CN110414389A - A kind of object detection method of the fast area search based on deep learning - Google Patents
A kind of object detection method of the fast area search based on deep learning Download PDFInfo
- Publication number
- CN110414389A CN110414389A CN201910631921.0A CN201910631921A CN110414389A CN 110414389 A CN110414389 A CN 110414389A CN 201910631921 A CN201910631921 A CN 201910631921A CN 110414389 A CN110414389 A CN 110414389A
- Authority
- CN
- China
- Prior art keywords
- target
- layer
- network
- deep learning
- search based
- 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
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- 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
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of object detection method of the fast area search based on deep learning, is related to algorithm of target detection field.The object in addition to train field cannot be identified to solve existing train control system, it cannot be remembered and be learnt for the object other than train field, and some emergency cases can not be handled, it can not judge that whether there are obstacles in front of train driving, the problem of generation so as to cause accident.Build the target detection network of the fast area search based on deep learning；The image for having obstacle target to be identified is obtained, and it is labeled, is divided into test sample collection and training sample set；Training sample set is input to target detection network and carries out propagated forward calculating, calculates error and by error back propagation, adjusts network weight；Network Recognition accuracy rate is assessed using test sample collection.If being unsatisfactory for requiring, if meeting the requirements, network parameter is saved, training is completed.The present invention is suitable for carrying out detection positioning to the target optical imagery.
Description
Technical field
The present invention relates to algorithm of target detection fields, and in particular to a kind of mesh of the fast area search based on deep learning
Mark detection method.
Background technique
With the fast development of railway construction in China, train running speed is continuously improved, and the requirement to traffic safety is not yet
Disconnected to improve, existing column are controlled and ATP system using nonpassage of signal, close to response means guarantees traffic safety, but train control system
It can only guarantee to monitor this cooperative target of traveling train, for invading limit event with the noncooperative targets such as sudden, unpredictable
Generation cannot give warning in advance.Especially, it should be noted that construction of railways operation recruitment is engaged in, mechanical car leads under the conditions of construction operation
Reason condition ground signal system is in closing or failure state, cannot play the role of monitoring train.In existing Train Control
Object of the identification in addition to train field in system, such as pedestrian, automobile, animal etc., and cannot be to paroxysmal situation
It makes and timely judging, lead to the generation of accident.
In conclusion existing train control system cannot identify the object in addition to train field, train cannot be directed to
Object other than field is remembered and is learnt, and can not handle some emergency cases, whether can not judge train driving front
There are barriers, so as to cause accident generation.
Summary of the invention
The present invention is to solve existing train control system to identify object in addition to train field, cannot be for column
Object other than vehicle field is remembered and is learnt, and can not handle some emergency cases, can not judge that train driving front is
No there are barriers, the problem of generation so as to cause accident, and propose a kind of mesh of fast area search based on deep learning
Mark detection method.
A kind of object detection method of fast area search based on deep learning of the invention, specific algorithm are as follows:
Step 1: building the target detection network of the fast area search based on deep learning；
Step 2: obtaining the image for having obstacle target to be identified, and it is labeled, is divided into test sample
Collection and training sample set；
Step 3: training sample set, which is input to target detection network, carries out propagated forward calculating, error is calculated and by error
Backpropagation adjusts network weight；
Step 4: assessing Network Recognition accuracy rate using test sample collection.If being unsatisfactory for requiring, network parameter is saved,
Training is completed；
Step 5: for detecting the target of region of search after the completion of training；
Further, target detection network structure in the step one is for first 16 layers in target detection network structure
Feature extraction layer, structure are alternate 8 convolutional layers and 8 pond layer compositions, and layer output in pond connects two full articulamentums,
Full articulamentum output two different peculiar task layers of connection, one is prediction classification output, another sits for predicted position
Mark；
The main purpose of each convolutional layer and pond layer is feature extraction and combination to the image progress depth of input,
Convolution kernel number is respectively 64,128,256,256,256,256,256,128, and convolution kernel size is 3*3, can be protected in this way
Card extract details low-level image feature abundant and it be combined, a pond layer is all connected with after each convolutional layer, be used for pair
The feature extracted carries out dimensionality reduction, and optimization caused by preventing parameter excessive is difficult, and pond layer is having a size of 2*2, step-length 2；
Full articulamentum is used to carry out the depth characteristic extracted further global combination, the input nerve of full articulamentum 1
Member has 256, and output neuron number is 200.The input neuron number of full articulamentum 2 is 200, output neuron number
It is 128；
Two peculiar layers of task are absolute construction, are input with the output of full articulamentum 2, the peculiar layer 1 of task is 1 layer
Full articulamentum, input neuron number are 128, and output neuron number is target class number to be identified, are used for predicted detection
Target type out, the multi-layer perception (MLP) that the peculiar layer 2 of task is one 3 layers, input neuron are 128, output neuron 3,
Output result is target start position, bounding box length, bounding box width；
Training sample set is input to target detection network and carries out propagated forward calculating, meter in the further step three
It calculates error and by error back propagation, adjusts the specific of network weight and calculate that steps are as follows:
If training sample set is combined into For sample data set,For
I-th of training sample, NsIt is total for training sample,For with XsCorresponding classification space,It is instructed for i-th
Practice sampleClass label,For the position of the sample in the picture, if f (x) is target detection network model in the present invention, x generation
Table input sample,yiIndicate the classification of prediction, piIndicate the position in the corresponding image of classification of prediction, root
It is predicted that value and true tag calculate loss function
WhereinFor Classification Loss,For Area Prediction loss, λ is that weight factor is lost for control area prediction in overall loss
Specific gravity in function, Classification Loss part d () indicate true tagY is predicted with classificationiBetween difference, calculation
It is not unique, Ω | | ω | |2For L2 regular terms, for preventing disaggregated model over-fitting.
