CN106919949A - A kind of real-time vehicle matching process based on convolutional neural networks - Google Patents
A kind of real-time vehicle matching process based on convolutional neural networks Download PDFInfo
- Publication number
- CN106919949A CN106919949A CN201710050763.0A CN201710050763A CN106919949A CN 106919949 A CN106919949 A CN 106919949A CN 201710050763 A CN201710050763 A CN 201710050763A CN 106919949 A CN106919949 A CN 106919949A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- convolutional neural
- neural networks
- information
- information flow
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of real-time vehicle matching process based on convolutional neural networks, including model data storehouse is built, and design convolutional neural networks;Convolutional neural networks training is carried out using model data storehouse, the vehicle feature of every kind of vehicle in optimal convolutional neural networks and model data storehouse is obtained;The topological structure of Storm is built, information flow upper strata is data source input module, and information flow middle level is the data handling component for arranging optimal convolutional neural networks, and information laminar sublayer is the data handling component for arranging SVM classifier;The live video stream of collection is issued information flow middle level data handling component by data source input module, and vehicle feature is extracted by convolutional neural networks;Information flow bottom data processing assembly is matched using SVM classifier to the vehicle feature that information flow middle level sends, and returns to matching result.With svm classifier method and Storm frameworks be combined the Feature Extraction Technology of convolutional neural networks by the present invention, improves the accuracy rate and efficiency of vehicle matching.
Description
Technical field
The present invention relates to vehicle automatic Matching, and in particular to a kind of real-time vehicle matching based on convolutional neural networks
Method.
Background technology
Vehicle automatic Matching is the important component of intelligent transportation system, by carrying out automatic of vehicle to vehicle
Match somebody with somebody, data can be provided for traffic administration, charge, scheduling, statistics, be one of study hotspot and difficult point of intelligent transportation field;
At present, the vehicle vehicle Auto-matching accuracy rate of China is also difficult to meet use requirement, improves algorithm to vehicle matching rate and grinds
Study carefully imperative.
Deep learning is the climax of current machine learning development, and convolutional neural networks are used as one kind side in deep learning
Method, has preferable effect in fields such as object identification, image procossings, and convolutional neural networks have can automatically learn image spy
Levy, reduce manual intervention, the characteristics of image quality of extraction advantage high, it can be considered to be applied in vehicle Auto-matching
In technology, to improve the accuracy rate of vehicle matching.But, convolutional neural networks can consume substantial amounts of GPU (Graphics
Processing Unit, graphic process unit) resource, may not reached when amount of calculation is excessive in practical application scene in real time
Effect.
The content of the invention
The technical problems to be solved by the invention are how convolutional neural networks to be applied in vehicle automatic Matching,
And reach the problem that vehicle matches real-time effect.
In order to solve the above-mentioned technical problem, the technical solution adopted in the present invention is to provide a kind of based on convolutional neural networks
Real-time vehicle matching process, comprise the following steps:
Step S10, structure model data storehouse, and it is designed for the convolutional neural networks of vehicle cab recognition;
Step S20, convolutional neural networks are trained using model data storehouse, obtain optimal convolutional neural networks with
And in model data storehouse every kind of vehicle vehicle feature;
Step S30, the topological structure for building Storm, its information flow upper strata are data source input module spout, information flow
Middle level is the data handling component bolt for arranging optimal convolutional neural networks, and information laminar sublayer is to arrange SVM classifier
Data handling component bolt;
The vehicle live video stream to be matched for gathering is issued information flow middle level by step S40, data source input module spout
Data handling component bolt, data handling component bolt extract vehicle feature by convolutional neural networks;
The vehicle that step S50, information flow bottom data processing assembly bolt are sent using SVM classifier to information flow middle level
Feature is matched, and returns to matching result.
In the above-mentioned methods, information of the model data storehouse comprising various, the information of every kind of vehicle has plurality of pictures.
