two frame is connected and composed into timing side, the key point configuration node collection V in all input frames,
All directed edges constitute side collection E, obtain space-time diagram G=(V, E) according to above-mentioned rule, remain skeleton key point naturally
Spatial information, and the motion profile of key point is showed in the form of timing side;
Specifically, node set V={ v in figureti| t=1,2...T, i=1,2...N } include the institute on crucial point sequence
Some artis, wherein T indicates video frame number, and N indicates the number of all key points of human body, is set as 18;When building space-time diagram
When, the feature vector F (v of t frame in key point, i-th artisti) be made of the coordinate information and confidence level of key point
's;There are two subsets to form by the set E on side, is the link E of each video frame frame intrinsic articulation point respectivelys={ vtivtj|(i,j)∈
P } and different video frame interframe link Et={ vtiv(t+1)i, wherein P indicates the set of all key points of human body, and i, j divide
It is not two any joints in set of keypoints.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811328308.3A CN109492581B (en) | 2018-11-09 | 2018-11-09 | Human body action recognition method based on TP-STG frame |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811328308.3A CN109492581B (en) | 2018-11-09 | 2018-11-09 | Human body action recognition method based on TP-STG frame |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109492581A true CN109492581A (en) | 2019-03-19 |
CN109492581B CN109492581B (en) | 2023-07-18 |
Family
ID=65695523
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811328308.3A Active CN109492581B (en) | 2018-11-09 | 2018-11-09 | Human body action recognition method based on TP-STG frame |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109492581B (en) |
Cited By (37)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110059761A (en) * | 2019-04-25 | 2019-07-26 | 成都睿沿科技有限公司 | A kind of human body behavior prediction method and device |
CN110070063A (en) * | 2019-04-29 | 2019-07-30 | 北京字节跳动网络技术有限公司 | Action identification method, device and the electronic equipment of target object |
CN110096950A (en) * | 2019-03-20 | 2019-08-06 | 西北大学 | A kind of multiple features fusion Activity recognition method based on key frame |
CN110110686A (en) * | 2019-05-14 | 2019-08-09 | 中国石油大学(华东) | Based on the human motion recognition methods for losing double-current convolutional neural networks more |
CN110119703A (en) * | 2019-05-07 | 2019-08-13 | 福州大学 | The human motion recognition method of attention mechanism and space-time diagram convolutional neural networks is merged under a kind of security protection scene |
CN110147743A (en) * | 2019-05-08 | 2019-08-20 | 中国石油大学(华东) | Real-time online pedestrian analysis and number system and method under a kind of complex scene |
CN110147729A (en) * | 2019-04-16 | 2019-08-20 | 深圳壹账通智能科技有限公司 | User emotion recognition methods, device, computer equipment and storage medium |
CN110188599A (en) * | 2019-04-12 | 2019-08-30 | 哈工大机器人义乌人工智能研究院 | A kind of human body attitude behavior intellectual analysis recognition methods |
CN110222611A (en) * | 2019-05-27 | 2019-09-10 | 中国科学院自动化研究所 | Human skeleton Activity recognition method, system, device based on figure convolutional network |
CN110490035A (en) * | 2019-05-17 | 2019-11-22 | 上海交通大学 | Human skeleton action identification method, system and medium |
CN110516686A (en) * | 2019-07-11 | 2019-11-29 | 西安电子科技大学 | The mosquito recognition methods of three color RGB images |
CN110598569A (en) * | 2019-08-20 | 2019-12-20 | 江西憶源多媒体科技有限公司 | Action recognition method based on human body posture data |
CN110648331A (en) * | 2019-09-02 | 2020-01-03 | 众安信息技术服务有限公司 | Detection method for medical image segmentation, medical image segmentation method and device |
CN110660017A (en) * | 2019-09-02 | 2020-01-07 | 北京航空航天大学 | Dance music recording and demonstrating method based on three-dimensional gesture recognition |
CN110837778A (en) * | 2019-10-12 | 2020-02-25 | 南京信息工程大学 | Traffic police command gesture recognition method based on skeleton joint point sequence |
