CN113642432A - Method for identifying human body posture by convolutional neural network based on covariance matrix transformation - Google Patents
Method for identifying human body posture by convolutional neural network based on covariance matrix transformation Download PDFInfo
- Publication number
- CN113642432A CN113642432A CN202110869054.1A CN202110869054A CN113642432A CN 113642432 A CN113642432 A CN 113642432A CN 202110869054 A CN202110869054 A CN 202110869054A CN 113642432 A CN113642432 A CN 113642432A
- Authority
- CN
- China
- Prior art keywords
- data
- matrix
- neural network
- covariance matrix
- convolutional
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a method for identifying human body postures by a convolutional neural network based on covariance matrix transformation, which relates to the application field of artificial intelligence, the method comprises collecting sensor time sequence related to human body posture by using smart phone, normalizing raw data, segmenting data set by sliding window and fixing action attribute label, then a convolution neural network system is built by utilizing covariance matrix transformation, the data correlation is eliminated by utilizing the inverse transformation of the covariance matrix, before each layer of convolution, covariance operation is added to avoid data correlation, a training data set is used for training network parameters, a test data set is used for verifying network precision, finally a network model is solidified to generate pb files, the pb files are transferred to an Android client through Android studio, the mobile device may be enabled to identify the state the user is currently in using its own accelerometer and gyroscope.
Description
Technical Field
The invention belongs to the field of attitude sensor data identification, and particularly relates to a human body attitude identification method for wearable equipment by using a convolutional neural network based on covariance matrix transformation.
Background
With the progress of science and technology and the development of society, artificial intelligence reaches the unprecedented development height, has been integrated into the aspect of life, and for intelligent equipment, better detection of human activities and states is a necessary research direction, and the daily behaviors of people can be improved while the life of people is facilitated. More and more sensors are integrated in wearable equipment, and the wearable equipment is applied to various fields: physical therapy, motion recognition, social interaction and the like, sensors acquire time sequence data as research objects, reasonable prediction is carried out on different models, and information related to the human body state can be obtained. The posture research mainly comprises two types of modes based on video images and sensor data, and essentially, the learning neural network can automatically identify the user state under certain sample training through big data input. Compared with gesture recognition based on videos or pictures, gesture recognition based on sensor data has gradually become a mainstream method due to the advantages of strong anti-interference capability, unlimited environment, low required cost and the like. In most sensor-based gesture recognition methods, a user must wear a plurality of dedicated motion sensors and attach to a specific part of the body, and although these methods have been able to achieve high recognition performance, their limitations on location are not well suited for long-term monitoring of the human body. With the development of technology, various embedded sensors (such as accelerometers, gyroscopes, magnetometers, and the like) carried by smart phones can be used to extract a large amount of valuable sensor data, and have been successfully used for human body gesture recognition.
Machine learning has been studied since the rise of computers in the 20 th century, convolutional neural networks have not entered the view of the public until the deep learning in the current artificial intelligence is proposed, the recognition accuracy of gestures has been greatly improved from the initial manual feature extraction to the automatic extraction of useful features using fully-connected layers to the current convolutional layers, from support vector machines, decision trees, bayesian models to the current deep learning models, and the data processing capability of computers has been qualitatively improved with the improvement of computing power, and is applied to various pattern recognition and image recognition. Human body posture recognition belongs to a typical recognition problem, the development of the process is promoted by the wide application of a Convolutional Neural Network (CNN) in human body posture recognition (HAR), the most advanced performance is shown, the application range is expanded, and more researchers begin to invest in the field. Since the sensor data collected by the smart phone is ordered in time, each layer of convolution in the convolution network will cause the correlation between the data to increase gradually, so that the problem is also one of the problems existing in the current convolution layer.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to solve the problem of data correlation, and provides a method for identifying human body postures by using a convolutional neural network based on covariance matrix transformation, which solves the problems of identification precision attenuation and operation duration caused by correlation to a certain extent.
The technical scheme is as follows: a method for using a convolutional neural network based on covariance matrix transformation for human body posture recognition comprises the following steps according to the specific implementation aspect of the invention:
step1, acquiring human posture and motion signal data by using sensors in wearable equipment such as a smart phone and the like to obtain different types of motion attributes, such as walking, going upstairs and downstairs, running, falling and the like;
step2, carrying out data preprocessing, data normalization, 1 norm processing and fixed Step length sliding window processing on the collected action signals, dividing according to data attributes and adding corresponding action labels;
step3, building a convolutional neural network with a covariance matrix, setting hyper-parameters and data batches, sending the processed data as input vectors into network training, and updating the weight value of a convolutional kernel by utilizing gradient descent back propagation to achieve optimal network performance;
step4, detecting the generalization ability of the model by using the public data set, transferring the trained network model to the Android client, and detecting the state of the human body in real time by using the acceleration sensor of the smart phone.
In Step1, in the data acquisition process, the sampling frequency is 20Hz, and 3-axis acceleration sensor and gyroscope data are adopted, namely the acceleration direction and the rotation angle of the body, so as to obtain sensor data with a time sequence.
The data preprocessing part in Step2 carries out 1 norm processing on the data, eliminates noise data, compresses all data between 0 and 1, does not change sensor information contained in the data, then carries out sliding window processing to enable the data to be labeled, increases the data volume, and divides the processed data into a training set of 70 percent and a testing set of 30 percent:
three convolutional layers and two fully-connected layers are built in the convolutional neural network built by Step3, the size of each convolutional layer is set to be 3x3, the sliding Step is 1, 64 layers of convolution-128 layers of convolution-256 layers of convolution-fully-connected layers (Softmax) are sequentially arranged, the core of the whole network is based on two-dimensional discrete convolution, and the basic convolution formula is as follows:
wherein A is a matrix of a previous layer, K is a convolution kernel, B is a feature output after convolution, and when the convolution layer formula is applied to a neural network, the convolution layer formula is as follows:
wherein output is input of the next layer, each convolution kernel W corresponds to an offset b, each layer is multiplied by input X through the convolution kernel, then the characteristic output of the previous layer is obtained through nonlinear activation function, and covariance matrix transformation is added to each layer to eliminate data correlation, specifically: the basic formula of covariance matrix transformation is:
for the obtained sensor data matrix XM*NWhere N is the number of samples, M is the number of features, μ is the matrix mean of X, one motion sample N times, each time containing M features, is a plurality of random tensors X1……XnFor mutual interference between different actions, a covariance matrix is calculated, the correlation of the actions in the convolution network is eliminated by utilizing the reversible property of the matrix, and for a mathematical formula of inverse operation, the approximate square root of a covariance matrix D is calculated:
where COV is a covariance transformation, the obtained sensor data may be transformed by a covariance matrix to eliminate the correlation effect of both data direction and channel direction, and the following is a specific operation procedure.
When the human body posture data is collected and sent into the convolutional layer, the characteristics contained in the data are not influenced through matrix transformation, and the transformation process of the input data X can be regarded as an identity matrix with covariance:
where I is the identity matrix, indicating that this step does not change the action profile. In the operation process, covariance is applied to the convolution process to process data, sensor data with action information is firstly subjected to matrix inverse transformation to greatly reduce the correlation among the data and eliminate interference among different actions, then the convolution operation is carried out, multiplied by w (convolution kernel) and added with a bias matrix b to reduce the offset of the data, and the complete expression of the process is as follows:
y＝(X-μ)*D*W+b
since the neural network processes data with three convolutional layers, the covariance operation in the i-th layer is represented as DiAt this time, the input data of each layer is simplified into a convolution operation between the data of the previous layer and the current layer. Thus, the next winding of layer Xi+1Becomes:
Xi+1＝fi*wi*Di*Xi
wherein the matrix operation modes are all right multiplication matrix operation, XiIs input data from the i-th layer, DiIs to perform a covariance matrix operation on the input, WiIs a weight matrix in each convolutional layer, fiThe activation function is used for enabling the network to have the nonlinear classification capability, after the convolution operation, the relu activation function is adopted to transmit the information carried by the sensor data to the next layer:
f(x)＝max(0,x)。
wherein the matrix operation modes are all right multiplication matrix operation, XiIs input data from the i-th layer, DiIs to perform a covariance matrix operation on the input, WiIs a weight matrix in each convolutional layer, fiThe activation function is used for enabling the network to have the nonlinear classification capability, and the relu activation function is adopted to transmit the information carried by the sensor data to the next layer:
f(x)＝max(0,x)。
the core idea of this patent can be simplified to reconstruct the convolutional layer, before using the two-dimensional discrete convolution formula, the data matrix is inverse covariance transformed to eliminate the previous time series influence. This transformation is taken for all three layers of main convolution to build a complete convolutional neural network.
Has the advantages that: compared with the prior art, the technical scheme of the invention has the following beneficial technical effects:
the method comprises the steps of firstly utilizing a built-in sensor of a smart phone to sample, enabling data to come from a real human motion process, facilitating migration to a smart phone client at a later stage, then carrying out 1 norm processing on original data to enable the data to be more reasonably concentrated between 0 and 1, then eliminating the problem of data correlation by using matrix transformation to build a convolutional neural network based on covariance matrix transformation, reducing the reduction of identification precision of the data due to time sequence, and finally migrating a fixed network model to a mobile phone through Android Studio, expanding the use range and detecting human behaviors and the state in real time.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of the present invention;
FIG. 3 is a plot of a small batch of waveform of the raw triaxial acceleration data of the present invention;
FIG. 4 is a graph of accuracy variation corresponding to the number of training sessions in accordance with the present invention;
FIG. 5 is a graph of a confusion matrix for a test data set of the present invention;
FIG. 6 is a diagram of an identification interface for the migration of the present invention to the android terminal.
Detailed Description
The data collection process and network construction of the present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.
As shown in a flow chart, the invention provides a method for recognizing human body postures by using a convolutional neural network based on covariance matrix transformation, which comprises the following steps:
the method comprises the following steps: the testee all has the same model smart mobile phone (the model is iphone11), places the smart mobile phone in left and right wrist top through fixed cell-phone package, and then the testee carries out daily action (walks, runs, goes up and down stairs and jump etc.), and three-axis sensor data when the testee moved can be recorded to acceleration sensor and the gyroscope in the smart mobile phone, and 300 actions are all gathered in advance to every group data.
The data acquisition scheme is shown in the attached drawing, and the left hand and the right hand of the same tested person are provided with smart phones.
Step two: carrying out data preprocessing, data normalization, 1 norm processing and fixed step length sliding window processing on the collected action signals, and dividing and adding corresponding action labels according to data attributes;
step three: building a convolutional neural network with a covariance matrix, setting hyper-parameters and data batches, sending processed data serving as input vectors into network training, and updating the weight value of a convolutional kernel by utilizing gradient descent back propagation to achieve optimal network performance;
step four: the generalization capability of the model is detected by using the public data set, the trained network model is migrated to the Android client, and the state of the human body can be detected in real time by using an acceleration sensor of the smart phone.
In the first step, in the data acquisition process, the sampling frequency is 20Hz, the data of the 3-axis acceleration sensor and the gyroscope are adopted, and the data of the body acceleration direction and the rotation angle are respectively obtained to obtain the sensor data with the time sequence. For example, when a tester goes upstairs with a smart phone, an accelerometer arranged in the smart phone can record upward acceleration on the basis of walking so as to distinguish upstairs from flat road walking, for a neural network, different fine features can be extracted from a convolution layer to obtain a feature map about going upstairs, multilayer convolution further deepens a feature part, training is carried out through certain training data, and the network can automatically recognize human body postures by utilizing the active learning capacity of the neural network.
And the data preprocessing part in the second step is used for performing 1 norm processing on the data, eliminating noise data, compressing all the data between 0 and 1 without changing sensor information contained in the data, such as special acceleration signals carried by going upstairs and downstairs, then performing sliding window processing to enable the data to be provided with labels, increasing the data volume, and dividing the processed data into 70% of training sets and 30% of testing sets.
The data is processed to be a four-dimensional tensor, and simultaneously contains data, characteristics and channel information.
The two-dimensional convolutional neural network constructed in the third step mainly comprises a convolutional layer (convolutional layer), a Pooling layer (Pooling layer) and a Fully-connected layer (Fully-connected layer), and a complete convolutional neural network model can be constructed by coupling and superposing the layers. In the invention, a network builds three convolutional layers and two fully-connected layers, the reconstructed convolutional layers are provided with covariance matrix transformation for eliminating data correlation, the size of a convolution kernel in the network is set to be 3x3, the sliding step length is 1, and 64 layers of convolution, 128 layers of convolution, 256 layers of convolution and fully-connected layers (Softmax) are arranged in sequence, specifically: the basic formula of covariance matrix transformation is:
for the obtained sensor data matrix XM*NWhere N is the number of samples, M is the number of features, μ is the matrix mean of X, one motion sample N times, each time containing M features, is a plurality of random tensors X1……XnFor mutual interference between different actions, a covariance matrix is calculated, the correlation of the actions in the convolution network is eliminated by utilizing the reversible property of the matrix, and for a mathematical formula of inverse operation, the approximate square root of a covariance matrix D is calculated:
where COV is a covariance transformation, the obtained sensor data may be transformed by a covariance matrix to eliminate the correlation effect of both data direction and channel direction, and the following is a specific operation procedure.
When the human body posture data is collected and sent into the convolutional layer, the characteristics contained in the data are not influenced through matrix transformation, and the transformation process of the input data X can be regarded as an identity matrix with covariance:
where I is the identity matrix, indicating that this step does not change the action profile. In the operation process, covariance is applied to the convolution process to process data, sensor data with action information is firstly subjected to matrix inverse transformation to greatly reduce the correlation among the data and eliminate interference among different actions, then the convolution operation is carried out, multiplied by w (convolution kernel) and added with a bias matrix b to reduce the offset of the data, and the complete expression of the process is as follows:
y＝(X-μ)*D*W+b
since the neural network processes data with three convolutional layers, willThe covariance operation in layer i is denoted as DiAt this time, the input data of each layer is simplified into a convolution operation between the data of the previous layer and the current layer. Thus, the next winding of layer Xi+1Becomes:
Xi+1＝fi*wi*Di*Xi
wherein the matrix operation modes are all right multiplication matrix operation, XiIs input data from the i-th layer, DiIs to perform a covariance matrix operation on the input, WiIs a weight matrix in each convolutional layer, fiThe activation function is used for enabling the network to have the nonlinear classification capability, after the convolution operation, the relu activation function is adopted to transmit the information carried by the sensor data to the next layer:
f(x)＝max(0,x)。
the core idea of the patent can be simplified to rebuild the convolution layer, the original convolution calculation formula is changed, and before the two-dimensional discrete convolution formula is used, the data matrix is subjected to inverse covariance transformation to eliminate the previous time sequence influence. This transformation is taken for all three layers of main convolution to build a complete convolutional neural network.
The implementation is based on a deep learning module under Pytorch in the software writing phase.
The common data sets in the fourth step include a UCI-HAR data set, a PAMAP2 data set, a UNIMIB-SHAR data set, and a WISDM data set. The data sets can be downloaded on a UCI human posture recognition website, are the data sets which are most quoted in the industry, and can evaluate the performance and the generalization capability of the model.
The migration to the Android client side is characterized in that firstly, the network is enabled to reach the optimal precision through the training of a data set, at the moment, the model weight parameter is the optimal value, the network model is solidified and a pb file is generated, the network structure and the weight map are stored, the pb file can be imported into an Android program package by applying Android Studio, and the migration to a specific smart phone can be realized after compiling by building an Android app applet.
The human body posture recognition method for the wearable device based on the convolutional neural network can recognize six action postures of jumping, walking, going upstairs, going downstairs, standing and sitting.
FIG. 1 is a flow chart of an object of the present invention, which includes two types of processes, data acquisition and motion recognition, in which the data acquisition is performed from an original sensor, the data is preprocessed and then input to a convolutional neural network for model training, and an ideal model obtained after training is applied to a mobile intelligent terminal, so as to realize human body posture discrimination.
Fig. 2 is a structural diagram of a convolutional neural network based on covariance matrix transformation for human posture recognition. Including the convolutional layer and the final classification layer, only two convolutional layers are shown. The graph also comprises a reconstructed convolutional layer internal structure with a covariance matrix, namely after data are input, firstly, the data are subjected to correlation elimination through covariance matrix transformation, and then, ordinary two-dimensional discrete convolution is carried out convolution operation, wherein a weighted value comes from main learning through back propagation, and finally, the network is subjected to nonlinear activation through an activation function RELU, so that the network is changed from linear to nonlinear, a new convolutional layer is constructed and coupled to a convolutional neural network, and the last fully-connected layer is mainly used for the last classification task and generally selects a Softmax function.
FIG. 3 is a plot of a small batch of waveforms of raw sensor triaxial acceleration data. The down-sampling frequency of the motion sensor is artificially fixed at 30 Hz. The real-time acceleration of the three axes is shown.
In order to train a convolutional neural network, the invention divides the acquired data samples into two categories: training samples and test samples. And the training sample is used as the input of the convolutional neural network for model training, and the test sample is used as a reference basis for identifying the accuracy rate. In this embodiment, 70% of the data set is used as the training set, and 30% of the data set is used as the test set.
FIG. 4 is a graph of the accuracy change of the neural network model after 200epochs training.
The embodiment of the invention can reduce the use of computing resources and has stronger generalization capability and the like. Through experimental comparison, the implementation can be obviously superior to the traditional convolutional neural network in precision and the speed is basically even.
FIG. 5 is a diagram of the confusion matrix for the present implementation test set, which is a technique used to summarize the performance of the classification algorithm. If the number of samples in each class is not equal, or there are more than two classes in the dataset, then misleading may occur if only the classification accuracy is used as the criterion. Computing the confusion matrix allows a better understanding of the behavior of the classification model and the types of errors it makes. In the figure, it can be seen that the horizontal axis represents the predicted result, the vertical axis represents the true labeled result, and the main diagonal is the same number of samples as the predicted result and the true result.
By analyzing the confusion matrix, the recognition precision conditions of the convolutional neural network model to different actions can be obtained, so that the network parameters can be modified. The method uses the identification of the public data set Pamap2, and when the classification precision of the final model meets the requirement of practical application, the model can be transplanted to an intelligent terminal for use.
Fig. 6 is an identification interface diagram transplanted to an android terminal in the invention, and the mobile phone is Huawei mata30 pro.
It should be noted that the human body posture recognition model in the embodiment of the present invention may be integrated in an intelligent mobile terminal, where the intelligent terminal may specifically be a terminal such as a smart phone or a smart watch, and the intelligent terminal is not limited herein, and in a real-time detection process using the smart phone, what is given by the model is based on the trained network last Softmax classification recognition accuracy, that is, the higher the score of the network for each predicted value of the action, the more the network tends to be determined as the current action.
As can be seen from the above, the human posture recognition device in the embodiment of the present invention acquires acceleration data of the intelligent terminal, then performs time series signal 1 norm and normalization processing on the acquired data, and inputs the preprocessed data into a human posture recognition neural network model, thereby obtaining a human posture recognition result. Because the human body posture recognition model is obtained based on the training of the convolutional neural network of the preset training set, the current human body posture can be recognized by inputting the acceleration data obtained in real time into the trained human body posture recognition model, and the human body posture recognition based on the acceleration data and a non-visual means is realized.
In the above embodiments, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.
The above embodiments are only for illustrating the technical idea of the present invention, and the protection scope of the present invention is not limited thereby, and any modifications made on the basis of the technical scheme according to the technical idea of the present invention fall within the protection scope of the present invention.
Claims (3)
1. A convolutional neural network based on covariance matrix transformation is used for a human body posture recognition method, and is characterized by comprising the following steps:
acquiring human posture action signal data through a sensor in wearable equipment to obtain different types of action attributes, wherein the acquisition process uses a phyphox software to acquire the sensor data in real time, the sampling frequency is 20Hz when the data are acquired, 3-axis acceleration sensor and gyroscope data are acquired, and the acceleration magnitude and the rotation angle of each direction of a body are respectively acquired to obtain sensor data with a time sequence;
step two, carrying out data preprocessing, data normalization, 1 norm processing and fixed step length sliding window processing on the collected action signals, and dividing and adding corresponding action labels according to data attributes;
inputting human body posture action signal data into a convolutional neural network model, taking action attributes of different categories as output, and training to obtain the convolutional neural network model for autonomously judging the current posture of the human body; the convolutional neural network model is a convolutional neural network with a covariance matrix, a hyper-parameter and a data batch are set during training, sensor data with a time sequence are sent to a network for training, and the weight value of a convolutional kernel is updated by utilizing gradient descent back propagation to achieve the optimal network performance;
and step four, transferring the trained network model to an Android client, and detecting the state of the human body in real time by using an acceleration sensor of the smart phone.
2. The method for recognizing the attitude of the convolutional neural network based on covariance matrix transformation as claimed in claim 1, wherein in the second step, 1 norm processing is performed on the data, noise data is eliminated, all data are compressed between 0 and 1, sensor information contained in the data is not changed, then sliding window processing is performed to label the data, the data volume is increased, and the processed data is divided into 70% of training set and 30% of testing set.
3. The method for recognizing the attitude of the convolutional neural network based on covariance matrix transformation as claimed in claim 1, wherein in step three, three convolutional layers and two fully-connected layers are built in the convolutional neural network, the size of each convolutional kernel is set to 3x3, the sliding step is 1, and 64 convolutional layers-128 convolutional layers-256 convolutional layers-fully-connected layers are sequentially arranged, wherein the convolutional layers are added with covariance matrix transformation to eliminate data correlation, and the covariance matrix transformation basic formula is as follows:
for the obtained sensor data matrix XM*NWhere N is the number of samples, M is the number of features, μ is the matrix mean of X, one motion sample N times, each time containing M features, is a plurality of random tensors X1……XnFor mutual interference between different actions, a covariance matrix is calculated, the correlation of the actions in the convolution network is eliminated by utilizing the reversible property of the matrix, and for a mathematical formula of inverse operation, an approximate square root of a covariance matrix D is calculated:
the COV is covariance transformation, and the obtained sensor data can be converted by a covariance matrix, so that the correlation influence of the data direction and the channel direction is eliminated, and the specific process is as follows:
when the human body posture data is collected and sent into the convolutional layer, the characteristics contained in the data are not influenced through matrix transformation, and the transformation process of the input data X can be regarded as an identity matrix with covariance:
wherein, I is an identity matrix, which indicates that the action characteristic is not changed in this step, in the operation process, covariance is applied to the convolution process to process data, the sensor data with action information is firstly subjected to matrix inverse transformation to reduce the correlation between data and eliminate the interference between different actions, and then the convolution operation is performed, multiplied by w, which is a convolution kernel, and added with a bias matrix b to reduce the offset of the data, and the complete expression of this process is:
y＝(X-μ)*D*W+b
since the neural network processes data with three convolutional layers, the covariance operation in the i-th layer is represented as DiAt this time, the input data of each layer is simplified into the convolution operation between the data of the previous layer and the current layer, and the next convolution layer Xi+1Becomes:
Xi+1＝fi*wi*Di*Xi
f(x)＝max(0,x)
wherein, the matrix operation modes are all right multiplication matrix operation, XiIs input data from the i-th layer, DiIs to perform a covariance matrix operation on the input, WiIs a weight matrix in each convolutional layer, fiThe activation function is used for enabling the network to have the nonlinear classification capability, after the convolution operation, the relu activation function is adopted to transmit the information carried by the sensor data to the next layer, and finally the information passes through the linear classifierThe sensor data can be classified, so that the current action of the human body can be automatically identified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110869054.1A CN113642432A (en) | 2021-07-30 | 2021-07-30 | Method for identifying human body posture by convolutional neural network based on covariance matrix transformation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110869054.1A CN113642432A (en) | 2021-07-30 | 2021-07-30 | Method for identifying human body posture by convolutional neural network based on covariance matrix transformation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113642432A true CN113642432A (en) | 2021-11-12 |
Family
ID=78419020
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110869054.1A Pending CN113642432A (en) | 2021-07-30 | 2021-07-30 | Method for identifying human body posture by convolutional neural network based on covariance matrix transformation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113642432A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114943324A (en) * | 2022-05-26 | 2022-08-26 | 中国科学院深圳先进技术研究院 | Neural network training method, human motion recognition method and device, and storage medium |
| CN116226727A (en) * | 2023-05-05 | 2023-06-06 | 济宁政韵信息科技有限公司 | Motion recognition system based on AI |
-
2021
- 2021-07-30 CN CN202110869054.1A patent/CN113642432A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114943324A (en) * | 2022-05-26 | 2022-08-26 | 中国科学院深圳先进技术研究院 | Neural network training method, human motion recognition method and device, and storage medium |
| CN114943324B (en) * | 2022-05-26 | 2023-10-13 | 中国科学院深圳先进技术研究院 | Neural network training method, human motion recognition method and device, and storage medium |
| CN116226727A (en) * | 2023-05-05 | 2023-06-06 | 济宁政韵信息科技有限公司 | Motion recognition system based on AI |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110309861B (en) | Multi-modal human activity recognition method based on generation of confrontation network | |
| Gil-Martín et al. | Improving physical activity recognition using a new deep learning architecture and post-processing techniques | |
| CN110610158A (en) | Human body posture identification method and system based on convolution and gated cyclic neural network | |
| Zhang et al. | Human activity recognition based on time series analysis using U-Net | |
| Su et al. | HDL: Hierarchical deep learning model based human activity recognition using smartphone sensors | |
| Lu et al. | A multichannel CNN-GRU model for human activity recognition | |
| CN112464738B (en) | Improved naive Bayes algorithm user behavior identification method based on mobile phone sensor | |
| CN113642432A (en) | Method for identifying human body posture by convolutional neural network based on covariance matrix transformation | |
| CN111950437A (en) | Gait recognition method and device based on deep learning model and computer equipment | |
| CN108958482B (en) | Similarity action recognition device and method based on convolutional neural network | |
| CN111723662B (en) | Human body posture recognition method based on convolutional neural network | |
| CN109726662A (en) | Multi-class human posture recognition method based on convolution sum circulation combination neural net | |
| CN111753683A (en) | Human body posture identification method based on multi-expert convolutional neural network | |
| CN111178288A (en) | Human body posture recognition method and device based on local error layer-by-layer training | |
| Yuan et al. | Smartphone-based activity recognition using hybrid classifier | |
| Zhang et al. | Sensors-based human activity recognition with convolutional neural network and attention mechanism | |
| Hernandez et al. | A deep learning framework for construction equipment activity analysis | |
| Alvee et al. | Application of machine learning classifiers for predicting human activity | |
| CN111291804A (en) | Multi-sensor time series analysis model based on attention mechanism | |
| CN113705339B (en) | Cross-user human behavior recognition method based on antagonism domain adaptation strategy | |
| Wenzheng | Human activity recognition based on acceleration sensor and neural network | |
| CN112861679B (en) | Transfer learning method and system for behavior recognition | |
| CN114601455A (en) | Motion recognition method based on two-stage neural network | |
| CN112766420A (en) | Human behavior identification method based on time-frequency domain information | |
| Al-Wesabi et al. | Design of optimal deep learning based human activity recognition on sensor enabled internet of things environment |
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 |