CN108921013A - A kind of visual scene identifying system and method based on deep neural network - Google Patents
A kind of visual scene identifying system and method based on deep neural network Download PDFInfo
- Publication number
- CN108921013A CN108921013A CN201810472012.2A CN201810472012A CN108921013A CN 108921013 A CN108921013 A CN 108921013A CN 201810472012 A CN201810472012 A CN 201810472012A CN 108921013 A CN108921013 A CN 108921013A
- Authority
- CN
- China
- Prior art keywords
- scene
- training
- weather
- network
- sample
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Traffic Control Systems (AREA)
- Image Processing (AREA)
Abstract
A kind of visual scene identifying system based on deep neural network, including:Vehicle-mounted vision system, for acquiring vehicle forward-looking vision image;Off-line training module is labeled for carrying out sample collection to from the vehicle forward-looking vision image that vehicle-mounted vision system acquires to vision input using depth convolutional neural networks, generates sample label, and carry out neural network parameter substep training;The depth convolutional neural networks are made of the three branch taxonomy networks for sharing two layers of shallow-layer convolution feature, pass through sample and training mission training network parameter;And on-line analysis module, real-time scene analysis is carried out to the sample after off-line training module training using Web compression and pipelined-flash analysis strategy, when exporting the day of road scene locating for vehicle-mounted vision system, weather and scene abnormality.
Description
Technical field
The present invention relates to field of vehicle safety, in particular to a kind of visual scene identifying system based on deep neural network
And method.
Background technique
Nowadays intelligence becomes the important development direction of automobile industry, the cognition technology of view-based access control model sensor increasingly at
It is ripe, and the application in vehicle active safety field is more extensive.For the visual system, for different road environments, from
Image Acquisition all can do corresponding adjustment, thus accurately identification vision input from parameter and tactful level to application layer algorithm
In scene information there is very strong application value and meaning, also, the abnormity diagnosis of visual scene can further enhance and be
The robustness and fault-tolerance of system.
Existing vision system does not have perfect such rudimentary algorithm mostly, when judgement is mostly based on system when existing day
Clock, weather judgement then do not have High relevancy based on the aiding sensors such as rainfall, such method and vision system input in itself,
It is thus poor in some border condition effects.The abnormity diagnosis of the visual scene piecemeal brightness statistics that then view-based access control model inputs, threshold
Value is difficult to set, and robustness is poor.
Summary of the invention
When judgement is mostly based on system when it is an object of the present invention to solve above-mentioned existing day of the existing technology
Clock, weather judgement then do not have High relevancy based on the aiding sensors such as rainfall, such method and vision system input in itself,
The thus problem poor in some border condition effects, provides a kind of visual scene identifying system based on deep neural network
And method.
The technical solution adopted by the present invention to solve the technical problems is:A kind of visual scene based on deep neural network
Identifying system, including:Vehicle-mounted vision system, for acquiring vehicle forward-looking vision image;Off-line training module, for utilizing depth
Convolutional neural networks carry out sample collection to from the vehicle forward-looking vision image that vehicle-mounted vision system acquires to vision input, into
Rower note, generates sample label, and carries out neural network parameter substep training;The depth convolutional neural networks are by sharing two layers
Three branch taxonomy networks of shallow-layer convolution feature are constituted, and pass through sample and training mission training network parameter;And on-line analysis
Module carries out real-time scene to the sample after off-line training module training using Web compression and pipelined-flash analysis strategy
Analysis, when exporting the day of road scene locating for vehicle-mounted vision system, weather and scene abnormality.
The present invention utilize multitask depth convolutional neural networks, when to day locating for visual scene, weather and scene it is different
Normal situation identified, input it is stronger with vision system correlation itself, thus for the bottom of vision system acquire and on
The required more acurrate and robustness of configuration parameter setting of layer application is stronger, so as to further promote the sense of vision system itself
Know ability and robustness, on the one hand output can provide related prior information for Image Acquisition and using algorithm, such as exposure ginseng
On the other hand number, the selection of detection classifier and image procossing threshold value etc. can provide image input exception for vision system
Common diagnosis, such as image is fuzzy, camera lens blocks.
Further, the depth convolutional neural networks by share two layers of shallow-layer convolution feature three branch taxonomy network structures
At, including:It when identify network:It when identification network inputs be shared shallow-layer feature, export as daytime, dusk and night three
Classify when kind day;Weather identifies network:Weather identifies that network inputs are shared shallow-layer feature, export for fine day, the cloudy, rainy day,
The five kinds of weather classification of snowy day and greasy weather;And scene anomalous identification network:Scene anomalous identification network inputs are shared shallow-layer
Feature exports as scene is normal, scene is blocked, Scene Blur, scene are excessively dark and five kinds of scene overexposure abnormal scene classifications.
