CN107463927A - A kind of deceleration driven detection method and device based on convolutional neural networks - Google Patents
A kind of deceleration driven detection method and device based on convolutional neural networks Download PDFInfo
- Publication number
- CN107463927A CN107463927A CN201710860765.6A CN201710860765A CN107463927A CN 107463927 A CN107463927 A CN 107463927A CN 201710860765 A CN201710860765 A CN 201710860765A CN 107463927 A CN107463927 A CN 107463927A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- layer
- road surface
- target
- 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
- G06V20/56—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
-
- E—FIXED CONSTRUCTIONS
- E01—CONSTRUCTION OF ROADS, RAILWAYS, OR BRIDGES
- E01F—ADDITIONAL WORK, SUCH AS EQUIPPING ROADS OR THE CONSTRUCTION OF PLATFORMS, HELICOPTER LANDING STAGES, SIGNS, SNOW FENCES, OR THE LIKE
- E01F9/00—Arrangement of road signs or traffic signals; Arrangements for enforcing caution
- E01F9/50—Road surface markings; Kerbs or road edgings, specially adapted for alerting road users
- E01F9/529—Road surface markings; Kerbs or road edgings, specially adapted for alerting road users specially adapted for signalling by sound or vibrations, e.g. rumble strips; specially adapted for enforcing reduced speed, e.g. speed bumps
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Structural Engineering (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Civil Engineering (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Architecture (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of deceleration driven detection method based on convolutional neural networks, obtain target information of road surface, according to constructing in advance and be trained the target convolutional neural networks of acquisition, detection target information of road surface simultaneously obtains object detection results, according to object detection results, determine whether include deceleration driven in target information of road surface.The technical scheme provided using the embodiment of the present invention, target convolutional neural networks are obtained by constructing and being trained so that after target convolutional neural networks detect to target information of road surface, determine whether include deceleration driven in target information of road surface.Because the extraction of target convolution neural network characteristics is strong, the ability of statistical learning is strong, improves the verification and measurement ratio to deceleration driven.The invention also discloses a kind of deceleration driven detection means based on convolutional neural networks, has relevant art effect.
Description
Technical field
The present invention relates to detection technique field, is detected more particularly to a kind of deceleration driven based on convolutional neural networks
Method and device.
Background technology
In vehicle traveling process, deceleration driven can be detected, so as to steady current in driver's driving conditions.
The detection of deceleration strip has become the pith of the environment sensing of intelligent automobile.
In research at home and abroad, the research to deceleration strip detection is especially few, and existing scheme presence is more obvious
The defects of.In the prior art, the surface of road is assessed based on radar and laser, determines the fluctuating situation of road, work as road
When the changing pattern of face difference in height meets standard curve, determine that vehicle has deceleration strip in front of advancing.However, due in the program
Standard curve be the deceleration strip pattern based on locality and propose, therefore the program only identification to local specific deceleration strip
Rate is higher.Also, with road safety and the requirement of traffic control, deceleration strip is on the increase, and pattern is also more and more, the program
When being detected to the deceleration strip in varying environment, correct recognition rata is relatively low, can not meet demand.It should be noted that this Shen
Please file the correct recognition rata for deceleration driven is referred to as verification and measurement ratio.
In addition, the program is also higher to the false drop rate of similar object, and deceleration strip is diurnally carried out under normal condition
Detection, do not consider the detection of night and the deceleration strip under Changes in weather.
In summary, the verification and measurement ratio for deceleration driven how is effectively improved, is that current those skilled in the art are anxious
The technical problem that need to be solved.
The content of the invention
It is an object of the invention to provide a kind of deceleration driven detection method and device based on convolutional neural networks, to carry
The high verification and measurement ratio to deceleration driven.
In order to solve the above technical problems, the present invention provides following technical scheme, this method includes：
Obtain target information of road surface；
According to constructing in advance and being trained the target convolutional neural networks of acquisition, detect the target information of road surface and obtain
To object detection results；
According to the object detection results, determine whether include deceleration driven in the target information of road surface.
Preferably, the target convolutional neural networks are obtained by following steps：
Construction exports the initial convolutional neural networks of the label information of deceleration driven for inputting information；
Determine the training parameter of the initial convolutional neural networks；
According to the training parameter, using the training sample that training sample is concentrated as input to the initially convolution nerve net
Network is trained；
The weights of the initial convolutional neural networks are updated in the training process, obtain the target convolutional neural networks；
Accordingly, it is described to detect the target information of road surface and obtain object detection results, including：
Detect the target information of road surface and obtain the object detection results of the label information comprising deceleration driven.
Preferably, initial each layer of convolutional neural networks is followed successively by：Input layer, the first convolutional layer, the first amendment are linear
Layer, the first pond layer, the second convolutional layer, second amendment linear layer, the second pond layer, the 3rd convolutional layer, the 3rd correct linear layer,
Volume Four lamination, the 4th amendment linear layer, the first full articulamentum, the 5th amendment linear layer, the second full articulamentum, softmax layers
And classification layer.
Preferably, the training parameter includes：Criticize sample size, rounds and learning rate.
Preferably, it is characterised in that obtain the training sample set by following steps：
Gather information of road surface and obtain original training set；
The original training set is expanded to obtain exptended sample collection；
Exptended sample collection progress dimension normalization is handled to obtain the training sample set.
