CN107194346A - A kind of fatigue drive of car Forecasting Methodology - Google Patents

A kind of fatigue drive of car Forecasting Methodology Download PDF

Info

Publication number
CN107194346A
CN107194346A CN201710356372.1A CN201710356372A CN107194346A CN 107194346 A CN107194346 A CN 107194346A CN 201710356372 A CN201710356372 A CN 201710356372A CN 107194346 A CN107194346 A CN 107194346A
Authority
CN
China
Prior art keywords
mrow
msubsup
convolutional neural
neural networks
face
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
Application number
CN201710356372.1A
Other languages
Chinese (zh)
Inventor
曾智勇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fujian Normal University
Original Assignee
Fujian Normal University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Fujian Normal University filed Critical Fujian Normal University
Priority to CN201710356372.1A priority Critical patent/CN107194346A/en
Publication of CN107194346A publication Critical patent/CN107194346A/en
Pending legal-status Critical Current

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
    • G06V20/59—Context or environment of the image inside of a vehicle, e.g. relating to seat occupancy, driver state or inner lighting conditions
    • G06V20/597—Recognising the driver's state or behaviour, e.g. attention or drowsiness
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
    • B60W40/08—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models related to drivers or passengers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • G—PHYSICS
    • G08—SIGNALLING
    • G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
    • G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
    • G08B21/02—Alarms for ensuring the safety of persons
    • G08B21/06—Alarms for ensuring the safety of persons indicating a condition of sleep, e.g. anti-dozing alarms
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W40/00—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models
    • B60W40/08—Estimation or calculation of non-directly measurable driving parameters for road vehicle drive control systems not related to the control of a particular sub unit, e.g. by using mathematical models related to drivers or passengers
    • B60W2040/0818—Inactivity or incapacity of driver
    • B60W2040/0827—Inactivity or incapacity of driver due to sleepiness

Abstract

The present invention discloses a kind of fatigue drive of car Forecasting Methodology, build set gradually one to level Four convolutional neural networks, input picture obtains candidate face window and its corresponding bounding box regression vector by one-level convolutional neural networks, then by the candidate window of one or two grades of convolutional neural networks merging high superposeds；Remaining candidate window is by three-level convolutional neural networks and utilizes human face characteristic point label information Forecasting recognition human eye area；Split eye areas again according to eye feature point, and input deep vision characteristic model of the level Four convolutional neural networks by deep learning Algorithm for Training eye image；The video of camera collection passes sequentially through CNN1, CNN2, CNN3, CNN4, differentiates the closure state of eyes；Driver fatigue visual assessment parameter PERCLOS is calculated, when PERCLOS values are more than 40%, then judge that driver starts fatigue or in fatigue state, exports pre-warning signal.The method of the present invention can not only detect the fatigue state of driver under the conditions of various illumination, posture, expression, and testing result robustness is high, effectively overcomes the influence that the factors such as illumination, posture, expression are detected to driver fatigue.

Description

A kind of fatigue drive of car Forecasting Methodology
Technical field
The present invention relates to computer vision and image processing field, more particularly to a kind of fatigue drive of car Forecasting Methodology.
