CN109662830A - A kind of language blind guiding stick, the deep neural network optimization method based on the walking stick - Google Patents
A kind of language blind guiding stick, the deep neural network optimization method based on the walking stick Download PDFInfo
- Publication number
- CN109662830A CN109662830A CN201910048854.XA CN201910048854A CN109662830A CN 109662830 A CN109662830 A CN 109662830A CN 201910048854 A CN201910048854 A CN 201910048854A CN 109662830 A CN109662830 A CN 109662830A
- Authority
- CN
- China
- Prior art keywords
- language
- neural network
- module
- blind guiding
- deep neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 54
- 241001166076 Diapheromera femorata Species 0.000 title claims abstract description 31
- 238000000034 method Methods 0.000 title claims abstract description 27
- 238000005457 optimization Methods 0.000 title claims abstract description 20
- 230000006870 function Effects 0.000 claims abstract description 24
- 230000003993 interaction Effects 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 18
- 238000012549 training Methods 0.000 claims description 16
- 238000004891 communication Methods 0.000 claims description 14
- 238000013139 quantization Methods 0.000 claims description 10
- 238000009966 trimming Methods 0.000 claims description 7
- 230000006835 compression Effects 0.000 claims description 6
- 238000007906 compression Methods 0.000 claims description 6
- 230000005484 gravity Effects 0.000 claims description 6
- 230000002452 interceptive effect Effects 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 230000004888 barrier function Effects 0.000 claims description 4
- 238000001514 detection method Methods 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 4
- 238000005259 measurement Methods 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 claims description 2
- 210000005036 nerve Anatomy 0.000 claims 1
- 238000005265 energy consumption Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000004438 eyesight Effects 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000192 social effect Effects 0.000 description 1
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61F—FILTERS IMPLANTABLE INTO BLOOD VESSELS; PROSTHESES; DEVICES PROVIDING PATENCY TO, OR PREVENTING COLLAPSING OF, TUBULAR STRUCTURES OF THE BODY, e.g. STENTS; ORTHOPAEDIC, NURSING OR CONTRACEPTIVE DEVICES; FOMENTATION; TREATMENT OR PROTECTION OF EYES OR EARS; BANDAGES, DRESSINGS OR ABSORBENT PADS; FIRST-AID KITS
- A61F9/00—Methods or devices for treatment of the eyes; Devices for putting-in contact lenses; Devices to correct squinting; Apparatus to guide the blind; Protective devices for the eyes, carried on the body or in the hand
- A61F9/08—Devices or methods enabling eye-patients to replace direct visual perception by another kind of perception
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61H—PHYSICAL THERAPY APPARATUS, e.g. DEVICES FOR LOCATING OR STIMULATING REFLEX POINTS IN THE BODY; ARTIFICIAL RESPIRATION; MASSAGE; BATHING DEVICES FOR SPECIAL THERAPEUTIC OR HYGIENIC PURPOSES OR SPECIFIC PARTS OF THE BODY
- A61H3/00—Appliances for aiding patients or disabled persons to walk about
- A61H3/06—Walking aids for blind persons
- A61H3/061—Walking aids for blind persons with electronic detecting or guiding means
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61H—PHYSICAL THERAPY APPARATUS, e.g. DEVICES FOR LOCATING OR STIMULATING REFLEX POINTS IN THE BODY; ARTIFICIAL RESPIRATION; MASSAGE; BATHING DEVICES FOR SPECIAL THERAPEUTIC OR HYGIENIC PURPOSES OR SPECIFIC PARTS OF THE BODY
- A61H3/00—Appliances for aiding patients or disabled persons to walk about
- A61H3/06—Walking aids for blind persons
- A61H3/068—Sticks for blind persons
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61H—PHYSICAL THERAPY APPARATUS, e.g. DEVICES FOR LOCATING OR STIMULATING REFLEX POINTS IN THE BODY; ARTIFICIAL RESPIRATION; MASSAGE; BATHING DEVICES FOR SPECIAL THERAPEUTIC OR HYGIENIC PURPOSES OR SPECIFIC PARTS OF THE BODY
- A61H2201/00—Characteristics of apparatus not provided for in the preceding codes
- A61H2201/50—Control means thereof
- A61H2201/5007—Control means thereof computer controlled
- A61H2201/501—Control means thereof computer controlled connected to external computer devices or networks
- A61H2201/5012—Control means thereof computer controlled connected to external computer devices or networks using the internet
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Veterinary Medicine (AREA)
- Public Health (AREA)
- Animal Behavior & Ethology (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Epidemiology (AREA)
- Pain & Pain Management (AREA)
- Physical Education & Sports Medicine (AREA)
- Rehabilitation Therapy (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Vascular Medicine (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Ophthalmology & Optometry (AREA)
- Probability & Statistics with Applications (AREA)
- Heart & Thoracic Surgery (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