In the above-mentioned methods, step S20 specifically includes following steps:
Using the information of vehicle not of the same race in model data storehouse as the input data source of convolutional neural networks, convolution god is used
Convolutional neural networks training is carried out through network default parameter;
According to training intermediate result, default parameters initial weight, training speed, iterations are constantly adjusted, directly
To obtaining optimal convolutional neural networks network parameter;
The information of vehicle not of the same race in traversal model data storehouse, every kind of car is extracted using the optimal convolutional neural networks of parameter
The vehicle feature of type, and preserve.
In the above-mentioned methods, the convolutional neural networks for vehicle cab recognition include four layers of convolution, respectively three layers pond layer,
Three layers of full articulamentum, a laminated and layer and softmax return grader layer.
In the above-mentioned methods, information flow middle level and information laminar sublayer arrange optimal convolutional Neural including two or more respectively
The data handling component bolt of the network and data handling component bolt for arranging SVM classifier;
Information flow upper layer data source input module according to principle of selecting the best qualified, by live video stream be transmitted to information flow middle level GUP compared with
Strong data handling component bolt;The vehicle feature of extraction is sent to information laminar sublayer by information flow middle level according to principle of selecting the best qualified
GUP stronger data handling component bolt.
In the above-mentioned methods, matching is carried out to the vehicle feature that information flow middle level sends and specifically includes following steps:
What the data handling component bolt receive information streams middle level different pieces of information processing assembly bolt of information laminar sublayer sent
Vehicle feature, and to receive each vehicle feature be combined, judges combine after feature whether be same vehicle, if
It is that the feature after combination is carried out into vehicle with the vehicle feature in model data storehouse matches, returns to matching result;If it is not,
Each vehicle feature and the vehicle feature in model data storehouse are carried out into vehicle respectively to match, matching result is returned.
In the above-mentioned methods, in step s 40,
Each data source input module spout is decoded as multiple frames by the live video stream from streaming media server is gathered
Image, then each two field picture is issued the data handling component bolt in information flow middle level;
Information flow middle level data handling component bolt extracts each two field picture using the convolutional neural networks for training thereon
In vehicle feature, and be combined.
The present invention will be based on Feature Extraction Technology, svm classifier method and the Storm real-time processing platforms of convolutional neural networks
It is combined, realizes that the real-time vehicle based on convolutional neural networks is matched, has the advantages that:
(1) feature extraction is carried out by convolutional neural networks, is effectively improved the accuracy rate and efficiency of feature extraction;
(2) characteristic matching is carried out using SVM, effectively raises the accuracy rate of vehicle matching;
(3) using the matching of Storm parallelizations vehicle, the purpose of real-time matching is reached, effectively raises the efficiency of matching.
Brief description of the drawings
A kind of flow chart of real-time vehicle matching process based on convolutional neural networks that Fig. 1 is provided for the present invention;
Fig. 2 is the network structure of convolutional neural networks in the present invention;
Fig. 3 is the topology diagram of distributed real-time processing framework Storm in the present invention.
Specific embodiment
The present invention makes full use of convolutional neural networks to learn the advantage and cloud computing technology in real time at present of characteristics of image
Feature, convolutional neural networks are combined with real-time cloud computing technology, and the vehicle feature that convolutional neural networks are applied to vehicle is carried
Take, reduce manual intervention, extract high-quality characteristics of image, improve the accuracy rate of vehicle matching;And utilize real-time cloud computing
In carry out the distributed real-time processing framework Storm of big data real-time processing, solve convolutional neural networks and consume substantial amounts of GPU moneys
Source, the problem of live effect in practical application scene may not be reached when amount of calculation is excessive, improve the efficiency of vehicle matching, this
It is a kind of new trial for being different from traditional vehicle matching system in terms of the vehicle automatic Matching of intelligent transportation.
The present invention is described in detail with reference to Figure of description and specific embodiment.
As shown in figure 1, a kind of real-time vehicle matching process based on convolutional neural networks that the present invention is provided, including it is following
Step:
Step S10, the video flowing historical data structure model data storehouse from streaming media server using storage, and design
For the convolutional neural networks of vehicle cab recognition, in the present invention, model data storehouse includes the information of various, every kind of vehicle
Information has plurality of pictures.
