CN109740651A - Limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks - Google Patents
Limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks Download PDFInfo
- Publication number
- CN109740651A CN109740651A CN201811581807.3A CN201811581807A CN109740651A CN 109740651 A CN109740651 A CN 109740651A CN 201811581807 A CN201811581807 A CN 201811581807A CN 109740651 A CN109740651 A CN 109740651A
- Authority
- CN
- China
- Prior art keywords
- data
- neural networks
- convolutional neural
- recognition methods
- norm
- 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
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 31
- 238000012545 processing Methods 0.000 title claims abstract description 31
- 238000000034 method Methods 0.000 title claims abstract description 22
- 230000009466 transformation Effects 0.000 title claims abstract description 11
- 238000012549 training Methods 0.000 claims abstract description 29
- 230000001133 acceleration Effects 0.000 claims abstract description 18
- 238000012360 testing method Methods 0.000 claims abstract description 9
- 230000000694 effects Effects 0.000 claims abstract description 8
- 238000001514 detection method Methods 0.000 claims abstract description 5
- 230000000875 corresponding effect Effects 0.000 claims description 4
- 210000002569 neuron Anatomy 0.000 claims description 4
- 230000008676 import Effects 0.000 claims description 2
- 230000009471 action Effects 0.000 abstract description 8
- 238000012544 monitoring process Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000011161 development Methods 0.000 description 3
- 230000036544 posture Effects 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 210000005036 nerve Anatomy 0.000 description 2
- 230000000386 athletic effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
Landscapes
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks that the invention discloses a kind of, includes the following steps: step 1, acquires 3-axis acceleration data by movable sensor, and stamp respective labels to every kind of Activity Type；Step 2, the processing of 1- norm is carried out to collected data, and divides training set and test set sample；Step 3, the training that above-mentioned processed data are carried out with convolutional neural networks, obtains most suitable weight and bigoted value, and .pb file is generated after arrangement；Step 4, file is entered and left into mobile intelligent terminal, obtains accurate human body limb motion detection effect.Recognition speed of the present invention is fast, and recognition accuracy is high.As long as the data bulk and type of acquisition are enough, the action classification that can will classify is expanded to more；It has important practical application of significance in terms of highway monitoring, human body attitude.
Description
Technical field
The present invention relates to the wearable intelligent monitoring methods of artificial intelligence field, more particularly to one kind to be based on 1- norm data
The limbs recognition methods of processing transformation and convolutional neural networks.
Background technique
Human body attitude identification is the hot spot of artificial intelligence and pattern identification research field.As intelligent wearable device is studied
Continuous development, based on wearable sensors human body attitude identification has become important field of research, such as athletic posture
Detection, smart home, intelligent medical assistant etc..But the posture form of human body be it is varied, even if same appearance
State can all make a big difference because of the difference of individual, how realize that a high-precision identification model becomes for people
The topic pursued.
In general, can be had more in human body to keep higher accuracy of identification and dispose multiple sensor devices on joint.Although
This method can intuitively find the acceleration signature of various movements, but user is required to carry multiple sensings in practical application
Device is very inconvenient.How using it is less in addition only with one group of sensor carry out high-accuracy human body attitude identify be one
It is a very actual to study a question.
Human body attitude identification is carried out using the built-in sensors of smart phone or smartwatch, it is early existing both at home and abroad much to grind
Study carefully application, most Intelligent bracelet wrist-watches and mobile phone have the application APP of gesture recognition on the market at present.Such human body attitude
The recognition methods overwhelming majority is threshold detection method, i.e., by judging that sensor is original or treated, whether data are more than or less than
Preset good threshold is come type of action of classifying.This method calculates simply, and the memory for occupying Intelligent mobile equipment is few, but with
This simultaneously, disadvantage is also apparent from: different product accuracy rate is irregular, and the action classification that can be identified is also extremely limited.This
On the one hand the reason of being each company research staff technological gap, prior one side the reason is that such method limitation.It needs
The action classification to be identified is more, and such algorithm constructs more complicated.
