CN106886755A - A kind of intersection vehicles system for detecting regulation violation based on Traffic Sign Recognition - Google Patents
A kind of intersection vehicles system for detecting regulation violation based on Traffic Sign Recognition Download PDFInfo
- Publication number
- CN106886755A CN106886755A CN201710037410.7A CN201710037410A CN106886755A CN 106886755 A CN106886755 A CN 106886755A CN 201710037410 A CN201710037410 A CN 201710037410A CN 106886755 A CN106886755 A CN 106886755A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- bounding box
- candidate
- convolutional
- coordinate
- 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
- G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- 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
- G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/017—Detecting movement of traffic to be counted or controlled identifying vehicles
- G08G1/0175—Detecting movement of traffic to be counted or controlled identifying vehicles by photographing vehicles, e.g. when violating traffic rules
Abstract
This patent discloses a kind of intersection vehicles peccancy detection and discrimination method based on deep learning, methods described comprises the following steps：Step one, collection intersection video detection image；The identification of step 2, traffic sign and vehicle；Step 3, the tracking of vehicle；Step 4, vehicle peccancy detection and identification system.The present invention can automatically pick out the situation violating the regulations of intersection vehicle by based on the traffic mark board and the automatic recognition and tracking of vehicle in crossing monitor video, and the system flase drop is few, efficiency high.
Description
Technical field
This patent belongs to technical field of intelligent traffic, is disobeyed in particular to a kind of intersection vehicles based on deep learning
Chapter examines car and identification system.
Technical background
The punishment of parking offense can be punished by manual site in the prior art, and by the non-at-scene place of camera
The mode penalized is carried out.In the prior art, the image by way of on road is clapped by being arranged on the camera in roadside, by intelligence
Know situation about being driven against traffic regulations in identification image otherwise, and it is a kind of common means to intercept corresponding evidence.
But, driving against traffic regulations for intersection in the prior art, for example no turns, and place turns around, not by regulation
Lanes etc., easily find in law enforcement at the scene, but are difficult to be realized by way of Intelligent Recognition.
Because different road its passing rules for setting are different, and same intersection sometimes passes through rule
It is then that can be changed according to the consideration of various factors.And the camera of intersection is located at for simple road conditions, example
The identification such as made a dash across the red light is accurate；And for complicated road conditions, such as road intersection not according to the rules
Lanes, parking offense, no turns, and place turns around etc. identification is inaccurate, or identification degree is not high, to traveling people or
Person is that vehicle supervision department brings puzzlement.
The content of the invention
This patent is based on the demand of the prior art and proposes, this patent technical problem to be solved is to carry
For a kind of intersection vehicles peccancy detection based on deep learning and identification system, in order to Intelligent Measurement in intersection
Vehicle peccancy situation.
In order to solve the above problems, the technical scheme that this patent is provided includes：
A kind of intersection vehicles peccancy detection and discrimination method based on deep learning, methods described comprise the following steps：
Step one, collection intersection video detection image
Video image is gathered using the video capture device in intersection；
The identification of step 2, traffic sign and vehicle
In this step, for the identification of traffic sign and vehicle, its process includes：
Step S201 carries out feature extraction to input picture
Automatic signature extraction is carried out to input picture using convolutional neural networks profit；The convolutional neural networks include feature
Extract and characteristic pattern cascades two parts；
1) feature extraction, the convolutional neural networks include five layers convolutional layers, wherein the first convolutional layer includes 1_1 convolution
Layer；Second convolutional layer includes 2_1 convolutional layers, 2_2 convolutional layers and 2_3 convolutional layers；3rd convolutional layer include 3_1 convolutional layers, 3_2 volumes
Lamination, 3_3 convolutional layers and 3_4 convolutional layers；Volume Four lamination includes 4_1 convolutional layers, C4_2 convolutional layers, 4_3 convolutional layers and 4_4 volumes
Lamination；5th convolutional layer includes 5_1 convolutional layers, 5_2 convolutional layers, 5_3 convolutional layers and 5_4 convolutional layers；Wherein the first convolutional layer
Action type uses C.ReLU, and receptive field size is 7*7 pixels, and output channel number is 32.The action type of the second convolutional layer is adopted
With C.ReLU, receptive field size is 3*3 pixels, and output channel number is 64；The action type of the 3rd convolutional layer uses C.ReLU, sense
It is 3*3 pixels by wild size, output channel number is 128；The action type of Volume Four lamination and the 5th convolutional layer is used
Inception；
2) characteristic pattern cascade, the convolution characteristic layer of the different scale in different convolution stages is cascaded up, and make use of high level
The detail textures information of low-level feature is additionally contemplates that while the semantic information of feature, 3_4 convolutional layers are carried out to use maximum
Chi Hua, 5_4 convolutional layer are up-sampled using bilinear interpolation algorithm, and then both are combined by 1x1's with 4_4 convolutional layers
The multiple dimensioned output characteristic of 512 passages is generated after convolution as the input of following candidate region network；
The detection and localization of S202 traffic signs and vehicle
The candidate region of traffic sign and vehicle is included from positioning from the image after feature extraction, afterwards by candidate region
Export the Classification and Identification module of traffic sign and vehicle；Including：
1) use candidate region network extraction candidate region, input picture by after feature extraction network, at last
Small network is slided in the convolution Feature Mapping of convolutional layer output, the small network of slip is connected to input convolution Feature Mapping entirely
In the spatial window of n*n, to generate candidate region frame.
2) loss function of candidate region network
In order to train candidate region network, a binary label is distributed (respectively to each candidate frame in this step
Corresponding situation is：It is target, is not target).Positive label is distributed to two class candidate frames：(1) with certain real estate bounding box
There is maximum IoU (two common factor areas in region and union area ratio) value；(2) with the IoU values of any real estate bounding box
Candidate frame more than 0.7.It is all negative labels of candidate frame distribution with real estate IoU values less than 0.3.Candidate region network
Loss function is defined as：
Wherein, i is once the candidate frame index of selection in batch iteration；piFor candidate frame is the probability of target, if waited
It is positive label to select frame, its corresponding real estate label pi *It is 1, otherwise, pi *It is 0；tiRepresent 4 parameters of the bounding box of prediction
Change coordinate vector, ti *It is the coordinate vector of corresponding real estate bounding box.
Classification Loss CclsIt is directed to the logarithm loss of two classifications of target and non-targeted：
Lost for returning, usedTo calculate, wherein S (x) is：
For returning, this step is using 4 parameters of coordinate：
tx=(x-xa)/wa ty=(y-ya)/ha
tw=log (w/wa) th=log (h/ha)
Wherein, x, y, w, h refer to (x, y) coordinate, width, the height at bounding box center, variable x, xa,x*Respectively predict
Bounding box, candidate frame, the x coordinate of real estate bounding box, variable y, ya,y*It is the bounding box respectively predicted, candidate frame, true
The y-coordinate of region bounding box；Variable w, wa,w*Bounding box, candidate frame, the w coordinates of real estate bounding box respectively predicted；
Variable h, ha,h*Bounding box, candidate frame, the h coordinates of real estate bounding box respectively predicted；NclsTake 256, NregTake 2400,
λ takes 10；tx,ty,tw,thRepresent 4 parametrization coordinate vectors of the bounding box of prediction, tx *,ty *,tw *,th *It is corresponding true
The coordinate vector of real region bounding box.