Compared with the prior art, the invention has the following beneficial effects:
One, the shortcomings that the present invention overcomes the prior arts, it is fixed using detect to target interested in optical imagery
Position, to realize that whether there are obstacles in judgement front.
Two, algorithm acquisition of the present invention largely has the image data of target to be identified, and is manually marked to it
Then note is built the fast area search target detection network based on deep learning, is instructed using the image largely marked
Practice, can be realized under complex background condition after the completion of training, accurate detection of obstacles.
Three, the shortcomings that the present invention overcomes the prior arts realizes and accurately judges in front of train driving with the presence or absence of obstacle
Object avoids that accident occurs；
Detailed description of the invention
Fig. 1 is the heretofore described fast area search algorithm of target detection block diagram based on deep learning.
Specific embodiment
Specific embodiment 1: embodiment is described with reference to Fig. 1, one kind described in present embodiment is based on deep learning
Fast area search object detection method, specific algorithmic procedure is as follows:
Step 1: building the target detection network of the fast area search based on deep learning；
Step 2: obtaining the image for having obstacle target to be identified, and it is labeled, is divided into test sample
Collection and training sample set；
Step 3: training sample set, which is input to target detection network, carries out propagated forward calculating, error is calculated and by error
Backpropagation adjusts network weight；
Step 4: assessing Network Recognition accuracy rate using test sample collection.If being unsatisfactory for requiring, network parameter is saved,
Training is completed；
Step 5: for detecting the target of region of search after the completion of training.
Specific embodiment 2: embodiment is described with reference to Fig. 1, present embodiment is to described in specific embodiment one
Detection algorithm further restriction, a kind of fast area search based on deep learning described in present embodiment
Object detection method, target detection network structure in the step one are characterized for first 16 layers in target detection network structure
Extract layer, structure are alternate 8 convolutional layers and 8 pond layer compositions, and layer output in pond meets two full articulamentums, Quan Lian
Layer output two different peculiar task layer of connection are connect, one is prediction classification output, another is used for predicted position coordinate.
Specific embodiment 3: embodiment is described with reference to Fig. 1, present embodiment is to described in specific embodiment two
Detection algorithm further restriction, a kind of fast area search based on deep learning described in present embodiment
Object detection method, the main purpose of each convolutional layer and pond layer is to carry out depth to the image of input
Feature extraction and combination, convolution kernel number are respectively 64,128,256,256,256,256,256 and 128, and convolution kernel size is equal
For 3*3, it can guarantee to extract details low-level image feature abundant in this way and it is combined, be all connected with after each convolutional layer
One pond layer, for carrying out dimensionality reduction to the feature extracted, optimization caused by preventing parameter excessive is difficult, pond layer size
For 2*2, step-length 2.
Specific embodiment 4: embodiment is described with reference to Fig. 1, present embodiment is to described in specific embodiment three
Detection algorithm further restriction, a kind of fast area search based on deep learning described in present embodiment
Object detection method, the full articulamentum is used to carry out the depth characteristic extracted further global combination, complete to connect
The input neuron of layer 1 has 256, and output neuron number is 200.The input neuron number of full articulamentum 2 is 200
A, output neuron number is 128；
Two peculiar layers of task are absolute construction, are input with the output of full articulamentum 2, the peculiar layer 1 of task is one layer
Full articulamentum, input neuron number be 128, output neuron number be target class number to be identified, for predicting
The target type detected, the multi-layer perception (MLP) that the peculiar layer 2 of task is one three layers, input neuron are 128, output mind
It is 3 through member, output result is target start position, bounding box length, bounding box width.
