CN108256498A - A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN - Google Patents
A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN Download PDFInfo
- Publication number
- CN108256498A CN108256498A CN201810103573.5A CN201810103573A CN108256498A CN 108256498 A CN108256498 A CN 108256498A CN 201810103573 A CN201810103573 A CN 201810103573A CN 108256498 A CN108256498 A CN 108256498A
- Authority
- CN
- China
- Prior art keywords
- training
- layer
- power driven
- driven vehicle
- non power
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- 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)
- 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
The present invention proposes a kind of method that can solve non power driven vehicle target detection problems, by the way that the mininet structure C affnet in EdgeBoxes algorithm fusion Fast R CNN is obtained new network structure, with reference to the non power driven vehicle data sample of VOC forms, training is iterated to new network structure and obtains non power driven vehicle target detection model, the classification problem that the target detection problems realification in road traffic scene is bicycle bicycle and electric vehicle evbike.This method is from target suggestion, effectively overcome the changeable difficulty brought of non power driven vehicle appearance using deep learning thought, and can training data driving under adaptively construction feature describe, it is capable of the non power driven vehicle target detected in road traffic scene of efficiently and accurately so that non power driven vehicle target detection more fast facilitates.
Description
Technical field
The present invention relates to a kind of object detection method, specifically a kind of non-machine based on EdgeBoxes and Fast R-CNN
Motor-car object detection method.
Background technology
The object detection and recognition of road traffic image hands over rule for improving friendship method, ensures that safety etc. has important meaning
Justice.Especially in recent years, bicycle, the fast development for taking out industry are shared, the non power driven vehicles such as bicycle increase sharply on road, road
Traffic safety problem is also increasingly severeer, since this kind of non power driven vehicle target is small, mobile flexible, in addition currently to such non-machine
The research of motor-car target detection is relatively fewer, and traffic monitoring faces huge challenge.Therefore, based on non power driven vehicle target signature
Information detects in actual traffic image and positions non power driven vehicle target, perfect to traffic accident fix duty, traffic video etc. to have
Important booster action.
Non power driven vehicle target in road traffic image can generate due to illumination, visual angle and driver are blocked when variation
How variation, ensure good detection result, traditional object detection method has done basis under the premise of existing for many variations
Property attempt, but still there are detection efficiency is low and the problems such as poor reliability.Such as HOG, SIFT traditional machine learning method leads to
Extraction target signature is crossed, and the feature extracted is inputted into the graders such as support vector machines, iterator AdaBoost and is divided
Class identifies：But entirely characteristic extraction procedure is sufficiently complex, and substantially these features for being characterized in hand-designed, needs to be directed to
Different image scenes selects suitable feature, is unfavorable for practical engineering application, generalization ability is poor.It is emerging with deep learning
It rises, the method for some deep learnings is applied to object detection field quickly, wherein the most prominent with depth convolutional neural networks CNN
Go out.Pioneers of the R-CNN in 2013 as deep learning target detection application field carries out conventional machines study and deep learning
Innovative combination, then occurs as the optimization of the target detections network such as SPP-NET, Fast R-CNN are incorporated within 2015
The advantages of R-CNN and SPP-Net so that under changeable road traffic image, quickly and accurately detect non power driven vehicle
Target is possibly realized.But the detection result of Fast R-CNN, it is largely dependent upon sample image target and suggests Object
The extraction quantity of Proposals, the extraction of a large amount of OP take time and effort, and aggravate the burden of model training, to network and hardware
It is more demanding.
Invention content
To solve the above-mentioned problems, the present invention proposes a kind of target merged based on EdgeBoxes and Fast R-CNN
Detection method, and applied in the target detection of non power driven vehicle.For the target proposed extraction stage, EdgeBoxes algorithms
It can ensure when extracting lesser amt, there is relatively high average recall rate, by controlling the extraction quantity of OP, reduce detection
The computation complexity of process.Core CNN in deep learning algorithm Fast R-CNN has geometric transformation, deformation, illumination etc.
Invariance to a certain degree effectively overcomes the changeable difficulty brought of non power driven vehicle appearance, and can be under training data driving certainly
Adaptively construction feature describes.
