CN110348494A - A kind of human motion recognition method based on binary channels residual error neural network - Google Patents
A kind of human motion recognition method based on binary channels residual error neural network Download PDFInfo
- Publication number
- CN110348494A CN110348494A CN201910567065.7A CN201910567065A CN110348494A CN 110348494 A CN110348494 A CN 110348494A CN 201910567065 A CN201910567065 A CN 201910567065A CN 110348494 A CN110348494 A CN 110348494A
- Authority
- CN
- China
- Prior art keywords
- residual error
- neural network
- channel
- binary channels
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of human motion recognition methods based on binary channels residual error neural network, short channel/long-channel in binary channels residual error neural network intercepts the data of corresponding length according to the sliding time window length of setting from the sensor time series data of input；The data of interception successively carry out feature extraction by a convolutional layer and a residual error layer, obtain the output feature of short channel/long-channel；The output feature of short channel and long-channel inputs an articulamentum jointly and is spliced, a full articulamentum is inputted again to be integrated, obtain a vector, a possibility that vector dimension is equal to tag along sort sum, and the sensor time series data tag along sort of the corresponding input of the element value of kth dimension is k size；Network is trained based on sample data；The sensor time series data input of unknown tag along sort is trained into network again, determines its human action tag along sort.The present invention is able to achieve high-precision human action identification.
Description
Technical field
The invention belongs to human action identification technology fields, and in particular to a kind of people based on binary channels residual error neural network
Body action identification method.
Background technique
With the development of mobile device and wearable sensing equipment and universal, allow one to pass through built-in sensors
(such as accelerometer, gyroscope, magnetometer etc.) real-time perception human action.Human action identifies in theoretical research and practice all
With important value, health detection, smart home, human-computer interaction can be widely applied to.The main target of human action identification
It is to realize the automatic detection and identification to human action by analyzing the data that sensor acquires.Human action identification
Key technology it is living including the use of the internet of things equipment such as portable sensor or camera acquisition physical activity data, processing human body
Dynamic data, feature extraction and Classification and Identification and etc..In order to enable a computer to identify different classes of human action, benefit is needed
Human action data is classified and identified with some linearly or nonlinearly classifiers.How to efficiently extract to have and well sentence
Other power is characterized in the major issue solved required for human action identification.
Traditional action identification method lays particular emphasis on the effective motion characteristic of manual extraction, very due to the correlation between feature
There is factors, the professions that manual feature extracting method needs related fields such as the noise for influencing accuracy of identification in feature and know in height
Know.If lacking the professional knowledge of related fields, the quality of feature extraction is difficult to be protected, and carries out identification meeting with this feature
Substantially reduce accuracy of identification.In addition, the feature of manual extraction is all shallow-layer, it is difficult to extract the further features to complexity.Therefore,
More systematic method is found to attract wide attention to obtain the feature with good judgement index.In recent years, machine learning is led
Domain achieves significant progress, these technologies allow model automatic learning characteristic from data, wherein it is great it is representative be base
In the feature extracting method of deep learning.Present deep learning method can be divided into two classes: full Connection Neural Network and convolution
Neural network.Full Connection Neural Network (for example, comprising an input layer, two or more hidden layers and an output
Layer) advanced abstract characteristics can be extracted by superposition hidden layer.Convolutional neural networks can be by from part to the overall situation
Weight is shared to model entire sequence, then extracts the abstract characteristics of deeper by a series of convolution operation and passes through
Original active signal is handled to capture potential feature.
With the development of deep learning, human action identification is carried out using deep neural network and has become human-computer interaction
Research hotspot.Computer can efficiently identify human action by learning a large amount of sensing data.In order to enable a computer to
Enough identify different human actions, key is to characterize to act using feature is differentiated to classify in turn.How effectively
Extracting, there is good judgement index to be characterized in the sixty-four dollar question solved required for human action identification.Though being based on depth
The method of habit achieves good achievement in human action identification, but due to the complexity of human action, currently
All there is certain deficiencies for recognition methods, cannot achieve high-precision human action identification.Therefore it needs to design a kind of new
Feature extracting method, to realize high-precision human action identification.