Step S20, by the use of the vehicle information in model data storehouse as the input of convolutional neural networks, by convolutional Neural
The convolutional neural networks off-line training module of network carries out the training of convolutional neural networks, obtain optimal convolutional neural networks and
Extract the vehicle feature of the every kind of vehicle in model data storehouse.
In the present invention, step S20 specifically includes following steps:
Step S21, the information input convolutional neural networks by the vehicle not of the same race in model data storehouse, use convolutional Neural
Network default parameter carries out convolutional neural networks training;
Step S22, according to training intermediate result, default parameters initial weight, training speed, iterations are carried out constantly
Adjustment, until obtaining optimal convolutional neural networks network parameter, now convolutional neural networks can be obtained most with highest efficiency
Good recognition effect, the recognition accuracy of the convolutional neural networks after training reaches 95%;
The information of vehicle not of the same race, is carried using the optimal convolutional neural networks of parameter in step S23, traversal model data storehouse
The vehicle feature of every kind of vehicle is taken, and is preserved.
Step S30, the topological structure for building distributed real-time processing framework Storm, as shown in figure 3, its information flow upper strata
It is data source input module spout;Information flow middle level is the data handling component bolt for arranging optimal convolutional neural networks, is used
In the convolutional neural networks characteristic extracting module of the vehicle feature for extracting vehicle to be matched;Information laminar sublayer (is propped up to arrange SVM
Hold vector machine) the data handling component bolt of grader, the characteristic matching module for carrying out vehicle matching.
Step S40, data source input module spout will be collected in the vehicle live video stream to be matched of streaming media server
Information flow middle level data handling component bolt is issued, it is special that data handling component bolt extracts vehicle by convolutional neural networks
Levy.
In the present invention, information flow middle level and information laminar sublayer arrange optimal convolutional Neural net including two or more respectively
The data handling component bolt of the network and data handling component bolt for arranging SVM classifier;
Because convolutional neural networks can consume substantial amounts of GPU resource, the scheduling of resource based on GPU is devised in Storm
Algorithm, can consume GPU excessive to enter by the GPU behaviours in service of each data handling component of monitor in real time Storm bolt
Journey is assigned on the data handling component bolt of stronger GPU, and solving convolutional neural networks can consume asking for substantial amounts of GPU resource
Live video stream is transmitted to information flow middle level GUP stronger by topic, i.e. information flow upper layer data source input module according to principle of selecting the best qualified
Data handling component bolt;Information flow middle level according to principle of selecting the best qualified, by the vehicle feature of extraction be sent to information laminar sublayer GUP compared with
Strong data handling component bolt.So as to avoid not reached because amount of calculation is excessive asking for live effect in practical application scene
Topic, realizes the live effect of vehicle matching.
In the present invention, each data source input module spout will gather the live video stream from streaming media server
Multiple two field pictures are decoded as, then each two field picture is issued the data handling component bolt in information flow middle level;
The data handling component bolt in information flow middle level extracts each frame figure using the convolutional neural networks for training thereon
Vehicle feature as in, and be combined.
The vehicle that step S50, information flow bottom data processing assembly bolt are sent using SVM classifier to information flow middle level
Feature is matched, and returns to matching result.
In step s 50, matching is carried out to the vehicle feature that information flow middle level sends and specifically includes following steps:
What the data handling component bolt receive information streams middle level different pieces of information processing assembly bolt of information laminar sublayer sent
Vehicle feature, and to receive each vehicle feature be combined, judges combine after feature whether be same vehicle, if
It is that the feature after combination is carried out into vehicle with the vehicle feature in model data storehouse matches, returns to matching result;If it is not,
Each vehicle feature and the vehicle feature in model data storehouse are carried out into vehicle respectively to match, matching result is returned.
In the present invention, model data storehouse is built using historical data, uses the convolutional Neural as shown in Figure 2 for designing
Network (including four layers of convolution, respectively three layers pond layer, three layers of full articulamentum, a laminated and layer and last layer are softmax
Return grader layer), training convolutional neural networks.
Real-time vehicle matching process based on convolutional neural networks of the invention, the feature based on convolutional neural networks is carried
Take technology to be combined with svm classifier method and Storm real-time processing platforms, improve the accuracy rate and efficiency of vehicle matching;Filling
Divide using on the premise of computing resource, find optimal network, vehicle is matched.