CN111008601A (en) * | 2019-12-06 | 2020-04-14 | 江西洪都航空工业集团有限责任公司 | Fighting detection method based on video |
CN111062364A (en) * | 2019-12-28 | 2020-04-24 | 青岛理工大学 | Deep learning-based assembly operation monitoring method and device |
CN111461078A (en) * | 2020-05-18 | 2020-07-28 | 江苏电力信息技术有限公司 | Anti-fishing monitoring method based on computer vision technology |
CN111611927A (en) * | 2020-05-21 | 2020-09-01 | 长沙明本信息科技有限公司 | Method for identifying unsafe behaviors of coal mine workers based on human body postures |
CN111723741A (en) * | 2020-06-19 | 2020-09-29 | 江苏濠汉信息技术有限公司 | Temporary fence movement detection alarm system based on visual analysis |
CN111783692A (en) * | 2020-07-06 | 2020-10-16 | 广东工业大学 | Action recognition method and device, electronic equipment and storage medium |
CN111931549A (en) * | 2020-05-20 | 2020-11-13 | 浙江大学 | Human skeleton action prediction method based on multitask non-autoregressive decoding |
CN112084861A (en) * | 2020-08-06 | 2020-12-15 | 中国科学院空天信息创新研究院 | Model training method and device, electronic equipment and storage medium |
CN112434655A (en) * | 2020-12-07 | 2021-03-02 | 安徽大学 | Gait recognition method based on adaptive confidence map convolution network |
CN112613490A (en) * | 2021-01-08 | 2021-04-06 | 云从科技集团股份有限公司 | Behavior recognition method and device, machine readable medium and equipment |
CN113033458A (en) * | 2021-04-09 | 2021-06-25 | 京东数字科技控股股份有限公司 | Action recognition method and device |
CN113052061A (en) * | 2021-03-22 | 2021-06-29 | 中国石油大学(华东) | Speed skating athlete motion identification method based on human body posture estimation |
CN112016528B (en) * | 2020-10-20 | 2021-07-20 | 成都睿沿科技有限公司 | Behavior recognition method and device, electronic equipment and readable storage medium |
CN113283381A (en) * | 2021-06-15 | 2021-08-20 | 南京工业大学 | Human body action detection method suitable for mobile robot platform |
WO2021179898A1 (en) * | 2020-03-11 | 2021-09-16 | 深圳市商汤科技有限公司 | Action recognition method and apparatus, electronic device, and computer-readable storage medium |
CN113536997A (en) * | 2021-07-01 | 2021-10-22 | 深圳中智明科智能科技有限公司 | Intelligent security system and method based on image recognition and behavior analysis |
US11270147B1 (en) | 2020-10-05 | 2022-03-08 | International Business Machines Corporation | Action-object recognition in cluttered video scenes using text |
CN114642858A (en) * | 2022-02-18 | 2022-06-21 | 上海大学 | Rehabilitation training device for senile asthenia patients |
CN114842560A (en) * | 2022-07-04 | 2022-08-02 | 广东瑞恩科技有限公司 | Computer vision-based construction site personnel dangerous behavior identification method |
US11423252B1 (en) | 2021-04-29 | 2022-08-23 | International Business Machines Corporation | Object dataset creation or modification using labeled action-object videos |
CN116259003A (en) * | 2023-01-06 | 2023-06-13 | 苏州同企人工智能科技有限公司 | Construction category identification method and system in construction scene |
CN116363566A (en) * | 2023-06-02 | 2023-06-30 | 华东交通大学 | Target interaction relation recognition method based on relation knowledge graph |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106803913A (en) * | 2017-03-10 | 2017-06-06 | 武汉东信同邦信息技术有限公司 | A kind of detection method and its device of the action that taken the floor for Auto-Sensing student |
CN107563381A (en) * | 2017-09-12 | 2018-01-09 | 国家新闻出版广电总局广播科学研究院 | The object detection method of multiple features fusion based on full convolutional network |
CN108154176A (en) * | 2017-12-22 | 2018-06-12 | 北京工业大学 | A kind of 3D human body attitude algorithm for estimating for single depth image |
CN108710868A (en) * | 2018-06-05 | 2018-10-26 | 中国石油大学(华东) | A kind of human body critical point detection system and method based under complex scene |
-
2018
- 2018-11-09 CN CN201811328308.3A patent/CN109492581B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106803913A (en) * | 2017-03-10 | 2017-06-06 | 武汉东信同邦信息技术有限公司 | A kind of detection method and its device of the action that taken the floor for Auto-Sensing student |
CN107563381A (en) * | 2017-09-12 | 2018-01-09 | 国家新闻出版广电总局广播科学研究院 | The object detection method of multiple features fusion based on full convolutional network |
CN108154176A (en) * | 2017-12-22 | 2018-06-12 | 北京工业大学 | A kind of 3D human body attitude algorithm for estimating for single depth image |