Further, the off-line training module includes sample collection and mark unit, neural network parameter substep training
Unit;The sample collection and mark unit extract discrete time series training sample for acquiring vehicle forward-looking vision image offline
This carries out sample expansion using spatial alternation;Balanced each task sample distribution of all categories, is labeled, and generates sample label;
The task includes:It when identification, weather identification and identification;The classification includes:It when classification, weather classification and scene
Abnormal class;The sample label of classification when for day:0- daytime, 1- dusk, 2- night;For the sample label of weather classification:
0- fine day, 1- is cloudy, the 2- rainy day, 3- snowy day, the 4- greasy weather;And the sample label for scene abnormal class:0- is normal, 1-
It blocks, 2- is fuzzy, and 3- is excessively bright, and 4- is excessively dark;
The neural network parameter substep training unit, for classification task training, using cross entropy as loss function,
Sharing feature layer parameter is trained first, and each task is identical to weight update contribution coefficient, i.e.,:
Loss=1/3*L_time+1/3*L_weather+1/3*L_abnormal;Then solidify sharing feature layer convolution
Parameter, each task update each branching networks weight coefficient by respective loss function.
Further, the on-line analysis module, including Web compression unit and pipelined-flash analytical unit;The network
Compression unit is evaluated and tested for the resulting neural network parameter of offline network training to be carried out quantization and rarefaction using test set
The output accuracy loss of network after compression, and determine whether to model after quantization according to data precision re -training after quantization;Institute
Pipelined-flash analytical unit is stated, for preferentially executing to the every frame of scene abnormality detection, weather and weather detection are using every frame or jump
Frame executive mode;In front end applications, by the way of dual-thread or multi-threaded parallel, a thread executes scene anomalous identification,
Another thread timesharing alternately day when with weather identify, convolutional neural networks hardware acceleration unit request priority is with scene
Anomalous identification Network Priority.
The visual scene recognition methods based on deep neural network that the present invention also provides a kind of, passes through vehicle-mounted vision system
Vehicle forward-looking vision image is acquired, offline network training is carried out to visual pattern using depth convolutional neural networks, then carry out
Line scene analysis, when exporting the day of road scene locating for vehicle-mounted vision system, weather and scene abnormality;The depth volume
Product neural network is made of the three branch taxonomy networks for sharing two layers of shallow-layer convolution feature, passes through sample and training mission training
Network parameter.
Further, the depth convolutional neural networks by share two layers of shallow-layer convolution feature three branch taxonomy network structures
At, including:It when identify network:It when identification network inputs be shared shallow-layer feature, export as daytime, dusk and night three
Classify when kind day;Weather identifies network:Weather identifies that network inputs are shared shallow-layer feature, export for fine day, the cloudy, rainy day,
The five kinds of weather classification of snowy day and greasy weather;And scene anomalous identification network:Scene anomalous identification network inputs are shared shallow-layer
Feature exports as scene is normal, scene is blocked, Scene Blur, scene are excessively dark and five kinds of scene overexposure abnormal scene classifications.
Further, the offline network training includes:Sample collection and mark, neural network parameter substep training;Institute
Sample collection is stated to refer to mark:Offline acquisition vehicle forward-looking vision image, is extracted discrete time series training sample, is become using space
It changes, carries out sample expansion;Balanced each task sample distribution of all categories, is labeled, and generates sample label;The task includes:
It when identification, weather identification and identification;The classification includes:It when classification, weather classification and scene abnormal class;For
It when classification sample label:0- daytime, 1- dusk, 2- night;For the sample label of weather classification:0- fine day, 1- is cloudy,
2- rainy day, 3- snowy day, 4- greasy weather;And the sample label for scene abnormal class:0- is normal, and 1- is blocked, and 2- is fuzzy, 3-
Cross bright, 4- is excessively dark;The neural network parameter substep training refers to:Using classification task training, loss function uses cross entropy,
Sharing feature layer parameter is trained first, and each task is identical to weight update contribution coefficient, i.e.,:Loss=1/3*L_time+1/3*
L_weather+1/3*L_abnormal;
Then solidify sharing feature layer deconvolution parameter, each task updates each branching networks weight system by respective loss function
Number.
Further, the online scene analysis, including Web compression and pipelined-flash analysis strategy;The Web compression
Refer to:The resulting neural network parameter of offline network training is subjected to quantization and rarefaction, wherein quantized data bit length and dilute
Thinization degree is configuration parameter;It is lost, and is determined whether to after quantization using network output accuracy after test set evaluation and test compression
Model is according to data precision re -training after quantization;The pipelined-flash analysis strategy refers to:The every frame of scene abnormality detection is preferential
It executes, weather and weather detection are using every frame or frame-skipping executive mode;In front end applications, using dual-thread or multi-threaded parallel
Mode, thread execute scene anomalous identification, another thread timesharing alternately day when identified with weather, convolutional Neural net
Network hardware acceleration unit request priority is with scene anomalous identification Network Priority.