Background technology
With expanding economy, the owning amount sharp increase of automobile.According to the scheme of seminar of Development Research Center of the State Council Predict the outcome, China's automobile guarantee-quantity reached 56,690,000 in 2010, reached within 2016 194000000, the year two thousand twenty will exceed 2.5 hundred million.Highway mileage increasingly increases, and the average speed of highway communication has large increase, and this is undoubtedly to economic hair Exhibition serves facilitation.But traffic accident simultaneously also accordingly increases, the safety of life and property to the people brings very big prestige The side of body.The data statistics provided according to Chinese Ministry of Public Security's traffic control department, road traffic accident occurs for China 66.7 ten thousand times within 2003.It is dead 10.44 ten thousand people, direct economic loss is up to 33.7 hundred million yuan.AIAE American institute of automobile engineers is counted, and the U.S. there are about 260,000 friendships every year Interpreter's event.Substantial amounts of fatigue driving road traffic accident causes incalculable damage to the life and property of people.Fatigue The fatigue state of driver can be detected by driving detecting system, and provide early warning in advance, allow the driver that there is fatigue state to exist Service area rest or urgent stopped in place of safety nearby is rested, it is to avoid the generation of various Traffic Accidents.This technology is related to And the technology of the multiple fields such as computer vision, image procossing, pattern-recognition, artificial intelligence, neutral net, machine learning.Closely Nian Lai, in terms of fatigue detecting, researcher both domestic and external sets out from different angles proposes many fruitful methods, root According to the difference of starting point, sum up, wherein representational method can substantially be divided into two classes：One class is the fatigue of feature based Detection method, such as document 1：W.O.Lee,E.C.Lee,K.R.Park,Blink detection robust to various The document 2 of facial poses, J.Neurosci.Methods 193 (2) (2010) 356-372.：K.Arai, R.Mardiyanto,Real time blinking detection based on Gabor filter, Int.J.Hum.Comput.Interact.1 (3) (2010) 33-40. and document 3：L.Lu,X.Ning,M.Qian,Y.Zhao, Close eye detected based on synthesized gray projection,in:Advances in Described by Multimedia, Software Engineering and Computing, vol.2,2012, pp.345-351., Such method mainly discriminates whether fatigue, this kind of side by extracting the shape facility of iris and eyelid or the intensity profile of eyes The validity of method is easily influenceed by accuracy and the ambient lighting change positioned.Another kind of is based on apparent fatigue detecting side Method, such as document 4：Y.Wu,T.Lee,Q.Wu,H.Liu,An eye state recognition method for drowsinessdetection,in:Proceedings of Vehicular Technology Conference,2010, Pp.1-5. documents 5：L.Zhou,H.Wang,Open/closed eye recognition by local binary increasingintensity patterns,in:Proceedings of IEEE Conference on Robotics, Automationand Mechatronics, 2011, what pp.7-11. was introduced, such method passes through the shape for extracting description target The middle level features of shape and texture, such as LBP, Gabor characteristic, and utilize advanced Machine learning tools, such as SVM, Adaboost Deng obtaining the result of robustness under good image-forming condition.However, above-mentioned two classes method is easily by illumination variation, image Fuzzy, eyes are blocked the influence of the accuracy positioned with eyes, so they are not suitable under the image-forming condition of challenge Fatigue detecting.
Last century the fifties end, Rosenblatt proposes artificial neural network theories for the first time.1984 Hopfield has extracted complete image from images fragment, so as to excite the research boom of artificial neural network, then Researcher proposes many neural network models, such as BP, RBF, RNN, and to last century the nineties, artificial neural network is in figure Develop on a large scale very much as having in treatment theory, and combine other mathematical tools such as statistics, Optimum Theory, morphology etc., formed Than more complete theoretical system.Image processing method based on neutral net has good Fundamentals of Mathematics, the deep theoretical back of the body Scape, certain achievement is achieved in terms of Image Edge-Detection, segmentation, reparation, cluster, classification and prediction.
Inspired by depth hierarchical structure in brain, and to can by feature extraction and classify the two processes it is organic Combination active demand, grind the person of making internal disorder or usurp and the concept of deep learning be incorporated into the research of artificial neural network, but computer is hard The appearance of development restriction, big data of part and the limitation of neutral net make this work become abnormal difficult.Researcher makes In the case of random initializtion network weight with standard, very poor training and extensive result are generally all obtained, so that obtaining not To good depth network structure.2006, Hinton et al. was using successively greedy algorithm and Back Propagation Algorithm realize one The depth network structure for being called deep belief network (Deep Belief Networks, DBNs) is planted, experiment knot well is achieved Really.
Modern physiology, which is studied, to be shown, the image processing process phase in the understanding and convolutional neural networks of the vision system of people Unanimously.LeCun et al. is designed based on stochastic gradient descent algorithm and be trained depth convolutional neural networks LeNet, is shown Relative to the leading performance of other methods.2012, the depth convolutional network model of Hinton project trainings contained 6,000 ten thousand Parameter, it from ImageNet go to school acquistion to character representation classification capacity far surpass engineer feature, with very strong Generalization ability, it is possible to be successfully applied in other data sets and task, such as object detection, tracking and retrieve.