Obviously, those skilled in the art can carry out various changes and modification without deviating from essence of the invention to the present invention
God and scope.So, if these modifications of the invention and modification belong to the scope of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to comprising these changes and modification.
Claims (7)
1. a kind of real-time vehicle matching process based on convolutional neural networks, it is characterised in that comprise the following steps:
Step S10, structure model data storehouse, and it is designed for the convolutional neural networks of vehicle cab recognition;
Step S20, convolutional neural networks are trained using model data storehouse, obtain optimal convolutional neural networks and car
The vehicle feature of every kind of vehicle in type database;
Step S30, the topological structure for building Storm, its information flow upper strata are data source input module spout, information flow middle level
To arrange the data handling component bolt of optimal convolutional neural networks, information laminar sublayer is the data for arranging SVM classifier
Processing assembly bolt;
The vehicle live video stream to be matched for gathering is issued layer data in information flow by step S40, data source input module spout
Processing assembly bolt, data handling component bolt extract vehicle feature by convolutional neural networks;
The vehicle feature that step S50, information flow bottom data processing assembly bolt are sent using SVM classifier to information flow middle level
Matched, and returned to matching result.
2. the method for claim 1, it is characterised in that information of the model data storehouse comprising various, every kind of vehicle
Information have plurality of pictures.
3. method as claimed in claim 2, it is characterised in that step S20 specifically includes following steps:
Using the information of vehicle not of the same race in model data storehouse as the input data source of convolutional neural networks, convolutional Neural net is used
Network default parameters carries out convolutional neural networks training;
According to training intermediate result, default parameters initial weight, training speed, iterations are constantly adjusted, until
To optimal convolutional neural networks network parameter;
The information of vehicle not of the same race in traversal model data storehouse, every kind of vehicle is extracted using the optimal convolutional neural networks of parameter
Vehicle feature, and preserve.
4. the method for claim 1, it is characterised in that the convolutional neural networks for vehicle cab recognition include four layers of volume
Product, respectively three layers pond layer, three layers of full articulamentum, a laminated and layer and softmax return grader layer.
5. the method for claim 1, it is characterised in that information flow middle level and information laminar sublayer include two or more respectively
The data handling component bolt for the arranging optimal convolutional neural networks and data handling component bolt for arranging SVM classifier;
Live video stream is transmitted to information flow middle level GUP stronger by information flow upper layer data source input module according to principle of selecting the best qualified
Data handling component bolt;Information flow middle level according to principle of selecting the best qualified, by the vehicle feature of extraction be sent to information laminar sublayer GUP compared with
Strong data handling component bolt.
6. method as claimed in claim 5, it is characterised in that to the vehicle feature that information flow middle level sends match specific
Comprise the following steps:
The vehicle that the data handling component bolt receive information streams middle level different pieces of information processing assembly bolt of information laminar sublayer sends
Feature, and to receive each vehicle feature be combined, judges combination after feature whether be same vehicle, if it is, general
Feature after combination carries out vehicle and matches with the vehicle feature in model data storehouse, returns to matching result;If it is not, by each
Vehicle feature carries out vehicle and matches respectively with the vehicle feature in model data storehouse, returns to matching result.