Substantial effect of the invention:The present invention carries out scene information knowledge to input picture using depth convolutional neural networks
, can not identify environmental information locating for current vehicle effectively, including when day, weather etc., this type of information can pass through optimization phase
Parameter configuration is closed effectively to be promoted the collection image quality of vision system and apply algorithm operational efficiency and effect.Meanwhile this
It invents proposed method and the camera lens in vision input scene can also be blocked, obscured and brightness identifies extremely, energy
Enough trouble diagnosibilities and fault-tolerance for effectively promoting vision system.And network structure of the invention has ductility, can
By adding other classification task branches to enrich scene classification result.
Detailed description of the invention
Fig. 1 is a kind of system structure total figure of the invention;
Fig. 2 is a kind of multitask scene Recognition deep neural network architecture diagram of the invention.
Specific embodiment
Below by specific embodiment, and in conjunction with attached drawing, technical scheme of the present invention will be further explained in detail.
The present invention is based on the inputs of in-vehicle camera vision, sentence to driving environment locating for vehicle and vision system scene state
It is disconnected, configurations information is provided for related algorithm, system input is the input of forward sight in-vehicle camera, when exporting as weather conditions, day
Situation and camera block situation etc., a kind of visual scene identifying system based on deep neural network are provided, such as Fig. 1 institute
Show, including:Vehicle-mounted vision system, for acquiring vehicle forward-looking vision image;Off-line training module is used for from vehicle-mounted vision system
Sample collection is carried out in the vehicle forward-looking vision image of acquisition, and is labeled, generates sample label, and carry out neural network ginseng
Number substep training, obtains depth convolutional neural networks;On-line analysis module, based on the input of vehicle-mounted vision system, using offline
The resulting depth convolutional neural networks of training module training, carry out real-time scene analysis;And output module, for exporting vehicle
When carrying the day of road scene locating for vision system, weather and scene abnormality.
Off-line training module includes sample collection and mark unit, neural network parameter substep training unit.Sample collection
Discrete time series training sample is extracted for acquiring vehicle forward-looking vision image offline with mark unit, using spatial alternation, into
Row sample expands;Balanced each task sample distribution of all categories, is labeled, and generates sample label;The task includes:It when know
Not, weather identification and identification;The classification includes:It when classification, weather classification and scene abnormal class;Class when for day
Other sample label:0- daytime, 1- dusk, 2- night;For the sample label of weather classification:0- fine day, 1- is cloudy, 2- rain
It, 3- snowy day, the 4- greasy weather;And the sample label for scene abnormal class:0- is normal, and 1- is blocked, and 2- is fuzzy, and 3- is excessively bright,
4- is excessively dark;The neural network parameter substep training unit, for classification task training, using cross entropy as loss function,
Sharing feature layer parameter is trained first, and each task is identical to weight update contribution coefficient, i.e.,:
Loss=1/3*L_time+1/3*L_weather+1/3*L_abnormal;Then solidify sharing feature layer convolution
Parameter, each task update each branching networks weight coefficient by respective loss function.
On-line analysis module, including Web compression unit and pipelined-flash analytical unit;The Web compression unit, is used for
The resulting neural network parameter of offline network training is subjected to quantization and rarefaction, network is defeated after compressing using test set evaluation and test
Loss of significance out, and determine whether to model after quantization according to data precision re -training after quantization;The pipelined-flash analysis
Unit, for preferentially executing to the every frame of scene abnormality detection, weather and weather detection are using every frame or frame-skipping executive mode;Front end
In, by the way of dual-thread or multi-threaded parallel, a thread executes scene anomalous identification, another thread timesharing is handed over
It is identified for when carrying out day with weather, convolutional neural networks hardware acceleration unit request priority is excellent with scene anomalous identification network
First.
A kind of visual scene recognition methods based on deep neural network acquires vehicle forward sight view by vehicle-mounted vision system
Feel image, offline network training is carried out to the vehicle forward-looking vision image of acquisition, obtains depth convolutional neural networks;Based on vehicle-mounted
The input of vision system carries out online scene analysis using the resulting depth convolutional neural networks of off-line training module training, defeated
Out when the day of road scene locating for vehicle-mounted vision system, weather and scene abnormality.
1, multitask deep neural network framework:Depth convolutional neural networks of the present invention are by sharing two layers of shallow-layer
Three branch taxonomy networks of convolution feature are constituted, and the network architecture is as shown in Fig. 2, each branching networks structure is multiplexed, by not same
This and training mission training heterogeneous networks parameter, specific branching networks structure are as follows:
Network is identified at 1.1 days:It when identification network inputs be shared shallow-layer feature, export as daytime, dusk and night
When (include tunnel) three kinds of days and similar classification when day.