A kind of lightweight deep neural network optimization method the invention discloses language blind guiding stick, based on the walking stick, belong to intelligent blind-guiding technical field, including wand body, the wheel that is set to the handle on wand body top and is set to wand body bottom end, the blind guiding stick further include: be embedded in the microprocessor ARM of wand body, deep neural network module, wide-angle camera, array radar, locating module, motor and the power module connected with the microprocessor ARM, the motor are connect with wheel；The present invention provides a kind of language blind guiding stick based on neural network and array radar, the walking stick combines the image recognition of deep neural network and the obstacle identification function of radar array, it is capable of providing multiple avoidance planning, the walking stick utilizes the language identification function of deep neural network simultaneously, demand to blind person user proposes carries out Recognition feedback, can easily realize the interaction between walking stick and user, easy to use, it is convenient to operate, securely and reliably.
Description
Technical field
The invention belongs to intelligent blind-guiding technical fields, and in particular to a kind of language blind guiding stick, the light weight based on the walking stick
Grade deep neural network optimization method.
Background technique
Currently, China is one of the country that blind person is most in the world.Since China human mortality is numerous, people's lives environment is more multiple
It is miscellaneous.In recent years, with the raising of education degree, changed as blind person's thought of disability crowd, be more intended to safeguard
Self-respect selects independent life style.But the blind person for losing eyesight be faced in daily life can not identify it is basic
Daily necessity, can not identify traffic lights, the public domains such as blind way and can not avoidance the problems such as.Guide equipment currently on the market
It is leading that be all with blind person itself be, by groping to find out surrounding enviroment for blind person itself, since surrounding enviroment are increasingly sophisticated simultaneously
Effective feedback cannot be provided to blind person.Therefore establishing a set of can capture basic living articles image and can be carried out image analysis
Processing, roadblock evade, the blind guiding system of interactive voice, motion feedback, for solves the problems, such as blind person's basic living daily life have weigh
The social effect and market prospects wanted.
Summary of the invention
In view of the deficiencies of the prior art, the first purpose of this invention is to provide a kind of language blind guiding stick, to blind person
The demand that user proposes carries out Recognition feedback, can easily realize the interaction between walking stick and user, easy to use, operation
It is convenient, securely and reliably.
Second object of the present invention is to provide the lightweight deep neural network based on the language blind guiding stick excellent
Change method reduces demand of the deep neural network to computing resource and operation energy consumption by this method, improves deep neural network
Computational efficiency, promote deployment of the deep neural network on embedded device.
In order to achieve the above object, the present invention the following technical schemes are provided:
This language blind guiding stick provided by the invention including wand body, is set to the handle on wand body top and is set to
The wheel of wand body bottom end, the blind guiding stick further include: be embedded in what the microprocessor ARM and the microprocessor ARM of wand body were connected
Deep neural network module, wide-angle camera, array radar, locating module, motor and power module, the motor and wheel connect
It connects；
The wide-angle camera is used to acquire the information of road surface of walking stick front end, and the image data of acquisition is sent to micro- place
Manage device ARM；Array radar is used for the precision ranging of barrier；Locating module is for being accurately positioned；Motor is used for driving wheel；Electricity
Source module is used to provide electric energy to entire blind guiding stick；Deep neural network module is used for the picture number acquired to wide-angle camera
According to being analyzed, the scenes such as the traffic lights, crossing and stair of traffic intersection are identified, obstacle recognition is carried out according to image；
Microprocessor ARM is used to coordinate and control the relevant operation of walking stick, according to the position of the distance measurement result of array radar, locating module
Information carries out accurate avoidance planning, then driving motor and wheel, realizes leading the way automatically for blind guiding stick, blind person only need to be with
The direction of walking stick traction achieve that the automatic guidance in path.