Preferably, it is described to be expanded the original training set to obtain exptended sample collection, including：
One in the original training set or multiple samples are subjected to flip horizontal, obtain exptended sample collection.
Preferably, it is described to be expanded the original training set to obtain exptended sample collection, including：
One in the original training set or multiple samples are subjected to color adjustment, obtain exptended sample collection.
Preferably, it is described to be expanded the original training set to obtain exptended sample collection, including：
One in the original training set or multiple samples are subjected to noise addition, obtain exptended sample collection.
Preferably, when it is determined that including deceleration driven in the target information of road surface, in addition to：
Export prompt message.
A kind of deceleration driven detection means based on convolutional neural networks, the device include：
Target information of road surface obtains module, for obtaining target information of road surface；
Object detection results obtain module, are constructed in advance for basis and are trained the target convolution nerve net of acquisition
Network, detect the target information of road surface and obtain object detection results；
Deceleration driven determining module, it is in the target information of road surface for according to the object detection results, determining
It is no to include deceleration driven.
The technical scheme provided using the embodiment of the present invention, target information of road surface is obtained, according to constructing and carry out in advance
The target convolutional neural networks obtained are trained, target information of road surface is detected and obtains object detection results, examined afterwards according to target
Result is surveyed, determines whether include deceleration driven in target information of road surface.
By the target convolutional neural networks for constructing and being trained acquisition so that target convolutional neural networks are to target road
After face information is detected, determine whether include deceleration driven in target information of road surface.Compared to prior art, mesh is utilized
The method that convolutional neural networks carry out deceleration driven detection is marked, because the extraction of target convolution neural network characteristics is strong, statistics
The ability of habit is strong, higher to the verification and measurement ratio of deceleration driven.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
There is the required accompanying drawing used in technology description to be briefly described, it should be apparent that, drawings in the following description are only this
Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, can be with
Other accompanying drawings are obtained according to these accompanying drawings.
Fig. 1 is a kind of implementing procedure figure of the deceleration driven detection method based on convolutional neural networks in the present invention；
Fig. 2 is the contrast schematic diagram of the verification and measurement ratio under different learning rates in the present invention；
Fig. 3 is the schematic diagram that the verification and measurement ratio in the present invention under different batches of sample sizes changes with rounds；
Fig. 4 is the visualization feature figure of target convolutional neural networks in a kind of embodiment of the present invention；
Fig. 5 is the schematic diagram of part sample image in original training set of the present invention；
Fig. 6 is by the schematic diagram of the image after initial sample progress flip horizontal in the present invention；
Fig. 7 is the schematic diagram that initial sample is carried out to the image after HSV color adjustment in the present invention；
Fig. 8 is the schematic diagram that initial sample is carried out to the image after noise addition in the present invention；
Fig. 9 is a kind of structural representation of the deceleration driven detection means based on convolutional neural networks in the present invention.
Embodiment
The core of the present invention is to provide a kind of deceleration driven detection method based on convolutional neural networks, by constructing simultaneously
It is trained and obtains target convolutional neural networks, the detection of deceleration driven is carried out using target convolutional neural networks, is improved
For the verification and measurement ratio of deceleration driven.
In order that those skilled in the art more fully understand the present invention program, with reference to the accompanying drawings and detailed description
The present invention is described in further detail.Obviously, described embodiment is only part of the embodiment of the present invention, rather than
Whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art are not making creative work premise
Lower obtained every other embodiment, belongs to the scope of protection of the invention.
Fig. 1 is refer to, for a kind of implementation stream of the deceleration driven detection method based on convolutional neural networks in the present invention
Cheng Tu, this method may comprise steps of：
S101：Obtain target information of road surface.
Target information of road surface can be obtained by image capture device, and target information of road surface can be pictorial information.For example, can
Shot with travelling the road surface in front to vehicle using vehicle-mounted camera.Target information of road surface can be periodically received,
The video information of target information of road surface can be included with real-time reception target information of road surface, such as real-time reception.Certainly, above scheme
Selection, have no effect on the present invention implementation.
After target information of road surface is obtained, step S102 operation can be carried out.
S102：According to constructing in advance and being trained the target convolutional neural networks of acquisition, detection target information of road surface is simultaneously
Obtain object detection results.
It is low to the verification and measurement ratio of deceleration driven due in the prior art, and false drop rate height also be present, it can only be examined on daytime
The problems such as survey, inventor consider the powerful ability in feature extraction of neutral net and statistical learning ability, utilize convolutional Neural
Network detects to deceleration driven, to improve the detectability to deceleration driven.
After target convolutional neural networks are constructed, target convolutional neural networks are trained, according to target convolution god
Through network, detect target information of road surface and obtain object detection results.
In a kind of embodiment of the present invention, target convolutional neural networks are obtained by following steps：
Step 1：Construction exports the initial convolutional neural networks of the label information of deceleration driven for inputting information；
Step 2：It is determined that the training parameter of initial convolutional neural networks；
Step 3：According to training parameter, using the training sample that training sample is concentrated as inputting to initial convolution nerve net
Network is trained；
Step 4：The weights of initial convolutional neural networks are updated in the training process, obtain target convolutional neural networks.
For the ease of description, aforementioned four step is merged into explanation.