CN108710868A (en) * | 2018-06-05 | 2018-10-26 | 中国石油大学(华东) | A kind of human body critical point detection system and method based under complex scene |
Non-Patent Citations (4)
Title |
---|
RYO ARAI等: "Hierarchical human motion recognition by using motion capture system", 《2018 INTERNATIONAL WORKSHOP ON ADVANCED IMAGE TECHNOLOGY (IWAIT)》 * |
XIAOYANG WANG等: "Learning dynamic Bayesian network discriminatively for human activity recognition", 《IEEE XPLORE》 * |
刘长征等: "复杂背景下基于定位的人体动作识别算法", 《实验室研究与探索》 * |
常晓龙等: "基于深度学习的运动目标实时识别与定位", 《计算机系统应用》 * |
Cited By (62)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110096950A (en) * | 2019-03-20 | 2019-08-06 | 西北大学 | A kind of multiple features fusion Activity recognition method based on key frame |
CN110096950B (en) * | 2019-03-20 | 2023-04-07 | 西北大学 | Multi-feature fusion behavior identification method based on key frame |
CN110188599A (en) * | 2019-04-12 | 2019-08-30 | 哈工大机器人义乌人工智能研究院 | A kind of human body attitude behavior intellectual analysis recognition methods |
CN110147729A (en) * | 2019-04-16 | 2019-08-20 | 深圳壹账通智能科技有限公司 | User emotion recognition methods, device, computer equipment and storage medium |
CN110059761A (en) * | 2019-04-25 | 2019-07-26 | 成都睿沿科技有限公司 | A kind of human body behavior prediction method and device |
WO2020220809A1 (en) * | 2019-04-29 | 2020-11-05 | 北京字节跳动网络技术有限公司 | Action recognition method and device for target object, and electronic apparatus |
CN110070063A (en) * | 2019-04-29 | 2019-07-30 | 北京字节跳动网络技术有限公司 | Action identification method, device and the electronic equipment of target object |
JP2021531589A (en) * | 2019-04-29 | 2021-11-18 | 北京字節跳動網絡技術有限公司Beijing Bytedance Network Technology Co., Ltd. | Motion recognition method, device and electronic device for target |
JP7181375B2 (en) | 2019-04-29 | 2022-11-30 | 北京字節跳動網絡技術有限公司 | Target object motion recognition method, device and electronic device |
GB2598015A (en) * | 2019-04-29 | 2022-02-16 | Beijing Bytedance Network Tech Co Ltd | Action recognition method and device for target object, and electronic apparatus |
GB2598015B (en) * | 2019-04-29 | 2023-07-05 | Beijing Bytedance Network Tech Co Ltd | Action recognition method and device for target object, and electronic apparatus |
CN110119703B (en) * | 2019-05-07 | 2022-10-04 | 福州大学 | Human body action recognition method fusing attention mechanism and spatio-temporal graph convolutional neural network in security scene |
CN110119703A (en) * | 2019-05-07 | 2019-08-13 | 福州大学 | The human motion recognition method of attention mechanism and space-time diagram convolutional neural networks is merged under a kind of security protection scene |
CN110147743A (en) * | 2019-05-08 | 2019-08-20 | 中国石油大学(华东) | Real-time online pedestrian analysis and number system and method under a kind of complex scene |
CN110110686A (en) * | 2019-05-14 | 2019-08-09 | 中国石油大学(华东) | Based on the human motion recognition methods for losing double-current convolutional neural networks more |
CN110490035A (en) * | 2019-05-17 | 2019-11-22 | 上海交通大学 | Human skeleton action identification method, system and medium |
CN110222611B (en) * | 2019-05-27 | 2021-03-02 | 中国科学院自动化研究所 | Human skeleton behavior identification method, system and device based on graph convolution network |
CN110222611A (en) * | 2019-05-27 | 2019-09-10 | 中国科学院自动化研究所 | Human skeleton Activity recognition method, system, device based on figure convolutional network |
CN110516686A (en) * | 2019-07-11 | 2019-11-29 | 西安电子科技大学 | The mosquito recognition methods of three color RGB images |
CN110516686B (en) * | 2019-07-11 | 2021-03-23 | 西安电子科技大学 | Mosquito recognition method of three-color RGB image |
CN110598569A (en) * | 2019-08-20 | 2019-12-20 | 江西憶源多媒体科技有限公司 | Action recognition method based on human body posture data |
CN110598569B (en) * | 2019-08-20 | 2022-03-08 | 江西憶源多媒体科技有限公司 | Action recognition method based on human body posture data |
CN110648331B (en) * | 2019-09-02 | 2023-01-17 | 众安信息技术服务有限公司 | Detection method for medical image segmentation, medical image segmentation method and device |
CN110660017A (en) * | 2019-09-02 | 2020-01-07 | 北京航空航天大学 | Dance music recording and demonstrating method based on three-dimensional gesture recognition |