Summary of the invention
Technical problem solved by the invention is, in view of the deficiencies of the prior art, it proposes a kind of based on binary channels residual error
The human motion recognition method of neural network can efficiently extract the space-time characteristic of data, realize high-precision human action
Identification, carries out more accurate human action classification.
In order to achieve the above object, technical solution provided by the present invention are as follows:
A kind of human motion recognition method based on binary channels residual error neural network, comprising the following steps:
Step 1: obtaining the sensor time series sample data for having recorded human action feature and its corresponding human action class
Not；
Step 2: building binary channels residual error neural network model；
The binary channels residual error neural network model includes a short channel, a long-channel, an articulamentum and one
Full articulamentum；Short channel and long-channel respectively include a sequentially connected convolutional layer and residual error layer；One is connected after two residual error layers
A articulamentum connects a full articulamentum after articulamentum；
The sliding time window length of short channel and long-channel is set, is denoted as T1 and T2, T1 < T2, so that short logical respectively
The data length of road interception is less than the data length of long-channel interception；
The Data Data treatment process of binary channels residual error neural network model are as follows: channel/long-channel short first is according to setting
Sliding time window length, from the sensor time series data of input intercept corresponding length data；Short channel/long-channel is cut
The data taken are first pre-processed by a convolutional layer, then carry out depth characteristic extraction by a residual error layer, obtain short channel/length
The output feature in channel；The output feature of short channel and long-channel inputs an articulamentum jointly and is spliced, and obtains more smart
Thin motion characteristic；The motion characteristic of articulamentum output inputs full articulamentum again and is integrated, and exports as a result, the output result is
One vector, dimension are equal to tag along sort sum, the sensor time series data of the corresponding input of the element value of kth dimension corresponding the
A possibility that k anthropoid movement size；
Step 3: the sensor time series sample data and its corresponding human action tag along sort obtained based on step 1 is to double
Channel residual error neural network model is trained；
Step 4: by the sensor time series data input of unknown tag along sort through the trained binary channels residual error nerve of step 3
Network model, a possibility that determining its correspondence maximum human action classification, to realize that human action identifies.
Further, identical (the i.e. short channel and long logical of the structure of two residual error layers in binary channels residual error neural network model
The structure of residual error layer is identical in road)；Each residual error layer includes the multiple residual error modules being sequentially connected in series；Each residual error module include according to
Secondary concatenated multiple sub- residual blocks；Every sub- residual block includes the multiple convolutional layers being sequentially connected in series；It is last in every sub- residual block
After the output of one convolutional layer is added with the input of the sub- residual block, then through an activation primitive, obtain the defeated of the sub- residual block
Out.
Further, each residual error layer includes 4 residual error modules being sequentially connected in series；Each residual error modular structure is as follows:
First residual error module includes 3 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes being sequentially connected in series
3 convolutional layers, the convolution kernel size of 3 convolutional layers is respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 64,64 and
256；
Second residual error module includes 4 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes being sequentially connected in series
3 convolutional layers, the convolution kernel size of 3 convolutional layers is respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 512,512 and
256；
Third residual error module includes 6 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes being sequentially connected in series
3 convolutional layers, the convolution kernel size of 3 convolutional layers is respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 256,256 and
1024；
4th residual error module includes 3 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes being sequentially connected in series
3 convolutional layers, the convolution kernel size of 3 convolutional layers is respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 128,128 and
512。
Further, the output result of full articulamentum is denoted as (z1, z2..., zK), it is total that dimension K is equal to tag along sort
Number, k-th of dimension correspond to tag along sort k,
Softmax classifier is provided with after full articulamentum, output of the softmax classifier based on full articulamentum as a result,
Softmax formula is used to calculate the sensor time series data x tag along sort of current input binary channels residual error neural network model as k
Probability p (k | x)；Softmax formula is as follows:
IfForIn maximum value, then model determines that training sample x is theIt is anthropoid
Movement.
Further, in the training process of binary channels residual error neural network model, loss function uses following cross entropy
Loss function:
Wherein,Indicate penalty values, the degree of closeness for reality output and desired output；P (k | x) binary channels residual error mind
The probability that the training sample x tag along sort for passing through softmax classifier reality output through network model is k；Q (k | x) indicate the phase
The training sample x tag along sort for hoping output is the probability of k, if the true tag along sort of training sample x is k, q (k | x)=1,
Otherwise q (k | x)=0.