Initial convolutional neural networks can be constructed according to the instruction or operation of user, the initial convolutional neural networks pin
The label information of deceleration driven, to the input information of the initial convolutional neural networks, can be exported to inputting.
In a kind of embodiment of the present invention, initial convolutional neural networks are 16 layers of convolutional neural networks,
Each layer is followed successively by：Input layer, the first convolutional layer, the first amendment linear layer, the first pond layer, the second convolutional layer, the second modified line
Property layer, the second pond layer, the 3rd convolutional layer, the 3rd amendment linear layer, Volume Four lamination, the 4th amendment linear layer, first connect entirely
Connect layer, the 5th amendment linear layer, the second full articulamentum, softmax layers and classification layer.
In the above-mentioned initial convolutional neural networks of the present invention are 16 layers of embodiment.
Specifically, the first layer of the initial convolutional neural networks is input layer, the input of input layer can be sized to
227 × 227 × 3 so that input layer can receive the image for the target information of road surface that size is 227 × 227.Target information of road surface
It can be RGB image.
The second layer of the initial convolutional neural networks is the first convolutional layer, can be by the convolution mask size of the first convolutional layer
11 × 11 are set to, port number can be set to 30, to carry out convolution to the image inputted to the second layer using 30 convolution masks.
Step-length in convolution process, which is set, can be set to 4 × 4.Using above-mentioned set-up mode, the output of the first convolutional layer for 30 55 ×
55 characteristic image, the input as third layer.
Third layer is the first amendment linear layer, and the amendment linear function that the first amendment linear layer is set can be to the second layer
Output valve is adjusted so that negative is mapped as 0, and nonnegative number keeps constant.The setting of first amendment linear layer can avoid
In the training process to initial convolutional neural networks, right value update is slow caused by the gradient disappearance of backpropagation, study
The problem of ability reduces, the first amendment linear layer can't change input to the input size of this layer.
The 4th layer be connected with the first linear modification level is the first pond layer, and the method to set up of the first pond layer can use
Maximum pond method.Specifically, pond region can be sized into 3 × 3, convolution step-length is set to 2 × 2, to unit Chi Hua areas
Pixel in domain takes output of the maximum as the first pond layer, and the output of the first pond layer is the spy of 30 27 × 27 sizes
Levy image.It should be noted that the present invention can also use other pond methods, the implementation of the present invention is had no effect on.
Layer 5 is the second convolutional layer, the characteristic image that the characteristic image of layer 5 output inputs from the 4th layer to layer 5
Collective effect obtains.Specifically, the pixel of each characteristic image of layer 5 output is the 4th layer of institute inputted to layer 5
There is sum of the characteristic image in the same area convolution results.The convolution mask size of layer 5 can be set to 5 × 5, and port number can be with
40 are set to, step-length can be set to 1.Using above-mentioned set-up mode, the output of layer 5 is the characteristic pattern that 40 sizes are 23 × 23
Picture.
Layer 6 is the second amendment linear layer, and the parameter that the parameter setting of layer 6 is referred to third layer is configured,
Such as it can use and be configured with the identical parameter of third layer.
Layer 7 is the second pond layer, and the parameter that the parameter setting of layer 7 is referred to the 4th layer is configured.
8th layer is the 3rd convolutional layer, and the 8th layer of convolution mask size can be set to 3 × 3, and port number can be set to 30,
Step-length can be set to 1.
9th layer is the 3rd amendment linear layer, and the parameter that parameter setting is referred to third layer is configured.Tenth layer is
Volume Four lamination.
The parameter that tenth layer of parameter setting is referred to the 8th layer is configured.
Eleventh floor is the 4th amendment linear layer, and the parameter that the parameter setting of eleventh floor is referred to third layer is set
Put.
Floor 12 is the first full articulamentum, and the neuron of the first full articulamentum can be set to 60, using above-mentioned setting
Mode, the output of Floor 12 are the characteristic vector that dimension is 60.
The 13rd layer be connected with Floor 12 is the 5th amendment linear layer, and the 13rd layer of parameter setting is referred to the
Three layers of parameter is configured.
13rd layer is connected entirely with the 14th layer, and the 14th layer is the second full articulamentum, the 14th layer of neuron
2 can be set to.
15th layer is softmax layers, and the neuron of softmax layers can be set to 2.For input to initial nerve net
The input information of network, the 15th layer can export the probabilistic information that deceleration driven is included in testing result.
16th layer is classification layer, and classification layer can be according to the probabilistic information of the 15th layer of output, it is determined that for input letter
The label information of the deceleration driven of breath.
Certainly, in above-mentioned initial convolutional neural networks are 16 layers of embodiment, target information of road surface passes through
After the detection of target nerve network, the object detection results of the label information comprising deceleration driven are obtained.Further according to target
Label information in testing result, determine whether include deceleration driven in target information of road surface.For example, the layer that will can classify is defeated
The label information gone out is set to+1 and -1, after target information of road surface is inputted to target convolutional neural networks, when classification layer output
Label information be+1 when, determine include deceleration driven in target information of road surface, when classification layer output label information be -1
When, determine not including deceleration driven in target information of road surface.Certainly, in other embodiments of the present invention, may be used also
With use other tagged manners, and can further in object detection results label information, determine road speed
The type of band, have no effect on the implementation of the present invention.For example, label information is arranged to+1 ,+2 and+3, when the mark of classification layer output
When label information is+1, determine not including deceleration driven in target information of road surface, when the label information for layer output of classifying is+2,
Determine to include deceleration driven in target information of road surface, and can also further determine that the type of deceleration driven is the first kind
Type.When the label information for layer output of classifying is+3, determine to include deceleration driven in target information of road surface, and can also enter
One step determines that the type of deceleration driven is Second Type.
It should be noted that in a kind of embodiment of the present invention, initial convolutional neural networks can be 15
The convolutional neural networks of layer, the setting of 15 layers of each layer of initial convolutional neural networks are referred to 16 layers of initial convolution god
One to 15 layer through network is configured, i.e., last layer is softmax layers.In this kind of embodiment of the present invention
In, target convolutional neural networks are obtained constructing 15 layers of initial convolutional neural networks and being trained, detection target road surface
After information acquisition object detection results, carried in obtained object detection results and road speed is included in target information of road surface
The probabilistic information of band.After object detection results are obtained, target can be determined according to the probabilistic information in object detection results
Whether deceleration driven is included in information of road surface.For example, when the probabilistic information in object detection results is more than 50%, mesh is determined
Deceleration driven is included in mark information of road surface.
After initial convolutional neural networks are constructed, it may be determined that the training parameter of initial convolutional neural networks.It can control
Variable processed determines training parameter.Fig. 2 can be referred to, coordinate system C1 abscissa represents the rounds of detection, and ordinate represents
Verification and measurement ratio, when crowd sample size mini-batch is set to 40, verification and measurement ratio when learning rate is 0.01 is higher than learning rate
Verification and measurement ratio when 0.05, that is to say, that Detection results when learning rate is 0.01 are better than the Detection results that learning rate is 0.05.
Can be so that as shown in fig.3, coordinate system C2 abscissa represents the rounds detected, ordinate represents verification and measurement ratio, when learning rate is set
For 0.01 when, the verification and measurement ratio that crowd sample size mini-batch is set to 45 is set to 40 better than crowd sample size mini-batch
Verification and measurement ratio, and can as seen from Figure 3, rounds take verification and measurement ratio when 12 to be higher than same batch sample size mini-batch
Other rounds verification and measurement ratio., can be by batch sample number in training parameter in a kind of embodiment of the present invention
Amount mini-batch is set to 45, and rounds are set to 12, and learning rate is set to 0.01.
The training parameter of initial convolutional neural networks can be determined by control variate method, according to training parameter, will be trained
Training sample in sample set is trained as input to initial convolutional neural networks, updates initial convolution in the training process
The weights of neutral net, and obtain target convolutional neural networks.For example, from training sample the sample for choosing 80% can be concentrated to make
For training sample, initial convolutional neural networks are trained using training sample.It should be noted that in one kind of the present invention
In embodiment, the initial weight of initial convolutional neural networks can be 0 in average, and standard deviation is 0.01 Gaussian Profile
Middle generation, initial biasing can be set to 0.
In a kind of embodiment of the present invention, training sample set is 4207 RGB images, wherein, wrapped in image
It is 3415 containing deceleration driven, is 792 not comprising deceleration driven.Can using training sample concentrate 80% as
Training sample is trained to initial convolutional neural networks.The weights of initial convolutional neural networks are updated in the training process, are obtained
To after target convolutional neural networks, the visualization feature figure of last layer of full articulamentum of target nerve network can be exported,
Such as can be the visualization view of the 14th layer of second full articulamentum in a kind of above-mentioned embodiment.It see Fig. 4,
Wherein Fig. 4 (a) is the visualization feature figure of deceleration region, and Fig. 4 (b) is the visualization feature figure of non-deceleration strip, it may be determined that
Target convolutional neural networks have extracted the yellow cord feature that deceleration strip is different from non-deceleration strip.Certainly, target volume is being obtained
, can be with the verification and measurement ratio to deceleration driven of test target convolutional neural networks after product neutral net.
In a kind of embodiment of the present invention, using the 80% of training sample concentration as training sample, it will train
Remaining 20% in sample set is used as test sample, and the verification and measurement ratio for obtaining the deceleration driven of test sample is 95.44, to surveying
The average detected time of sample sheet is 1.26ms.
In a kind of embodiment of the present invention, using the 70% of training sample concentration as training sample, mesh is obtained
After marking convolutional neural networks, the residue 30% concentrated to training sample detects, and can be repeated 3 times, each time from training
The detection sample determined in sample set is different, and the average detected rate for obtaining 3 detections is 92.63%, the average detected time
For 1.27ms, illustrate to use the solution of the present invention, good verification and measurement ratio is carried to road speed.
S103：According to object detection results, determine whether include deceleration driven in target information of road surface.
After object detection results are obtained, it can determine whether wrapped in target information of road surface according to object detection results
Containing deceleration driven.For example, in a kind of embodiment of the present invention, when the label information in object detection results is+1
When, determine to include deceleration driven in target information of road surface.
Certainly, after it is determined that including deceleration driven in target information of road surface, prompt message can be exported.For example, can
To send information of voice prompt to driver or send prompt message by the flicker of indicator lamp, so that automobile balance passes through
Road deceleration strip.
The technical scheme provided using the embodiment of the present invention, target information of road surface is obtained, according to constructing and carry out in advance
The target convolutional neural networks obtained are trained, target information of road surface is detected and obtains object detection results, examined afterwards according to target
Result is surveyed, determines whether include deceleration driven in target information of road surface.
By the target convolutional neural networks for constructing and being trained acquisition so that target convolutional neural networks are to target road
After face information is detected, determine whether include deceleration driven in target information of road surface.Compared to prior art, mesh is utilized
The method that convolutional neural networks carry out deceleration driven detection is marked, because the extraction of target convolution neural network characteristics is strong, statistics
The ability of habit is strong, higher to the verification and measurement ratio of deceleration driven.
In a kind of embodiment of the present invention, training sample set is obtained by following steps：
First step：Gather information of road surface and obtain original training set；
Second step：Original training set is expanded to obtain exptended sample collection；
3rd step：Exptended sample collection progress dimension normalization is handled to obtain training sample set.
Can take pictures or video by way of gather information of road surface obtain original training set, see shown in Fig. 5, be
The exemplary plot of the initial sample image in part in the present invention.After original training set is obtained, original training set can be expanded
Fill to obtain exptended sample collection, exptended sample collection progress dimension normalization is handled to obtain training sample set.
In a kind of embodiment of the present invention, one in original training set can be chosen or multiple samples enter
Row flip horizontal and scale processing, obtain exptended sample collection.Specifically, the symmetrical pixel in image can be carried out
Exchange, expanded original training set using flip horizontal, for same deceleration driven in being lived with simulation, vehicle is from two
Individual direction passes through the situation of the deceleration strip respectively.Fig. 6 is see, for initial sample to be carried out to the figure after flip horizontal in the present invention
The schematic diagram of picture.For example, the size of the sample image P in original training set is (M, N), by what is obtained after image P flip horizontals
Image is Pe, obtain the image P after flip horizontaleMathematic(al) representation be：
Pe(:, m)=P (:,N-1-m)；
Wherein, Pe(:, m) represent flip horizontal image PeM row pixels, the value be equal to sample image P N-1-m
The pixel value of row.
Because the input layer of target convolutional neural networks is limited by size the RGB image of input, exptended sample is being obtained
After collection, exptended sample collection progress dimension normalization is handled to obtain training sample set.Specifically, cubic interpolation method can be used
Carry out dimension normalization processing.For example, the pixel for carrying out the sample image of dimension normalization processing is (xi,yj), the pixel
Corresponding pixel value is f (xi,yj), the pixel of the size of sample image 4 × 4 is weighted summation, yardstick normalizing can be obtained
One pixel value f'(x, y of the image after change), interpolation formula is：
Wherein, W (x-xi)、W(y-yj) can be obtained by bicubic function BiCubic function W (x), W (x) can be represented
For：
Wherein, parameter a is the coefficient of bicubic function, can generally be set to -0.5 or -0.75.
In a kind of embodiment of the present invention, one in original training set can be chosen or multiple samples enter
Row color adjusts, and obtains exptended sample collection.
Color adjustment can be that HSV colors adjust, and the RGB image that video camera is shot can be converted into HSV images, passed through
The H (Hue, form and aspect) of HSV images is adjusted, S (Saturation, saturation degree), V (Value, colourity), different visions can be obtained
The image of performance, to simulate the image of the different color of video camera shooting and different light levels under varying environment.For example, choosing
A RGB image in original training set is taken to carry out HSV color adjustment, R, G, B maximum during max can be set as the RGB image
Value, min are the minimum value of R, G, B in the RGB image, and the RGB image is converted to the mathematic(al) representation of HSV images to be：
V=max；
By taking if max=min as an example, represent when max is equal to min, form and aspect H value is 0 °.By taking if max=0 as an example, table
When showing that max is equal to 0, saturation degree S is equal to 0.After the HSV images of the RGB image are obtained, in a kind of specific implementation of the present invention
In mode, saturation degree S and colourity V can be adjusted, to change the bright-dark degree of the RGB image and imaged color, method of adjustment
Equation below can be used：
S=Sj；V=VkJ, k ∈ (0.4,2.5)；
Wherein, j and k is to obey equally distributed stochastic variable respectively.In adjustment HSV images, multiple different HSV are obtained
After image, the HSV images after adjustment can be converted back RGB image, see shown in Fig. 7.HSV images are converted back into RGB
Image, can be by substituting into following mathematic(al) representations by H, S and V of HSV images：
P=V × (1-S)；
Q=V × (1-f × S)；
T=V × (1- (1-f) × S)；
In a kind of embodiment of the present invention, one in original training set can be chosen or multiple samples enter
Row noise adds, and obtains exptended sample collection.
For example, noise addition can be carried out to a RGB image in original training set, each pixel can be added
Enter a random quantity for resulting from Gaussian Profile, the different degrees of noise being subject in shooting process with analog video camera is done
Disturb, see Fig. 8.The random quantity for resulting from Gaussian Profile is z, and the Gaussian distribution formula used can be for：
Wherein, Gaussian Profile mean μ can be set to 0, and standard deviation sigma can be set to 0.1.
Certainly, the sample image in original training set can be divided into three parts by the present invention, and Part I carries out sample graph
The flip horizontal of picture, Part II carry out the color adjustment of sample image, and Part III carries out the noise addition of sample image, this
Invention can also carry out expansion processing to a part of sample image in original training set, and another part does not do expansion processing,
The implementation of the present invention is not influenceed.
The present invention a kind of embodiment in, initial sample covered it is a variety of in the case of deceleration driven
Image, such as fine day and rainy day, the image of the deceleration driven at daytime and night is right in this kind of embodiment
Initial sample is expanded, including the HSV colors adjustment of the flip horizontal of sample image, sample image and making an uproar for sample image
Sound adds so that target convolutional neural networks can detect to the road conditions of complexity, improve target convolutional neural networks
Predictive ability and generalization ability.
Corresponding to above method embodiment, the embodiment of the present invention additionally provides a kind of road based on convolutional neural networks
Deceleration strip detection means, a kind of deceleration driven detection means based on convolutional neural networks described below with it is above-described
A kind of deceleration driven detection method based on convolutional neural networks can be mutually to should refer to.
It is shown in Figure 9, for a kind of structure of the deceleration driven detection means based on convolutional neural networks in the present invention
Schematic diagram, the device can include with lower module：
Target information of road surface obtains module 901, for obtaining target information of road surface；
Object detection results obtain module 902, are constructed in advance for basis and are trained the target convolutional Neural of acquisition
Network, detect target information of road surface and obtain object detection results；
Deceleration driven determining module 903, for according to object detection results, determining whether included in target information of road surface
Deceleration driven.
The device provided using the embodiment of the present invention, target information of road surface is obtained, according to constructing and be trained in advance
The target convolutional neural networks of acquisition, detect target information of road surface and obtain object detection results, afterwards according to target detection knot
Fruit, determine whether include deceleration driven in target information of road surface.
By the target convolutional neural networks for constructing and being trained acquisition so that target convolutional neural networks are to target road
After face information is detected, determine whether include deceleration driven in target information of road surface.Compared to prior art, mesh is utilized
The method that convolutional neural networks carry out deceleration driven detection is marked, because the extraction of target convolution neural network characteristics is strong, statistics
The ability of habit is strong, higher to the verification and measurement ratio of deceleration driven.
In a kind of embodiment of the present invention, in addition to target convolutional neural networks obtain module, are used for：
Construction exports the initial convolutional neural networks of the label information of deceleration driven for inputting information；
It is determined that the training parameter of initial convolutional neural networks；
According to training parameter, initial convolutional neural networks are instructed using the training sample that training sample is concentrated as input
Practice；
The weights of initial convolutional neural networks are updated in the training process, obtain target convolutional neural networks.
Object detection results obtain module 902, specifically for detection target information of road surface and obtain including deceleration driven
Label information object detection results.
In a kind of embodiment of the present invention, initial each layer of convolutional neural networks is followed successively by：Input layer, the first volume
Lamination, the first amendment linear layer, the first pond layer, the second convolutional layer, the second amendment linear layer, the second pond layer, the 3rd convolution
Layer, the 3rd amendment linear layer, Volume Four lamination, the 4th amendment linear layer, the first full articulamentum, the 5th amendment linear layer, second
Full articulamentum, softmax layers and classification layer.
In a kind of embodiment of the present invention, training parameter includes：Criticize sample size, rounds and study
Rate.
In a kind of embodiment of the present invention, in addition to：
Original training set obtains module, and original training set is obtained for gathering information of road surface；
Exptended sample collection obtains module, for being expanded original training set to obtain exptended sample collection；
Training sample set obtains module, for handling to obtain training sample set exptended sample collection progress dimension normalization.
In a kind of embodiment of the present invention, exptended sample collection obtains module, specifically for by original training set
In one or multiple samples carry out flip horizontal, obtain exptended sample collection.
In a kind of embodiment of the present invention, exptended sample collection obtains module, specifically for regard to original training set
In one or multiple samples carry out color adjustment, obtain exptended sample collection.
In a kind of embodiment of the present invention, exptended sample collection obtains module, specifically for by original training set
In one or multiple samples carry out noise addition, obtain exptended sample collection.
In a kind of embodiment of the present invention, in addition to：
Prompt message output module, for when it is determined that including deceleration driven in target information of road surface, output prompting to be believed
Breath.
Each embodiment is described by the way of progressive in this specification, what each embodiment stressed be with it is other
The difference of embodiment, between each embodiment same or similar part mutually referring to.For dress disclosed in embodiment
For putting, because it is corresponded to the method disclosed in Example, so description is fairly simple, related part is referring to method part
Explanation.
Professional further appreciates that, with reference to the unit of each example of the embodiments described herein description
And algorithm steps, can be realized with electronic hardware, computer software or the combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software, the composition and step of each example are generally described according to function in the above description.These
Function is performed with hardware or software mode actually, application-specific and design constraint depending on technical scheme.Specialty
Technical staff can realize described function using distinct methods to each specific application, but this realization should not
Think beyond the scope of this invention.
Directly it can be held with reference to the step of method or algorithm that the embodiments described herein describes with hardware, processor
Capable software module, or the two combination are implemented.Software module can be placed in random access memory (RAM), internal memory, read-only deposit
Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology
In any other form of storage medium well known in field.
Specific case used herein is set forth to the principle and embodiment of the present invention, and above example is said
It is bright to be only intended to help and understand technical scheme and its core concept.It should be pointed out that for the common of the art
For technical staff, under the premise without departing from the principles of the invention, some improvement and modification can also be carried out to the present invention, these
Improve and modification is also fallen into the protection domain of the claims in the present invention.
Claims (10)
- A kind of 1. deceleration driven detection method based on convolutional neural networks, it is characterised in that including：Obtain target information of road surface；According to constructing in advance and being trained the target convolutional neural networks of acquisition, detect the target information of road surface and obtain mesh Mark testing result；According to the object detection results, determine whether include deceleration driven in the target information of road surface.
- 2. according to the method for claim 1, it is characterised in that obtain the target convolution nerve net by following steps Network：Construction exports the initial convolutional neural networks of the label information of deceleration driven for inputting information；Determine the training parameter of the initial convolutional neural networks；According to the training parameter, the initial convolutional neural networks are entered using the training sample that training sample is concentrated as input Row training；The weights of the initial convolutional neural networks are updated in the training process, obtain the target convolutional neural networks；Accordingly, it is described to detect the target information of road surface and obtain object detection results, including：Detect the target information of road surface and obtain the object detection results of the label information comprising deceleration driven.
- 3. according to the method for claim 2, it is characterised in that initial each layer of convolutional neural networks is followed successively by：Input Layer, the first convolutional layer, first amendment linear layer, the first pond layer, the second convolutional layer, second correct linear layer, the second pond layer, 3rd convolutional layer, the 3rd amendment linear layer, Volume Four lamination, the 4th amendment linear layer, the first full articulamentum, the 5th amendment are linear Layer, the second full articulamentum, softmax layers and classification layer.
- 4. according to the method for claim 2, it is characterised in that the training parameter includes：Batch sample size, rounds with And learning rate.
- 5. according to the method described in any one of claim 2 to 4, it is characterised in that obtain the training sample by following steps This collection：Gather information of road surface and obtain original training set；The original training set is expanded to obtain exptended sample collection；Exptended sample collection progress dimension normalization is handled to obtain the training sample set.
- 6. according to the method for claim 5, it is characterised in that described to be expanded the original training set Sample set, including：One in the original training set or multiple samples are subjected to flip horizontal, obtain exptended sample collection.
- 7. according to the method for claim 5, it is characterised in that described to be expanded the original training set Sample set, including：One in the original training set or multiple samples are subjected to color adjustment, obtain exptended sample collection.
- 8. according to the method for claim 5, it is characterised in that described to be expanded the original training set Sample set, including：One in the original training set or multiple samples are subjected to noise addition, obtain exptended sample collection.
- 9. according to the method described in any one of Claims 1-4, it is characterised in that it is determined that being wrapped in the target information of road surface When containing deceleration driven, in addition to：Export prompt message.
- A kind of 10. deceleration driven detection means based on convolutional neural networks, it is characterised in that including：Target information of road surface obtains module, for obtaining target information of road surface；Object detection results obtain module, for according to constructing in advance and being trained the target convolutional neural networks of acquisition, examining Survey the target information of road surface and obtain object detection results；Deceleration driven determining module, for according to the object detection results, determining whether wrapped in the target information of road surface Containing deceleration driven.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710860765.6A CN107463927A (en) | 2017-09-21 | 2017-09-21 | A kind of deceleration driven detection method and device based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710860765.6A CN107463927A (en) | 2017-09-21 | 2017-09-21 | A kind of deceleration driven detection method and device based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107463927A true CN107463927A (en) | 2017-12-12 |
Family
ID=60552961
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710860765.6A Pending CN107463927A (en) | 2017-09-21 | 2017-09-21 | A kind of deceleration driven detection method and device based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107463927A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108860101A (en) * | 2018-04-20 | 2018-11-23 | 深圳市路畅智能科技有限公司 | For the safety assistant driving method of deceleration strip |
| CN109359565A (en) * | 2018-09-29 | 2019-02-19 | 广东工业大学 | A kind of deceleration driven detection method and system |
| CN110349119A (en) * | 2019-05-27 | 2019-10-18 | 北京邮电大学 | Pavement disease detection method and device based on edge detection neural network |
| WO2020151438A1 (en) * | 2019-01-25 | 2020-07-30 | 京东方科技集团股份有限公司 | Neural network processing method and evaluation method, and data analysis method and device |
| CN111738040A (en) * | 2019-06-25 | 2020-10-02 | 北京京东尚科信息技术有限公司 | Deceleration strip identification method and system |
| CN112171057A (en) * | 2020-09-10 | 2021-01-05 | 五邑大学 | Quality detection method and device based on laser welding and storage medium |
| CN113780196A (en) * | 2021-09-15 | 2021-12-10 | 江阴市浩华新型复合材料有限公司 | Abnormal data real-time reporting system |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104537393A (en) * | 2015-01-04 | 2015-04-22 | 大连理工大学 | Traffic sign recognizing method based on multi-resolution convolution neural networks |