Depth model has a powerful learning ability, efficient feature representation ability, from Pixel-level initial data to abstract Semantic concept successively extract information.This causes it to have protrusion in terms of the global characteristics and contextual information that extract image Advantage.This brings new thinking to solve computer vision problem such as image segmentation and critical point detection etc..With the figure of face Exemplified by segmentation.In order to predict that each pixel belongs to which face's organ (eyes, nose, mouth, hair), the common practice be A small region, texture feature extraction (such as local binary patterns) are taken around the pixel, then support is utilized based on this feature The shallow Models such as vector machine are classified.Because regional area is limited comprising information content, classification error is often produced, therefore will be to segmentation Image afterwards adds the constraint such as smooth and shape prior.Even if in the case of in fact there is partial occlusion, human eye can also root It is blocked the mark at place according to the information estimation in the other regions of face.This means the information of global and context is sentenced for local Disconnected is very important, and these information just lost in the method based on local feature from the most incipient stage.Deep learning It can learn to express for the layered characteristic of facial image, it is first that model implicitly adds shape in high dimensional data transfer process Test.The bottom can portray local edge and textural characteristics from original pixels learning filters；By to various edge filters Device is combined, and middle level wave filter can describe different types of human face；The top overall situation for describing whole face Feature.In highest hidden layer, each neuron represents an attributive classification device.
It is good for the image classification effect based on deep learning, while big data and computer hardware develop rapidly promotion The fact that deep learning classifying quality, there is the method that researcher carries out recognition of face using big data, but due to the target of eyes It is smaller, be related to block, expression shape change, light change etc. complex situations, so at present this kind of method can not be applicable the fatigue of eyes Detection, the fatigue detection method of the cascade deep convolutional neural networks based on multi-task learning is not seen still.
The content of the invention
It is an object of the invention to overcome the deficiencies of the prior art and provide the automobile that a kind of robust is efficient, Detection results are good Fatigue driving Forecasting Methodology.
The technical solution adopted by the present invention is：
A kind of fatigue drive of car Forecasting Methodology, it comprises the following steps：
Step 1, any piece image f (x, y) given for training dataset, multiple dimensioned setting image is with structural map picture Pyramid is simultaneously used as the input of following cascade framework；
Step 2, multiple dimensioned training image collection is inputted into one-level convolutional neural networks CNN1, obtains one-level candidate face window And its corresponding one-level bounding box regression vector；One-level candidate window is corrected using one-level bounding box regression vector, and Merge the one-level candidate window of high superposed using non-maxima suppression NMS；
Step 3, corrected one-level candidate window is fully entered to two grades of convolutional neural networks CNN2, obtains two grades Candidate face window and its corresponding secondary packet enclose box regression vector；Secondary packet is enclosed into box regression vector to two grades of candidate windows again It is secondary to be corrected, and application non-maxima suppression NMS merges two grades of high candidate windows of degree of overlapping；
Step 4, three-level convolutional neural networks CNN3 will be fully entered by the two grades of candidate windows corrected again, utilized Human face characteristic point in human face characteristic point label information prognostic chart picture, and export 5 characteristic point positions of face；
Step 5,2 human eye characteristic point positions in 5 human face characteristic points, are partitioned into 24 centered on characteristic point × 24 eye areas, the eye areas image set for splitting alignment is input in level Four convolutional neural networks CNN4, eye opening is used With eye closing label information, human eye deep vision characteristic model is obtained by stochastic gradient descent algorithm SGD；
Step 6, the test image that camera is gathered is passed sequentially through into CNN1, CNN2, CNN3, CNN4, currently tested The closure state of video frame images eyes；
Step 7, driver fatigue visual assessment parameter PERCLOS is calculated.Obtained 3 seconds according to camera acquisition rate first The totalframes that video bag contains in time, the frame number of the human eye closure included in video for 3 seconds is calculated then according to step 6. PERCLOSE calculation formula is：
When PERCLOS values are more than 40%, then judge that driver starts fatigue or in fatigue state, output alarm shape State.
Further, two are represented as according to the object functions learnt of one-level convolutional neural networks CNN1 in the step 2 Class classification problem, the two classifications are mankind's mark and non-face mark respectively；For each sample xiUsing intersection entropy loss It is used as object function：
Wherein, piIt is the expression sample x that neutral net is producediIt is the probability of human eye,Face is represented for 1, be that 0 expression is non- Face,Represent sample xiIt is the loss of face.CNN1 training goal is exactly to pass through the neutral net CNN1 sides of making The Classification Loss of journey (1) is minimized.
Further, a recurrence is expressed as according to the learning objective of two grades of convolutional neural networks CNN2 in the step 3 Problem, for each sample xiObject function is used as using Euclidean loss：
Wherein,For sample xiPrediction bounding box and real bounding box loss,For sample xiPrediction surround Box,For the real bounding box of sample.CNN2 training goal is exactly by training neutral net CNN2 to make the people of equation (2) Face and non-face prediction minimization of loss.
Further, a recurrence is expressed as according to the learning objective of three-level convolutional neural networks CNN3 in the step 4 Problem, for each sample xiObject function is used as using Euclidean loss：
Wherein,For sample xiPrediction face characteristic point coordinates and real face characteristic point coordinates lose, For sample xiPrediction face characteristic point coordinates,For the real face characteristic point coordinates of sample.CNN3 training goal is just It is by training neutral net CNN3 to make the face characteristic point prediction minimization of loss of equation (3).
Further, according to using the error function Triplet Loss based on metric learning in the step 5 as god Object function through network.If concentrating one sample of random selection (to be designated as Anchor from training data), then respectively from One sample of random selection is concentrated (to be designated as Positive with Anchor same class and inhomogeneous training data) and Negative (is designated as), (Anchor, Positive, a Negative) triple is constituted, then level Four convolutional neural networks CNN4 learning objectives are to make triple classification minimization of loss：
In equation (4), triple Classification Loss is respectively to the gradient of three sample characteristics：
Wherein, L represents triple Classification Loss,Respectively represent random sample, positive example sample and The feature of negative example sample, α is representedWithThe distance between withWithThe distance between minimum interval.
Further, step 2, step 3, step 4 and step 5 use stochastic gradient descent algorithm iterative learning target letter Count to classify or predict face and non-face candidate window and bounding box regression vector, prediction human face characteristic point, differentiate eyes Closure state is opened, the characteristic pattern, pond characteristic pattern and activation characteristic pattern of each layer of network are obtained by propagated forward；After recycling Weight and the biasing of each layer of network are updated to propagation algorithm, until network convergence or termination.
Further, when stochastic gradient descent algorithm iterative learning object function method is applied to the one-level convolution of step 2 During neutral net CNN1, the input of the corresponding convolutional neural networks of step 2 is the multi-Scale Pyramid image of training dataset；
When stochastic gradient descent algorithm iterative learning object function method is applied to two grades of convolutional neural networks of step 3 During CNN2, the input of the corresponding convolutional neural networks of step 3 trains obtained multiple dimensioned facial image for step 2；
When stochastic gradient descent algorithm iterative learning object function method is applied to the three-level convolutional neural networks of step 4 During CNN3, the input of the corresponding convolutional neural networks of step 4 trains obtained multiple dimensioned facial image for step 3；
When stochastic gradient descent algorithm iterative learning object function method is applied to the level Four convolutional neural networks of step 5 During CNN4, the input of the corresponding convolutional neural networks of step 5 constitutes ternary to split obtained eye image and non-ocular image Group training setIn every piece image.
Further, when stochastic gradient descent algorithm iterative learning object function method is applied to the level Four convolution of step 5 During neutral net CNN4, the activation value that step 5 obtains output layer is the activation value of correspondence tripleWith
The specific calculating process of step 5 is：For output layer n-thlLayer, is calculated respectively according to formula (5), (6), (7) Partial derivative of the triple Classification Loss to the activation value of tripleAnd accumulate per piece image It is used as the median of back-propagating；For each layer l=n before output layerl-1,nl-2,..., 2 each layers, calculate each layer respectivelyAnd by each layer per piece image The median as back-propagating is accumulated, the weight and biasing coefficient of neutral net are updated until net using above-mentioned median Network is restrained.
Further, the image input step to be tested 2 that camera is gathered, step 3, step 4, step 5 are trained In network, the closure classification information of human eye is directly obtained.
The present invention uses above technical scheme, and advantage is compared with existing fatigue detection method：Accuracy of detection is high, in real time Property it is good, robustness is high, effectively overcomes the influence of the factors such as light change, Changes in weather.The method of the present invention can not only The fatigue state of ordinary ray servant is detected, but also the fatigue state of dark and obstruction conditions servant can be detected.