1.2 weathers identify network:Weather identifies that network inputs are shared shallow-layer feature, export for fine day, cloudy (cloudy day),
Five kinds of rainy day, snowy day and greasy weather weather classification.
1.3 scene anomalous identification networks:Scene anomalous identification network inputs are shared shallow-layer feature, are exported as scene just
Often, scene is blocked, Scene Blur, scene are excessively dark and five kinds of scene overexposure abnormal scene classifications.
The training of 2 offline networks:Forward sight Driving Scene data sample is acquired, and makees corresponding mark, divides three classes identification mission point
Each branching networks parameter is not trained:
2.1 sample collections and mark:Vehicle-mounted forward sight Driving Scene data are acquired offline, extract discrete time series training sample
100000, balanced each task sample distribution of all categories, artificial mark generates sample label.Label substance includes:It when classification
(0- daytime, 1- dusk, 2- night), weather classification (0- fine day, 1- cloudy (cloudy day), 2- rainy day, 3- snowy day, 4- greasy weather) and
Scene abnormal class (0- is normal, and 1- is blocked, and 2- is fuzzy, and 3- is excessively bright, and 4- is excessively dark).Using image gamut, several how spatial alternations,
It carries out sample expansion (if collecting sample further expands, this step can be omitted).
The substep training of 2.2 network parameters:Since training is classification task, loss function uses cross entropy.It trains first
Sharing feature layer parameter, each task are identical to weight update contribution coefficient, i.e.,:
Loss=1/3*L_time+1/3*L_weather+1/3*L_abnormal
Then solidify sharing feature layer deconvolution parameter, each task updates each branching networks weight system by respective loss function
Number.
3, online scene analysis:It is inputted based on vehicle-mounted vision, the resulting depth convolutional neural networks of training in utilization 2, into
The analysis of row real-time scene, including Web compression and pipelined-flash analysis strategy.
3.1 Web compression:The resulting neural network parameter of training in 2 is quantified into (8/16) and rarefaction (20%-
50%), wherein quantized data bit length and rarefaction degree are configurable parameter, defeated using network after test set evaluation and test compression
Loss of significance out, and determine whether to quantization rear mold type according to data precision re -training after quantization.
3.2 pipelined-flash analysis strategies:Scene abnormality detection priority is higher, and every frame is needed preferentially to execute, weather and day
It waits that detection relative result renewal frequency is lower, the mode every frame or frame-skipping execution can be used, in front end applications, dual-thread can be used
(or multithreading) parallel form, thread execute scene anomalous identification, when another thread timesharing is alternately done day and weather
Identification, convolutional neural networks hardware acceleration unit request priority is with scene anomalous identification Network Priority.Embodiment described above
A kind of only preferable scheme of the invention, is not intended to limit the present invention in any form, without departing from claim institute
There are also other variants and remodeling under the premise of the technical solution of record.
Claims (8)
1. a kind of visual scene identifying system based on deep neural network, which is characterized in that including:
Vehicle-mounted vision system, for acquiring vehicle forward-looking vision image;
Off-line training module, for utilizing depth convolutional neural networks to the vehicle forward-looking vision figure acquired from vehicle-mounted vision system
Sample collection is carried out as in, is labeled, sample label is generated, and carries out neural network parameter substep training;The depth volume
Product neural network is made of the three branch taxonomy networks for sharing two layers of shallow-layer convolution feature, passes through sample and training mission training net
Network parameter;And
On-line analysis module, using Web compression and pipelined-flash analysis strategy to the sample after off-line training module training
Real-time scene analysis is carried out, when exporting the day of road scene locating for vehicle-mounted vision system, weather and scene abnormality.
2. a kind of visual scene identifying system based on deep neural network according to claim 1, which is characterized in that institute
Depth convolutional neural networks are stated to be made of the three branch taxonomy networks for sharing two layers of shallow-layer convolution feature, including:
It when identify network:It when identification network inputs be shared shallow-layer feature, export as three kinds of daytime, dusk and night days
When classify;
Weather identifies network:Weather identifies that network inputs are shared shallow-layer feature, export for fine day, the cloudy, rainy day, snowy day and
Five kinds of weather classification of greasy weather;And
Scene anomalous identification network:Scene anomalous identification network inputs are shared shallow-layer feature, are exported as scene is normal, scene hides
Gear, Scene Blur, scene be excessively dark and five kinds of scene overexposure abnormal scene classifications.