7. method as claimed in claim 5, it is characterised in that in step s 40,
Each data source input module spout is decoded as multiple frame figures by the live video stream from streaming media server is gathered
Picture, then each two field picture is issued the data handling component bolt in information flow middle level;
Information flow middle level data handling component bolt is extracted in each two field picture using the convolutional neural networks for training thereon
Vehicle feature, and be combined.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710050763.0A CN106919949A (en) | 2017-01-23 | 2017-01-23 | A kind of real-time vehicle matching process based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710050763.0A CN106919949A (en) | 2017-01-23 | 2017-01-23 | A kind of real-time vehicle matching process based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106919949A true CN106919949A (en) | 2017-07-04 |
Family
ID=59454085
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710050763.0A Pending CN106919949A (en) | 2017-01-23 | 2017-01-23 | A kind of real-time vehicle matching process based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106919949A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110032131A (en) * | 2018-01-12 | 2019-07-19 | 中科院微电子研究所昆山分所 | Electric vehicle state monitoring processing system and monitoring system based on Storm |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104657748A (en) * | 2015-02-06 | 2015-05-27 | 中国石油大学(华东) | Vehicle type recognition method based on convolutional neural network |
CN104992147A (en) * | 2015-06-09 | 2015-10-21 | 中国石油大学(华东) | License plate identification method of deep learning based on fast and slow combination cloud calculation environment |
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN105678214A (en) * | 2015-12-21 | 2016-06-15 | 中国石油大学(华东) | Vehicle flow statistical method based on convolutional neural network vehicle model recognition in cloud environment |
CN106203330A (en) * | 2016-07-08 | 2016-12-07 | 西安理工大学 | A kind of vehicle classification method based on convolutional neural networks |
-
2017
- 2017-01-23 CN CN201710050763.0A patent/CN106919949A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN104657748A (en) * | 2015-02-06 | 2015-05-27 | 中国石油大学(华东) | Vehicle type recognition method based on convolutional neural network |
CN104992147A (en) * | 2015-06-09 | 2015-10-21 | 中国石油大学(华东) | License plate identification method of deep learning based on fast and slow combination cloud calculation environment |
CN105678214A (en) * | 2015-12-21 | 2016-06-15 | 中国石油大学(华东) | Vehicle flow statistical method based on convolutional neural network vehicle model recognition in cloud environment |
CN106203330A (en) * | 2016-07-08 | 2016-12-07 | 西安理工大学 | A kind of vehicle classification method based on convolutional neural networks |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110032131A (en) * | 2018-01-12 | 2019-07-19 | 中科院微电子研究所昆山分所 | Electric vehicle state monitoring processing system and monitoring system based on Storm |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106845415B (en) | Pedestrian fine identification method and device based on deep learning | |
CN106874883A (en) | A kind of real-time face detection method and system based on deep learning | |
CN112653899B (en) | Network live broadcast video feature extraction method based on joint attention ResNeSt under complex scene | |
CN106372648A (en) | Multi-feature-fusion-convolutional-neural-network-based plankton image classification method | |
CN104217073B (en) | A kind of visual layout's method of network community gravitation guiding | |
CN109829443A (en) | Video behavior recognition methods based on image enhancement Yu 3D convolutional neural networks | |
CN108830252A (en) | A kind of convolutional neural networks human motion recognition method of amalgamation of global space-time characteristic | |
CN110390246A (en) | A kind of video analysis method in side cloud environment | |
CN106791927A (en) | A kind of video source modeling and transmission method based on deep learning | |
CN107426542A (en) | The hardware structure of Real-Time Arltificial Intelligence video system based on edge calculations | |
CN104063686B (en) | Crop leaf diseases image interactive diagnostic system and method | |
CN108090403A (en) | Face dynamic identification method and system based on 3D convolutional neural network | |
CN110430395A (en) | Video data AI processing system and processing method | |
CN108921942B (en) | Method and device for 2D (two-dimensional) conversion of image into 3D (three-dimensional) | |
CN104092988A (en) | Method, device and system for managing passenger flow in public place | |
CN106650795A (en) | Sorting method of hotel room type images | |
CN110245550A (en) | A kind of face noise data collection CNN training method based on overall cosine distribution | |
CN105574896B (en) | A kind of efficient background modeling method towards high-resolution video | |
CN104156729B (en) | A kind of classroom demographic method | |
CN109712127A (en) | A kind of electric transmission line fault detection method for patrolling video flowing for machine | |
CN107730453A (en) | Picture quality method for improving | |
CN112749663B (en) | Agricultural fruit maturity detection system based on Internet of things and CCNN model | |
CN109523558A (en) | A kind of portrait dividing method and system | |
CN110264407A (en) | Image Super-resolution model training and method for reconstructing, device, equipment and storage medium | |
CN116665092A (en) | Method and system for identifying sewage suspended matters based on IA-YOLOV7 |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170704 |