Deep learning has good development prospect in pattern-recognition.Deep learning is ground originating from artificial neural network
Study carefully.Wherein convolutional neural networks are the neural networks containing convolutional layer.Convolutional neural networks are in computer vision field by pole
Big concern, it not only can handle multidimensional data, but also more more aobvious than conventional method effect on classification is built.However convolution is refreshing
The characteristic of complex calculation through network determine it is more high to hardware-dependent degree, for common property computer be even more hardly possible
To realize the normal operation of network, how to reduce the desirability to hardware device and subtract under the premise of guaranteeing computational accuracy quality
Light computer load becomes primarily to solve the problems, such as.
Summary of the invention
Goal of the invention: in view of the above-mentioned problems, the object of the present invention is to provide one kind based on the data processing of 1- norm transformation and
The limbs recognition methods of convolutional neural networks, big to get rid of Computing load, operation time is long, recognition accuracy is low asks
Topic.
Technical solution: a kind of limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks, including
Following steps:
Step 1,3-axis acceleration data are acquired by movable sensor, and respective labels is stamped to every kind of Activity Type；
Step 2, the processing of 1- norm is carried out to collected data, and the data of processing is divided into training set and test
Collection；
Step 3, the training that above-mentioned processed data are carried out with convolutional neural networks obtains most suitable weight and bigoted
Value generates .pb file after arrangement；
Step 4, file is entered and left into mobile intelligent terminal, obtains accurate human body limb motion detection effect.
Specifically, sample frequency is set as 30Hz-40Hz in the step 1.
In the step 2, using the m% in data as training sample, n% is as test sample, the m%+n%=
100%, and 60%≤m%≤80%, 20%≤n%≤40%.It further, further include that data are carried out in the step 2
Again the processing of arranged in sequence after null value rejecting and rejecting.
The step 3 specifically includes following content:
3.1, establish 3 layers of convolutional neural networks model；
3.2, it imports training sample and adjusts convolutional neural networks model parameter, obtain the model of high-accuracy.
Preferably, in the convolutional neural networks model, first layer convolution kernel having a size of (1,3), pond layer having a size of (1,
8)；Second layer convolution kernel is having a size of (Isosorbide-5-Nitrae), pond layer size (1,7)；Third layer convolution kernel is having a size of (1,5), pond layer size
(1,6)；Convolutional layer and pond layer step-length are disposed as (1,1,1,1), and full articulamentum neuron is 1024, are finally sentenced by score value
It is disconnected to belong to corresponding Activity Type.
The utility model has the advantages that compared to the prior art, the present invention has following marked improvement: 1- norm data processing transformation can incite somebody to action
Three axis of data characteristics is soft each other, and by realizing that high-precision differentiates after convolutional neural networks training；The present invention passes through to sensing
The data of device acquisition carry out the processing of 1- norm and convert the hardware requirement that can be reduced for computer, and energy several times shorten network training
Time；For the present invention when guaranteeing that data do not lose motion characteristic, handling data quickly and efficiently avoids biography
The drawbacks of system data processing；1- norm algorithm is simple and effective in the present invention, and becomes convenient for the relationship between data after observation processing
Change；The present invention can adapt to era development using Android smartphone and smartwatch etc. after convolutional neural networks training
Trend.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is the principle of the present invention figure；
Fig. 3 is the small batch waveform diagram of raw sensor 3-axis acceleration data of the present invention；
Fig. 4 is that small batch variation waveform diagram is corresponded to after 1- norm of the present invention is handled；
Fig. 5 is the corresponding penalty values variation diagram of frequency of training of the present invention；
Fig. 6 is the precision figure after training of the embodiment of the present invention 300 times.
Specific embodiment
Below with reference to the drawings and specific embodiments, technical solution of the present invention and beneficial effect are described in detail.
A kind of limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks, includes the following steps:
Step 1, volunteer is recruited, and wears movable sensor, movable sensor can be put into pocket, record volunteer is not
With the 3-axis acceleration data under limb action, 3-axis acceleration data are acquired by movable sensor, and to every kind of activity class
Type stamps respective labels, and it is txt-formatted file that data, which save form,；
Step 2, it to three number of axle of acquisition according to traversing, while removing wherein sensor and failing correctly to record and occur
Ergodic data is carried out the processing of 1- norm, training set and test set is splitted data into after normalized by null value.