Preferably, the language blind guiding stick further includes that nine axle sensors connected with the microprocessor ARM, network are logical
Believe module, nine axle sensors are used for fall detection, and network communication module is used for the data communication of blind guiding stick；It is falling
When, microprocessor ARM generates warning message in conjunction with the locating module, is sent to designated contact by network communication module.
Preferably, the network communication module is 4G module.
Preferably, the locating module is GPS positioning module or Beidou positioning module.
Preferably, the language blind guiding stick further includes the voice module connected with the microprocessor ARM, voice module
For carrying out interactive voice with blind person, and acquire the voice data of blind person, receive user for blind guiding stick operational order simultaneously
It is uploaded to microprocessor ARM.
Voice module can identify the control voice command of blind person, control the relevant operation of walking stick, can also identify blind person
Voice inquirement order relevant information, such as synoptic climate information etc. are searched for by 4G module；On the other hand, walking stick works
Relevant information feed back to user also by voice module, such as turning left when navigation is turned right, traffic road when going across the road
Lamp state etc..
Preferably, the language blind guiding stick further includes bluetooth module, and bluetooth module is connect with voice module, and blind person passes through
The bluetooth module of the bluetooth headset and blind guiding stick of wearing band carries out interactive voice.
The inventive concept total as one, the present invention also provides a kind of, and the lightweight deep neural network based on the walking stick is excellent
Change method, comprising the following steps:
Step 1: the structure and network parameter of initialization deep neural network；
Step 2: carrying out structure trimming and training to deep neural network using different Loss functions；
Step 3: realizing that network quantization and weight are shared in the form of product quantization to deep neural network；
Step 4: lossless compression is carried out using weight of the entropy coding to deep neural network, the depth after obtaining final optimization pass
Spend neural network parameter；
Step 5: the deep neural network parameter after optimization is downloaded in the deep neural network module of walking stick.
In a specific embodiment, in the step 2, specifically:
2.1) trimming threshold value, degree of rarefication, similarity and the accuracy parameter of setting network；
2.2) using the training network of Loss function 1, the network parameter and node for being lower than threshold value in training process are then deleted,
Until reaching the degree of rarefication of setting, wherein Loss function 1 is defined as follows:
W in formula (1)jThe weight matrix of jth layer is represented,Indicating prediction output, y indicates reality output,It indicates
Predict error, | | Wj||pIndicate weight matrix WjP norm, the sparsity of Controlling model, α are come as penalty termjFor adjusting
The specific gravity of sparsity penalty term；
2.3) using the training network of Loss function 2, adjustment network parameter is until reach the similarity of setting, wherein Loss
Function 2 is defined as follows:
W in formula (2)jRepresent the weight matrix of jth layer, rjRepresent WjLine number, cjRepresent WjColumns,Indicate that prediction is defeated
Out, y indicates reality output,Indicate prediction error, | Wj(a,b)-Wj(a ', b ') | indicate weight matrix WjIn each parameter
Between difference, the diversity of Controlling model is carried out as penalty term, the diversity of model parameter is lower, and the effect of subsequent cluster is got over
It is good.βjFor adjusting the specific gravity of diversity penalty term；
2.4) using the training network of Loss function 3, adjusting parameter is until reach the accuracy of setting, wherein Loss function 3
It is defined as follows:
In formula (3)Indicating prediction output, y indicates reality output,Indicate prediction error.
In a specific embodiment, in the step 3, specifically:
3.1) cluster centre is initialized using Chaos dynamical equation logistic sequence, logistic sequence definition is such as
Under:
C (t+1)=4c (t) (1-c (t)) 1≤t≤k (4)
C (t) is t-th of cluster centre value in formula (4), and k is cluster centre number；
3.2) by the weight matrix W of jth layerjS matrix is resolved by column:
3.3) eachIt is middle to be clustered using k-means method:
In formula (6)It representsZ row,N-th of cluster centre is represented, k is cluster centre number, rjFor WjRow
Number；
3.4) after clustering, network weight is quantified as cluster centre index and cluster centre value.