The Classification and Identification of S203 traffic signs and vehicle
The main task of the Classification and Identification of traffic sign and vehicle be the traffic sign that finally navigates to for previous step and
The candidate region of vehicle carries out feature extraction and Classification and Identification, so as to obtain the type information of traffic sign and vehicle, wherein
Including：
1) identification of traffic sign and vehicle is carried out using region convolutional network, for obtaining containing traffic mark in S202
The candidate region of will, traffic sign is carried out using a kind of with candidate region network share convolution region convolutional neural networks algorithm
Identification；
2) singular value decomposition, the recognition speed of Traffic Sign Recognition System is accelerated using singular value decomposition；If W is one
Size is the weight matrix of u × v, it is possible to use singular value decomposition is obtained, such as formula：
W≈U∑tVT
In formula：U：Size is a vector of u × t；∑t：Size is the diagonal matrix of t × t；VT：Size for t × v to
Amount；By | WTW- λ I |=0 tries to achieve WTThe eigenvalue λ of W；If the transposition of matrix W is WT, by formula (WTW)vi=λiviTry to achieve WTW's
Characteristic vector v；According to formulaSingular value σ is obtained, and U then can be according to formulaObtain：U, ∑t, VTThree
Matrix multiple is close to matrix W；
3) loss function of region convolutional neural networks, region convolutional neural networks can export K+1 classes to input area
Target, comprising background, probability and the bounding box coordinate after returning.Loss function is defined for each training candidate region
For：
L(p,u,tu,t*)=Lcls(p,u)+[u≥1]Lloc(tu,t*)
Wherein, Lcls(p, u)=- lgpuIt is the logarithm loss of real goal classification u corresponding to candidate region；For k+1
Classification indicates p=(p0,p1,....,pk)；[u >=1] is indicative function, when candidate region is background, u=0；Lloc(tu,t*)
For the recurrence of bounding box coordinate is lost,
Function S in formula is：Wherein have for each in k class targetst*It is the parameter coordinate vector of the corresponding real goal bounding box in candidate region；For in k class targets
Predict the x coordinate of bounding box；It is the y-coordinate of the prediction bounding box in k class targets；It is the prediction bounding box in k class targets
W coordinates；It is the h coordinates of the prediction bounding box in k class targets；
Step 3, the tracking of vehicle
Traffic sign in image and after the identification of vehicle, is painted by the vehicle to recognizing using track algorithm
Make the enforcement track of vehicle；Using error least square and wave filterMake its response in target maximum, wherein
G represents response output, and f represents input picture, and h represents Filtering Template,Represent convolution operation；Fast Fourier is carried out to above formula
Become convolution operation of changing commanders and become dot product operation, that is, become：Write a Chinese character in simplified form into：G
=FH*Therefore, the task of tracking is exactly to find H*, H*It is the conjugation of H, i.e.,：During tracking, model is used
Formula is：Wherein, ⊙ represents Hadamard products,
ObtainWherein w and v are the indexes of each element in H；It is asked
Local derviation, and make local derviation be 0, i.e.,：Obtain
Finally obtaining H is：The as model formation of wave filter；Following template renewal mode：
Bt=η Ft·Ft *+(1-η)Bt-1
The model formation of wave filter is divided into two parts of molecule and denominator to be updated respectively, parameter more is that experience is normal
Number η.Wherein AtAnd At-1What is represented respectively is the molecule of present frame and previous frame；Wherein BtAnd Bt+1What is represented respectively is present frame
With the denominator of previous frame；During tracking, the image of template above and present frame is done associative operation, the corresponding knot that will be obtained
In fruit that maximum corresponding coordinate as target in the position of present frame；Wherein, HtIt is the masterplate of t wave filter, GtTable
Show t response output, Gt* it is GtConjugation；FtRepresent t input picture, Ft* it is FtConjugation.
Step 4, vehicle peccancy detection and identification system
The specification of the legal traveling in the intersection is obtained by the identification of traffic mark board, then is processed by vehicle identification and tracking
The driving trace of each vehicle in the intersection is obtained afterwards；By comparing the information of traffic mark board representative and the traveling rail of vehicle
Mark can clearly pick out which vehicle peccancy traveling, it is possible to the video of this segment record vehicle peccancy is preserved as
The foundation of the vehicle supervision departments such as delay penalty law enforcement.
The advantage of the invention is that：
1st, the present invention is by based on the traffic mark board and the automatic recognition and tracking of vehicle in crossing monitor video, entering
And the situation violating the regulations of intersection vehicle can be automatically picked out, the system flase drop is few, efficiency high, without setting special taking the photograph again
It is applied widely as head, can avoid expending substantial amounts of human and material resources financial resources.
2nd, the present invention makes full use of the advantage of image convolution, reduces illumination variation, the mould that color fading, motion are caused
The influence of the factor to image recognition such as paste, complicated background, partial occlusion, improves antijamming capability, and recognition accuracy is high, by mistake
Discrimination is low.
3. C.ReLu (x) activation primitives are applied in the present invention in the convolution of low layer, it is possible to reduce network parameter, make port number
Half is reduced, the output before neuronal activation and its opposite number are carried out into cascade makes output become double, that is, reach original output
Quantity, this cause without loss accuracy in the case of, calculating speed lifted one times.
4. Inception is applied in the present invention in high-rise network, mainly by controlling convolution kernel size to be lifted
To the performance of Small object identification, while the port number of input can be reduced, play a part of dimensionality reduction, computational complexity drops significantly
It is low.
5. the characteristic pattern cascade advantage in the present invention is the abstract characteristics of image of multilayer, and its resolution ratio is suitable to detection and localization, counts
Calculate efficiency high.
Brief description of the drawings
Fig. 1 is the overall flow figure of method in this specific embodiment；
Fig. 2 is Traffic Sign Recognition and Vehicular system framework；
The overall structure figure of Fig. 3 recognizers；
Fig. 4 is C.ReLu (x) and standardized operational flowchart in this specific embodiment；
Fig. 5 is the pie graph of Inception in this specific embodiment；
In Fig. 6 this specific embodiment small Principles of Network schematic diagram is slided in convolution Feature Mapping.
Specific embodiment
The present invention is stated in detail with reference to the accompanying drawings and examples.It is pointed out that the specific embodiment
Only to the citing of this patent optimal technical scheme, the limitation to this patent protection domain can not be interpreted as.
A kind of intersection vehicles peccancy detection and discrimination method based on deep learning are provided in this specific embodiment,
Methods described is broadly divided into：The identification of the collection of crossing monitor video, traffic sign and vehicle, the tracking of vehicle, vehicle is disobeyed
Chapter is detected and identification.Its flow is as shown in Figure 1.
Step one, collection intersection video detection image
The existing picture pick-up device acquisition monitoring of each intersection being arranged in urban road can be utilized in this step
Video, generally can all set the first-class video capture device of shooting in the intersection of road.Kept away by existing video capture device
The increase of cost is exempted from.But existing video detecting device directly can not intelligently detect the parking offense feelings at crossing
Condition.Therefore also need to process the video that intersection collects.The video image detected in intersection includes vehicle
Information, road information, traffic signal information etc..
The identification of step 2, traffic sign and vehicle
The parking offense information of vehicle can not be directly given in the video image for collecting, in order to detect parking offense
Information is, it is necessary to by the position of vehicle and the species of vehicle, and traffic sign position in road and the species of traffic sign are entered
Row identification.These identification informations are obtained, the transport information of correlation is obtained while computer or similar device intelligence can be easy to,
For the Intelligent Recognition of parking offense provides basis.
The recognizer of traffic sign and vehicle it is critical only that a kind of more effective area-of-interest exacting method and one
Kind more rapidly, accurately recognition methods, it quick detection and can identify the traffic sign and vehicle in high-resolution pictures.
The identifying system framework of traffic sign and vehicle is as shown in Figure 2 in implementation method.