CN110648331A (en) * | 2019-09-02 | 2020-01-03 | 众安信息技术服务有限公司 | Detection method for medical image segmentation, medical image segmentation method and device |
CN110837778B (en) * | 2019-10-12 | 2023-08-18 | 南京信息工程大学 | Traffic police command gesture recognition method based on skeleton joint point sequence |
CN110837778A (en) * | 2019-10-12 | 2020-02-25 | 南京信息工程大学 | Traffic police command gesture recognition method based on skeleton joint point sequence |
CN111008601A (en) * | 2019-12-06 | 2020-04-14 | 江西洪都航空工业集团有限责任公司 | Fighting detection method based on video |
CN111062364B (en) * | 2019-12-28 | 2023-06-30 | 青岛理工大学 | Method and device for monitoring assembly operation based on deep learning |
CN111062364A (en) * | 2019-12-28 | 2020-04-24 | 青岛理工大学 | Deep learning-based assembly operation monitoring method and device |
WO2021179898A1 (en) * | 2020-03-11 | 2021-09-16 | 深圳市商汤科技有限公司 | Action recognition method and apparatus, electronic device, and computer-readable storage medium |
CN111461078A (en) * | 2020-05-18 | 2020-07-28 | 江苏电力信息技术有限公司 | Anti-fishing monitoring method based on computer vision technology |
CN111931549A (en) * | 2020-05-20 | 2020-11-13 | 浙江大学 | Human skeleton action prediction method based on multitask non-autoregressive decoding |
CN111931549B (en) * | 2020-05-20 | 2024-02-02 | 浙江大学 | Human skeleton motion prediction method based on multi-task non-autoregressive decoding |
CN111611927A (en) * | 2020-05-21 | 2020-09-01 | 长沙明本信息科技有限公司 | Method for identifying unsafe behaviors of coal mine workers based on human body postures |
CN111723741A (en) * | 2020-06-19 | 2020-09-29 | 江苏濠汉信息技术有限公司 | Temporary fence movement detection alarm system based on visual analysis |
CN111783692A (en) * | 2020-07-06 | 2020-10-16 | 广东工业大学 | Action recognition method and device, electronic equipment and storage medium |
CN112084861A (en) * | 2020-08-06 | 2020-12-15 | 中国科学院空天信息创新研究院 | Model training method and device, electronic equipment and storage medium |
US11270147B1 (en) | 2020-10-05 | 2022-03-08 | International Business Machines Corporation | Action-object recognition in cluttered video scenes using text |
GB2614170B (en) * | 2020-10-05 | 2023-12-13 | Ibm | Action-object recognition in cluttered video scenes using text |
WO2022074483A1 (en) * | 2020-10-05 | 2022-04-14 | International Business Machines Corporation | Action-object recognition in cluttered video scenes using text |
GB2614170A (en) * | 2020-10-05 | 2023-06-28 | Ibm | Action-object recognition in cluttered video scenes using text |
US11928849B2 (en) | 2020-10-05 | 2024-03-12 | International Business Machines Corporation | Action-object recognition in cluttered video scenes using text |
CN112016528B (en) * | 2020-10-20 | 2021-07-20 | 成都睿沿科技有限公司 | Behavior recognition method and device, electronic equipment and readable storage medium |
CN112434655B (en) * | 2020-12-07 | 2022-11-08 | 安徽大学 | Gait recognition method based on adaptive confidence map convolution network |
CN112434655A (en) * | 2020-12-07 | 2021-03-02 | 安徽大学 | Gait recognition method based on adaptive confidence map convolution network |
CN112613490A (en) * | 2021-01-08 | 2021-04-06 | 云从科技集团股份有限公司 | Behavior recognition method and device, machine readable medium and equipment |
CN112613490B (en) * | 2021-01-08 | 2022-02-01 | 云从科技集团股份有限公司 | Behavior recognition method and device, machine readable medium and equipment |
CN113052061A (en) * | 2021-03-22 | 2021-06-29 | 中国石油大学(华东) | Speed skating athlete motion identification method based on human body posture estimation |
CN113033458B (en) * | 2021-04-09 | 2023-11-07 | 京东科技控股股份有限公司 | Action recognition method and device |
WO2022213857A1 (en) * | 2021-04-09 | 2022-10-13 | 京东科技控股股份有限公司 | Action recognition method and apparatus |
CN113033458A (en) * | 2021-04-09 | 2021-06-25 | 京东数字科技控股股份有限公司 | Action recognition method and device |
US11423252B1 (en) | 2021-04-29 | 2022-08-23 | International Business Machines Corporation | Object dataset creation or modification using labeled action-object videos |
CN113283381B (en) * | 2021-06-15 | 2024-04-05 | 南京工业大学 | Human body action detection method suitable for mobile robot platform |