Compared with the existing technology, the present invention has following advantageous effects:
The present invention provides a kind of language blind guiding stick based on neural network and array radar, which combines depth mind
The obstacle identification function of image recognition and radar array through network is capable of providing multiple avoidance planning, while walking stick benefit
With the language identification function of deep neural network, the demand to blind person user proposes carries out Recognition feedback, can be easily real
Interaction between existing walking stick and user, easy to use, it is convenient to operate, securely and reliably.
The present invention is based on the lightweight deep neural network optimization method of language blind guiding stick, the present invention is directed to large-scale depth
Resource and limited energy problem of the learning network when disposing on embedded device, put forward a kind of neural network side simplified
Case, this method first optimize the structure of neural network in terms of degree of rarefication, similarity and accuracy three, are guaranteeing standard
Under the premise of exactness, using degree of rarefication as target, reduce network weight and node, while again using similarity as target, reduces power
The otherness of value, for weight optimization and shared establish good basis.Then, the form that the present invention is quantified with product is to weight
Quantified and shared, be further reduced the difference of weight and indicate expense, is laid a good foundation for third step lossless compression；
Finally, the present invention carries out entropy coding to the weight after quantization, the data volume of weight is further compressed, can be set embedded
Fitting depth learning model is disposed in standby lesser storage resource, simultaneously because network structure is simplified, when the network operation
Energy consumption is also further reduced.
Detailed description of the invention
Fig. 1 provides a kind of structural schematic diagram of language blind guiding stick for the present invention.
Fig. 2 is the functional block diagram of language blind guiding stick of the present invention.
Fig. 3 is that the present invention is based on the flow charts of the lightweight deep neural network optimization method of the walking stick.
Fig. 4 is the flow chart that structure is trimmed and trained.
Fig. 5 is the flow chart of network quantization and weight sharing method.
Specific embodiment
The technical scheme in the embodiments of the invention will be clearly and completely described below, it is clear that described implementation
Example is only a part of the embodiment of the present invention, rather than whole embodiments, based on the embodiments of the present invention, ordinary skill
Personnel's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
Invention is further explained with attached drawing combined with specific embodiments below:
As shown in Fig. 1~2, this language blind guiding stick provided by the invention, including wand body 101, it is set to wand body top
Handle 102 and be set to the wheel 103 of wand body bottom end, the blind guiding stick further include: be embedded in the microprocessor ARM of wand body
2 connected with microprocessor ARM deep neural network module 3, wide-angle camera 4, array radar 5, locating module 6, motor 7
With power module 8, motor 7 is connect with wheel 103；In view of being convenient for motor driven wheel, the installation site of motor is close to wheel；
Wide-angle camera 4 is used to acquire the information of road surface of walking stick front end, and the image data of acquisition is sent to micro process
Device ARM；Array radar 5 is used for the precision ranging of barrier；Locating module 6 is for being accurately positioned；Motor 7 is used for driving wheel；
Power module 8 is used to provide electric energy to entire blind guiding stick；Deep neural network module 3 is used for the figure acquired to wide-angle camera
As data are analyzed, the scenes such as the traffic lights, crossing and stair of traffic intersection are identified, barrier knowledge is carried out according to image
Not；Microprocessor ARM 2 is used to coordinate and control the relevant operation of walking stick, and according to the distance measurement result of array radar 5, positioning mould
The location information of block 6 carries out accurate avoidance planning, then driving motor 7 and wheel 103, realizes drawing automatically for blind guiding stick
Road, the direction that blind person need to only follow walking stick to draw achieve that the automatic guidance in path.
In a particular embodiment, the model Exynos 4412 of microprocessor ARM.
In order to carry out fall detection to walking stick and blind person, language blind guiding stick further includes nine connected with microprocessor ARM
Axle sensor 9, network communication module 10, nine axle sensors 9 are used for fall detection, and network communication module 10 is for blind guiding stick
Data communication；When falling, microprocessor ARM combination locating module 6 generates warning message, passes through network communication module 10
It is sent to designated contact.