In this step, the identification for traffic sign and vehicle mainly includes three below part：The spy of input picture
Levy the Classification and Identification of detection and localization, traffic sign and the vehicle of extraction, traffic sign and vehicle.It is above-mentioned to be identified by as follows specifically
The step of realize：Main feature extraction, the traffic realized using convolutional neural networks for input picture in this step
The Classification and Identification of the detection and localization, traffic sign and vehicle of mark and vehicle.In the present embodiment, the convolutional neural networks bag
Include multiple convolutional layers to be set up with pond layer, more effective feature can be extracted from image, for the inspection of ensuing candidate region
Survey and Classification and Identification, its course of work is as shown in Figure 3.Because convolution operation has translation invariant shape, the characteristic pattern of convolution generation
In not only contain the characteristic information of object, also include the positional information of object.
Step S201 carries out feature extraction to input picture
Automatic signature extraction is carried out to input picture using convolutional neural networks profit in this step.Avoid artificial extraction
The trouble of feature, simultaneously because the translation invariant characteristic of convolution, not only contains the feature of object in the characteristic pattern of convolution generation
Information, also includes the positional information of object.The feature extraction network for using in this step includes feature extraction and characteristic pattern
Cascade two parts.
1) feature extraction
For feature extraction, in this step, the network of first five layer of the convolutional neural networks of the feature extraction for using
Result is as shown in table 1 below.
1 preceding 5 layer of convolutional layer network settings of table
The action type of wherein the first convolutional layer uses C.ReLU, and receptive field size is 7*7 (pixel), and output channel number is
32.The action type of the second convolutional layer uses C.ReLU, and receptive field size is 3*3 (pixel), and output channel number is 64；Volume three
The action type of lamination uses C.ReLU, and receptive field size is 3*3 (pixel), and output channel number is 128.The C.ReLU operations
C.ReLu (x) activation primitives, C.ReLu (x) activation primitives are used in type to be used for the convolution of low layer, can reduce network parameter, make
Port number reduces half, and the output before neuronal activation and its opposite number are carried out into cascade makes output become double, that is, reach original
Come the quantity for exporting, this causes in the case of without loss accuracy, to lift calculating speed.The mathematic(al) representation of C.ReLu is such as
Under：
C.ReLu (wx+b)=ReLU (wx+b)+ReLU (- (wx+b))
Wherein w is weight, and x is the input of neuron, and b is the biasing of neuron.
In this step, it is preferable that the activation primitive is in active coating, standardization (batch is used before active coating
Normalization, BN) accelerate network convergence, improve model accuracy.In the convolutional neural networks of this specific embodiment,
Before BN should act on Nonlinear Mapping, i.e., the output wx+b of convolutional layer is done and standardized.I.e. in the convolution of this specific embodiment
In neutral net, the input of neuron is x, then obtains (wx+b), reuses BN, then with activation primitive, obtains the nerve
The output of unit.BN allows activation primitive in linearly interval by normalization, reduces the sensitivity to being input into, and increases gradient,
Being conducive to model carries out gradient decline, prevents gradient disperse.
The input of BN is x={ x1,x2,...,xmAnd given parameters γ, β, so having：
μ is average, σ2It is variance,To be processed into the result of standardized normal distribution
Wherein, ε is the constant more than 0.The final output of BN is：
So, it is as shown in Figure 4 using C.ReLu (x) and standardized operating process.
In 4-5 convolutional layers in this step, using Inception action types, in 4-5 layers of Inception operation class
Type lifts the performance to Small object identification by controlling convolution kernel size.1x1 convolution kernels in Inception can not only increase
Screening network it is non-linear, while the receptive field (not losing resolution ratio) of preceding layer can be kept, therefore to the detection of wisp
There is effect well.The port number of input can be reduced after addition 1x1 convolution, plays a part of dimensionality reduction, computational complexity is significantly
Reduce.The composition of Inception in this embodiment is as shown in Figure 5.
2) characteristic pattern cascade
In order to the convolution characteristic layer of the different scale in different convolution stages is cascaded up, the semanteme of high-level characteristic is make use of
The detail textures information of low-level feature is additionally contemplates that while information, carries out 3_4 convolutional layers to use maximum in this step
Chi Hua, 5_4 convolutional layer are up-sampled using bilinear interpolation algorithm, and then both are combined by 1x1's with 4_4 convolutional layers
The multiple dimensioned output characteristic of 512 passages is generated after convolution as the input of following candidate region network.This method is to target
The accuracy and positioning precision of detection are all benefited, and the identifying system of traffic sign and vehicle exactly with classification know by target positioning
It is not combined.Characteristic pattern subtended network configuration such as table 2.Characteristic superiority after cascade is that multilayer is abstract, and resolution ratio is adapted to detect for, meter
Calculate efficiency high.The characteristic pattern subtended network configuration table of table 2
The detection and localization of S202 traffic signs and vehicle
The detection and localization of traffic sign and vehicle as whole identifying system one of committed step, its main task is
Candidate region of the positioning comprising traffic sign and vehicle from input picture, afterwards by candidate region output to traffic sign and car
Classification and Identification module.
(1) candidate region network extraction candidate region is used
Input picture, in order to generate candidate region frame, is exported by after feature extraction network in last convolutional layer
Small network is slided in convolution Feature Mapping, the small network of slip is connected to the spatial window of the nxn of input convolution Feature Mapping entirely
On mouth.Sliding window operation is carried out on convolutional layer characteristic pattern, and dimension compared with sliding window is used on the original image
Reduce.Each sliding window is mapped on a low-dimensional vector.This vector is exported to two layers of the full connection of peer：Bag
Enclose box and return layer and bounding box classification layer.Wicket is slided in Feature Mapping layer set up a small network；Classification：Target and non-mesh
Mark；Return：The position of bounding box.The position of sliding window provides the location information of relative picture.The recurrence of bounding box is provided
Location information after the adjustment of relative sliding window, its principle is as shown in Figure 6.
In the position of each sliding window, while using multiple dimensioned candidate frame mechanism, various yardsticks and many have been corresponded to
Length-width ratio is planted, k candidate frame is predicted.So returning layer has 4k output, i.e., the k codes co-ordinates of candidate frame.Classification layer leads to
Cross the probability Estimation 2k score of output for carrying out target and non-targeted using softmax functions to every k candidate frame.Classification is given
Possibility size of each candidate frame comprising target.The k candidate frame coordinate offset amount that layer gives is returned, by bounding box
The target area coordinates position that recurrence is predicted.
Wherein, softmax functions are that multiple scalars are mapped as into a probability distribution.For k scalar x1,x2,...,
xk, softmax functions are defined as：
Finally, a series of selected by frame, candidate regions with target score are obtained.