The utility model has the advantages that
Residual error neural network of the invention can be realized the automatic study of advanced features, while can prevent in depth conditions
Lower depth convolutional network performance degradation.Different from only with the single pass human motion recognition method based on deep learning, sheet
Invention is an asymmetric network, possesses two paths, respectively short channel and long-channel, and mainly acquisition acts number in short channel
According to space characteristics, long-channel mainly obtains the temporal characteristics of data.The present invention can efficiently handle information flow, can be from biography
The feature distribution of different human body movement is captured in sensor data, and learns effective human action feature automatically.The reality of data set
Test the result shows that, present invention accuracy rate with higher.The present invention can provide a New view for human body action recognition, new to think
Dimension facilitates the development of human action identification.
Detailed description of the invention
Fig. 1 is Inspiration Sources figure of the invention
Fig. 2 is the structural schematic diagram of binary channels residual error neural network model
Specific embodiment
For the purpose for making the embodiment of the present invention, technical solution and advantage are clearer, below in conjunction in inventive embodiments
Attached drawing, technical scheme in the embodiment of the invention is clearly and completely described.Described ground embodiment is ground of the invention
A part of the embodiment, rather than whole embodiments.
Fig. 1 is Inspiration Sources figure of the invention.Creation Inspiration Sources of the invention are in Biological Knowledge, as shown in Figure 1, when
When biological vision processing system handles a picture, Parvocellular cell and Magnocellular cell in brain will
The meeting effect of generation simultaneously, Parvocellular cell receptive field is smaller, can extract a large amount of space characteristics in picture.
Magnocellular cell receptive field is bigger, can extract a large amount of temporal characteristics in picture.Based on this, the present invention is proposed
A kind of human motion recognition method based on binary channels residual error neural network model, binary channels residual error neural network model include
Two parts, are referred to as short channel and long-channel.Short channel is intercepted using shorter sliding time window and is handled sensor
Data, can efficiently extract the space characteristics (similar to the Parvocellular cell in brain) of data, and long-channel uses
Longer sliding time window intercepts and handles sensing data, and the temporal characteristics that can efficiently extract data (are similar to big
Magnocellular cell in brain).Therefore, the present invention can efficiently extract time and the space characteristics of movement, so as to
Enough realize high-precision human action identification.The most method based on deep learning is all based on single pass at present, is not had
There are while considering time and the space characteristics of action data, the present invention is mentioned relative to these existing methods based on deep learning
The fine degree that high motion characteristic extracts.
The invention discloses a kind of human motion recognition method based on binary channels residual error neural network model, including it is following
Step:
Step 1: obtaining the sensor time series sample data and its corresponding human action mark for having recorded human action feature
Label；
The present embodiment chooses the public data collection OPPORTUNITY of human action identification as experimental data set.
OPPORTUNITY is widely used in the experimental study of human action identification, representative.
In OPPORTUNITY data set, ADL is the activation record of tester in its natural state, and Drill is to a system
20 repetitions of column movement.ADL1, ADL2, ADL3 and Drill are as training set, ADL4 in selection OPPORTUNITY data set
With ADL5 as test set.
Step 2: building binary channels residual error neural network model；
As shown in Fig. 2, binary channels residual error neural network model includes a short channel, a long-channel, an articulamentum
(Concate layers) and full articulamentum；Short channel and long-channel respectively include a sequentially connected convolutional layer and residual error layer；Two
It is connected to an articulamentum after a residual error layer, a full articulamentum is connected to after articulamentum；
The sliding time window length of short channel and long-channel is set, is denoted as T1 and T2, T1 < T2, so that short logical respectively
The data length of road interception is less than the data length of long-channel interception；Sliding time window is at the time of sensor time series data
Unit is managed, for intercepting the data of corresponding length from the sensor time series data of input；Raw sensor time series data be D ×
The data of T, wherein D indicates quantity (the corresponding sensor dimension of one measurement dimension of each sensor, the biography of sensor passage
It can be the data that each measurement dimension of single sensor measures in sensor time series data, be also possible to what multiple sensors measured
Data), D value is 63 in the present embodiment；T indicates time dimension, i.e. data collection point number in sensor time series data.Training
Collection and the data dimension of test set interception are 63 × T.
In the present embodiment, short channel is identical with the network structure of residual error layer in long-channel.Each residual error layer includes successively going here and there
Four residual error modules (n1, n2, n3, n4) of connection；Each residual error module includes the multiple sub- residual blocks being sequentially connected in series；Every height is residual
Poor block includes the multiple convolutional layers being sequentially connected in series；The output of the last one convolutional layer and the sub- residual block in every sub- residual block
After input is added, then through an activation primitive, obtain the output of the sub- residual block.
Specifically, network parameter is as shown in table 1 in the present embodiment:
1 network parameter table of table
First module of binary channels residual error neural network model is convolutional layer, includes one layer of convolution, there is 64 5*1 sizes
Convolution kernel, step-length 1, activation primitive ReLU, pond mode are maximum pond, are used for data prediction.Second module is residual
Poor layer, is divided into 4 big residual error modules, includes 3 groups of 3 convolutional layers in n1 residual error module, and convolution kernel is respectively 1*1,3*1,1*1,
The number of convolution kernel is respectively 64,64,256.K indicates the quantity of convolution kernel in Fig. 2, and S indicates the size of convolution kernel.Feature
Depth extraction be exactly constantly to carry out convolution operation, transmit low-level image feature in convolutional layer always, improve the utilization of feature
Rate can extract more validity features.Third module is articulamentum, is first carried out the output feature of short channel and long-channel complete
The average pond of office, is then attached operation and splices the feature that two channels are extracted.4th module is full articulamentum, will
The data characteristics that front obtains very high level conceptual after articulamentum is handled is integrated, and is exported as a result, the output result
For one-dimensional vector, it is denoted as (z1, z2..., zK), dimension K is equal to tag along sort sum, and k-th of dimension corresponds to tag along sort k,
The output result of full articulamentum is classified by softmax classifier.For each training sample x, pass through
Softmax formula calculates the probability that its tag along sort is k, i.e. p (k | x)；The formula of softmax is as follows:
IfForIn maximum value, then model determines that training sample x is theIt is anthropoid
Movement.
Step 3: training binary channels residual error neural network model.
Research learns that feature extracting method of the invention can show when sliding time window length T takes 32,64 and 96
Good performance out.Therefore in the present embodiment experimentation, T1 and T2 are taken as 32 and 64,32 and 96 and 64 Hes respectively
96, carry out three groups of experiments.In three groups of experimentations, respectively according to the training in different sliding time window length and step 1
Collection is trained network model.By residual error neural network, short channel can obtain a large amount of space characteristics, and long-channel can obtain
A large amount of temporal characteristics.It is merged the feature that short channel and long-channel export to obtain fine-grained data characteristics.
In order to make neural network learning to the feature of more resolving power, binary channels residual error nerve is judged by loss function
The reality output of network model and the degree of closeness of desired output.The present embodiment loss function uses following intersection entropy loss letter
Number:
Wherein,Indicate penalty values, the degree of closeness for reality output and desired output；Q (k | x) indicate desired output
Training sample x tag along sort be k probability, if the true tag along sort of training sample x is k, q (k | x)=1, otherwise q (k
| x)=0.
Step 4: human action classification being carried out to test set with network model, obtains experimental result.
Test set in step 1 is input in the network model that training finishes, carries out Performance Evaluation.Because
Sample size is uneven between OPPORTUNITY data set class, includes a large amount of Null class (empty class), so general classification is smart
Degree is not the best-evaluated index of recognition effect.The present embodiment is using weighted F1-score (weighting F1 score) FwAs
Evaluation index, FwIt is influenced simultaneously by accuracy and recall rate, calculation formula is expressed as follows:
Wherein, wk=Nk/Ntotal, NkIt indicates really as the number of the sample of the anthropoid movement of kth, NtotalIt is the total of sample
Number, pkIt indicates the accuracy of kth class, i.e., really accounts for model for the sample of the anthropoid movement of kth and determine the anthropoid movement of its kth
All sample specific gravity, rkIt indicates the recall rate of kth class, i.e., is really the sample of the anthropoid movement of kth, and model determines its kth
The sample of anthropoid movement accounts for the specific gravity of true all samples for the anthropoid movement of kth.