In a particular embodiment, network communication module is 4G module.
In a particular embodiment, locating module is GPS positioning module.
In order to carry out interactive voice with blind person, language blind guiding stick further includes the voice module connected with microprocessor ARM
11, voice module is used to acquire the voice data of blind person, receives user for the operational order of blind guiding stick and is uploaded to micro- place
Manage device ARM.
Voice module can identify the control voice command of blind person, control the relevant operation of walking stick, can also identify blind person
Voice inquirement order relevant information, such as synoptic climate information etc. are searched for by 4G module；On the other hand, walking stick works
Relevant information feed back to user also by voice module, such as turning left when navigation is turned right, traffic road when going across the road
Lamp state etc..
In a particular embodiment, voice module plays chip (type by voice recognition chip (model: LD3320) and voice
Number: YX6200-16S) chip forms, and voice module 11 is installed on the top of wand body 101, can hear that voice broadcast is believed convenient for blind person
Breath, while can identify the phonetic order of blind person.
Under more noisy environment, prompt information is not heard for the ease of blind person, blind guiding stick further includes bluetooth module 12,
Bluetooth module is connect with voice module, and blind person carries out voice friendship by the bluetooth module of the bluetooth headset and blind guiding stick of wearing band
Mutually, it avoids because being put the case where causing user to answer not in time outside prompt information.
As shown in figure 3, the present invention also provides a kind of lightweight deep neural network optimization method based on the walking stick, including
Following steps:
Step 1: the structure and network parameter of initialization deep neural network；
Step 2: carrying out structure trimming and training to deep neural network using different Loss functions；
Step 3: realizing that network quantization and weight are shared in the form of product quantization to deep neural network；
Step 4: lossless compression is carried out using weight of the entropy coding to deep neural network, the depth after obtaining final optimization pass
Spend neural network parameter；
Step 5: the deep neural network parameter after optimization is downloaded in the deep neural network module of walking stick.
In embodiments of the present invention, step 1 specifically: the structure of neural network is initialized (i.e. according to the complexity of task
How many a hidden layers) and network parameter (i.e. the weight of hidden layer).
In embodiments of the present invention, in step 2, structure trimming and training process as shown in figure 4, specifically:
2.1) trimming threshold value, degree of rarefication, similarity and the accuracy parameter of setting network；
2.2) using the training network of Loss function 1, the network parameter and node for being lower than threshold value in training process are then deleted,
Until reaching the degree of rarefication of setting, wherein Loss function 1 is defined as follows:
W in formula (1)jThe weight matrix of jth layer is represented,Indicating prediction output, y indicates reality output,It indicates
Predict error, | | Wj||pIndicate weight matrix WjP norm, the sparsity of Controlling model, α are come as penalty termjFor adjusting
The specific gravity of sparsity penalty term；
2.3) using the training network of Loss function 2, adjustment network parameter is until reach the similarity of setting, wherein Loss
Function 2 is defined as follows:
W in formula (2)jRepresent the weight matrix of jth layer, rjRepresent WjLine number, cjRepresent WjColumns,Indicate that prediction is defeated
Out, y indicates reality output,Indicate prediction error, | Wj(a,b)-Wj(a ', b ') | indicate weight matrix WjIn each parameter
Between difference, the diversity of Controlling model is carried out as penalty term, the diversity of model parameter is lower, and the effect of subsequent cluster is got over
It is good.βjFor adjusting the specific gravity of diversity penalty term；
2.4) using the training network of Loss function 3, adjusting parameter is until reach the accuracy of setting, wherein LossFunction 3
It is defined as follows:
In formula (3)Indicating prediction output, y indicates reality output,Indicate prediction error.