(2) loss function of candidate region network
In order to train candidate region network, a binary label is distributed (respectively to each candidate frame in this step
Corresponding situation is：It is target, is not target).Positive label is distributed to two class candidate frames：(1) with certain real estate bounding box
There is maximum IoU (two common factor areas in region and union area ratio) value；(2) with the IoU values of any real estate bounding box
Candidate frame more than 0.7.It is all negative labels of candidate frame distribution with real estate IoU values less than 0.3.Candidate region network
Loss function is defined as：
Wherein, i is once the candidate frame index of selection in batch iteration；piFor candidate frame is the probability of target, if waited
It is positive label to select frame, its corresponding real estate label pi *It is 1, otherwise, pi *It is 0；tiRepresent 4 parameters of the bounding box of prediction
Change coordinate vector, ti *It is the coordinate vector of corresponding real estate bounding box；
Classification Loss CclsIt is directed to the logarithm loss of two classifications of target and non-targeted：
Wherein, i is once the candidate frame index of selection in batch iteration；piFor candidate frame is the probability of target, if waited
It is positive label to select frame, its corresponding real estate label pi *It is 1, otherwise, pi *It is 0；
Lost for returning, usedTo calculate, wherein S (x) is：
For returning, this step is using 4 parameters of coordinate：
tx=(x-xa)/wa ty=(y-ya)/ha
tw=log (w/wa) th=log (h/ha)
Wherein, x, y, w, h refer to (x, y) coordinate, width, the height at bounding box center, variable x, xa,x*Respectively predict
Bounding box, candidate frame, the x coordinate of real estate bounding box, variable y, ya,y*It is the bounding box respectively predicted, candidate frame, true
The y-coordinate of region bounding box；Variable w, wa,w*Bounding box, candidate frame, the w coordinates of real estate bounding box respectively predicted；
Variable h, ha,h*Bounding box, candidate frame, the h coordinates of real estate bounding box respectively predicted；NclsTake 256, NregTake 2400,
λ takes 10；tx,ty,tw,thRepresent 4 parametrization coordinate vectors of the bounding box of prediction, tx *, ty *, tw *, th *It is corresponding true
The coordinate vector of real region bounding box.
The Classification and Identification of S203 traffic signs and vehicle
The main task of the Classification and Identification of traffic sign and vehicle be the traffic sign that finally navigates to for previous step and
The candidate region of vehicle carries out feature extraction and Classification and Identification, and so as to obtain the type information of traffic sign and vehicle, this is
The final step of identifying system, is also one of most important task in algorithm.
(1) identification of traffic sign and vehicle is carried out using region convolutional network
For the candidate region containing traffic sign obtained in S202, in the Classification and Identification stage of traffic sign, this hair
A kind of bright identification that traffic sign is carried out using and candidate region network share convolution region convolutional neural networks algorithm.Training
When, first by the data set input network with mark, generate convolution characteristic pattern.Candidate region then is mapped into this to share
Characteristic pattern in, obtain corresponding characteristic information.Then, bilinear interpolation algorithm is carried out to this feature information, obtains one greatly
The small provincial characteristics pondization for 7*7 is schemed.And then, one 4096 characteristic vector of dimension is further obtained by full connection, this to
Amount is exactly the final feature of each candidate region that convolutional network is proposed.Finally, this characteristic vector is separately input to
Softmax classifies with bounding box recurrence, and target classification and the position of candidate region are judged using non-maxima suppression method.It
Afterwards, loss function is obtained using the difference between judgment value and actual mark value, then using back-propagation algorithm and boarding steps
Degree descent method is optimized to network parameter structure, finally gives output network.In experiment, testing image is input to network
In, directly export the classification of traffic sign and positioned.
(2) singular value decomposition
Full articulamentum in fast area network can consume the plenty of time.And Eigenvalues Decomposition is especially suitable for using in mathematics
In the feature for extracting square formation.Because matrix is not square formation in full articulamentum, singular value decomposition now can be used to obtain matrix
The feature of itself.Therefore this paper Selection utilizations singular value decomposition accelerates the recognition speed of Traffic Sign Recognition System.
If W is the weight matrix that a size is u × v, it is possible to use singular value decomposition is obtained, such as formula：
W≈U∑tVT
In formula：U：Size is a vector of u × t
∑t：Size is the diagonal matrix of t × t
VT：Size is the vector of t × v
By | WTW- λ I |=0 tries to achieve WTThe eigenvalue λ of W.
If the transposition of matrix W is WT, W is tried to achieve by equation belowTThe characteristic vector v of W.
(WTW)vi=λivi
In addition to this it is possible to singular value σ is obtained according to formula (3-36), and U can then be obtained according to equation below：
U, ∑t, VTThree matrix multiples are close to matrix W, and the area sum of these three matrixes will be far smaller than original
Matrix W, from the point of view of storage viewpoint, matrix area is smaller, and amount of storage is just smaller.Original matrix W is represented to compression stroke, we
Leave three matrixes：U, ∑t, VT.The identification for reducing 30% using the method for singular value decomposition herein calculates time-consuming.
(3) loss function of region convolutional neural networks
After region convolutional neural networks can be to the probability of input area output K+1 classes target (including background) and recurrence
Bounding box coordinate.Defining loss function for each training candidate region is：
L(p,u,tu,t*)=Lcls(p,u)+[u≥1]Lloc(tu,t*)
Wherein, Lcls(p, u)=- lgpuIt is the logarithm loss of real goal classification u corresponding to candidate region.
P=(p are indicated for k+1 classifications0,p1,....,pk).[u >=1] is indicative function, when candidate region is background,
U=0.Lloc(tu,t*) it is the recurrence loss of bounding box coordinate,
Function S in formula is：
Wherein have for each in k class targetst*It is the corresponding real goal in candidate region
The parameter coordinate vector of bounding box；It is the x coordinate of the prediction bounding box in k class targets；For the prediction in k class targets is surrounded
The y-coordinate of box；It is the w coordinates of the prediction bounding box in k class targets；It is the h coordinates of the prediction bounding box in k class targets.
Step 3, the tracking of vehicle
Traffic sign in image and after the identification of vehicle, is painted by the vehicle to recognizing using track algorithm
Make the enforcement track of vehicle.
The present invention utilizes error least square and wave filter (Minimum Output Sum of Squared Error
Filter, MOSSE), make its response in target maximum, then equation below：
Wherein g represents response output, and f represents input picture, and h represents Filtering Template,Represent convolution operation.Responded
Output g, only need to determine filter template h.The calculating of above formula is to carry out convolutional calculation, when this calculating in a computer is consumed very
Big, therefore Fast Fourier Transform (FFT) (FFT) is carried out to above formula herein.Consequently, it is possible to convolution operation becomes by fast Fourier
Dot product operation has been reformed into after changing, amount of calculation has been greatly reduced.Above formula becomes following form：
Describe for convenience, above formula is written as form：H=FH*Therefore, the task of tracking is exactly to find H*, H*For
The conjugation of H, i.e.,：
During actual tracking, it is contemplated that the influence of the factor such as variable cosmetic of target, while considering m of target
Image is used as reference, so as to improve the robustness of filter template, this model of MOSSE is proposed in this patent, and its model is public
Under formula enters：
Wherein, ⊙ represents Hadamard products, for point-to-point multiplying between matrix or vector, the operation of above formula
All it is that Element-Level is other, it is only necessary to make the MOSSE minimums of each element (w and v are the indexes of each element in H) therein i.e.
Can.Therefore above formula can be exchanged into following form：
Expect minimumOnly need to seek above formula local derviation, and make local derviation be 0, i.e.,：
Obtain：
Finally obtaining H is：
Above formula is the model formation of wave filter.
There is more preferable robustness in order to allow the ectocines such as wave filter pair and deformation, illumination, following template is taken
Update mode：
Bt=η Ft·Ft *+(1-η)Bt-1
Wherein, HtIt is the masterplate of t wave filter, GtRepresent t response output, Gt* it is GtConjugation；FtWhen representing t
Carve input picture, Ft* it is FtConjugation.
The model formation of wave filter is divided into two parts of molecule and denominator to be updated respectively, parameter more is that experience is normal
Number η.Wherein AtAnd At-1What is represented respectively is the molecule of present frame and previous frame；Wherein BtAnd Bt-1What is represented respectively is present frame
With the denominator of previous frame.
During tracking, it is only necessary to the image of template above and present frame is done associative operation, the corresponding knot that will be obtained
In fruit that maximum corresponding coordinate as target in the position of present frame.This is the equal of to be translated on two dimensional surface
Our filter template.