The experimental result obtained based on OPPORTUNITY data set is as shown in table 2:
Table 2: experimental result table
The experimental results showed that in evaluation index FwUnder, the binary channels residual error neural network model classification of the present embodiment training
Accuracy rate can reach 90% or more, can classify to fast accurate；When short channel time window length takes 32, the long-channel time
When window length takes 96, classifying quality is best.Compare three groups of experimental results it is found that the length chosen of time window is to the precision of classification
Influence is smaller, so the present invention proposes that model is the network model of a stabilization and robust.
Claims (5)
1. a kind of human motion recognition method based on binary channels residual error neural network, which comprises the following steps:
Step 1: obtaining the sensor time series sample data for having recorded human action feature and its corresponding human action classification；
Step 2: building binary channels residual error neural network model；
The binary channels residual error neural network model connects entirely including a short channel, a long-channel, an articulamentum and one
Connect layer；Short channel and long-channel respectively include a sequentially connected convolutional layer and residual error layer；A company is connected after two residual error layers
Layer is connect, connects a full articulamentum after articulamentum；
The sliding time window length of short channel and long-channel is set, is denoted as T1 and T2, T1 < T2 respectively, so that short channel is cut
The data length taken is less than the data length of long-channel interception；
The Data Data treatment process of binary channels residual error neural network model are as follows: channel/long-channel short first is according to the cunning of setting
Dynamic time window length, intercepts the data of corresponding length from the sensor time series data of input；Short channel/long-channel interception
Data are first pre-processed by a convolutional layer, then carry out depth characteristic extraction by a residual error layer, obtain short channel/long-channel
Output feature；The output feature of short channel and long-channel inputs an articulamentum jointly and is spliced, and obtains more fine
Motion characteristic；The motion characteristic of articulamentum output inputs full articulamentum again and is integrated, and output is as a result, the output result is one
Vector, dimension are equal to tag along sort sum, and the sensor time series data of the corresponding input of the element value of kth dimension corresponds to kth class
A possibility that human action size；
Step 3: the sensor time series sample data and its corresponding human action tag along sort obtained based on step 1 is to binary channels
Residual error neural network model is trained；
Step 4: the sensor time series data of unknown classification is inputted through the trained binary channels residual error neural network mould of step 3
Type realizes human action identification.
2. the human motion recognition method according to claim 1 based on binary channels residual error neural network, which is characterized in that
The structure of two residual error layers is identical in binary channels residual error neural network model；Each residual error layer includes the multiple residual errors being sequentially connected in series
Module；Each residual error module includes the multiple sub- residual blocks being sequentially connected in series；Every sub- residual block includes the multiple volumes being sequentially connected in series
Lamination；After the output of the last one convolutional layer is added with the input of the sub- residual block in every sub- residual block, then through an activation
Function obtains the output of the sub- residual block.
3. the human motion recognition method according to claim 2 based on binary channels residual error neural network, which is characterized in that
Each residual error layer includes 4 residual error modules being sequentially connected in series；Each residual error modular structure is as follows:
First residual error module includes 3 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes 3 be sequentially connected in series
Convolutional layer, the convolution kernel size of 3 convolutional layers are respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 64,64 and 256；
Second residual error module includes 4 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes 3 be sequentially connected in series
Convolutional layer, the convolution kernel size of 3 convolutional layers are respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 512,512 and 256；
Third residual error module includes 6 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes 3 be sequentially connected in series
Convolutional layer, the convolution kernel size of 3 convolutional layers are respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 256,256 and 1024；
4th residual error module includes 3 sub- residual blocks being sequentially connected in series, wherein every sub- residual block includes 3 be sequentially connected in series
Convolutional layer, the convolution kernel size of 3 convolutional layers are respectively 1*1,3*1 and 1*1, and convolution kernel number is respectively 128,128 and 512.