CN113283381A (en) * | 2021-06-15 | 2021-08-20 | 南京工业大学 | Human body action detection method suitable for mobile robot platform |
CN113536997A (en) * | 2021-07-01 | 2021-10-22 | 深圳中智明科智能科技有限公司 | Intelligent security system and method based on image recognition and behavior analysis |
CN114642858A (en) * | 2022-02-18 | 2022-06-21 | 上海大学 | Rehabilitation training device for senile asthenia patients |
CN114842560A (en) * | 2022-07-04 | 2022-08-02 | 广东瑞恩科技有限公司 | Computer vision-based construction site personnel dangerous behavior identification method |
CN116259003A (en) * | 2023-01-06 | 2023-06-13 | 苏州同企人工智能科技有限公司 | Construction category identification method and system in construction scene |
CN116259003B (en) * | 2023-01-06 | 2023-11-10 | 苏州同企人工智能科技有限公司 | Construction category identification method and system in construction scene |
CN116363566B (en) * | 2023-06-02 | 2023-10-17 | 华东交通大学 | Target interaction relation recognition method based on relation knowledge graph |
CN116363566A (en) * | 2023-06-02 | 2023-06-30 | 华东交通大学 | Target interaction relation recognition method based on relation knowledge graph |
Also Published As
Publication number | Publication date |
---|---|
CN109492581B (en) | 2023-07-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109492581A (en) | A kind of human motion recognition method based on TP-STG frame | |
CN110425005B (en) | Safety monitoring and early warning method for man-machine interaction behavior of belt transport personnel under mine | |
CN110147743B (en) | Real-time online pedestrian analysis and counting system and method under complex scene | |
CN108710868B (en) | Human body key point detection system and method based on complex scene | |
CN107330920B (en) | Monitoring video multi-target tracking method based on deep learning | |
CN107133569A (en) | The many granularity mask methods of monitor video based on extensive Multi-label learning | |
CN111488804A (en) | Labor insurance product wearing condition detection and identity identification method based on deep learning | |
CN107330396A (en) | A kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study | |
CN111666843A (en) | Pedestrian re-identification method based on global feature and local feature splicing | |
CN109902573A (en) | Multiple-camera towards video monitoring under mine is without mark pedestrian's recognition methods again | |
CN110321873A (en) | Sensitization picture recognition methods and system based on deep learning convolutional neural networks | |
CN105528794A (en) | Moving object detection method based on Gaussian mixture model and superpixel segmentation | |
CN105243356B (en) | A kind of method and device that establishing pedestrian detection model and pedestrian detection method | |
CN106778474A (en) | 3D human body recognition methods and equipment | |
CN111191667A (en) | Crowd counting method for generating confrontation network based on multiple scales | |
CN109903339B (en) | Video group figure positioning detection method based on multi-dimensional fusion features | |
CN110390289A (en) | Based on the video security protection detection method for censuring understanding | |
CN115527269B (en) | Intelligent human body posture image recognition method and system | |
Janku et al. | Fire detection in video stream by using simple artificial neural network | |
CN106611158A (en) | Method and equipment for obtaining human body 3D characteristic information | |
CN110458022A (en) | It is a kind of based on domain adapt to can autonomous learning object detection method | |
CN110135237A (en) | A kind of gesture identification method | |
CN114943873B (en) | Method and device for classifying abnormal behaviors of staff on construction site | |
Luo et al. | A slight smoke perceptual network | |
CN112597842B (en) | Motion detection facial paralysis degree evaluation system based on artificial intelligence |
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 | ||
CB03 | Change of inventor or designer information |
Inventor after: Gong Faming Inventor after: Ji Xiaofeng Inventor after: Ma Yuhui Inventor after: Gong Wenjuan Inventor before: Gong Faming Inventor before: Ma Yuhui Inventor before: Gong Wenjuan |
|
CB03 | Change of inventor or designer information | ||
GR01 | Patent grant | ||
GR01 | Patent grant |