Step 4, vehicle peccancy detection and identification system
The specification of the legal traveling in the intersection is obtained by the identification of traffic mark board, then is processed by vehicle identification and tracking
The driving trace of each vehicle in the intersection is obtained afterwards.By comparing the information of traffic mark board representative and the traveling rail of vehicle
Mark can clearly pick out which vehicle peccancy traveling, it is possible to the video of this segment record vehicle peccancy is preserved as
The foundation of the vehicle supervision departments such as delay penalty law enforcement.For example, identifying the traffic mark board for representing that no turns (no_turn)
While, if the driving trace of certain vehicle of intersection shows that the vehicle is the traveling that turns around, show that the vehicle exists
Drive against traffic regulations.
The present invention by based on the traffic mark board and the automatic recognition and tracking of vehicle in crossing monitor video, and then
The situation violating the regulations of intersection vehicle can be automatically picked out, the system flase drop is few, efficiency high, without setting special shooting again
Head, it is applied widely, can avoid expending substantial amounts of human and material resources financial resources.
Claims (2)
1. a kind of intersection vehicles peccancy detection and discrimination method based on deep learning, it is characterised in that methods described includes
Following steps：
Step one, collection intersection video detection image
Video image is gathered using the video capture device in intersection；
The identification of step 2, traffic sign and vehicle
In this step, for the identification of traffic sign and vehicle, its process includes：
Step S201 carries out feature extraction to input picture
Automatic signature extraction is carried out to input picture using convolutional neural networks profit；The convolutional neural networks include feature extraction
Two parts are cascaded with characteristic pattern；
1) feature extraction, the convolutional neural networks include five layers convolutional layers, wherein the first convolutional layer includes 1_1 convolutional layers；
Second convolutional layer includes 2_1 convolutional layers, 2_2 convolutional layers and 2_3 convolutional layers；3rd convolutional layer includes 3_1 convolutional layers, 3_2 convolution
Layer, 3_3 convolutional layers and 3_4 convolutional layers；Volume Four lamination includes 4_1 convolutional layers, C4_2 convolutional layers, 4_3 convolutional layers and 4_4 convolution
Layer；5th convolutional layer includes 5_1 convolutional layers, 5_2 convolutional layers, 5_3 convolutional layers and 5_4 convolutional layers；The wherein behaviour of the first convolutional layer
C.ReLU is used as type, receptive field size is 7*7 pixels, and output channel number is 32.The action type of the second convolutional layer is used
C.ReLU, receptive field size is 3*3 pixels, and output channel number is 64；The action type of the 3rd convolutional layer uses C.ReLU, impression
Wild size is 3*3 pixels, and output channel number is 128；The action type of Volume Four lamination and the 5th convolutional layer is used
Inception；
2) characteristic pattern cascade, the convolution characteristic layer of the different scale in different convolution stages is cascaded up, and make use of high-level characteristic
Semantic information while be additionally contemplates that the detail textures information of low-level feature, 3_4 convolutional layers are carried out using maximum pond,
5_4 convolutional layers are up-sampled using bilinear interpolation algorithm, and then both are combined by the convolution of 1x1 with 4_4 convolutional layers
The multiple dimensioned output characteristic of 512 passages is generated afterwards as the input of following candidate region network；
The detection and localization of S202 traffic signs and vehicle
The candidate region of traffic sign and vehicle is included from positioning from the image after feature extraction, afterwards candidate region is exported
To traffic sign and the Classification and Identification module of vehicle；Including：
1) use candidate region network extraction candidate region, input picture by after feature extraction network, in last convolution
Small network is slided in the convolution Feature Mapping of layer output, the small network of slip is connected to the n*n of input convolution Feature Mapping entirely
Spatial window on, to generate candidate region frame.
2) loss function of candidate region network
In order to train candidate region network, distribute a binary label to each candidate frame in this step and (correspond to respectively
Situation be：It is target, is not target).Positive label is distributed to two class candidate frames：(1) have most with certain real estate bounding box
Big IoU (two common factor areas in region and union area ratio) value；(2) the IoU values with any real estate bounding box are more than
0.7 candidate frame.It is all negative labels of candidate frame distribution with real estate IoU values less than 0.3.The loss of candidate region network
Function is defined as：
Wherein, i is once the candidate frame index of selection in batch iteration；piFor candidate frame is the probability of target, if candidate frame is
Positive label, its corresponding real estate label pi *It is 1, otherwise, pi *It is 0；tiRepresent 4 parametrization coordinates of the bounding box of prediction
Vector, ti *It is the coordinate vector of corresponding real estate bounding box；
Classification Loss CclsIt is directed to the logarithm loss of two classifications of target and non-targeted：
Wherein, i is once the candidate frame index of selection in batch iteration；piFor candidate frame is the probability of target, if candidate frame is
Positive label, its corresponding real estate label pi *It is 1, otherwise, pi *It is 0；
Lost for returning, usedTo calculate, wherein S (x) is：
For returning, this step is using 4 parameters of coordinate：
tx=(x-xa)/wa ty=(y-ya)/ha
tw=log (w/wa) th=log (h/ha)
Wherein, x, y, w, h refer to (x, y) coordinate, width, the height at bounding box center, variable x, xa,x*The encirclement respectively predicted
Box, candidate frame, the x coordinate of real estate bounding box, variable y, ya,y*Bounding box, candidate frame, the real estate respectively predicted
The y-coordinate of bounding box；Variable w, wa,w*Bounding box, candidate frame, the w coordinates of real estate bounding box respectively predicted；Variable
h,ha,h*Bounding box, candidate frame, the h coordinates of real estate bounding box respectively predicted；NclsTake 256, Nreg2400, λ is taken to take
10；tx,ty,tw,thRepresent 4 parametrization coordinate vectors of the bounding box of prediction, tx *,ty *,tw *,th *It is corresponding true area
The coordinate vector of domain bounding box；
The Classification and Identification of S203 traffic signs and vehicle
The main task of the Classification and Identification of traffic sign and vehicle is the traffic sign and vehicle finally navigated to for previous step
Candidate region carry out feature extraction and Classification and Identification, so as to obtain the type information of traffic sign and vehicle, including：
1) identification of traffic sign and vehicle is carried out using region convolutional network, for what is obtained in S202 containing traffic sign
Candidate region, using a kind of knowledge that traffic sign is carried out with candidate region network share convolution region convolutional neural networks algorithm
Not；
2) singular value decomposition, the recognition speed of Traffic Sign Recognition System is accelerated using singular value decomposition；If W is a size
It is the weight matrix of u × v, it is possible to use singular value decomposition is obtained, such as formula：
W≈U∑tVT
In formula：U：Size is a vector of u × t；∑t：Size is the diagonal matrix of t × t；VT：Size is the vector of t × v；
By | WTW- λ I |=0 tries to achieve WTThe eigenvalue λ of W；If the transposition of matrix W is WT, by formula (WTW)vi=λiviTry to achieve WTThe feature of W
Vector v；According to formulaSingular value σ is obtained, and U then can be according to formulaObtain：U, ∑t, VTThree matrixes
It is multiplied close to matrix W；
3) loss function of region convolutional neural networks, region convolutional neural networks can export K+1 class targets to input area,
Comprising background, probability and the bounding box coordinate after returning.Defining loss function for each training candidate region is：
L(p,u,tu,t*)=Lcls(p,u)+[u≥1]Lloc(tu,t*)
Wherein, Lcls(p, u)=- lg puIt is the logarithm loss of real goal classification u corresponding to candidate region；For k+1 classifications
Indicate p=(p0,p1,....,pk)；[u >=1] is indicative function, when candidate region is background, u=0；Lloc(tu,t*) it is bag
The recurrence loss of box coordinate is enclosed,
Function S in formula is：Wherein have for each in k class targetst*It is the parameter coordinate vector of the corresponding real goal bounding box in candidate region；For in k class targets
Predict the x coordinate of bounding box；It is the y-coordinate of the prediction bounding box in k class targets；It is the prediction bounding box in k class targets
W coordinates；It is the h coordinates of the prediction bounding box in k class targets；
Step 3, the tracking of vehicle
Traffic sign in image and after the identification of vehicle, is drawn out by the vehicle to recognizing using track algorithm
The enforcement track of vehicle；Using error least square and wave filterMake its response in target maximum, wherein g tables
Show response output, f represents input picture, and h represents Filtering Template,Represent convolution operation；Fast Fourier change is carried out to above formula
Convolution operation of changing commanders becomes dot product operation, that is, become：Write a Chinese character in simplified form into：G=
F·H*Therefore, the task of tracking is exactly to find H*, H*It is the conjugation of H, i.e.,：It is public using model during tracking
Formula is：Wherein, ⊙ represents Hadamard products,
ObtainWherein w and v are the indexes of each element in H；Local derviation is asked to it,
And make local derviation be 0, i.e.,：Obtain
Finally obtaining H is：The as model formation of wave filter；Following template renewal mode：
The model formation of wave filter is divided into two parts of molecule and denominator to be updated respectively, parameter more is empirical η.