In embodiments of the present invention, in step 3, network quantization and weight sharing method as shown in figure 5, specifically:
3.1) cluster centre is initialized using Chaos dynamical equation logistic sequence, logistic sequence definition is such as
Under:
C (t+1)=4c (t) (1-c (t)) 1≤t≤k (4)
C (t) is t-th of cluster centre value in formula (4), and k is cluster centre number；
3.2) by the weight matrix W of jth layerjS matrix is resolved by column:
3.3) eachIt is middle to be clustered using k-means method:
In formula (6)It representsZ row,N-th of cluster centre is represented, k is cluster centre number, rjFor WjRow
Number；
3.4) after clustering, network weight is quantified as cluster centre index and cluster centre value.
In embodiments of the present invention, step 4 specifically: the weight of neural network is subjected to lossless compression using entropy coding,
Deep neural network parameter after obtaining final optimization pass.
In embodiments of the present invention, step 5 specifically: the neural network parameter after optimization is downloaded to the depth of walking stick
In neural network module, to realize depth analysis and judgement on walking stick.
It should be understood that the part that this specification does not elaborate belongs to the prior art.
The method of the present invention is directed to resource and limited energy of the large-scale deep learning network when disposing on embedded device and asks
Topic, puts forward a kind of neural network scheme simplified, this method is first to mind in terms of degree of rarefication, similarity and accuracy three
Structure through network optimizes, and under the premise of guaranteeing accuracy, using degree of rarefication as target, reduces network weight and section
Point, while again using similarity as target, reduces the otherness of weight, for weight optimization and shared establish good basis.So
Afterwards, the form that the present invention is quantified with product quantifies and shares to weight, is further reduced the difference of weight and indicates expense,
It lays a good foundation for third step lossless compression；Finally, the present invention carries out entropy coding to the weight after quantization, further compress
The data volume of weight can dispose fitting depth learning model, simultaneously on embedded device in lesser storage resource
Due to simplifying for network structure, the energy consumption when network operation is also further reduced.
Claims (9)
1. a kind of language blind guiding stick, including wand body, the wheel that is set to the handle on wand body top and is set to wand body bottom end,
It is characterized in that, the walking stick further include: be embedded in the depth nerve that the microprocessor ARM of wand body is connected with the microprocessor ARM
Network module, wide-angle camera, array radar, locating module, motor and power module, the motor are connect with wheel；
The wide-angle camera is used to acquire the information of road surface of walking stick front end, and the image data of acquisition is sent to microprocessor
ARM；Array radar is used for the precision ranging of barrier；Locating module is for being accurately positioned；Motor is used for driving wheel；Power supply mould
Block is used to provide electric energy to entire blind guiding stick；The image data that deep neural network module is used to acquire wide-angle camera into
Row analysis, identifies the scenes such as the traffic lights, crossing and stair of traffic intersection, carries out obstacle recognition according to image；Micro- place
Reason device ARM is used to coordinate and control the relevant operation of walking stick, is believed according to the position of the distance measurement result of array radar, locating module
Breath, carries out accurate avoidance planning, then driving motor and wheel, realizes leading the way automatically for blind guiding stick.
2. language blind guiding stick according to claim 1, which is characterized in that the language blind guiding stick further include with it is described
Nine axle sensors, the network communication module of microprocessor ARM connection, nine axle sensors are used for fall detection, network communication module
Data communication for blind guiding stick；When falling, microprocessor ARM generates warning message in conjunction with the locating module,
Designated contact is sent to by network communication module.
3. language blind guiding stick according to claim 2, which is characterized in that the network communication module is 4G module.
4. language blind guiding stick according to claim 1 or 2, which is characterized in that the locating module is GPS positioning module
Or Beidou positioning module.
5. language blind guiding stick according to claim 1 or 2, which is characterized in that the language blind guiding stick further include and
The voice module of the microprocessor ARM connection, voice module is used for and blind person carries out interactive voice, and acquires the voice of blind person
Data receive user for the operational order of blind guiding stick and are uploaded to microprocessor ARM.
6. language blind guiding stick according to claim 5, which is characterized in that the language blind guiding stick further includes bluetooth mould
Block, bluetooth module are connect with voice module, and blind person carries out language by the bluetooth module of the bluetooth headset and blind guiding stick of wearing band
Sound interaction.