4. the human motion recognition method according to claim 1 based on binary channels residual error neural network, which is characterized in that
The output result of full articulamentum is denoted as (z1, z2..., zK), dimension K is equal to tag along sort sum, and k-th of dimension is corresponding to divide
Class label k,
Softmax classifier is provided with after full articulamentum, output of the softmax classifier based on full articulamentum as a result, using
The sensor time series data x tag along sort that softmax formula calculates current input binary channels residual error neural network model is the general of k
Rate p (k | x)；Softmax formula is as follows:
IfFor p (k | x),In maximum value, then model determines that training sample x is theIt is anthropoid dynamic
Make.
5. the human motion recognition method according to claim 1 based on binary channels residual error neural network, which is characterized in that
In the training process of binary channels residual error neural network model, loss function uses following cross entropy loss function:
Wherein, l indicates penalty values, the degree of closeness for reality output and desired output；P (k | x) binary channels residual error nerve net
The probability that the training sample x tag along sort that network model passes through softmax classifier reality output is k；Q (k | x) indicate that expectation is defeated
Training sample x tag along sort out is the probability of k, if the true tag along sort of training sample x is k, q (k | x)=1, otherwise q
(k | x)=0.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910567065.7A CN110348494A (en) | 2019-06-27 | 2019-06-27 | A kind of human motion recognition method based on binary channels residual error neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910567065.7A CN110348494A (en) | 2019-06-27 | 2019-06-27 | A kind of human motion recognition method based on binary channels residual error neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110348494A true CN110348494A (en) | 2019-10-18 |
Family
ID=68176804
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910567065.7A Pending CN110348494A (en) | 2019-06-27 | 2019-06-27 | A kind of human motion recognition method based on binary channels residual error neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110348494A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111652170A (en) * | 2020-06-09 | 2020-09-11 | 电子科技大学 | Secondary radar signal processing method based on two-channel residual error deep neural network |
| CN112419268A (en) * | 2020-11-23 | 2021-02-26 | 广东电网有限责任公司广州供电局 | Method, device, equipment and medium for detecting image defects of power transmission line |
| CN112433207A (en) * | 2020-11-06 | 2021-03-02 | 浙江理工大学 | Human body identity recognition method based on two-channel convolutional neural network |
| CN113111886A (en) * | 2021-04-19 | 2021-07-13 | 太原科技大学 | Traffic scene image semantic segmentation method based on double-layer residual error network |
| WO2022161026A1 (en) * | 2021-01-28 | 2022-08-04 | Oppo广东移动通信有限公司 | Action recognition method and apparatus, and electronic device and storage medium |
| WO2022165675A1 (en) * | 2021-02-03 | 2022-08-11 | 深圳市锐明技术股份有限公司 | Gesture recognition method and apparatus, terminal device, and readable storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170098308A1 (en) * | 2015-10-02 | 2017-04-06 | Varian Medical Systems, Inc. | Systems and methods for detecting camera defect caused by exposure to radiation |
| CN107230200A (en) * | 2017-05-15 | 2017-10-03 | 东南大学 | A kind of method for extracting rotor coil contour feature |
| CN107506740A (en) * | 2017-09-04 | 2017-12-22 | 北京航空航天大学 | A kind of Human bodys' response method based on Three dimensional convolution neutral net and transfer learning model |
| CN108021889A (en) * | 2017-12-05 | 2018-05-11 | 重庆邮电大学 | A kind of binary channels infrared behavior recognition methods based on posture shape and movable information |
| CN109446923A (en) * | 2018-10-10 | 2019-03-08 | 北京理工大学 | Depth based on training characteristics fusion supervises convolutional neural networks Activity recognition method |
-
2019
- 2019-06-27 CN CN201910567065.7A patent/CN110348494A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170098308A1 (en) * | 2015-10-02 | 2017-04-06 | Varian Medical Systems, Inc. | Systems and methods for detecting camera defect caused by exposure to radiation |
| CN107230200A (en) * | 2017-05-15 | 2017-10-03 | 东南大学 | A kind of method for extracting rotor coil contour feature |
| CN107506740A (en) * | 2017-09-04 | 2017-12-22 | 北京航空航天大学 | A kind of Human bodys' response method based on Three dimensional convolution neutral net and transfer learning model |