Wherein AtAnd At-1What is represented respectively is the molecule of present frame and previous frame；Wherein BtAnd Bt-1What is represented respectively is present frame and upper
The denominator of one frame；During tracking, the image of template above and present frame is done associative operation, in the accordingly result that will be obtained
That maximum corresponding coordinate is as target in the position of present frame；Wherein, HtIt is the masterplate of t wave filter, GtRepresent t
Time of day response is exported, Gt* it is GtConjugation；FtRepresent t input picture, Ft* it is FtConjugation.
Step 4, vehicle peccancy detection and identification system
The specification of the legal traveling in the intersection is obtained by the identification of traffic mark board, then by being obtained after vehicle identification and tracking treatment
The driving trace of each vehicle in the intersection；By comparing the information of traffic mark board representative and the driving trace energy of vehicle
Enough which vehicle peccancy travelings that clearly picks out, it is possible to preserve the video of this segment record vehicle peccancy as violating the regulations
The foundation of the vehicle supervision departments such as fine law enforcement.
2. a kind of intersection vehicles peccancy detection and discrimination method based on deep learning according to claim 1, it is special
Levy and be,
Each sliding window is mapped on a low-dimensional vector；This vector is exported to two layers of the full connection of peer：Surround
Box returns layer and bounding box classification layer；
Wicket is slided in Feature Mapping layer set up a small network；Classification：Target and non-targeted；Return：The position of bounding box.
The position of sliding window provides the location information of relative picture；After the recurrence of bounding box provides the adjustment of relative sliding window
Location information；In the position of each sliding window, while using multiple dimensioned candidate frame mechanism, corresponded to various yardsticks and
Various length-width ratios, predict k candidate frame.So returning layer has 4k output, i.e., the k codes co-ordinates of candidate frame；
Classification layer exports 2k by carrying out target using softmax functions to every k candidate frame with the probability Estimation of non-targeted
Score；Classification gives possibility size of each candidate frame comprising target；Return the k candidate frame coordinate offset that layer gives
Amount, the target area coordinates position predicted is returned by bounding box；
Wherein, soft max functions are that multiple scalars are mapped as into a probability distribution.For k scalar x1,x2,...,xk,
Soft max functions are defined as：Finally, obtain it is a series of it is being selected by frame, obtain with target
The candidate region divided.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710037410.7A CN106886755A (en) | 2017-01-19 | 2017-01-19 | A kind of intersection vehicles system for detecting regulation violation based on Traffic Sign Recognition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710037410.7A CN106886755A (en) | 2017-01-19 | 2017-01-19 | A kind of intersection vehicles system for detecting regulation violation based on Traffic Sign Recognition |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106886755A true CN106886755A (en) | 2017-06-23 |
Family
ID=59175741
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710037410.7A Pending CN106886755A (en) | 2017-01-19 | 2017-01-19 | A kind of intersection vehicles system for detecting regulation violation based on Traffic Sign Recognition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106886755A (en) |
Cited By (36)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107609491A (en) * | 2017-08-23 | 2018-01-19 | 中国科学院声学研究所 | A kind of vehicle peccancy parking detection method based on convolutional neural networks |
| CN107729895A (en) * | 2017-10-18 | 2018-02-23 | 吉林大学 | A kind of intelligent vehicle ADAS aims of systems detection method and device |
| CN108010025A (en) * | 2017-12-14 | 2018-05-08 | 浙江大学 | Switch and indicator lamp positioning and state identification method of screen cabinet based on RCNN |
| CN108090918A (en) * | 2018-02-12 | 2018-05-29 | 天津天地伟业信息系统集成有限公司 | A kind of Real-time Human Face Tracking based on the twin network of the full convolution of depth |
| CN108171752A (en) * | 2017-12-28 | 2018-06-15 | 成都阿普奇科技股份有限公司 | A kind of sea ship video detection and tracking based on deep learning |
| CN108509978A (en) * | 2018-02-28 | 2018-09-07 | 中南大学 | The multi-class targets detection method and model of multi-stage characteristics fusion based on CNN |
| CN108536287A (en) * | 2018-03-26 | 2018-09-14 | 深圳市深晓科技有限公司 | A kind of method and device indicating reading according to user |
| CN108898078A (en) * | 2018-06-15 | 2018-11-27 | 上海理工大学 | A kind of traffic sign real-time detection recognition methods of multiple dimensioned deconvolution neural network |
| CN108932850A (en) * | 2018-06-22 | 2018-12-04 | 安徽科力信息产业有限责任公司 | A kind of record motor vehicle runs at a low speed the method and device of illegal activities |
| CN108985145A (en) * | 2018-05-29 | 2018-12-11 | 同济大学 | The Opposite direction connection deep neural network model method of small size road traffic sign detection identification |
| CN109190687A (en) * | 2018-08-16 | 2019-01-11 | 新智数字科技有限公司 | A kind of nerve network system and its method for identifying vehicle attribute |
| CN109190442A (en) * | 2018-06-26 | 2019-01-11 | 杭州雄迈集成电路技术有限公司 | A kind of fast face detecting method based on depth cascade convolutional neural networks |
| CN109326124A (en) * | 2018-10-17 | 2019-02-12 | 江西洪都航空工业集团有限责任公司 | A kind of urban environment based on machine vision parks cars Activity recognition system |
| CN109344779A (en) * | 2018-10-11 | 2019-02-15 | 高新兴科技集团股份有限公司 | A kind of method for detecting human face under ring road scene based on convolutional neural networks |
| CN109376572A (en) * | 2018-08-09 | 2019-02-22 | 同济大学 | Real-time vehicle detection and trace tracking method in traffic video based on deep learning |
| CN109684901A (en) * | 2017-10-19 | 2019-04-26 | 富士通株式会社 | Image processing apparatus and image processing method |
| CN109685030A (en) * | 2018-12-29 | 2019-04-26 | 哈尔滨理工大学 | A kind of mug rim of a cup defects detection classification method based on convolutional neural networks |
| CN109740424A (en) * | 2018-11-23 | 2019-05-10 | 深圳市华尊科技股份有限公司 | Traffic violations recognition methods and Related product |
| CN109993056A (en) * | 2019-02-25 | 2019-07-09 | 平安科技（深圳）有限公司 | A kind of method, server and storage medium identifying vehicle violation behavior |