7. a kind of lightweight deep neural network optimization side based on language blind guiding stick described in any one of claim 1~6
Method, which comprises the following steps:
Step 1: the structure and network parameter of initialization deep neural network；
Step 2: carrying out structure trimming and training to deep neural network using different Loss functions；
Step 3: realizing that network quantization and weight are shared in the form of product quantization to deep neural network；
Step 4: lossless compression is carried out using weight of the entropy coding to deep neural network, the depth mind after obtaining final optimization pass
Through network parameter；
Step 5: the deep neural network parameter after optimization to be downloaded to the deep neural network module of the language blind guiding stick
In.
8. the lightweight deep neural network optimization method based on language blind guiding stick, feature exist according to claim 7
In, in the step 2, specifically:
2.1) trimming threshold value, degree of rarefication, similarity and the accuracy parameter of setting network；
2.2) using the training network of Loss function 1, the network parameter and node for being lower than threshold value in training process are then deleted, until
Reach the degree of rarefication of setting, wherein Loss function 1 is defined as follows:
W in formula (1)jThe weight matrix of jth layer is represented,Indicating prediction output, y indicates reality output,Indicate that prediction misses
Difference, | | Wj||pIndicate weight matrix WjP norm, the sparsity of Controlling model, α are come as penalty termjFor adjusting sparsity
The specific gravity of penalty term；
2.3) using the training network of Loss function 2, adjustment network parameter is until reach the similarity of setting, wherein Loss function 2
It is defined as follows:
W in formula (2)jRepresent the weight matrix of jth layer, rjRepresent WjLine number, cjRepresent WjColumns,Indicate prediction output, y
Indicate reality output,Indicate prediction error, | Wj(a,b)-Wj(a ', b ') | indicate weight matrix WjIn between each parameter
Difference carrys out the diversity of Controlling model as penalty term, and the diversity of model parameter is lower, and the effect of subsequent cluster is better,
βjFor adjusting the specific gravity of diversity penalty term；
2.4) using the training network of Loss function 3, adjusting parameter is until reach the accuracy of setting, wherein Loss function 3 defines
It is as follows:
In formula (3)Indicating prediction output, y indicates reality output,Indicate prediction error.
9. the lightweight deep neural network optimization method based on language blind guiding stick, feature exist according to claim 7
In, in the step 3, specifically:
3.1) cluster centre is initialized using Chaos dynamical equation logistic sequence, logistic sequence definition is as follows:
C (t+1)=4c (t) (1-c (t)) 1≤t≤k (4)
C (t) is t-th of cluster centre value in formula (4), and k is cluster centre number；
3.2) by the weight matrix W of jth layerjS matrix is resolved by column:
3.3) eachIt is middle to be clustered using k-means method:
In formula (6)It representsZ row,N-th of cluster centre is represented, k is cluster centre number, rjFor WjLine number；
3.4) after clustering, network weight is quantified as cluster centre index and cluster centre value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910048854.XA CN109662830B (en) | 2019-01-18 | 2019-01-18 | A kind of language blind guiding stick, the deep neural network optimization method based on the walking stick |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910048854.XA CN109662830B (en) | 2019-01-18 | 2019-01-18 | A kind of language blind guiding stick, the deep neural network optimization method based on the walking stick |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109662830A true CN109662830A (en) | 2019-04-23 |
| CN109662830B CN109662830B (en) | 2019-11-29 |
Family
ID=66151067
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910048854.XA Expired - Fee Related CN109662830B (en) | 2019-01-18 | 2019-01-18 | A kind of language blind guiding stick, the deep neural network optimization method based on the walking stick |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109662830B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111109786A (en) * | 2019-12-25 | 2020-05-08 | 李嘉伦 | Intelligent obstacle early warning crutch based on deep learning and early warning method thereof |
| CN113032535A (en) * | 2019-12-24 | 2021-06-25 | 中国移动通信集团浙江有限公司 | Visual question and answer method and device for assisting visually impaired people, computing equipment and storage medium |