Step 3, shape conversion is carried out to above-mentioned processed data, thus reach correctly entering for convolutional neural networks,
Definition and initialization weight and bias function, realize the update to weight and biasing in calling process, satisfied accurate when reaching
Pb file is generated after degree；
Step 4,3-axis acceleration sensor data are acquired using mobile intelligent terminal, after being pre-processed, is input to instruction
The convolutional neural networks model perfected, obtains human body attitude recognition result.
The present invention is based on the limbs recognition methods of 1- norm data processing transformation and convolutional neural networks, can walk to jump
It downstairs, stands, six kinds of movement postures of sitting down are identified upstairs on road.
Fig. 1 is the flow chart of target processing, and the three-dimensional acceleration time series of human motion is collected from movable sensor
Afterwards, initial convolutional neural networks are input to after Data Integration processing and carry out model training, ideal model after training is applied into shifting
On dynamic sensor, make it the discrimination that can be realized human action on mobile intelligent terminal.
Fig. 2 is convolutional neural networks structure chart, specifically includes that input layer, the long-pending and maximum pond layer of three-layer coil, one connects entirely
Connect layer and an output layer.
In the present embodiment, the sample frequency of movable sensor is preferably set to 32Hz, and under this frequency, Fig. 3 is original biography
The small batch waveform diagram of sensor 3-axis acceleration data, Fig. 4 are that small batch variation waveform diagram is corresponded to after 1- norm is handled；This implementation
Example defines every 4.0 seconds as a sample action, i.e., every 128 groups of data are a sample.Certainly, sample frequency can be according to reality
Demand self-setting is suitably worth, herein without limitation.
For training convolutional neural networks, collected sample is divided into two classes: training sample and test sample by the present invention.
Training sample carries out model training as the input of convolutional neural networks, and test sample considers foundation as recognition accuracy.
In the present embodiment, using the 70% of data set as training set, using the 30% of data set as test set.
Data only can just reach satisfactory accuracy by correct processing, and the present invention pays attention to the 1- norm of data
Processing.Its mathematic(al) representation are as follows:
From the above equation, we can see that mathematical relationship is the sum of each value absolute value, because input data is 3-axis acceleration data, because
This above-mentioned formula can convert are as follows:
By formula (2) it is seen that three number of axle are according to a data volume is converted into, because of each row of data in original txt file
It all include x transverse acceleration, y longitudinal acceleration, three number of axle evidence of z vertical direction, when by being traversed to every row element, and it is right
Three acceleration informations of every row carry out absolute value processing, are finally added, realize the fusion of feature, if conventionally not
The processing of 1- norm is carried out, the port number for inputting convolutional Neural just will increase three times, this considerably increases calculation amount, therefore the present invention
Calculating speed can largely be improved.
It is real by the way that after traversing all data sets, 3-axis acceleration data are converted into a digital representation known to expression formula
Triple channel is showed to single pass conversion.
1- norm data processing algorithm of the invention is also applicable by angular transducer, gyroscope, and the sensors such as GPS are constituted
Multiaxis data, respective formula can convert are as follows:
Wherein, xniMiddle n indicates that shared n kind variable, i.e. n number of axle evidence, i indicate corresponding the i-th row of every axis.From the above equation, we can see that
When number of sensors increases, it can be achieved that the transformation of n to 1, data are more complicated, the advantage of this algorithm can be more highlighted.
After the initial data 1- norm processing of sensor (there is still a need for normalizeds for data after processing), data are passed through
The feature extraction of convolutional neural networks, which finally obtains, accurately to be determined.Wherein the correspondence first layer convolution kernel of convolutional network having a size of
(1,3), pond layer share 128 convolution kernels having a size of (1,8), second layer convolution kernel having a size of (Isosorbide-5-Nitrae), pond layer size (1,
7) 128 convolution kernels, are shared, second layer convolution kernel shares 128 convolution kernels having a size of (1,5), pond layer size (1,6), will
Three layers of pond result arranged side by side are vertical merged, and convolutional layer and pond layer step-length are disposed as (1,1,1,1), full articulamentum neuron
It is 1024, learning rate 0.000001.
If amount of training data is not big enough, need to reuse data.100 data are input to nerve every time
Network is trained, every 100 recognition accuracies of measurement and cross entropy, and wherein the variation diagram of cross entropy is as shown in figure 5, figure
6 be the precision figure after the present embodiment training 300 times.
When the precision of prediction of trained convolutional neural networks meets design requirement, pb file can be generated, and this article
Part is extracted and is used on mobile terminal.If the convolutional neural networks of training are undesirable, need suitably to increase the nerve of each hidden layer
First number.If the method for the neuron number of above-mentioned each hidden layer of modification is little on recognition accuracy influence, appropriate increasing can be added
Add number of training.
It should be noted that the human body attitude identification device in the embodiment of the present invention specifically can integrate at intelligent mobile end
In end, above-mentioned intelligent terminal is specifically as follows the terminals such as smart phone, smartwatch, is not construed as limiting herein.
Therefore the human body attitude identification device in the embodiment of the present invention passes through the acceleration degree of acquisition intelligent terminal
According to being then based on 1- norm algorithm to acquisition data processing, and by pretreated data input trained human body attitude
Identification model obtains human body attitude recognition result.Since human body attitude identification model is based on preset training set convolution mind
It is obtained through network training, therefore, by inputting trained human body attitude identification model after pre-processing acceleration information, i.e.,
The identification to human body attitude can be achieved, to realize the human body attitude identification of the nonvisual means based on acceleration information.
Claims (6)
1. a kind of limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks, which is characterized in that including
Following steps:
Step 1,3-axis acceleration data are acquired by movable sensor, and respective labels is stamped to every kind of Activity Type；
Step 2, the processing of 1- norm is carried out to collected data, and the data of processing is divided into training set and test set；
Step 3, the training that above-mentioned processed data are carried out with convolutional neural networks, obtains most suitable weight and bigoted value, whole
.pb file is generated after reason；
Step 4, file is entered and left into mobile intelligent terminal, obtains accurate human body limb motion detection effect.
2. limbs recognition methods according to claim 1, it is characterised in that: in the step 1, sample frequency is set as
30Hz-40Hz。
3. limbs recognition methods according to claim 1, it is characterised in that: in the step 2, including carried out to data empty
Again the processing of arranged in sequence after value rejecting and rejecting.
4. limbs recognition methods according to claim 1 or 3, it is characterised in that: in the step 2, by the m% in data
As training sample, n% is as test sample, the m%+n%=100%, and 60%≤m%≤80%, 20%≤n%≤
40%.
5. limbs recognition methods as described in claim 1, it is characterised in that: the step 3 specifically includes following content:
3.1, establish 3 layers of convolutional neural networks model；
3.2, it imports training sample and adjusts convolutional neural networks model parameter, obtain the model of high-accuracy.
6. limbs recognition methods according to claim 5, it is characterised in that: in the convolutional neural networks model, first
Layer convolution kernel is having a size of (1,3), and pond layer is having a size of (1,8)；Second layer convolution kernel having a size of (Isosorbide-5-Nitrae), pond layer size (1,
7)；Third layer convolution kernel is having a size of (1,5), pond layer size (1,6)；Convolutional layer and pond layer step-length be disposed as (1,1,1,
1), full articulamentum neuron is 1024, finally belongs to corresponding Activity Type by score value judgement.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811581807.3A CN109740651A (en) | 2018-12-24 | 2018-12-24 | Limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811581807.3A CN109740651A (en) | 2018-12-24 | 2018-12-24 | Limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109740651A true CN109740651A (en) | 2019-05-10 |
Family
ID=66361116
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811581807.3A Pending CN109740651A (en) | 2018-12-24 | 2018-12-24 | Limbs recognition methods based on 1- norm data processing transformation and convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740651A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107145878A (en) * | 2017-06-01 | 2017-09-08 | 重庆邮电大学 | Old man's anomaly detection method based on deep learning |
| CN107958221A (en) * | 2017-12-08 | 2018-04-24 | 北京理工大学 | A kind of human motion Approach for Gait Classification based on convolutional neural networks |
| CN108062170A (en) * | 2017-12-15 | 2018-05-22 | 南京师范大学 | Multi-class human posture recognition method based on convolutional neural networks and intelligent terminal |
| US20180261071A1 (en) * | 2017-03-10 | 2018-09-13 | Turing Video, Inc. | Surveillance method and system based on human behavior recognition |
| CN108694408A (en) * | 2017-04-11 | 2018-10-23 | 西安邮电大学 | A kind of driving behavior recognition methods based on depth sparseness filtering convolutional neural networks |
-
2018
- 2018-12-24 CN CN201811581807.3A patent/CN109740651A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180261071A1 (en) * | 2017-03-10 | 2018-09-13 | Turing Video, Inc. | Surveillance method and system based on human behavior recognition |
| CN108694408A (en) * | 2017-04-11 | 2018-10-23 | 西安邮电大学 | A kind of driving behavior recognition methods based on depth sparseness filtering convolutional neural networks |
| CN107145878A (en) * | 2017-06-01 | 2017-09-08 | 重庆邮电大学 | Old man's anomaly detection method based on deep learning |
| CN107958221A (en) * | 2017-12-08 | 2018-04-24 | 北京理工大学 | A kind of human motion Approach for Gait Classification based on convolutional neural networks |
| CN108062170A (en) * | 2017-12-15 | 2018-05-22 | 南京师范大学 | Multi-class human posture recognition method based on convolutional neural networks and intelligent terminal |
Non-Patent Citations (1)
| Title |
|---|
| 刘渭滨等: "模式分类中的特征融合方法", 《北京邮电大学学报》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Sun et al. | Gesture recognition based on kinect and sEMG signal fusion | |
| Chen et al. | A novel gait pattern recognition method based on LSTM-CNN for lower limb exoskeleton | |
| CN112754468B (en) | Human body lower limb movement detection and identification method based on multi-source signals | |
| CN110610158A (en) | Human body posture identification method and system based on convolution and gated cyclic neural network | |
| CN108062170A (en) | Multi-class human posture recognition method based on convolutional neural networks and intelligent terminal | |
| Lu et al. | A multichannel CNN-GRU model for human activity recognition | |
| CN108334827B (en) | Gait identity authentication method based on intelligent shoe and intelligent shoe | |
| CN108345846A (en) | A kind of Human bodys' response method and identifying system based on convolutional neural networks | |
| CN107423730A (en) | A kind of body gait behavior active detecting identifying system and method folded based on semanteme | |
| CN110236550B (en) | Human gait prediction device based on multi-mode deep learning | |
| CN110232412B (en) | Human gait prediction method based on multi-mode deep learning | |
| CN111723662B (en) | Human body posture recognition method based on convolutional neural network | |
| CN109979161B (en) | Human body falling detection method based on convolution cyclic neural network | |
| CN109726662A (en) | Multi-class human posture recognition method based on convolution sum circulation combination neural net | |
| Luo et al. | Binarized neural network for edge intelligence of sensor-based human activity recognition | |
| Su et al. | HDL: Hierarchical deep learning model based human activity recognition using smartphone sensors | |
| CN109685148A (en) | Multi-class human motion recognition method and identifying system | |
| CN110659677A (en) | Human body falling detection method based on movable sensor combination equipment | |
| CN111178288B (en) | Human body posture recognition method and device based on local error layer-by-layer training | |
| CN110977961A (en) | Motion information acquisition system of self-adaptive power-assisted exoskeleton robot | |
| CN109976526A (en) | A kind of sign Language Recognition Method based on surface myoelectric sensor and nine axle sensors | |
| CN106295675A (en) | A kind of Activity recognition method of based on smart mobile phone of high accuracy | |
| CN110532898A (en) | A kind of physical activity recognition methods based on smart phone Multi-sensor Fusion | |
| Yang et al. | Inertial sensing for lateral walking gait detection and application in lateral resistance exoskeleton | |
| Chen et al. | An extended spatial transformer convolutional neural network for gesture recognition and self-calibration based on sparse sEMG electrodes |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190510 |
|
| RJ01 | Rejection of invention patent application after publication |