| CN110069961A (en) * | 2018-01-24 | 2019-07-30 | 北京京东尚科信息技术有限公司 | A kind of object detecting method and device |
| CN110096981A (en) * | 2019-04-22 | 2019-08-06 | 长沙千视通智能科技有限公司 | A kind of video big data traffic scene analysis method based on deep learning |
| CN110197589A (en) * | 2019-05-29 | 2019-09-03 | 杭州诚道科技股份有限公司 | A kind of illegal detection method of making a dash across the red light based on deep learning |
| CN110287756A (en) * | 2018-03-19 | 2019-09-27 | 株式会社东芝 | Identification device and Vehicular system |
| CN110490077A (en) * | 2019-07-18 | 2019-11-22 | 北京工业大数据创新中心有限公司 | A kind of intelligence is broken rules and regulations recognition methods and system |
| CN110633595A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device by utilizing bilinear interpolation |
| CN111160206A (en) * | 2019-12-24 | 2020-05-15 | 国汽(北京)智能网联汽车研究院有限公司 | Traffic environment element visual perception method and device |
| CN111199647A (en) * | 2018-11-16 | 2020-05-26 | 中电科新型智慧城市研究院有限公司 | Monitoring video detection method for continuous lane changing and illegal turning of road vehicles |
| CN111209923A (en) * | 2020-04-23 | 2020-05-29 | 北京慧智数据科技有限公司 | Deep learning technology-based muck truck cover or uncover identification method |
| CN111507162A (en) * | 2019-01-30 | 2020-08-07 | 斯特拉德视觉公司 | Blind spot warning method and device based on cooperation of communication between vehicles |
| CN111695403A (en) * | 2020-04-19 | 2020-09-22 | 东风汽车股份有限公司 | 2D and 3D image synchronous detection method based on depth perception convolutional neural network |
| CN112041908A (en) * | 2018-04-27 | 2020-12-04 | 上海趋视信息科技有限公司 | System and method for monitoring traffic sign violations |
| WO2021008039A1 (en) * | 2019-07-17 | 2021-01-21 | Zhejiang Dahua Technology Co., Ltd. | Systems and methods for object monitoring |
| WO2021142944A1 (en) * | 2020-01-13 | 2021-07-22 | 南京新一代人工智能研究院有限公司 | Vehicle behaviour recognition method and apparatus |
| WO2021185379A1 (en) * | 2020-03-20 | 2021-09-23 | 长沙智能驾驶研究院有限公司 | Dense target detection method and system |
| CN113593239A (en) * | 2021-08-12 | 2021-11-02 | 上海仙塔智能科技有限公司 | Method and server for monitoring violation among vehicles and vehicle |
| CN114677652A (en) * | 2022-05-30 | 2022-06-28 | 武汉博观智能科技有限公司 | Illegal behavior monitoring method and device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP0565864A1 (en) * | 1992-04-16 | 1993-10-20 | Inventio Ag | Artificially intelligent traffic modelling and prediction system |
| CN102902955A (en) * | 2012-08-30 | 2013-01-30 | 中国科学技术大学 | Method and system for intelligently analyzing vehicle behaviour |
| CN106326858A (en) * | 2016-08-23 | 2017-01-11 | 北京航空航天大学 | Road traffic sign automatic identification and management system based on deep learning |
-
2017
- 2017-01-19 CN CN201710037410.7A patent/CN106886755A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP0565864A1 (en) * | 1992-04-16 | 1993-10-20 | Inventio Ag | Artificially intelligent traffic modelling and prediction system |
| CN102902955A (en) * | 2012-08-30 | 2013-01-30 | 中国科学技术大学 | Method and system for intelligently analyzing vehicle behaviour |
| CN106326858A (en) * | 2016-08-23 | 2017-01-11 | 北京航空航天大学 | Road traffic sign automatic identification and management system based on deep learning |
Non-Patent Citations (5)
| Title |
|---|
| 凌志刚: "《无人机景象匹配辅助导航技术》", 31 August 2016 * |
| 徐常青等: "《数学实验与软件计算》", 31 January 2014 * |
| 董艳梅: ""基于相关滤波器的目标跟踪技术"", 《中国优秀硕士学位论文全文数据库》 * |
| 钟晓明等: ""基于快速区域卷积神经网络的交通标志识别算法研究"", 《2016中国汽车工程学会年会论文集》 * |
| 阎平凡等: "《人工神经网络与模拟进化计算》", 30 November 2000 * |
Cited By (49)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107609491A (en) * | 2017-08-23 | 2018-01-19 | 中国科学院声学研究所 | A kind of vehicle peccancy parking detection method based on convolutional neural networks |
| CN107609491B (en) * | 2017-08-23 | 2020-05-26 | 中国科学院声学研究所 | Vehicle illegal parking detection method based on convolutional neural network |
| CN107729895A (en) * | 2017-10-18 | 2018-02-23 | 吉林大学 | A kind of intelligent vehicle ADAS aims of systems detection method and device |
| CN109684901B (en) * | 2017-10-19 | 2023-06-06 | 富士通株式会社 | Image processing apparatus and image processing method |
| CN109684901A (en) * | 2017-10-19 | 2019-04-26 | 富士通株式会社 | Image processing apparatus and image processing method |
| CN108010025B (en) * | 2017-12-14 | 2022-05-13 | 浙江大学 | Switch and indicator lamp positioning and state identification method of screen cabinet based on RCNN |
| CN108010025A (en) * | 2017-12-14 | 2018-05-08 | 浙江大学 | Switch and indicator lamp positioning and state identification method of screen cabinet based on RCNN |
| CN108171752A (en) * | 2017-12-28 | 2018-06-15 | 成都阿普奇科技股份有限公司 | A kind of sea ship video detection and tracking based on deep learning |
| CN110069961A (en) * | 2018-01-24 | 2019-07-30 | 北京京东尚科信息技术有限公司 | A kind of object detecting method and device |
| CN108090918A (en) * | 2018-02-12 | 2018-05-29 | 天津天地伟业信息系统集成有限公司 | A kind of Real-time Human Face Tracking based on the twin network of the full convolution of depth |
| CN108509978A (en) * | 2018-02-28 | 2018-09-07 | 中南大学 | The multi-class targets detection method and model of multi-stage characteristics fusion based on CNN |
| CN108509978B (en) * | 2018-02-28 | 2022-06-07 | 中南大学 | Multi-class target detection method and model based on CNN (CNN) multi-level feature fusion |
| CN110287756A (en) * | 2018-03-19 | 2019-09-27 | 株式会社东芝 | Identification device and Vehicular system |
| CN108536287A (en) * | 2018-03-26 | 2018-09-14 | 深圳市深晓科技有限公司 | A kind of method and device indicating reading according to user |
| CN108536287B (en) * | 2018-03-26 | 2021-03-02 | 深圳市同维通信技术有限公司 | Method and device for reading according to user instruction |
| CN112041908A (en) * | 2018-04-27 | 2020-12-04 | 上海趋视信息科技有限公司 | System and method for monitoring traffic sign violations |
| US11594030B2 (en) | 2018-04-27 | 2023-02-28 | Shanghai Truthvision Information Technology Co., Ltd. | Systems and methods for monitoring traffic sign violation |
| CN108985145A (en) * | 2018-05-29 | 2018-12-11 | 同济大学 | The Opposite direction connection deep neural network model method of small size road traffic sign detection identification |
| CN108898078A (en) * | 2018-06-15 | 2018-11-27 | 上海理工大学 | A kind of traffic sign real-time detection recognition methods of multiple dimensioned deconvolution neural network |
| CN110633595B (en) * | 2018-06-21 | 2022-12-02 | 北京京东尚科信息技术有限公司 | Target detection method and device by utilizing bilinear interpolation |
| CN110633595A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device by utilizing bilinear interpolation |
| CN108932850A (en) * | 2018-06-22 | 2018-12-04 | 安徽科力信息产业有限责任公司 | A kind of record motor vehicle runs at a low speed the method and device of illegal activities |