| CN114191267A (en) * | 2021-12-06 | 2022-03-18 | 南通大学 | Light-weight intelligent method and system for assisting blind person in going out in complex environment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN202920579U (en) * | 2012-10-10 | 2013-05-08 | 成都掌握移动信息技术有限公司 | Intelligent seeing-eye vehicle based on infrared sensing technology |
| CN204319215U (en) * | 2014-12-04 | 2015-05-13 | 天津朗誉科技发展有限公司 | A kind of bland person walking stick |
| CN106389078A (en) * | 2016-11-24 | 2017-02-15 | 贵州大学 | Intelligent blind guiding glass system and blind guiding method thereof |
| CN107693314A (en) * | 2017-08-02 | 2018-02-16 | 华中科技大学 | A kind of cane formula, which is independently fallen, protects recovery walking aiding robot |
-
2019
- 2019-01-18 CN CN201910048854.XA patent/CN109662830B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN202920579U (en) * | 2012-10-10 | 2013-05-08 | 成都掌握移动信息技术有限公司 | Intelligent seeing-eye vehicle based on infrared sensing technology |
| CN204319215U (en) * | 2014-12-04 | 2015-05-13 | 天津朗誉科技发展有限公司 | A kind of bland person walking stick |
| CN106389078A (en) * | 2016-11-24 | 2017-02-15 | 贵州大学 | Intelligent blind guiding glass system and blind guiding method thereof |
| CN107693314A (en) * | 2017-08-02 | 2018-02-16 | 华中科技大学 | A kind of cane formula, which is independently fallen, protects recovery walking aiding robot |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113032535A (en) * | 2019-12-24 | 2021-06-25 | 中国移动通信集团浙江有限公司 | Visual question and answer method and device for assisting visually impaired people, computing equipment and storage medium |
| CN111109786A (en) * | 2019-12-25 | 2020-05-08 | 李嘉伦 | Intelligent obstacle early warning crutch based on deep learning and early warning method thereof |
| CN114191267A (en) * | 2021-12-06 | 2022-03-18 | 南通大学 | Light-weight intelligent method and system for assisting blind person in going out in complex environment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109662830B (en) | 2019-11-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109662830B (en) | A kind of language blind guiding stick, the deep neural network optimization method based on the walking stick | |
| CN103126862B (en) | Outdoor blind guiding robot based on global position system (GPS), general packet radio service (GPRS) and radio frequency identification devices (RFID) and navigational positioning method | |
| CN104772748B (en) | A kind of social robot | |
| US7801332B2 (en) | Controlling a system based on user behavioral signals detected from a 3D captured image stream | |
| CN111643324A (en) | Intelligent glasses for blind people | |
| CN110368275A (en) | A kind of blind-guidance robot and blind guiding system, blind-guiding method | |
| CN105563493A (en) | Height and direction adaptive service robot and adaptive method | |
| CN110728308A (en) | Interactive blind guiding system and method based on improved Yolov2 target detection and voice recognition | |
| CN110623820A (en) | Blind device is led to wearable intelligence | |
| CN109000655B (en) | Bionic indoor positioning and navigation method for robot | |
| Schwarze et al. | A camera-based mobility aid for visually impaired people | |
| CN109998873A (en) | A kind of wearable blindmen intelligent positioning and blind guiding system | |
| US11371859B1 (en) | Computer-implemented method, wearable device, computer program and computer readable medium for assisting the movement of a visually impaired user | |
| CN113520812B (en) | Four-foot robot blind guiding system and method | |
| CN105816303A (en) | GPS and visual navigation-based blind guiding system and method thereof | |
| CN107242964A (en) | Blind guiding system and method for work based on deep learning | |
| CN202134012U (en) | Meteorological popular science robot system | |
| Wang et al. | A survey of 17 indoor travel assistance systems for blind and visually impaired people | |
| CN205521430U (en) | High position adaptive service robot | |
| CN118285995A (en) | Visual auxiliary system based on cloud analysis | |
| CN113840588A (en) | Touch sensing system and use method thereof | |
| Hossain et al. | State of the art review on walking support system for visually impaired people | |
| Hersh et al. | Mobility: an overview | |
| CN110682297A (en) | Intelligent interaction system and method for indoor guiding robot | |
| CN209500102U (en) | A kind of intelligent wheelchair of interior independent navigation Internet of Things |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20191129 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |