CN108389227A - A kind of dimensional posture method of estimation based on multiple view depth perceptron frame - Google Patents
A kind of dimensional posture method of estimation based on multiple view depth perceptron frame Download PDFInfo
- Publication number
- CN108389227A CN108389227A CN201810171088.1A CN201810171088A CN108389227A CN 108389227 A CN108389227 A CN 108389227A CN 201810171088 A CN201810171088 A CN 201810171088A CN 108389227 A CN108389227 A CN 108389227A
- Authority
- CN
- China
- Prior art keywords
- network
- view
- dimensional
- perceptron
- encoder
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/50—Depth or shape recovery
- G06T7/55—Depth or shape recovery from multiple images
- G06T7/593—Depth or shape recovery from multiple images from stereo images
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
- G06T2207/10012—Stereo images
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of dimensional posture method of estimation based on multiple view depth perceptron frame proposed in the present invention, main contents include：Particular figure perceptron network, Multi-view Integration network, layering jump connection, data prediction, training and assessment, its process is, particular figure perceptron network extracts two-dimensional shapes and layering texture information from different views, particular figure perceptron network generates mapping for each view, by the high-resolution thermal map in each joint of hourglass network struction that encoder and decoder form, it is connected using the jump of hourglass network to realize Multi-view Integration network, Multi-view Integration network synthesizes the information from all available views, provides accurate 3 d pose.Hierarchy information is combined deduction three-dimensional structure by the present invention with estimation joint thermal map, can overcome the limitation directly measured with observation system, has higher accuracy to the estimation of dimensional posture.
Description
Technical field
The present invention relates to pose estimation fields, more particularly, to a kind of three-dimensional appearance based on multiple view depth perceptron frame
Gesture method of estimation.
Background technology
Human posture's estimation is the research hotspot of computer vision field in recent years.Computer system is by from image or regarding
Human posture's picture is extracted in frequency, the posture of human body is analyzed and counted, to judge the behavior of personage.Cause
This, human posture's estimation has extremely wide application.In unusual checking systematic difference, by being drawn to video monitoring
Personage's posture in face carries out detection and analysis in real time, when occurring the behaviors such as fight, steal in picture, system energy and
Shi Jilu simultaneously sends out alarm.In the application of sports posture analysis, the video analytic system pair of dimensional posture estimation is utilized
Video data when training athlete carries out analyzing processing, calculates relevant kinematic parameter and carries out specialty analysis, to for
Training next time provides specific aim guidance；It can also be used for competition tactics analysis, be carried out at analysis by the match video to opponent
Reason, can analyze the related data of opponent's posture, to assist sportsman to formulate related tactical concept.In body-sensing interactive game
In, by monitoring the posture and motion state of player in real time, convert player's body gesture before screen and action to
Data transfer is to games system, to realize human-computer interaction.The system for the capture 3 D human body posture developed in recent years, such as directly
Measure and observation system, these systems costly, and take it is longer, accuracy is not high.
The present invention proposes a kind of dimensional posture method of estimation based on multiple view depth perceptron frame, particular figure perception
Device network extracts two-dimensional shapes and layering texture information from different views, and particular figure perceptron network is directed to each view
Mapping is generated, by the high-resolution thermal map in each joint of hourglass network struction that encoder and decoder form, utilizes hourglass net
The jump of network connects to realize that Multi-view Integration network, Multi-view Integration network synthesize the information from all available views, carry
For accurate 3 d pose.Hierarchy information is combined deduction three-dimensional structure by the present invention with estimation joint thermal map, can be with
Overcome the limitation directly measured with observation system, there is higher accuracy to the estimation of dimensional posture.
Invention content
Costly for existing system, and longer, the not high problem of accuracy is taken, the purpose of the present invention is to provide
A kind of dimensional posture method of estimation based on multiple view depth perceptron frame, particular figure perceptron network is from different views
Extract two-dimensional shapes and layering texture information, particular figure perceptron network generates mapping for each view, by encoder and
The high-resolution thermal map in each joint of hourglass network struction of decoder composition is connected more to realize using the jump of hourglass network
View integrated network, Multi-view Integration network synthesize the information from all available views, provide accurate 3 d pose.