By the mininet structure C affnet in EdgeBoxes algorithm fusion Fast R-CNN, with reference to the non-of VOC forms
Motor vehicles data sample asks classification of the target detection problems realification in road traffic scene for bicycle and electric vehicle
Topic.Suitable area-of-interest is built using the target suggestion of EdgeBoxes algorithms extraction sample and sample inputs network together
Training is iterated, obtains non power driven vehicle target detection model, new samples test is carried out to model and analyzes test effect, with
Standby subsequent model generalization process uses.The fusion of EdgeBoxes and Fast R-CNN is so that the target detection of non power driven vehicle
With higher flexibility and generalization ability.Specifically include following steps：
Step 1 defines network structure, selectes the mininet structure C affenet networks in Fast R-CNN algorithms：
Step 1-1, the setting Caffenet network cores number of plies is 11 layers, specifically includes 1 11 × 11 convolutional layer, 15 ×
5 convolutional layers, 33 × 3 convolutional layers, 1 area-of-interest pond layer, 2 full articulamentums, 3 pond layers；
Step 1-2, Caffenet networks output layer is multitask loss function, and specific there are two output layers, and first defeated
Go out layer and probability Ps of each RoI in 2 classes, i.e. target classification are calculated as a result, second defeated by the recurrence of softmax algorithms
Go out the coordinate value that layer is responsible for calculating 2 class target detection blocks, i.e. detection block coordinate result；
Step 2 defines goal task, acquires target sample figure and data set makes：
Step 2-1, it is non power driven vehicle, i.e. bicycle and electric vehicle to define goal task；
Step 2-2, data set makes, and prepares the sample image of appropriate goal task, to ensure feature adequacy, sample graph
As the order of magnitude is set as 103-104, any numerical value M in the range of this can be as sample size, bicycle b and electric vehicle e distribution
Ratio is b:E, and b ≈ e, M=b+e；
Step 2-3, the sample image for obtaining step 2-2 carries out VOC formattings, obtains VOC data sets；
Step 2-4, the sample graph in the VOC data sets for obtaining step 2-3 carries out target mark, then uses
EdgeBoxes algorithms carry out the sample image after mark target and suggest OP extractions, using the extraction strategy of sliding window, by with
Lower parameter controls：α command deployment step-lengths, β control IoU threshold values, δ command deployment accuracy, minScore control minimum score
The maximum number magnitude of extraction OP is controlled with maxBoxes, sample OP information is obtained with this, carries out obtaining sample after centainly calculating
Area-of-interest RoIs；
Step 2-5, the VOC data sets for obtaining step 2-3, according to training>Test>The ratio of verification, and each ratio
Division principle of the sum of the item less than or equal to 10 is divided, and obtains training set, test set and verification collection；
Step 3, model training set pre-training parameter：
Step 3-1, initiation parameter setting is carried out to the Caffnet network structures defined in step 1, utilizes Caffenet
The layer parameter before the layer of training parameter initialization pre-training network RoI ponds under ImageNet, setting learning rate are
0.001, trained batchsize are 2, iterations I, order of magnitude 10^4, carry out network training；
Step 3-2, goal task is bicycle and electric vehicle, i.e. K=2, then adds a background classes, then input data layer class
Not Shuo num_class=2+1=3, similary RoI ponds layer data classification number num_class=3, branch of classifying to obtain in output layer
Divided data classification number num_class=3, bounding box return partial data classification number num_class=3 × 4=12；Step 4, mould
Type training, pre-training network：
Step 4-1, it will be instructed on training dataset that the Caffnet network structures defined in step 1 prepare in step 2-5
Practice, sample includes dynamic bicycle and electric vehicle, and training parameter is set by step 3-1；
Step 4-2, the output process of training a kind of type of multitask loss function to the RoI of each label and detection
Frame coordinate carries out recurrence calculating, and computational methods are：
L(p,u,tu, v) and=Lcls(p,u)+λ[u≥1]Lloc(tu,v) (1)
Lcls(p, u)=logPu (2)
Step 4-3, to hyper parameter λ=1 in step 4-2 setting formula (1), and bicycle class u=1, electric vehicle class u are marked
=2, LclsIt is type probability PuLogarithm loss, wherein PuIt is calculated with softmax, the output vector for enabling this layer is ai=
(a1,a2,......,an), according to softmax computational methods
Then
Pu=maxPi (4)