3. a kind of visual scene identifying system based on deep neural network according to claim 1, which is characterized in that institute
Stating off-line training module includes sample collection and mark unit, neural network parameter substep training unit;
The sample collection and mark unit acquire vehicle forward-looking vision image for offline, extract discrete time series training sample,
Using spatial alternation, sample expansion is carried out;Balanced each task sample distribution of all categories, is labeled, and generates sample label;
The task includes:It when identification, weather identification and identification;
The classification includes:It when classification, weather classification and scene abnormal class;
The sample label of classification when for day:0- daytime, 1- dusk, 2- night;
For the sample label of weather classification:0- fine day, 1- is cloudy, the 2- rainy day, 3- snowy day, the 4- greasy weather;And
For the sample label of scene abnormal class:0- is normal, and 1- is blocked, and 2- is fuzzy, and 3- is excessively bright, and 4- is excessively dark;
The neural network parameter substep training unit, for classification task training, using cross entropy as loss function, first
Training sharing feature layer parameter, each task are identical to weight update contribution coefficient, i.e.,:
Loss=1/3*L_time+1/3*L_weather+1/3*L_abnormal;
Then solidify sharing feature layer deconvolution parameter, each task updates each branching networks weight coefficient by respective loss function.
4. a kind of visual scene identifying system based on deep neural network according to claim 1, which is characterized in that institute
State on-line analysis module, including Web compression unit and pipelined-flash analytical unit;
The Web compression unit, for the resulting neural network parameter of offline network training to be carried out quantization and rarefaction, benefit
It is lost, and is determined whether to model after quantization according to data precision after quantization with the output accuracy of network after test set evaluation and test compression
Re -training;
The pipelined-flash analytical unit, for preferentially being executed to the every frame of scene abnormality detection, weather and weather detection using every
Frame or frame-skipping executive mode;In front end applications, by the way of dual-thread or multi-threaded parallel, it is abnormal that a thread executes scene
Identification, another thread timesharing alternately day when with weather identify, convolutional neural networks hardware acceleration unit request priority
With scene anomalous identification Network Priority.
5. a kind of visual scene recognition methods based on deep neural network, which is characterized in that acquired by vehicle-mounted vision system
Vehicle forward-looking vision image carries out offline network training to visual pattern using depth convolutional neural networks, then carries out in the field of line
Scape analysis, when exporting the day of road scene locating for vehicle-mounted vision system, weather and scene abnormality;
The depth convolutional neural networks by share two layers of shallow-layer convolution feature three branch taxonomy networks constitute, by sample with
And training mission training network parameter.
6. a kind of visual scene recognition methods based on deep neural network according to claim 5, which is characterized in that institute
Depth convolutional neural networks are stated to be made of the three branch taxonomy networks for sharing two layers of shallow-layer convolution feature, including:
It when identify network:It when identification network inputs be shared shallow-layer feature, export as three kinds of daytime, dusk and night days
When classify;
Weather identifies network:Weather identifies that network inputs are shared shallow-layer feature, export for fine day, the cloudy, rainy day, snowy day and
Five kinds of weather classification of greasy weather;And
Scene anomalous identification network:Scene anomalous identification network inputs are shared shallow-layer feature, are exported as scene is normal, scene hides
Gear, Scene Blur, scene be excessively dark and five kinds of scene overexposure abnormal scene classifications.
7. a kind of visual scene recognition methods based on deep neural network according to claim 5, which is characterized in that institute
Stating offline network training includes:Sample collection and the substep training of mark and neural network parameter;
The sample collection refers to mark:Offline acquisition vehicle forward-looking vision image, extracts discrete time series training sample, utilizes
Spatial alternation carries out sample expansion;Balanced each task sample distribution of all categories, is labeled, and generates sample label;
The task includes:It when identification, weather identification and identification;The classification includes:It when classification, weather classification and
Scene abnormal class;
The sample label of classification when for day:0- daytime, 1- dusk, 2- night;
For the sample label of weather classification:0- fine day, 1- is cloudy, the 2- rainy day, 3- snowy day, the 4- greasy weather;And
For the sample label of scene abnormal class:0- is normal, and 1- is blocked, and 2- is fuzzy, and 3- is excessively bright, and 4- is excessively dark;The nerve institute
Network parameter substep training is stated to refer to:Using classification task training, loss function uses cross entropy, trains sharing feature layer first
Parameter, each task are identical to weight update contribution coefficient, i.e.,:
Loss=1/3*L_time+1/3*L_weather+1/3*L_abnormal;Then solidify sharing feature layer deconvolution parameter,
Each task updates each branching networks weight coefficient by respective loss function.