To solve the above problems, the present invention provides a kind of dimensional posture estimation side based on multiple view depth perceptron frame
Method, main contents include：
(1) particular figure perceptron network；
(2) Multi-view Integration network；
(3) layering jump connection；
(4) data prediction；
(5) training and assessment.
Wherein, the dimensional posture method of estimation based on multiple view depth perceptron frame, this method is by two nets
Network forms：One " particular figure perceptron " network extracts two-dimensional shapes and layering texture information from different views；And it is " more
View is integrated " network can then synthesize the information from all available views, accurate 3 d pose is provided.
Wherein, the particular figure perceptron network, particular figure perceptron network extract abundant from each view
Information further includes for the three-dimensional layering texture information inferred in next step wherein including not only two-dimensional shapes；Each two dimension body
Posture is indicated that wherein J indicates the quantity of body joints by J thermal maps；IfFor the input RGB image of view i, It is mapped for s-th of textural characteristics of view i,To regard
Scheme j-th of joint thermal map of i；Then, as follows for particular figure perceptron network (f) mapping of i-th of view：
Intermediate parity is performed by the loss of pixel thermal map：
Wherein, ‖ ‖ are Euclidean distances,It is to be rendered from the true two-dimensional attitude demarcated by gaussian kernel function
, average value is equal to the truthful data demarcated and variance；Then the hourglass network being made of encoder and decoder is used.
Further, the hourglass network, the figure layer that encoder is merged using convolution sum handle input picture, generate
The characteristic pattern of low resolution, decoder handle low resolution characteristic pattern using up-sampling and convolutional layer, build the height in each joint
Resolution ratio thermal map；One of the key components of hourglass network are that jump connects, i.e., the Feature Mapping before each convergence-level, it
Be directly appended to corresponding part in decoder, prevent the loss of encoder middle high-resolution information；The layering of these networks
Abundant texture information is shared in jump connection in varing proportions；Therefore, it is suggested that by being supplied to Multi-view Integration network
They are used for more efficient three-dimensional reasoning；They allow to use more rich gradient signal, and can provide more
Three-dimensional cues, and it is not only the combination using thermal map and untreated input picture.
Wherein, the Multi-view Integration network, the information of the multiple views of Multi-view Integration system integrating synthesize three
Tie up Attitude estimation；The input of the network is the connection for the output of the particular figure perceptron network of N number of different views, and
Output is 3 d pose；Each dimensional posture skeletonOne group of joint coordination being defined as in three dimensions；Therefore
The mapping of Multi-view Integration network (g) is as follows：
By assuming that three-dimensional joint annotation can be used for training dataset, loss function can be defined as：
Wherein, pjWithIt is the three-dimensional coordinate of the truthful data demarcated and estimation joint j respectively；
It is proposed a kind of data-driven method from bottom to top, it directly generates three from the output of the specific perceptron network of vision
Tie up posture skeleton；Multi-view Integration network is designed to encoder.
Further, the encoder tests two kinds of encoder：First, encoder by a series of kernels and
The convolutional layer that step-length is 2 is constituted, and wherein the resolution ratio of Feature Mapping is all half in each layer；Secondly, one is similar to hourglass
The encoder of network first part is replaced including maximum tether layer and Standard convolution layer by a pile residual error study module；
First and second network architectures are referred to as simple encoder and half hourglass network；For two kinds of network architecture, coding
Device exports and then is forwarded to the full articulamentum that output size is 3 × J, for estimating 3 d pose skeleton and measuring trained damage
Lose function；As can be seen that half hourglass network that is benefited from residual error module and being periodically inserted into maximum pond layer can provide ratio
The more accurate 3 d pose of simple encoder network compares.
Wherein, layering jump connection is connected using the jump of hourglass network to realize Multi-view Integration network；
In the frame of proposition, each in four jump connections of the encoder section generation of hourglass network is at residual error module
Reason, and be added with the corresponding part in half hourglass network；In order to handle multiple view setting, each jump connection should be as network
It is connected between view before input.