Specific embodiment 5: embodiment is described with reference to Fig. 1, present embodiment is to described in specific embodiment one
Detection algorithm further restriction, a kind of fast area search based on deep learning described in present embodiment
Object detection method, training sample set is input to target detection network and carries out propagated forward calculating in the step three, calculates
Error and by error back propagation, adjusting the specific calculating of network weight, steps are as follows:
If training sample set is combined into For sample data set,For
I-th of training sample, NsIt is total for training sample,For with XsCorresponding classification space,It is instructed for i-th
Practice sampleClass label,For the position of the sample in the picture, if f (x) is target detection network model in the present invention, x generation
Table input sample,yiIndicate the classification of prediction, piIndicate the position in the corresponding image of classification of prediction, root
It is predicted that value and true tag calculate loss function
WhereinFor Classification Loss,For Area Prediction loss, λ is that weight factor is lost for control area prediction in overall loss
Specific gravity in function, Classification Loss part d () indicate true tagY is predicted with classificationiBetween difference, calculation
It is not unique, Ω | | ω | |2For L2 regular terms, for preventing disaggregated model over-fitting.
Claims (5)
1. a kind of object detection method of the fast area search based on deep learning, it is characterised in that: the target detection is calculated
Method, specific algorithmic procedure are as follows:
Step 1: building the target detection network of the fast area search based on deep learning；
Step 2: obtaining the image for having obstacle target to be identified, and it is labeled, be divided into test sample collection and
Training sample set；
Step 3: training sample set, which is input to target detection network, carries out propagated forward calculating, calculating error is simultaneously reversed by error
It propagates, adjusts network weight；
Step 4: assessing Network Recognition accuracy rate using test sample collection.If being unsatisfactory for requiring, if meeting the requirements, save
Network parameter, training are completed；
Step 5: for detecting the target of region of search after the completion of training.
2. a kind of object detection method of fast area search based on deep learning according to claim 1, feature
Be: target detection network structure in the step one is characterized extract layer for first 16 layers in target detection network structure,
Structure is alternate 8 convolutional layers and 8 pond layer compositions, and layer output in pond connects two full articulamentums, and full articulamentum output connects
Two different peculiar task layers are connect, one is prediction classification output, another is used for predicted position coordinate.
3. a kind of object detection method of fast area search based on deep learning according to claim 2, feature
Be: the main purpose of each convolutional layer and pond layer is to carry out the image of input the feature extraction and combination of depth, volume
Product core number is respectively 64,128,256,256,256,256,256 and 128, and convolution kernel size is 3*3, can be guaranteed in this way
It extracts details low-level image feature abundant and it is combined, a pond layer is all connected with after each convolutional layer, for mentioning
The feature got carries out dimensionality reduction, and optimization caused by preventing parameter excessive is difficult, and pond layer is having a size of 2*2, step-length 2.
4. a kind of object detection method of fast area search based on deep learning according to claim 3, feature exist
In: full articulamentum, which is used to carry out the depth characteristic extracted further global combination, the input neuron of full articulamentum 1, to be had
256, output neuron number is 200.The input neuron number of full articulamentum 2 is 200, output neuron number
It is 128；
Two peculiar layers of task are absolute construction, are input with the output of full articulamentum 2, and the peculiar layer 1 of task is one layer complete
Articulamentum, input neuron number are 128, and output neuron number is target class number to be identified, are used for predicted detection
Target type out, the multi-layer perception (MLP) that the peculiar layer 2 of task is one three layers, input neuron is 128, and output neuron is
3, output result is target start position, bounding box length, bounding box width.