8. a kind of visual scene recognition methods based on deep neural network according to claim 5, which is characterized in that institute
State online scene analysis, including Web compression and pipelined-flash analysis strategy;
The Web compression refers to:The resulting neural network parameter of offline network training is subjected to quantization and rarefaction, wherein amount
Change data bit length and rarefaction degree is configuration parameter;It is lost using network output accuracy after test set evaluation and test compression, and really
It is fixed whether to quantization rear mold type according to data precision re -training after quantization;
The pipelined-flash analysis strategy refers to:The every frame of scene abnormality detection preferentially executes, and weather and weather detection are using every frame
Or frame-skipping executive mode;In front end applications, by the way of dual-thread or multi-threaded parallel, a thread executes scene to be known extremely
Not, another thread timesharing alternately day when with weather identify, convolutional neural networks hardware acceleration unit request priority with
Scene anomalous identification Network Priority.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810472012.2A CN108921013B (en) | 2018-05-16 | 2018-05-16 | Visual scene recognition system and method based on deep neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810472012.2A CN108921013B (en) | 2018-05-16 | 2018-05-16 | Visual scene recognition system and method based on deep neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108921013A true CN108921013A (en) | 2018-11-30 |
CN108921013B CN108921013B (en) | 2020-08-18 |
Family
ID=64402563
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810472012.2A Active CN108921013B (en) | 2018-05-16 | 2018-05-16 | Visual scene recognition system and method based on deep neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108921013B (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109740551A (en) * | 2019-01-09 | 2019-05-10 | 贵州宽凳智云科技有限公司 | A kind of night Lane detection method and system based on computer vision |
CN109766873A (en) * | 2019-02-01 | 2019-05-17 | 中国人民解放军陆军工程大学 | A kind of pedestrian mixing deformable convolution recognition methods again |
CN109858372A (en) * | 2018-12-29 | 2019-06-07 | 浙江零跑科技有限公司 | A kind of lane class precision automatic Pilot structured data analysis method |
CN109934096A (en) * | 2019-01-22 | 2019-06-25 | 浙江零跑科技有限公司 | Automatic Pilot visual perception optimization method based on feature timing dependence |
CN110119768A (en) * | 2019-04-24 | 2019-08-13 | 苏州感测通信息科技有限公司 | Visual information emerging system and method for vehicle location |
CN110135261A (en) * | 2019-04-15 | 2019-08-16 | 北京易华录信息技术股份有限公司 | A kind of method and system of trained road anomalous identification model, road anomalous identification |
CN110186471A (en) * | 2019-05-06 | 2019-08-30 | 平安科技(深圳)有限公司 | Air navigation aid, device, computer equipment and storage medium based on history video |
CN110543600A (en) * | 2019-09-11 | 2019-12-06 | 上海携程国际旅行社有限公司 | Search ranking method, system, device and storage medium based on neural network |
CN111178253A (en) * | 2019-12-27 | 2020-05-19 | 深圳佑驾创新科技有限公司 | Visual perception method and device for automatic driving, computer equipment and storage medium |
CN111259719A (en) * | 2019-10-28 | 2020-06-09 | 浙江零跑科技有限公司 | Cab scene analysis method based on multi-view infrared vision system |
WO2020156028A1 (en) * | 2019-01-28 | 2020-08-06 | 南京航空航天大学 | Outdoor non-fixed scene weather identification method based on deep learning |
CN111950572A (en) * | 2019-05-14 | 2020-11-17 | 北京字节跳动网络技术有限公司 | Method, apparatus, electronic device and computer-readable storage medium for training classifier |
CN111950656A (en) * | 2020-08-25 | 2020-11-17 | 深圳思谋信息科技有限公司 | Image recognition model generation method and device, computer equipment and storage medium |
CN111967577A (en) * | 2020-07-29 | 2020-11-20 | 华北电力大学 | Energy internet scene generation method based on variational self-encoder |
CN112669316A (en) * | 2021-01-29 | 2021-04-16 | 南方电网调峰调频发电有限公司 | Power production abnormity monitoring method and device, computer equipment and storage medium |
CN113011216A (en) * | 2019-12-19 | 2021-06-22 | 合肥君正科技有限公司 | Multi-classification threshold self-adaptive occlusion detection method |
CN113252058A (en) * | 2021-05-24 | 2021-08-13 | 北京航迹科技有限公司 | IMU data processing method, system, device and storage medium |
CN115107832A (en) * | 2022-08-09 | 2022-09-27 | 中车唐山机车车辆有限公司 | Train headlamp control method, vehicle-mounted terminal, control system and storage medium |
CN115631482A (en) * | 2022-11-30 | 2023-01-20 | 广汽埃安新能源汽车股份有限公司 | Driving perception information acquisition method and device, electronic equipment and readable medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105069472A (en) * | 2015-08-03 | 2015-11-18 | 电子科技大学 | Vehicle detection method based on convolutional neural network self-adaption |
CN105184271A (en) * | 2015-09-18 | 2015-12-23 | 苏州派瑞雷尔智能科技有限公司 | Automatic vehicle detection method based on deep learning |
CN105539326A (en) * | 2015-12-24 | 2016-05-04 | 百利得汽车主动安全系统(苏州)有限公司 | Weather identification system and method based on vehicle-mounted camera |
CN105575119A (en) * | 2015-12-29 | 2016-05-11 | 大连楼兰科技股份有限公司 | Road condition climate deep learning and recognition method and apparatus |
CN107209856A (en) * | 2015-02-11 | 2017-09-26 | 高通股份有限公司 | Environment scene condition detection |
US20180040248A1 (en) * | 2016-08-04 | 2018-02-08 | Herman L. Bigham | Anti-Collision, Accident Prevention Car Safety Systems to Provide Visibility when Passing, Pulling or Backing out of Parking Spaces/Driveways and Entering or Crossing Intersections |
-
2018
- 2018-05-16 CN CN201810472012.2A patent/CN108921013B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107209856A (en) * | 2015-02-11 | 2017-09-26 | 高通股份有限公司 | Environment scene condition detection |
CN105069472A (en) * | 2015-08-03 | 2015-11-18 | 电子科技大学 | Vehicle detection method based on convolutional neural network self-adaption |
CN105184271A (en) * | 2015-09-18 | 2015-12-23 | 苏州派瑞雷尔智能科技有限公司 | Automatic vehicle detection method based on deep learning |
CN105539326A (en) * | 2015-12-24 | 2016-05-04 | 百利得汽车主动安全系统(苏州)有限公司 | Weather identification system and method based on vehicle-mounted camera |
CN105575119A (en) * | 2015-12-29 | 2016-05-11 | 大连楼兰科技股份有限公司 | Road condition climate deep learning and recognition method and apparatus |
US20180040248A1 (en) * | 2016-08-04 | 2018-02-08 | Herman L. Bigham | Anti-Collision, Accident Prevention Car Safety Systems to Provide Visibility when Passing, Pulling or Backing out of Parking Spaces/Driveways and Entering or Crossing Intersections |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109858372A (en) * | 2018-12-29 | 2019-06-07 | 浙江零跑科技有限公司 | A kind of lane class precision automatic Pilot structured data analysis method |
CN109740551A (en) * | 2019-01-09 | 2019-05-10 | 贵州宽凳智云科技有限公司 | A kind of night Lane detection method and system based on computer vision |
CN109934096A (en) * | 2019-01-22 | 2019-06-25 | 浙江零跑科技有限公司 | Automatic Pilot visual perception optimization method based on feature timing dependence |
CN109934096B (en) * | 2019-01-22 | 2020-12-11 | 浙江零跑科技有限公司 | Automatic driving visual perception optimization method based on characteristic time sequence correlation |
WO2020156028A1 (en) * | 2019-01-28 | 2020-08-06 | 南京航空航天大学 | Outdoor non-fixed scene weather identification method based on deep learning |
CN109766873A (en) * | 2019-02-01 | 2019-05-17 | 中国人民解放军陆军工程大学 | A kind of pedestrian mixing deformable convolution recognition methods again |
CN109766873B (en) * | 2019-02-01 | 2021-04-06 | 中国人民解放军陆军工程大学 | Pedestrian re-identification method based on hybrid deformable convolution |
CN110135261A (en) * | 2019-04-15 | 2019-08-16 | 北京易华录信息技术股份有限公司 | A kind of method and system of trained road anomalous identification model, road anomalous identification |
CN110119768A (en) * | 2019-04-24 | 2019-08-13 | 苏州感测通信息科技有限公司 | Visual information emerging system and method for vehicle location |
CN110186471A (en) * | 2019-05-06 | 2019-08-30 | 平安科技(深圳)有限公司 | Air navigation aid, device, computer equipment and storage medium based on history video |
CN111950572A (en) * | 2019-05-14 | 2020-11-17 | 北京字节跳动网络技术有限公司 | Method, apparatus, electronic device and computer-readable storage medium for training classifier |
CN110543600A (en) * | 2019-09-11 | 2019-12-06 | 上海携程国际旅行社有限公司 | Search ranking method, system, device and storage medium based on neural network |
CN111259719B (en) * | 2019-10-28 | 2023-08-25 | 浙江零跑科技股份有限公司 | Cab scene analysis method based on multi-view infrared vision system |
CN111259719A (en) * | 2019-10-28 | 2020-06-09 | 浙江零跑科技有限公司 | Cab scene analysis method based on multi-view infrared vision system |
CN113011216A (en) * | 2019-12-19 | 2021-06-22 | 合肥君正科技有限公司 | Multi-classification threshold self-adaptive occlusion detection method |
CN113011216B (en) * | 2019-12-19 | 2024-04-02 | 合肥君正科技有限公司 | Multi-classification threshold self-adaptive shielding detection method |
CN111178253A (en) * | 2019-12-27 | 2020-05-19 | 深圳佑驾创新科技有限公司 | Visual perception method and device for automatic driving, computer equipment and storage medium |
CN111178253B (en) * | 2019-12-27 | 2024-02-27 | 佑驾创新(北京)技术有限公司 | Visual perception method and device for automatic driving, computer equipment and storage medium |