| CN108021889A (en) * | 2017-12-05 | 2018-05-11 | 重庆邮电大学 | A kind of binary channels infrared behavior recognition methods based on posture shape and movable information |
| CN109446923A (en) * | 2018-10-10 | 2019-03-08 | 北京理工大学 | Depth based on training characteristics fusion supervises convolutional neural networks Activity recognition method |
Non-Patent Citations (1)
| Title |
|---|
| JUN LONG 等: "Asymmetric Residual Neural Network for Accurate Human Activity Recognition", 《ARXIV》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111652170A (en) * | 2020-06-09 | 2020-09-11 | 电子科技大学 | Secondary radar signal processing method based on two-channel residual error deep neural network |
| CN112433207A (en) * | 2020-11-06 | 2021-03-02 | 浙江理工大学 | Human body identity recognition method based on two-channel convolutional neural network |
| CN112433207B (en) * | 2020-11-06 | 2024-05-28 | 浙江理工大学 | Human body identity recognition method based on double-channel convolutional neural network |
| CN112419268A (en) * | 2020-11-23 | 2021-02-26 | 广东电网有限责任公司广州供电局 | Method, device, equipment and medium for detecting image defects of power transmission line |
| WO2022161026A1 (en) * | 2021-01-28 | 2022-08-04 | Oppo广东移动通信有限公司 | Action recognition method and apparatus, and electronic device and storage medium |
| WO2022165675A1 (en) * | 2021-02-03 | 2022-08-11 | 深圳市锐明技术股份有限公司 | Gesture recognition method and apparatus, terminal device, and readable storage medium |
| CN113111886A (en) * | 2021-04-19 | 2021-07-13 | 太原科技大学 | Traffic scene image semantic segmentation method based on double-layer residual error network |
| CN113111886B (en) * | 2021-04-19 | 2023-03-24 | 太原科技大学 | Traffic scene image semantic segmentation method based on double-layer residual error network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110348494A (en) | A kind of human motion recognition method based on binary channels residual error neural network | |
| Xie et al. | A deep-learning-based real-time detector for grape leaf diseases using improved convolutional neural networks | |
| Jiang et al. | Real-time detection of apple leaf diseases using deep learning approach based on improved convolutional neural networks | |
| He et al. | Application of deep learning in integrated pest management: a real‐time system for detection and diagnosis of oilseed rape pests | |
| CN106897738B (en) | A kind of pedestrian detection method based on semi-supervised learning | |
| CN104063719B (en) | Pedestrian detection method and device based on depth convolutional network | |
| CN109815785A (en) | A kind of face Emotion identification method based on double-current convolutional neural networks | |
| CN108229298A (en) | The training of neural network and face identification method and device, equipment, storage medium | |
| CN106651830A (en) | Image quality test method based on parallel convolutional neural network | |
| CN109583322A (en) | A kind of recognition of face depth network training method and system | |
| CN109497990B (en) | Electrocardiosignal identity recognition method and system based on canonical correlation analysis | |
| CN105303179A (en) | Fingerprint identification method and fingerprint identification device | |
| CN104408405B (en) | Face representation and similarity calculating method | |
| CN110674875A (en) | Pedestrian motion mode identification method based on deep hybrid model | |
| CN111954250B (en) | Lightweight Wi-Fi behavior sensing method and system | |
| Bu | Human motion gesture recognition algorithm in video based on convolutional neural features of training images | |
| CN111914643A (en) | Human body action recognition method based on skeleton key point detection | |
| CN109522961A (en) | A kind of semi-supervision image classification method based on dictionary deep learning | |
| CN106874929A (en) | A kind of pearl sorting technique based on deep learning | |
| CN108764486A (en) | A kind of feature selection approach and device based on integrated study | |
| CN105989336A (en) | Scene identification method based on deconvolution deep network learning with weight | |
| CN106295675A (en) | A kind of Activity recognition method of based on smart mobile phone of high accuracy | |
| CN104376611A (en) | Method and device for attendance of persons descending well on basis of face recognition | |
| Li et al. | Dating ancient paintings of Mogao Grottoes using deeply learnt visual codes | |
| CN109325513A (en) | A kind of image classification network training method based on magnanimity list class single image |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191018 |