Step 4-4, another task loss L of outputloc, it is detection block loss, passes through true u classes detection block coordinate
Value v=(vx,vy,vw,vh) and u classes prediction coordinate valueIt obtains, since background classes are without former label
Frame can ignore the L of background classesloc, detection block is returned and is lost, uses loss function
Wherein, when regressive object non-boundary, traditional loss function susceptibility in training is very high, needs very carefully
Ground regularized learning algorithm rate with prevent explosion gradient, andEliminate this sensibility：
Step 5, repetitive exercise obtain target detection model：
Step 5-1, repetitive exercise process carries out backpropagation, x by RoI ponds layeri(xi∈ R) it is i-th of activation, it is defeated
Enter to RoI ponds layer, enable yrjIt is output, i.e., the jth layer of r-th RoI exports, and RoI ponds layer calculates yrj=xi*(r,j), wherein i*
(r, j)=argmaxi'∈R(i,j)xi', i.e., so that xi'Maximum i, R (i, j) is output unit yrjIn child window where maximum pondization
The indexed set of input, single xiSeveral different output y can be assigned torj, RoI ponds layer is by calculating relative to each defeated
Enter variable xiLoss function partial derivative carry out backpropagation：
Step 5-2, repetitive exercise process introduces fine tuning strategy, and the training parameter of continuous trim step 3-1 settings is so handed over
For network training is carried out, until network convergence, obtains target detection model at this time；
Step 6, model measurement, the target detection model that the test set image input step 5-2 that step 2-5 is obtained is obtained
It is tested, is broadly divided into individual test and integrated testability two parts：
Step 6-1, the sample image income single image detection test in selected test set, and shown on image pattern
Detection block, to ensure the practical application and accuracy rate of detection, setting detection type probability P range is 0.5≤P≤1 during test
Numerical value；
Step 6-2, that step 2-5 is obtained the target detection model that whole test data set input step 5-2 obtain is whole
Target detection model is verified in test；
Preferably, sample values M is 4096 in step 2-2, and bicycle b is 2262, and electric vehicle e is 1834.
Preferably, according to training in step 2-5：Test：Verification=5：2：1 ratio cut partition sample M, obtains training set
2560, verification collection 1024,512 Zhang San part of test set.Preferably, iterations I is set as 40000 times in step 3-1
Carry out network repetitive exercise.Preferably, detection probability P is set as 0.7 in step 6-1, i.e., a display target probability is more than or equal to
0.7 target.
The present invention merges EdgeBoxes algorithms and deep learning method Fast R-CNN, obtains new network mould
Type structure, and it is applied to the target detection process of non power driven vehicle, it has obtained melting based on EdgeBoxes and Fast R-CNN
The non power driven vehicle object detection method of conjunction.Compared with conventional target detection method, the present invention had both maintained deep learning calculation
The advantage of method possesses higher Detection accuracy；Meanwhile using the extraction quantity of EdgeBoxes algorithms control OP, reduce net
The complexity of network training improves the detection efficiency of target detection；And the method after fusion is applied to non power driven vehicle
Blank of the traffic monitoring in terms of non power driven vehicle has been filled up in field, is improved for road traffic supervision video and is provided new side
Method.
Description of the drawings
Fig. 1 is the schematic network structure that Caffenet of the present invention merges EdgeBoxes algorithms, wherein the network core number of plies
It is 11 layers, specifically includes 1 11 × 11 convolutional layer, 15 × 5 convolutional layer, 33 × 3 convolutional layers, 1 area-of-interest pond
Layer, 2 full articulamentums, 3 pond layers carry out target proposed extraction using EdgeBoxes algorithms to the sample image after mark,
It is fused to the area-of-interest pond layer of Caffenet networks later, network output layer is multitask loss function, specifically there is two
A output layer is each responsible for output target classification result and target detection frame coordinate result；Fig. 2 be the present invention is based on
The non power driven vehicle object detection method flow chart of EdgeBoxes and Fast R-CNN fusions, detailed process step is by having
Body embodiment provides.
Specific embodiment
In order to deepen the understanding of the present invention, the present invention is described in further detail with reference to example.But this
Invention can be realized with different goal tasks, however it is not limited to example described herein.
The present invention is a kind of object detection method, as shown in Figure 2 based on the non-of EdgeBoxes and Fast R-CNN fusions
Motor vehicles object detection method flow chart, mainly includes the following steps that：Step 1 defines network structure, selectes Fast R-CNN
Mininet structure C affenet networks in algorithm：
Step 1-1, the setting Caffenet network cores number of plies is 11 layers, specifically includes 1 11 × 11 convolutional layer, 15 ×
5 convolutional layers, 33 × 3 convolutional layers, 1 area-of-interest pond layer, 2 full articulamentums, 3 pond layers；
Step 1-2, Caffenet networks output layer is multitask loss function, and specific there are two output layers, and first defeated
Go out layer and probability Ps of each RoI in 2 classes, i.e. target classification are calculated as a result, second defeated by the recurrence of softmax algorithms
Go out the coordinate value that layer is responsible for calculating 2 class target detection blocks, i.e. detection block coordinate result；
Step 2 defines goal task, acquires target sample figure and data set makes：
Step 2-1, it is non power driven vehicle, i.e. bicycle and electric vehicle to define goal task；
Step 2-2, data set makes, and prepares the sample image of appropriate goal task, to ensure feature adequacy, sample graph
As the order of magnitude is set as 103-104, any numerical value M in the range of this can be as sample size, bicycle b and electric vehicle e distribution
Ratio is b:E, and b ≈ e, M=b+e；
Step 2-3, the sample image for obtaining step 2-2 carries out VOC formattings, obtains VOC data sets；
Step 2-4, the sample graph in the VOC data sets for obtaining step 2-3 carries out target mark, then uses
EdgeBoxes algorithms carry out the sample image after mark target and suggest OP extractions, using the extraction strategy of sliding window, by with
Lower parameter controls：α command deployment step-lengths, β control IoU threshold values, δ command deployment accuracy, minScore control minimum score
The maximum number magnitude of extraction OP is controlled with maxBoxes, sample OP information is obtained with this, carries out obtaining sample after centainly calculating
Area-of-interest RoIs；
Step 2-5, the VOC data sets for obtaining step 2-3, according to training>Test>The ratio of verification, and each ratio
Division principle of the sum of the item less than or equal to 10 is divided, and obtains training set, test set and verification collection；
Step 3, model training set pre-training parameter：
Step 3-1, initiation parameter setting is carried out to the Caffnet network structures defined in step 1, utilizes Caffenet
The layer parameter before the layer of training parameter initialization pre-training network RoI ponds under ImageNet, setting learning rate are
0.001, trained batchsize are 2, iterations I, order of magnitude 10^4, carry out network training；
Step 3-2, goal task is bicycle and electric vehicle, i.e. K=2, then adds a background classes, then input data layer class
Not Shuo num_class=2+1=3, similary RoI ponds layer data classification number num_class=3, branch of classifying to obtain in output layer
Divided data classification number num_class=3, bounding box return partial data classification number num_class=3 × 4=12；Step 4, mould
Type training, pre-training network：
Step 4-1, it will be instructed on training dataset that the Caffnet network structures defined in step 1 prepare in step 2-5
Practice, sample includes dynamic bicycle and electric vehicle, and training parameter is set by step 3-1；
Step 4-2, the output process of training a kind of type of multitask loss function to the RoI of each label and detection
Frame coordinate carries out recurrence calculating, and computational methods are：
L(p,u,tu, v) and=Lcls(p,u)+λ[u≥1]Lloc(tu,v) (1)
Lcls(p, u)=logPu (2)
Step 4-3, to hyper parameter λ=1 in step 4-2 setting formula (1), and bicycle class u=1, electric vehicle class u are marked
=2, LclsIt is type probability PuLogarithm loss, wherein PuIt is calculated with softmax, the output vector for enabling this layer is ai=
(a1,a2,......,an), according to softmax computational methods
Then
Pu=maxPi (4)
Step 4-4, another task loss L of outputloc, it is detection block loss, passes through true u classes detection block coordinate
Value v=(vx,vy,vw,vh) and u classes prediction coordinate valueIt obtains, since background classes are without former label
Frame can ignore the L of background classesloc, detection block is returned and is lost, uses loss function
Wherein, when regressive object non-boundary, traditional loss function susceptibility in training is very high, needs very carefully
Ground regularized learning algorithm rate with prevent explosion gradient, andEliminate this sensibility：
Step 5, repetitive exercise obtain target detection model：
Step 5-1, repetitive exercise process carries out backpropagation, x by RoI ponds layeri(xi∈ R) it is i-th of activation, it is defeated
Enter to RoI ponds layer, enable yrjIt is output, i.e., the jth layer of r-th RoI exports, and RoI ponds layer calculates yrj=xi*(r,j), wherein i*
(r, j)=argmaxi'∈R(i,j)xi', i.e., so that xi'Maximum i, R (i, j) is output unit yrjIn child window where maximum pondization
The indexed set of input, single xiSeveral different output y can be assigned torj, RoI ponds layer is by calculating relative to each defeated
Enter variable xiLoss function partial derivative carry out backpropagation：
Step 5-2, repetitive exercise process introduces fine tuning strategy, and the training parameter of continuous trim step 3-1 settings is so handed over
For network training is carried out, until network convergence, obtains target detection model at this time；
Step 6, model measurement, the target detection model that the test set image input step 5-2 that step 2-5 is obtained is obtained
It is tested, is broadly divided into individual test and integrated testability two parts：
Step 6-1, the sample image income single image detection test in selected test set, and shown on image pattern
Detection block, to ensure the practical application and accuracy rate of detection, setting detection type probability P range is 0.5≤P≤1 during test
Numerical value；
Step 6-2, that step 2-5 is obtained the target detection model that whole test data set input step 5-2 obtain is whole
Target detection model is verified in test；
Preferably, sample values M is 4096 in step 2-2, and bicycle b is 2262, and electric vehicle e is 1834.