Wherein, the data prediction, in order to prepare training image, the down-sampled images from video；Each video bag
200 frames are included, the rate of 30 frame per second prevents overfitting only with odd-numbered frame；All images are adjusted to 256 × 256
Pixel, and be cut, so that main body is located at center；Three-dimensional joint annotation is provided by motion capture system.
Further, the described three-dimensional joint annotation, select 23 labels define including head, neck, left/right shoulder,
14 joints including left/right elbow, left/right wrist, left/right hip, left/right knee and left/right ankle, and using only the rail in these joints
Mark is used to train the joint of network；The coordination in each joint is normalized from zero to one in entire data set；After pretreatment, data
Structure by video each odd-numbered frame cutting image and the three-dimensional joint annotation of corresponding standardization form.
Wherein, the training and assessment propose a kind of two stage training strategy, use hourglass mould in the first stage
Type, and it is finely adjusted on weight lifting data set, learning rate 0.00025, by 5 periods；In second stage, by making
3 d pose skeleton is normalized with corresponding with dual-view image, from the beginning to Multi-view Integration model on weight lifting data set
It is trained；For 50 periods, the learning rate of model is 0.0005；
In order to assess the performance that single-view and dual-view is arranged in network, two experiments are carried out：First, network is for single
View setting is trained, and uses 90 degree and 135 degree of views respectively；Secondly, which is trained to total using two views
It is arranged with the dual-view as network inputs；In all experiments, the repetition of all objects and weight lifting task are as training number
According to collection, the repetition of all objects is as test data set.
Description of the drawings
Fig. 1 is a kind of system flow chart of the dimensional posture method of estimation based on multiple view depth perceptron frame of the present invention.
Fig. 2 is that a kind of multiple view of the dimensional posture method of estimation based on multiple view depth perceptron frame of the present invention is felt deeply and known
Device frame.
Fig. 3 is a kind of Multi-view Integration net of the dimensional posture method of estimation based on multiple view depth perceptron frame of the present invention
Network.
Specific implementation mode
It should be noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application can phase
It mutually combines, invention is further described in detail in the following with reference to the drawings and specific embodiments.
Fig. 1 is a kind of system flow chart of the dimensional posture method of estimation based on multiple view depth perceptron frame of the present invention.
Include mainly particular figure perceptron network, Multi-view Integration network, layering jump connection, data prediction, training and assessment.
Particular figure perceptron network extracts two-dimensional shapes and layering texture information, particular figure sense from different views
Know that device network generates mapping for each view, by the high score in each joint of hourglass network struction that encoder and decoder form
Resolution thermal map is connected using the jump of hourglass network to realize that Multi-view Integration network, the synthesis of Multi-view Integration network come from institute
The information for having available views provides accurate 3 d pose.
Wherein, in order to prepare training image, the down-sampled images from video；Each video includes 200 frames, 30 frame per second
Rate prevents overfitting only with odd-numbered frame；All images are adjusted to 256 × 256 pixels, and are cut, and make master
Body is located at center；Three-dimensional joint annotation is provided by motion capture system.
Selection 23 label come define including head, neck, left/right shoulder, left/right elbow, left/right wrist, left/right hip, a left side/
14 joints including right knee and left/right ankle, and the joint using only the track in these joints for training network；Entire data
The coordination in each joint is concentrated to be normalized from zero to one；After pretreatment, data structure by video each odd-numbered frame cutting
Image and the three-dimensional joint annotation composition of corresponding standardization.
It proposes a kind of two stage training strategy, uses hourglass model in the first stage, and to it on weight lifting data set
It is finely adjusted, learning rate 0.00025, by 5 periods；In second stage, return by using dual-view image and accordingly
One changes 3 d pose skeleton, is trained from the beginning to Multi-view Integration model on weight lifting data set；For 50 periods,
The learning rate of model is 0.0005；
In order to assess the performance that single-view and dual-view is arranged in network, two experiments are carried out：First, network is for single
View setting is trained, and uses 90 degree and 135 degree of views respectively；Secondly, which is trained to total using two views
It is arranged with the dual-view as network inputs；In all experiments, the repetition of all objects and weight lifting task are as training number
According to collection, the repetition of all objects is as test data set.