| CN105930830A (en) * | 2016-05-18 | 2016-09-07 | 大连理工大学 | Road surface traffic sign recognition method based on convolution neural network |
| CN106372577A (en) * | 2016-08-23 | 2017-02-01 | 北京航空航天大学 | Deep learning-based traffic sign automatic identifying and marking method |
| CN106815596A (en) * | 2016-12-08 | 2017-06-09 | 中国银联股份有限公司 | A kind of Image Classifier method for building up and device |
| CN106919915A (en) * | 2017-02-22 | 2017-07-04 | 武汉极目智能技术有限公司 | Map road mark and road quality harvester and method based on ADAS systems |
| CN107122776A (en) * | 2017-04-14 | 2017-09-01 | 重庆邮电大学 | A kind of road traffic sign detection and recognition methods based on convolutional neural networks |
-
2017
- 2017-09-21 CN CN201710860765.6A patent/CN107463927A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104537393A (en) * | 2015-01-04 | 2015-04-22 | 大连理工大学 | Traffic sign recognizing method based on multi-resolution convolution neural networks |
| CN105930830A (en) * | 2016-05-18 | 2016-09-07 | 大连理工大学 | Road surface traffic sign recognition method based on convolution neural network |
| CN106372577A (en) * | 2016-08-23 | 2017-02-01 | 北京航空航天大学 | Deep learning-based traffic sign automatic identifying and marking method |
| CN106815596A (en) * | 2016-12-08 | 2017-06-09 | 中国银联股份有限公司 | A kind of Image Classifier method for building up and device |
| CN106919915A (en) * | 2017-02-22 | 2017-07-04 | 武汉极目智能技术有限公司 | Map road mark and road quality harvester and method based on ADAS systems |
| CN107122776A (en) * | 2017-04-14 | 2017-09-01 | 重庆邮电大学 | A kind of road traffic sign detection and recognition methods based on convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| ANDREW G.HOWARD: "Some Improvements on Deep Convolutional Neural Network Based Image Classification", 《ARXIV》 * |
| 李怀: "基于集成卷积神经网络的人脸年龄识别研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108860101A (en) * | 2018-04-20 | 2018-11-23 | 深圳市路畅智能科技有限公司 | For the safety assistant driving method of deceleration strip |
| CN109359565A (en) * | 2018-09-29 | 2019-02-19 | 广东工业大学 | A kind of deceleration driven detection method and system |
| WO2020151438A1 (en) * | 2019-01-25 | 2020-07-30 | 京东方科技集团股份有限公司 | Neural network processing method and evaluation method, and data analysis method and device |
| CN110349119A (en) * | 2019-05-27 | 2019-10-18 | 北京邮电大学 | Pavement disease detection method and device based on edge detection neural network |
| CN110349119B (en) * | 2019-05-27 | 2020-05-19 | 北京邮电大学 | Pavement disease detection method and device based on edge detection neural network |
| CN111738040A (en) * | 2019-06-25 | 2020-10-02 | 北京京东尚科信息技术有限公司 | Deceleration strip identification method and system |
| CN112171057A (en) * | 2020-09-10 | 2021-01-05 | 五邑大学 | Quality detection method and device based on laser welding and storage medium |
| CN112171057B (en) * | 2020-09-10 | 2022-04-08 | 五邑大学 | Quality detection method and device based on laser welding and storage medium |
| CN113780196A (en) * | 2021-09-15 | 2021-12-10 | 江阴市浩华新型复合材料有限公司 | Abnormal data real-time reporting system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107463927A (en) | A kind of deceleration driven detection method and device based on convolutional neural networks | |
| CN104834933B (en) | A kind of detection method and device in saliency region | |
| CN111967393A (en) | Helmet wearing detection method based on improved YOLOv4 | |
| CN107220618B (en) | Face detection method and device, computer readable storage medium and equipment | |
| CN109584248A (en) | Infrared surface object instance dividing method based on Fusion Features and dense connection network | |
| CN103578119B (en) | Target detection method in Codebook dynamic scene based on superpixels | |
| CN112084869B (en) | Compact quadrilateral representation-based building target detection method | |
| CN104392228B (en) | Unmanned plane image object class detection method based on conditional random field models | |
| CN109670452A (en) | Method for detecting human face, device, electronic equipment and Face datection model | |
| CN106683091A (en) | Target classification and attitude detection method based on depth convolution neural network | |
| CN108710863A (en) | Unmanned plane Scene Semantics dividing method based on deep learning and system | |
| CN111881730A (en) | Wearing detection method for on-site safety helmet of thermal power plant | |
| CN107229929A (en) | A kind of license plate locating method based on R CNN | |
| CN108229575A (en) | For detecting the method and apparatus of target | |
| CN106997461A (en) | A kind of firework detecting method and device | |
| CN109815997A (en) | The method and relevant apparatus of identification vehicle damage based on deep learning | |
| CN108256544A (en) | Picture classification method and device, robot | |
| CN110399856A (en) | Feature extraction network training method, image processing method, device and its equipment | |
| CN110472611A (en) | Method, apparatus, electronic equipment and the readable storage medium storing program for executing of character attribute identification | |
| CN108154110A (en) | A kind of intensive people flow amount statistical method based on the detection of the deep learning number of people | |
| He et al. | A robust method for wheatear detection using UAV in natural scenes | |
| CN104102928B (en) | A kind of Classifying Method in Remote Sensing Image based on texture primitive | |
| CN109858547A (en) | A kind of object detection method and device based on BSSD | |
| CN107273832A (en) | Licence plate recognition method and system based on integrating channel feature and convolutional neural networks | |
| CN110766058A (en) | Battlefield target detection method based on optimized RPN (resilient packet network) |
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 |
Application publication date: 20171212 |
|
| RJ01 | Rejection of invention patent application after publication |