Preferably, according to training in step 2-5：Test：Verification=5：2：1 ratio cut partition sample M, obtains training set
2560, verification collection 1024,512 Zhang San part of test set.Preferably, iterations I is set as 40000 times in step 3-1
Carry out network repetitive exercise.Preferably, detection probability P is set as 0.7 in step 6-1, i.e., a display target probability is more than or equal to
0.7 target.
The present invention merges EdgeBoxes algorithms and deep learning method Fast R-CNN, obtains new network mould
Type structure, and it is applied to the target detection process of non power driven vehicle, it has obtained melting based on EdgeBoxes and Fast R-CNN
The non power driven vehicle object detection method of conjunction.Compared with conventional target detection method, the present invention had both maintained deep learning calculation
The advantage of method possesses higher Detection accuracy；Meanwhile using the extraction quantity of EdgeBoxes algorithms control OP, reduce net
The complexity of network training improves the detection efficiency of target detection；And the method after fusion is applied to non power driven vehicle
Blank of the traffic monitoring in terms of non power driven vehicle has been filled up in field, is improved for road traffic supervision video and is provided new side
Method.
Target suggestion is extracted with EdgeBoxes algorithms in step 2-4, and from the aspect of calculation amount, suitable control target
It is recommended that quantity, data set manufacturing process, the sample graph in data set carries out EdgeBoxes pretreatments, obtains OP information, first
Sample image after being marked using EdgeBoxes algorithms to target carries out OP extractions, using the extraction strategy of sliding window, by with
Lower parameter carrys out command deployment as a result, α command deployment step-lengths, β control IoU threshold values, δ command deployment accuracy, minScore controls
The maximum number magnitude of minimum score and maxBoxes control extractions OP.
The training process EdgeBoxes algorithms of step 4 and Fast R-CNN algorithm fusions, and Caffnet is selected to obtain newly
Network structure, the appropriate OP that will be extracted obtains RoIs after centainly calculating, then by sample image input network extraction
Convolution characteristic pattern and RoIs are then inputted the RoI ponds layer of network structure by convolution characteristic pattern together.
Algorithm after fusion is applied to non power driven vehicle object detection field by entire model training and test process, from
Row acquisition target sample, makes data set and is trained, tests, and introduces fine tuning strategy, constantly fine tuning instruction in training process
Practice parameter, so alternately network training, until network convergence, obtains being easy to extensive non power driven vehicle target detection mould
Type.
The present invention can be with the non power driven vehicle target detection problems in solving road traffic scene, by acquiring non-motor vehicle
Sample image from control sample image target suggested quantity, changes to target image using the algorithm of deep learning
Generation training, obtains non power driven vehicle target detection model, can solve the non power driven vehicle target under Ordinary Rd traffic scene
Test problems.This method, which is based on target suggestion, can preferably control validity feature data volume, effective using deep learning algorithm
Overcome the changeable difficulty brought of non power driven vehicle appearance, and can adaptively construction feature describes under training data driving,
Accurate and quickly classification can be carried out to target signature under new samples test.And this method can be appointed by defining target
The detection model of business training plurality of target, generalization ability is strong, is more suitable for actual requirement of engineering.Proposed by the present invention is a kind of base
In the non power driven vehicle object detection method of EdgeBoxes and Fast R-CNN, this method, which can be applied to, many to be needed to carry out
The field of target detection.