Fig. 2 is that a kind of multiple view of the dimensional posture method of estimation based on multiple view depth perceptron frame of the present invention is felt deeply and known
Device frame.Particular figure perceptron network extracts abundant information from each view, wherein including not only two-dimensional shapes, further includes
For the three-dimensional layering texture information inferred in next step；Each two dimension body gesture is indicated that wherein J indicates that body is closed by J thermal maps
The quantity of section；IfFor the input RGB image of view i,It is the of view i
S textural characteristics mapping,For j-th of joint thermal map of view i；Then, for i-th
Particular figure perceptron network (f) mapping of view is as follows：
Intermediate parity is performed by the loss of pixel thermal map：
Wherein, ‖ ‖ are Euclidean distances,It is to be rendered from the true two-dimensional attitude demarcated by gaussian kernel function
, average value is equal to the truthful data demarcated and variance；Then the hourglass network being made of encoder and decoder is used.
The figure layer that encoder is merged using convolution sum handles input picture, generates the characteristic pattern of low resolution, decoder
Low resolution characteristic pattern is handled using up-sampling and convolutional layer, builds the high-resolution thermal map in each joint；The pass of hourglass network
One of key component part is jump connection, i.e., the Feature Mapping before each convergence-level, they are directly appended in decoder
Corresponding part prevents the loss of encoder middle high-resolution information；The layering jump connection of these networks is shared in varing proportions
Abundant texture information；Therefore, it is suggested that they are used for more efficient three by being supplied to Multi-view Integration network
Tie up reasoning；They allow to use more rich gradient signal, and can provide more three-dimensional cues, and are not only using heat
The combination of figure and untreated input picture.
It is connected using the jump of hourglass network to realize Multi-view Integration network；In the frame of proposition, hourglass network
Encoder section generate four jump connection in each use residual error resume module, and with it is corresponding in half hourglass network
Part is added；In order to handle multiple view setting, each jump connection should connect before as network inputs between view
Come.
Fig. 3 is a kind of Multi-view Integration net of the dimensional posture method of estimation based on multiple view depth perceptron frame of the present invention
Network.The information of the multiple views of Multi-view Integration system integrating synthesizes 3 d pose estimation；The input of the network is for N number of
The connection of the output of the particular figure perceptron network of different views, and it is 3 d pose to export；Each dimensional posture skeletonOne group of joint coordination being defined as in three dimensions；Therefore the mapping of Multi-view Integration network (g) is as follows：
By assuming that three-dimensional joint annotation can be used for training dataset, loss function can be defined as：
Wherein, pjWithIt is the three-dimensional coordinate of the truthful data demarcated and estimation joint j respectively；
It is proposed a kind of data-driven method from bottom to top, it directly generates three from the output of the specific perceptron network of vision
Tie up posture skeleton；Multi-view Integration network is designed to encoder.
Test two kinds of encoder：First, encoder is made of a series of convolutional layer that kernels and step-length are 2,
The resolution ratio of middle Feature Mapping is all half in each layer；Secondly, an encoder for being similar to hourglass network first part,
Include that maximum tether layer and Standard convolution layer are replaced by a pile residual error study module；First and second network architectures claim respectively
For simple encoder and half hourglass network；For two kinds of network architecture, then encoder output is forwarded to output big
The small full articulamentum for being 3 × J, for estimating 3 d pose skeleton and measuring trained loss function；As can be seen that from residual error mould
It is benefited in block and the half hourglass network for being periodically inserted into maximum pond layer can provide more accurate than simple encoder network three
Dimension posture compares.
For those skilled in the art, the present invention is not limited to the details of above-described embodiment, in the essence without departing substantially from the present invention
In the case of refreshing and range, the present invention can be realized in other specific forms.In addition, those skilled in the art can be to this hair
Bright to carry out various modification and variations without departing from the spirit and scope of the present invention, these improvements and modifications also should be regarded as the present invention's
Protection domain.Therefore, the following claims are intended to be interpreted as including preferred embodiment and falls into all changes of the scope of the invention
More and change.
Claims (10)
1. a kind of dimensional posture method of estimation based on multiple view depth perceptron frame, which is characterized in that include mainly specific regard
Figure perceptron network (one)；Multi-view Integration network (two)；Layering jump connection (three)；Data prediction (four)；It trains and comments
Estimate (five).
2. based on the dimensional posture method of estimation based on multiple view depth perceptron frame described in claims 1, feature exists
In this method is made of two networks：One " particular figure perceptron " network extracts two-dimensional shapes from different views
With layering texture information；And " Multi-view Integration " network can then synthesize the information from all available views, provide accurately
3 d pose.
3. based on the particular figure perceptron network (one) described in claims 1, which is characterized in that particular figure perceptron net
Network extracts abundant information from each view, further includes for three-dimensional point inferred in next step wherein including not only two-dimensional shapes
Layer texture information；Each two dimension body gesture is indicated that wherein J indicates the quantity of body joints by J thermal maps；If
For the input RGB image of view i,It is mapped for s-th of textural characteristics of view i, For j-th of joint thermal map of view i；Then, for the particular figure perceptron of i-th of view
Network (f) mapping is as follows：
Intermediate parity is performed by the loss of pixel thermal map：
Wherein, ‖ ‖ are Euclidean distances,It is to be rendered from the true two-dimensional attitude demarcated by gaussian kernel function,
Average value is equal to the truthful data demarcated and variance；Then the hourglass network being made of encoder and decoder is used.
4. based on the hourglass network described in claims 3, which is characterized in that the figure layer that encoder is merged using convolution sum is located
Input picture is managed, the characteristic pattern of low resolution is generated, decoder handles low resolution characteristic pattern, structure using up-sampling and convolutional layer
Build the high-resolution thermal map in each joint；One of the key components of hourglass network be jump connection, i.e., each convergence-level it
Preceding Feature Mapping, they are directly appended to the corresponding part in decoder, prevent the loss of encoder middle high-resolution information；
Abundant texture information is shared in the layering jump connection of these networks in varing proportions；Therefore, it is suggested that by being supplied to
They are used for more efficient three-dimensional reasoning by Multi-view Integration network；They allow to use more rich gradient signal, and
More three-dimensional cues can be provided, and be not only the combination using thermal map and untreated input picture.
5. based on the Multi-view Integration network (two) described in claims 1, which is characterized in that Multi-view Integration system integrating
The information of multiple views is estimated to synthesize 3 d pose；The input of the network is the particular figure perception for N number of different views
The connection of the output of device network, and it is 3 d pose to export；Each dimensional posture skeletonIt is defined as three-dimensional space
Between in one group of joint coordination；Therefore the mapping of Multi-view Integration network (g) is as follows：
By assuming that three-dimensional joint annotation can be used for training dataset, loss function can be defined as：
Wherein, pjWithIt is the three-dimensional coordinate of the truthful data demarcated and estimation joint j respectively；
It is proposed a kind of data-driven method from bottom to top, it directly generates three-dimensional appearance from the output of the specific perceptron network of vision
State skeleton；Multi-view Integration network is designed to encoder.
6. based on the encoder described in claims 5, which is characterized in that test two kinds of encoder：First, encoder
A series of convolutional layer for being 2 by kernels and step-length is constituted, and wherein the resolution ratio of Feature Mapping is all half in each layer；Secondly,
One encoder for being similar to hourglass network first part, including maximum tether layer and Standard convolution layer by a pile residual error
Module is practised to be replaced；First and second network architectures are referred to as simple encoder and half hourglass network；For two kinds of networks
Architecture, then encoder output is forwarded to the full articulamentum that output size is 3 × J, for estimating 3 d pose skeleton
And measure trained loss function；As can be seen that half hourglass of maximum pond layer is benefited and is periodically inserted into from residual error module
Network can provide 3 d pose more accurate than simple encoder network and compare.