CN111967577A (en) * | 2020-07-29 | 2020-11-20 | 华北电力大学 | Energy internet scene generation method based on variational self-encoder |
CN111967577B (en) * | 2020-07-29 | 2024-04-05 | 华北电力大学 | Energy Internet scene generation method based on variation self-encoder |
CN111950656A (en) * | 2020-08-25 | 2020-11-17 | 深圳思谋信息科技有限公司 | Image recognition model generation method and device, computer equipment and storage medium |
CN111950656B (en) * | 2020-08-25 | 2021-06-25 | 深圳思谋信息科技有限公司 | Image recognition model generation method and device, computer equipment and storage medium |
CN112669316A (en) * | 2021-01-29 | 2021-04-16 | 南方电网调峰调频发电有限公司 | Power production abnormity monitoring method and device, computer equipment and storage medium |
CN112669316B (en) * | 2021-01-29 | 2023-05-30 | 南方电网调峰调频发电有限公司 | Power production abnormality monitoring method, device, computer equipment and storage medium |
CN113252058A (en) * | 2021-05-24 | 2021-08-13 | 北京航迹科技有限公司 | IMU data processing method, system, device and storage medium |
CN115107832A (en) * | 2022-08-09 | 2022-09-27 | 中车唐山机车车辆有限公司 | Train headlamp control method, vehicle-mounted terminal, control system and storage medium |
CN115631482B (en) * | 2022-11-30 | 2023-04-04 | 广汽埃安新能源汽车股份有限公司 | Driving perception information acquisition method and device, electronic equipment and readable medium |
CN115631482A (en) * | 2022-11-30 | 2023-01-20 | 广汽埃安新能源汽车股份有限公司 | Driving perception information acquisition method and device, electronic equipment and readable medium |
Also Published As
Publication number | Publication date |
---|---|
CN108921013B (en) | 2020-08-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108921013A (en) | A kind of visual scene identifying system and method based on deep neural network | |
Rajendran et al. | Real-time traffic sign recognition using YOLOv3 based detector | |
CN110443969B (en) | Fire detection method and device, electronic equipment and storage medium | |
DE112018004797T5 (en) | VIEWPOINT INVARIANT OBJECT DETECTION THROUGH SYNTHETIZING AND DOMAIN ADAPTATION | |
DE102017123002A1 (en) | MULTI-SENSOR PRECIPITATION CLASSIFICATION DEVICE AND METHOD | |
CN104200494B (en) | Real-time visual target tracking method based on light streams | |
CN103578119B (en) | Target detection method in Codebook dynamic scene based on superpixels | |
CN110807740B (en) | Image enhancement method and system for monitoring scene vehicle window image | |
CN107292386A (en) | Detected using the rainwater of the view-based access control model of deep learning | |
CN107133974A (en) | The vehicle type classification method that Gaussian Background modeling is combined with Recognition with Recurrent Neural Network | |
CN109961003A (en) | A kind of airborne auxiliary inspection device of embedded transmission line of electricity based on FPGA | |
CN104156734A (en) | Fully-autonomous on-line study method based on random fern classifier | |
CN109389136A (en) | Classifier training method | |
CN109543691A (en) | Ponding recognition methods, device and storage medium | |
DE112010002232T5 (en) | Semantic scene segmentation using random multinominal logit (RML) | |
Tourani et al. | Motion-based vehicle speed measurement for intelligent transportation systems | |
CN103344583A (en) | Praseodymium-neodymium (Pr/Nd) component content detection system and method based on machine vision | |
CN112489072B (en) | Vehicle-mounted video perception information transmission load optimization method and device | |
CN212009589U (en) | Video identification driving vehicle track acquisition device based on deep learning | |
CN111931715A (en) | Method and device for recognizing state of vehicle lamp, computer equipment and storage medium | |
CN106779089A (en) | Method of testing and device | |
Manzari et al. | Pyramid transformer for traffic sign detection | |
CN113177528A (en) | License plate recognition method and system based on multi-task learning strategy training network model | |
CN103294990B (en) | Select the visibility region in nighttime image to perform clear path detection | |
CN112232226A (en) | Method and system for detecting target object through discriminant model |
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 | ||
CP01 | Change in the name or title of a patent holder | ||
CP01 | Change in the name or title of a patent holder |
Address after: 310051 1st and 6th floors, no.451 Internet of things street, Binjiang District, Hangzhou City, Zhejiang Province Patentee after: Zhejiang Zero run Technology Co.,Ltd. Address before: 310051 1st and 6th floors, no.451 Internet of things street, Binjiang District, Hangzhou City, Zhejiang Province Patentee before: ZHEJIANG LEAPMOTOR TECHNOLOGY Co.,Ltd. |