Claims (5)
1. a kind of non power driven vehicle object detection method based on EdgeBoxes and Fast R-CNN, it is characterised in that：Including with
Lower step：
Step 1 defines network structure, selectes the mininet structure C affenet networks in Fast R-CNN algorithms：
Step 1-1, the setting Caffenet network cores number of plies is 11 layers, specifically includes 1 11 × 11 convolutional layer, 1 volume 5 × 5
Lamination, 33 × 3 convolutional layers, 1 area-of-interest pond layer, 2 full articulamentums, 3 pond layers；
Step 1-2, Caffenet networks output layer is multitask loss function, and specific there are two output layer, first output layers
Probability Ps of each RoI in 2 classes, i.e. target classification are calculated as a result, second output layer by the recurrence of softmax algorithms
It is responsible for calculating the coordinate value of 2 class target detection blocks, i.e. detection block coordinate result；
Step 2 defines goal task, acquires target sample figure and data set makes：
Step 2-1, it is non power driven vehicle, i.e. bicycle and electric vehicle to define goal task；
Step 2-2, data set makes, and prepares the sample image of appropriate goal task, to ensure feature adequacy, sample image number
Magnitude is set as 103-104, should in the range of any numerical value M can all be used as sample size, bicycle b and electric vehicle e allocation proportions
For b:E, and b ≈ e, M=b+e；
Step 2-3, the sample image for obtaining step 2-2 carries out VOC formattings, obtains VOC data sets；
Step 2-4, the sample graph in the VOC data sets for obtaining step 2-3 carries out target mark, then using EdgeBoxes
Algorithm carries out the sample image after mark target and suggests OP extractions, using the extraction strategy of sliding window, by following parameter Lai
Control：α command deployment step-lengths, β control IoU threshold values, δ command deployment accuracy, minScore control minimum score and
The maximum number magnitude of maxBoxes control extractions OP, sample OP information is obtained with this, carries out obtaining the sense of sample after centainly calculating
Interest region RoIs；
Step 2-5, the VOC data sets for obtaining step 2-3, according to training>Test>The ratio of verification, and each proportional it
It is divided with the division principle less than or equal to 10, obtains training set, test set and verification collection；
Step 3, model training set pre-training parameter：
Step 3-1, initiation parameter setting is carried out to the Caffnet network structures defined in step 1, is existed using Caffenet
The layer parameter before the layer of training parameter initialization pre-training network RoI ponds under ImageNet, setting learning rate is 0.001,
Trained batchsize is 2, iterations I, order of magnitude 10^4, carries out network training；
Step 3-2, goal task is bicycle and electric vehicle, i.e. K=2, then adds a background classes, then input data layer classification number
Num_class=2+1=3, similary RoI ponds layer data classification number num_class=3, score portion number of classifying in output layer
According to classification number num_class=3, bounding box returns partial data classification number num_class=3 × 4=12；
Step 4, model training, pre-training network：
Step 4-1, training, sample on the training dataset for preparing the Caffnet network structures defined in step 1 in step 2-5
This includes dynamic bicycle and electric vehicle, and training parameter is set by step 3-1；
Step 4-2, the output process of training sits the type and detection block of the RoI of each label with a kind of multitask loss function
Mark carries out recurrence calculating, and computational methods are：
L(p,u,tu, v) and=Lcls(p,u)+λ[u≥1]Lloc(tu,v) (1)
Lcls(p, u)=logPu (2)
Step 4-3, to hyper parameter λ=1 in step 4-2 setting formula (1), and bicycle class u=1, electric vehicle class u=2 are marked,
LclsIt is type probability PuLogarithm loss, wherein PuIt is calculated with softmax, the output vector for enabling this layer is ai=(a1,
a2,......,an), according to softmax computational methods
Then
Pu=maxPi (4)
Step 4-4, another task loss L of outputloc, it is detection block loss, passes through true u classes detection block coordinate value v=
(vx,vy,vw,vh) and u classes prediction coordinate valueIt obtains, it, can since background classes are without former indicia framing
To ignore the L of background classesloc, detection block is returned and is lost, uses loss function
Wherein, when regressive object non-boundary, traditional loss function susceptibility in training is very high, needs carefully to adjust very much
Whole learning rate with prevent explosion gradient, andEliminate this sensibility：
Step 5, repetitive exercise obtain target detection model：
Step 5-1, repetitive exercise process carries out backpropagation, x by RoI ponds layeri(xi∈ R) it is i-th of activation, it is input to
RoI ponds layer, enables yrjIt is output, i.e., the jth layer of r-th RoI exports, and RoI ponds layer calculatesWherein i* (r,
J)=argmaxi'∈R(i,j)xi', i.e., so that xi'Maximum i, R (i, j) is output unit yrjIt is defeated in child window where maximum pondization
The indexed set entered, single xiSeveral different output y can be assigned torj, RoI ponds layer is by calculating relative to each input
Variable xiLoss function partial derivative carry out backpropagation：
Step 5-2, repetitive exercise process introduces fine tuning strategy, the training parameter of continuous trim step 3-1 setting, so alternately into
Row network training, until network convergence, obtains target detection model at this time；
Step 6, model measurement, the target detection model that the test set image input step 5-2 that step 2-5 is obtained is obtained carry out
Test is broadly divided into individual test and integrated testability two parts：
Step 6-1, the sample image income single image detection test in selected test set, and detection is shown on image pattern
Frame, to ensure the practical application and accuracy rate of detection, setting detection type probability P range is the numerical value of 0.5≤P≤1 during test；
Step 6-2, step 2-5 is obtained into the target detection model integrated testability that whole test data set input step 5-2 obtain,
Verify target detection model.