7. based on the layering jump connection (three) described in claims 1, which is characterized in that connected using the jump of hourglass network
To realize Multi-view Integration network；In the frame of proposition, in four jump connections of the encoder section generation of hourglass network
Each use residual error resume module, and be added with the corresponding part in half hourglass network；In order to handle multiple view setting, often
A jump connection should connect before as network inputs between view.
8. based on the data prediction (four) described in claims 1, which is characterized in that in order to prepare training image, from video
Middle down-sampled images；Each video includes 200 frames, and the rate of 30 frame per second prevents overfitting only with odd-numbered frame；It is all
Image be adjusted to 256 × 256 pixels, and be cut, main body made to be located at center；Three-dimensional joint annotation is by motion capture system
It provides.
9. based on the three-dimensional joint annotation described in claims 8, which is characterized in that selection 23 marks to define including head
14 joints including portion, neck, left/right shoulder, left/right elbow, left/right wrist, left/right hip, left/right knee and left/right ankle, and only
It is used to train the joint of network using the track in these joints；The coordination in each joint normalizing from zero to one in entire data set
Change；After pretreatment, data structure by video each odd-numbered frame cutting image and the three-dimensional joint of corresponding standardization annotate
Composition.
10. based on described in claims 1 training and assessment (five), which is characterized in that propose a kind of two stage training war
Slightly, hourglass model is used in the first stage, and it is finely adjusted on weight lifting data set, learning rate 0.00025, by 5
A period；In second stage, by using dual-view image and corresponding normalization 3 d pose skeleton, on weight lifting data set
Multi-view Integration model is trained from the beginning；For 50 periods, the learning rate of model is 0.0005；
In order to assess the performance that single-view and dual-view is arranged in network, two experiments are carried out：First, network is directed to single-view
Setting is trained, and uses 90 degree and 135 degree of views respectively；Secondly, which is trained to jointly make using two views
It is arranged for the dual-view of network inputs；In all experiments, the repetitions of all objects and weight lifting task as training dataset,
The repetition of all objects is as test data set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810171088.1A CN108389227A (en) | 2018-03-01 | 2018-03-01 | A kind of dimensional posture method of estimation based on multiple view depth perceptron frame |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810171088.1A CN108389227A (en) | 2018-03-01 | 2018-03-01 | A kind of dimensional posture method of estimation based on multiple view depth perceptron frame |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108389227A true CN108389227A (en) | 2018-08-10 |
Family
ID=63069225
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810171088.1A Withdrawn CN108389227A (en) | 2018-03-01 | 2018-03-01 | A kind of dimensional posture method of estimation based on multiple view depth perceptron frame |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108389227A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109376673A (en) * | 2018-10-31 | 2019-02-22 | 南京工业大学 | Method for identifying unsafe behaviors of underground coal mine personnel based on human body posture estimation |
| CN110751039A (en) * | 2019-09-18 | 2020-02-04 | 平安科技（深圳）有限公司 | Multi-view 3D human body posture estimation method and related device |
| CN112257686A (en) * | 2020-12-16 | 2021-01-22 | 北京沃东天骏信息技术有限公司 | Training method and device for human body posture recognition model and storage medium |
| CN112542230A (en) * | 2020-12-23 | 2021-03-23 | 南开大学 | Multi-mode interactive segmentation method and system for medical image |
| CN114036969A (en) * | 2021-03-16 | 2022-02-11 | 上海大学 | 3D human body action recognition algorithm under multi-view condition |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106528826A (en) * | 2016-11-18 | 2017-03-22 | 广东技术师范学院 | Deep learning-based multi-view appearance patent image retrieval method |
| CN107609587A (en) * | 2017-09-11 | 2018-01-19 | 浙江工业大学 | A kind of multi-class multi views data creation method that confrontation network is generated based on depth convolution |
-