5. a kind of object detection method of fast area search based on deep learning according to claim 1, feature
Be: training sample set is input to target detection network and carries out propagated forward calculating in the step three, and calculating error simultaneously will
Error back propagation, adjust network weight it is specific calculating steps are as follows:
If training sample set is combined into For sample data set,It is i-th
A training sample, NsIt is total for training sample,For with XsCorresponding classification space,For i-th of training
SampleClass label,For the position of the sample in the picture, if f (x) is target detection network model in the present invention, x
Input sample is represented,yiIndicate the classification of prediction, piIndicate the position in the corresponding image of classification of prediction, root
It is predicted that value and true tag calculate loss function
WhereinFor Classification Loss,For Area Prediction loss, λ is that weight factor is lost for control area prediction in overall loss letter
Specific gravity in number, Classification Loss part d () indicate true tagY is predicted with classificationiBetween difference, calculation is not
Uniquely, Ω | | ω | |2For L2 regular terms, for preventing disaggregated model over-fitting.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910631921.0A CN110414389A (en) | 2019-07-12 | 2019-07-12 | A kind of object detection method of the fast area search based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910631921.0A CN110414389A (en) | 2019-07-12 | 2019-07-12 | A kind of object detection method of the fast area search based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110414389A true CN110414389A (en) | 2019-11-05 |
Family
ID=68361329
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910631921.0A Pending CN110414389A (en) | 2019-07-12 | 2019-07-12 | A kind of object detection method of the fast area search based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110414389A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109447033A (en) * | 2018-11-14 | 2019-03-08 | 北京信息科技大学 | Vehicle front obstacle detection method based on YOLO |
| CN109615007A (en) * | 2018-12-10 | 2019-04-12 | 天津工业大学 | Deep learning network objectives detection method based on particle filter |
| CN109886066A (en) * | 2018-12-17 | 2019-06-14 | 南京理工大学 | Fast target detection method based on the fusion of multiple dimensioned and multilayer feature |
| CN109977812A (en) * | 2019-03-12 | 2019-07-05 | 南京邮电大学 | A kind of Vehicular video object detection method based on deep learning |
-
2019
- 2019-07-12 CN CN201910631921.0A patent/CN110414389A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109447033A (en) * | 2018-11-14 | 2019-03-08 | 北京信息科技大学 | Vehicle front obstacle detection method based on YOLO |
| CN109615007A (en) * | 2018-12-10 | 2019-04-12 | 天津工业大学 | Deep learning network objectives detection method based on particle filter |
| CN109886066A (en) * | 2018-12-17 | 2019-06-14 | 南京理工大学 | Fast target detection method based on the fusion of multiple dimensioned and multilayer feature |
| CN109977812A (en) * | 2019-03-12 | 2019-07-05 | 南京邮电大学 | A kind of Vehicular video object detection method based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112700470B (en) | Target detection and track extraction method based on traffic video stream | |
| JP6599986B2 (en) | Hyperclass expansion and regularization deep learning for fine-grained image classification | |
| Ye et al. | Railway traffic object detection using differential feature fusion convolution neural network | |
| US9665802B2 (en) | Object-centric fine-grained image classification | |
| US20160140400A1 (en) | Atomic scenes for scalable traffic scene recognition in monocular videos | |
| CN106530831A (en) | System and method for monitoring and early warning of high-threat vehicles | |
| US20210403034A1 (en) | Systems and Methods for Optimizing Trajectory Planner Based on Human Driving Behaviors | |
| CN111415533B (en) | Bend safety early warning monitoring method, device and system | |
| CN112633176B (en) | Rail transit obstacle detection method based on deep learning | |
| CN108909624A (en) | A kind of real-time detection of obstacles and localization method based on monocular vision | |
| JP2018503160A (en) | Near-online multi-target tracking using aggregate local flow descriptor (ALFD) | |
| US11562556B1 (en) | Prediction error scenario mining for machine learning models | |
| CN113378741B (en) | Auxiliary sensing method and system for aircraft tractor based on multi-source sensor | |
| CN106291506A (en) | Vehicle target recognition methods based on single line cloud data machine learning and device | |
| CN115257784A (en) | Vehicle-road cooperative system based on 4D millimeter wave radar | |
| CN115993597A (en) | Visual radar perception fusion method and terminal equipment | |
| CN115662166B (en) | Automatic driving data processing method and automatic driving traffic system | |
| CN117133099A (en) | Automatic monitoring system for disaster of expressway high slope | |
| CN116611603A (en) | Vehicle path scheduling method, device, computer and storage medium | |
| CN113552883B (en) | Ground unmanned vehicle autonomous driving method and system based on deep reinforcement learning | |
| CN115713687A (en) | Method, system, and medium for determining dynamic parameters of an object | |
| CN113435404A (en) | Safe auxiliary driving method and system for shunting electric bus based on radar and image detection | |
| CN110414389A (en) | A kind of object detection method of the fast area search based on deep learning | |
| CN114333331B (en) | Method and system for identifying vehicle passing information and vehicle weight of multi-lane bridge | |
| CN114872741B (en) | Locomotive auxiliary automatic driving system and method based on safety guidance |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191105 |
|
| WD01 | Invention patent application deemed withdrawn after publication |