2. a kind of non power driven vehicle object detection method based on EdgeBoxes and Fast R-CNN as described in claim 1,
It is characterized in that：Sample values M is 4096 in step 2-2, and bicycle b is 2262.
3. a kind of non power driven vehicle object detection method based on EdgeBoxes and Fast R-CNN as described in claim 1,
It is characterized in that：According to training in step 2-5：Test：Verification=5：2：1 ratio cut partition sample number M, obtains training set 2560
It opens, verification collection 1024,512 Zhang San part of test set.
4. a kind of non power driven vehicle object detection method based on EdgeBoxes and Fast R-CNN as described in claim 1,
It is characterized in that：Iterations I is set as 40000 in step 3-1.
5. a kind of non power driven vehicle object detection method based on EdgeBoxes and Fast R-CNN as described in claim 1,
It is characterized in that：Detection probability P is set as 0.7 in step 6-1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810103573.5A CN108256498A (en) | 2018-02-01 | 2018-02-01 | A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810103573.5A CN108256498A (en) | 2018-02-01 | 2018-02-01 | A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108256498A true CN108256498A (en) | 2018-07-06 |
Family
ID=62743721
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810103573.5A Pending CN108256498A (en) | 2018-02-01 | 2018-02-01 | A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108256498A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109101908A (en) * | 2018-07-27 | 2018-12-28 | 北京工业大学 | Driving procedure area-of-interest detection method and device |
| CN109239102A (en) * | 2018-08-21 | 2019-01-18 | 南京理工大学 | A kind of flexible circuit board open defect detection method based on CNN |
| WO2020083073A1 (en) * | 2018-10-23 | 2020-04-30 | 苏州科达科技股份有限公司 | Non-motorized vehicle image multi-label classification method, system, device and storage medium |
| CN111723623A (en) * | 2019-03-22 | 2020-09-29 | 杭州海康威视数字技术股份有限公司 | Method and device for detecting platform |
| CN111967452A (en) * | 2020-10-21 | 2020-11-20 | 杭州雄迈集成电路技术股份有限公司 | Target detection method, computer equipment and readable storage medium |
| CN116189098A (en) * | 2023-04-23 | 2023-05-30 | 四川弘和通讯集团有限公司 | Method and device for identifying whether engine cover of vehicle is opened or not |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718912A (en) * | 2016-01-26 | 2016-06-29 | 浙江捷尚视觉科技股份有限公司 | Vehicle characteristic object detection method based on deep learning |
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| US20170124409A1 (en) * | 2015-11-04 | 2017-05-04 | Nec Laboratories America, Inc. | Cascaded neural network with scale dependent pooling for object detection |
| CN107016390A (en) * | 2017-04-11 | 2017-08-04 | 华中科技大学 | A kind of vehicle part detection method and system based on relative position |
-
2018
- 2018-02-01 CN CN201810103573.5A patent/CN108256498A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170124409A1 (en) * | 2015-11-04 | 2017-05-04 | Nec Laboratories America, Inc. | Cascaded neural network with scale dependent pooling for object detection |
| CN105718912A (en) * | 2016-01-26 | 2016-06-29 | 浙江捷尚视觉科技股份有限公司 | Vehicle characteristic object detection method based on deep learning |
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN107016390A (en) * | 2017-04-11 | 2017-08-04 | 华中科技大学 | A kind of vehicle part detection method and system based on relative position |
Non-Patent Citations (5)
| Title |
|---|
| C.LAWRENCE ZITNICK 等: "Edge boxes:Locating object proposals from edges", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
| ROSS GIRSHICK 等: "Fast R-CNN Object Detection with Caffe", 《TUTORIAL.CAFFE.BERKELEYVISION.ORG》 * |
| ROSS GIRSHICK 等: "Fast R-CNN", 《2015 IEEE INTERNATIONAL CONFERENCE ON CONFERENCE ON COMPUTER VISION》 * |
| ZHONG-QIU ZHAO 等: "PEDESTRIAN DETECTION BASED ON FAST R-CNN AND BATCH NORMALIZATION", 《INTERNATIONAL CONFERENCE ON INTELLIGENT COMPUTING》 * |