2018
- 2018-03-01 CN CN201810171088.1A patent/CN108389227A/en not_active Withdrawn
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106528826A (en) * | 2016-11-18 | 2017-03-22 | 广东技术师范学院 | Deep learning-based multi-view appearance patent image retrieval method |
| CN107609587A (en) * | 2017-09-11 | 2018-01-19 | 浙江工业大学 | A kind of multi-class multi views data creation method that confrontation network is generated based on depth convolution |
Non-Patent Citations (1)
| Title |
|---|
| RAHIL MEHRIZI: "Toward Marker-free 3D Pose Estimation in Lifting:A Deep Multi-view Solution", 《ARXIV:1802.01741V1》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109376673A (en) * | 2018-10-31 | 2019-02-22 | 南京工业大学 | Method for identifying unsafe behaviors of underground coal mine personnel based on human body posture estimation |
| CN109376673B (en) * | 2018-10-31 | 2022-02-25 | 南京工业大学 | Method for identifying unsafe behaviors of underground coal mine personnel based on human body posture estimation |
| CN110751039A (en) * | 2019-09-18 | 2020-02-04 | 平安科技（深圳）有限公司 | Multi-view 3D human body posture estimation method and related device |
| WO2021051526A1 (en) * | 2019-09-18 | 2021-03-25 | 平安科技（深圳）有限公司 | Multi-view 3d human pose estimation method and related apparatus |
| CN110751039B (en) * | 2019-09-18 | 2023-07-25 | 平安科技（深圳）有限公司 | Multi-view 3D human body posture estimation method and related device |
| CN112257686A (en) * | 2020-12-16 | 2021-01-22 | 北京沃东天骏信息技术有限公司 | Training method and device for human body posture recognition model and storage medium |
| CN112542230A (en) * | 2020-12-23 | 2021-03-23 | 南开大学 | Multi-mode interactive segmentation method and system for medical image |
| CN114036969A (en) * | 2021-03-16 | 2022-02-11 | 上海大学 | 3D human body action recognition algorithm under multi-view condition |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108389227A (en) | A kind of dimensional posture method of estimation based on multiple view depth perceptron frame | |
| Li et al. | [Retracted] Intelligent Sports Training System Based on Artificial Intelligence and Big Data | |
| Zhu et al. | Visual7w: Grounded question answering in images | |
| CN105631861B (en) | Restore the method for 3 D human body posture from unmarked monocular image in conjunction with height map | |
| CN104035557B (en) | Kinect action identification method based on joint activeness | |
| CN105787439A (en) | Depth image human body joint positioning method based on convolution nerve network | |
| CN103227888B (en) | A kind of based on empirical mode decomposition with the video stabilization method of multiple interpretational criteria | |
| Avola et al. | Deep temporal analysis for non-acted body affect recognition | |
| CN102521595A (en) | Method for extracting image region of interest based on eye movement data and bottom-layer features | |
| CN105912991A (en) | Behavior identification method based on 3D point cloud and key bone nodes | |
| Reveret et al. | 3D visualization of body motion in speed climbing | |
| CN109045664A (en) | Diving scoring method, server and system based on deep learning | |
| CN106447707A (en) | Image real-time registration method and system | |
| CN112001217A (en) | Multi-person human body posture estimation algorithm based on deep learning | |
| Nie et al. | [Retracted] The Construction of Basketball Training System Based on Motion Capture Technology | |
| CN103310191A (en) | Human body action identification method for motion information imaging | |
| Guo et al. | PhyCoVIS: A visual analytic tool of physical coordination for cheer and dance training | |
| Sarwar et al. | Skeleton Based Keyframe Detection Framework for Sports Action Analysis: Badminton Smash Case | |
| CN116503954A (en) | Physical education testing method and system based on human body posture estimation | |
| Li | [Retracted] Tennis Technology Recognition and Training Attitude Analysis Based on Artificial Intelligence Sensor | |
| Palanimeera et al. | Yoga posture recognition by learning spatial-temporal feature with deep learning techniques | |
| Shi et al. | Sport training action correction by using convolutional neural network | |
| Li et al. | Analytical Model of Action Fusion in Sports Tennis Teaching by Convolutional Neural Networks | |
| Ascenso | Development of a non-invasive motion capture system for swimming biomechanics | |
| Luo | A sports digital training system based on middle and bottom visual information |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180810 |
|
| WW01 | Invention patent application withdrawn after publication |