| CN109190442A (en) * | 2018-06-26 | 2019-01-11 | 杭州雄迈集成电路技术有限公司 | A kind of fast face detecting method based on depth cascade convolutional neural networks |
| CN109190442B (en) * | 2018-06-26 | 2021-07-06 | 杭州雄迈集成电路技术股份有限公司 | Rapid face detection method based on deep cascade convolution neural network |
| CN109376572A (en) * | 2018-08-09 | 2019-02-22 | 同济大学 | Real-time vehicle detection and trace tracking method in traffic video based on deep learning |
| CN109376572B (en) * | 2018-08-09 | 2022-05-03 | 同济大学 | Real-time vehicle detection and trajectory tracking method in traffic video based on deep learning |
| CN109190687A (en) * | 2018-08-16 | 2019-01-11 | 新智数字科技有限公司 | A kind of nerve network system and its method for identifying vehicle attribute |
| CN109344779A (en) * | 2018-10-11 | 2019-02-15 | 高新兴科技集团股份有限公司 | A kind of method for detecting human face under ring road scene based on convolutional neural networks |
| CN109326124A (en) * | 2018-10-17 | 2019-02-12 | 江西洪都航空工业集团有限责任公司 | A kind of urban environment based on machine vision parks cars Activity recognition system |
| CN111199647A (en) * | 2018-11-16 | 2020-05-26 | 中电科新型智慧城市研究院有限公司 | Monitoring video detection method for continuous lane changing and illegal turning of road vehicles |
| CN109740424A (en) * | 2018-11-23 | 2019-05-10 | 深圳市华尊科技股份有限公司 | Traffic violations recognition methods and Related product |
| CN109685030A (en) * | 2018-12-29 | 2019-04-26 | 哈尔滨理工大学 | A kind of mug rim of a cup defects detection classification method based on convolutional neural networks |
| CN111507162B (en) * | 2019-01-30 | 2023-11-07 | 斯特拉德视觉公司 | Blind spot warning method and device based on cooperation of inter-vehicle communication |
| CN111507162A (en) * | 2019-01-30 | 2020-08-07 | 斯特拉德视觉公司 | Blind spot warning method and device based on cooperation of communication between vehicles |
| CN109993056A (en) * | 2019-02-25 | 2019-07-09 | 平安科技（深圳）有限公司 | A kind of method, server and storage medium identifying vehicle violation behavior |
| WO2020173022A1 (en) * | 2019-02-25 | 2020-09-03 | 平安科技（深圳）有限公司 | Vehicle violation identifying method, server and storage medium |
| CN110096981A (en) * | 2019-04-22 | 2019-08-06 | 长沙千视通智能科技有限公司 | A kind of video big data traffic scene analysis method based on deep learning |
| CN110197589A (en) * | 2019-05-29 | 2019-09-03 | 杭州诚道科技股份有限公司 | A kind of illegal detection method of making a dash across the red light based on deep learning |
| WO2021008039A1 (en) * | 2019-07-17 | 2021-01-21 | Zhejiang Dahua Technology Co., Ltd. | Systems and methods for object monitoring |
| CN110490077B (en) * | 2019-07-18 | 2022-04-15 | 北京工业大数据创新中心有限公司 | Intelligent violation identification method and system |
| CN110490077A (en) * | 2019-07-18 | 2019-11-22 | 北京工业大数据创新中心有限公司 | A kind of intelligence is broken rules and regulations recognition methods and system |
| CN111160206A (en) * | 2019-12-24 | 2020-05-15 | 国汽(北京)智能网联汽车研究院有限公司 | Traffic environment element visual perception method and device |
| WO2021142944A1 (en) * | 2020-01-13 | 2021-07-22 | 南京新一代人工智能研究院有限公司 | Vehicle behaviour recognition method and apparatus |
| WO2021185379A1 (en) * | 2020-03-20 | 2021-09-23 | 长沙智能驾驶研究院有限公司 | Dense target detection method and system |
| CN111695403A (en) * | 2020-04-19 | 2020-09-22 | 东风汽车股份有限公司 | 2D and 3D image synchronous detection method based on depth perception convolutional neural network |
| CN111695403B (en) * | 2020-04-19 | 2024-03-22 | 东风汽车股份有限公司 | Depth perception convolutional neural network-based 2D and 3D image synchronous detection method |
| CN111209923A (en) * | 2020-04-23 | 2020-05-29 | 北京慧智数据科技有限公司 | Deep learning technology-based muck truck cover or uncover identification method |
| CN113593239A (en) * | 2021-08-12 | 2021-11-02 | 上海仙塔智能科技有限公司 | Method and server for monitoring violation among vehicles and vehicle |
| CN114677652A (en) * | 2022-05-30 | 2022-06-28 | 武汉博观智能科技有限公司 | Illegal behavior monitoring method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106886755A (en) | A kind of intersection vehicles system for detecting regulation violation based on Traffic Sign Recognition | |
| CN110188705B (en) | Remote traffic sign detection and identification method suitable for vehicle-mounted system | |
| CN104392463B (en) | Image salient region detection method based on joint sparse multi-scale fusion | |
| CN109614985A (en) | A kind of object detection method based on intensive connection features pyramid network | |
| CN108171112A (en) | Vehicle identification and tracking based on convolutional neural networks | |
| CN107730904A (en) | Multitask vehicle driving in reverse vision detection system based on depth convolutional neural networks | |
| CN107730906A (en) | Zebra stripes vehicle does not give precedence to the vision detection system of pedestrian behavior | |
| CN105809198B (en) | SAR image target recognition method based on depth confidence network | |
| CN109508715A (en) | A kind of License Plate and recognition methods based on deep learning | |
| CN106682569A (en) | Fast traffic signboard recognition method based on convolution neural network | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN105354568A (en) | Convolutional neural network based vehicle logo identification method | |
| CN108805093A (en) | Escalator passenger based on deep learning falls down detection algorithm | |
| CN107134144A (en) | A kind of vehicle checking method for traffic monitoring | |
| CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
| CN106778854A (en) | Activity recognition method based on track and convolutional neural networks feature extraction | |
| CN111460919B (en) | Monocular vision road target detection and distance estimation method based on improved YOLOv3 | |
| CN106372571A (en) | Road traffic sign detection and identification method | |
| CN107133974A (en) | The vehicle type classification method that Gaussian Background modeling is combined with Recognition with Recurrent Neural Network | |
| CN107016357A (en) | A kind of video pedestrian detection method based on time-domain convolutional neural networks | |
| CN105069468A (en) | Hyper-spectral image classification method based on ridgelet and depth convolution network | |
| CN108280397A (en) | Human body image hair detection method based on depth convolutional neural networks | |
| He et al. | A robust method for wheatear detection using UAV in natural scenes | |
| CN109978882A (en) | A kind of medical imaging object detection method based on multi-modal fusion | |
| CN109034024B (en) | Logistics vehicle type classification and identification method based on image target detection |
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: 20170623 |
|
| RJ01 | Rejection of invention patent application after publication |