| 曹诗雨 等: "基于Fast R-CNN的车辆目标检测", 《中国图象图形学报》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109101908A (en) * | 2018-07-27 | 2018-12-28 | 北京工业大学 | Driving procedure area-of-interest detection method and device |
| CN109101908B (en) * | 2018-07-27 | 2020-12-25 | 北京工业大学 | Method and device for detecting region of interest in driving process |
| CN109239102A (en) * | 2018-08-21 | 2019-01-18 | 南京理工大学 | A kind of flexible circuit board open defect detection method based on CNN |
| CN109239102B (en) * | 2018-08-21 | 2021-04-09 | 南京理工大学 | CNN-based flexible circuit board appearance defect detection method |
| WO2020083073A1 (en) * | 2018-10-23 | 2020-04-30 | 苏州科达科技股份有限公司 | Non-motorized vehicle image multi-label classification method, system, device and storage medium |
| CN111723623A (en) * | 2019-03-22 | 2020-09-29 | 杭州海康威视数字技术股份有限公司 | Method and device for detecting platform |
| CN111967452A (en) * | 2020-10-21 | 2020-11-20 | 杭州雄迈集成电路技术股份有限公司 | Target detection method, computer equipment and readable storage medium |
| CN111967452B (en) * | 2020-10-21 | 2021-02-02 | 杭州雄迈集成电路技术股份有限公司 | Target detection method, computer equipment and readable storage medium |
| CN116189098A (en) * | 2023-04-23 | 2023-05-30 | 四川弘和通讯集团有限公司 | Method and device for identifying whether engine cover of vehicle is opened or not |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108256498A (en) | A kind of non power driven vehicle object detection method based on EdgeBoxes and FastR-CNN | |
| CN106650806B (en) | A kind of cooperating type depth net model methodology for pedestrian detection | |
| CN107564025B (en) | Electric power equipment infrared image semantic segmentation method based on deep neural network | |
| CN108615010B (en) | Facial expression recognition method based on parallel convolution neural network feature map fusion | |
| CN107368787A (en) | A kind of Traffic Sign Recognition algorithm that application is driven towards depth intelligence | |
| CN107657279B (en) | Remote sensing target detection method based on small amount of samples | |
| CN104537393B (en) | A kind of traffic sign recognition method based on multiresolution convolutional neural networks | |
| CN108509978A (en) | The multi-class targets detection method and model of multi-stage characteristics fusion based on CNN | |
| CN108009509A (en) | Vehicle target detection method | |
| CN106971194A (en) | A kind of driving intention recognition methods based on the double-deck algorithms of improvement HMM and SVM | |
| CN108171112A (en) | Vehicle identification and tracking based on convolutional neural networks | |
| CN105825502B (en) | A kind of Weakly supervised method for analyzing image of the dictionary study based on conspicuousness guidance | |
| CN105512684A (en) | Vehicle logo automatic identification method based on principal component analysis convolutional neural network | |
| CN104657717B (en) | A kind of pedestrian detection method based on layering nuclear sparse expression | |
| CN103984948B (en) | A kind of soft double-deck age estimation method based on facial image fusion feature | |
| CN101964063B (en) | Method for constructing improved AdaBoost classifier | |
| CN105335716A (en) | Improved UDN joint-feature extraction-based pedestrian detection method | |
| CN108446662A (en) | A kind of pedestrian detection method based on semantic segmentation information | |
| CN109977980A (en) | A kind of method for recognizing verification code and device | |
| CN109034224A (en) | Hyperspectral classification method based on double branching networks | |
| CN104992165A (en) | Extreme learning machine based traffic sign recognition method | |
| CN110009648A (en) | Trackside image Method of Vehicle Segmentation based on depth Fusion Features convolutional neural networks | |
| CN112487862A (en) | Garage pedestrian detection method based on improved EfficientDet model | |
| CN107194938A (en) | Image outline detection method based on depth convolutional neural networks | |
| CN106919902A (en) | A kind of vehicle identification and trajectory track method based on CNN |
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: 20180706 |