Brief description of the drawings
The present invention is described in further details below in conjunction with the drawings and specific embodiments；
Fig. 1 is a kind of schematic flow sheet of fatigue drive of car Forecasting Methodology of the invention；
Fig. 2 is a kind of one-level convolutional neural networks CNN1 of fatigue drive of car Forecasting Methodology of the invention structural representation Figure；
Fig. 3 is a kind of two grades of convolutional neural networks CNN2 of fatigue drive of car Forecasting Methodology structural representation of the invention Figure；
Fig. 4 is a kind of three-level convolutional neural networks CNN3 of fatigue drive of car Forecasting Methodology of the invention structural representation Figure；
Fig. 5 is a kind of level Four convolutional neural networks CNN4 of fatigue drive of car Forecasting Methodology of the invention structural representation Figure.
Embodiment
As shown in one of Fig. 1-5, the invention discloses a kind of fatigue drive of car Forecasting Methodology, it comprises the following steps：
Step 1, any piece image f (x, y) given for training dataset, multiple dimensioned setting image is with structural map picture Pyramid is simultaneously used as the input of neutral net under following cascade framework；
In the present invention, it is 1.2 to produce multi-scale image pyramidal scale factor, 1.0,0.8.Can using scale factor The training of three yardsticks of generation and test chart image set.Network inputs when they learn and tested respectively as training.
In the present invention, training dataset presses 3 by positive example (face), negative example (non-face), part face and feature face image： 1：1：2 ratios are constituted.Negative example：Handed over benchmark test facial image and compare the image-region less than 0.3.Positive example：With benchmark test Facial image is handed over and compares the image-region more than 0.65.Part face：Hand over and compare in 0.4-0.65 with benchmark test facial image Between image-region.Eigenface：Include the face of 5 characteristic points.
Step 2, multiple dimensioned training dataset is inputted in convolutional neural networks CNN1, obtains candidate face window and its right The bounding box regression vector answered；Then above-mentioned candidate window is corrected using the bounding box regression vector of estimation, and applied Non-maxima suppression NMS merges the candidate window of high superposed；
Further, convolutional neural networks CNN1 is designed according to the characteristics of recognition of face in the step 2.CNN1's Architecture is as shown in Figure 2.The size of input picture is 12 × 12, and the size of 10 wave filters of first layer of neutral net is 3 × 3, learning rate is 0.0001, decays to 0.01, step-length is 1, and filter weight parameter is obtained using " xavier " algorithm, the layer Network is biased to 0,10 wave filters and presses formula (13) convolution with input picture respectively, obtains 10 characteristic patterns；Utilize maximum pond Method carries out pond to features described above figure, and pond window size is 2 × 2, and step-length is 1, obtains 10 new feature figures；Then use The activation of these characteristic patterns is 10 characteristic patterns by PReLU activation primitives.Second convolutional layer of neutral net is by 10 of activation Characteristic pattern and 16 3 × 3 wave filter convolution, obtain 16 characteristic patterns；Obtained using PReLU activation primitives activation features described above figure 16 new feature figures.In the 3rd convolutional layer of neutral net, newly activate 16 characteristic patterns and 32 3 × 3 wave filters are rolled up Product, obtains 32 characteristic patterns, them is activated using PReLU activation primitives, obtains 32 activation characteristic patterns.Finally, 32 activation Characteristic pattern and 21 × 1 wave filter convolution, obtain candidate's human eye window, Softmax graders identification candidate window is face area Domain and non-face region.Meanwhile, by 32 activation characteristic patterns again with 41 × 1 wave filter convolution, obtain bounding box corresponding Regression vector.32 activation characteristic patterns and 10 1 × 1 wave filter convolution are obtained to the coordinate of 5 characteristic points.
Step 3, corrected above-mentioned candidate window is fully entered into convolutional neural networks CNN2, obtains two grades of candidates Face window and its corresponding secondary packet enclose box regression vector；Box regression vector is enclosed using secondary packet to two grades of candidate windows again It is corrected, and application non-maxima suppression NMS merges two grades of high candidate windows of degree of overlapping；
Convolutional neural networks CNN2 architecture is as shown in Figure 3 in step 3.The corrected time that step 2 is obtained Select window to readjust for size 24 × 24 input picture, be input in convolutional neural networks CNN2.The first of neutral net The size of 28 wave filter of layer is 3 × 3, and learning rate is 0.0001, decays to 0.01, step-length is 1, and filter weight is utilized " Xavier " algorithms are obtained, and the layer network is biased to 0, and 28 wave filter of the layer press formula (13) convolution with input picture respectively, obtain 28 characteristic patterns；Then pond is carried out to features described above figure using maximum pond method, pond window size is 3 × 3, and step-length is 2, obtain 28 new feature figures；Then 28 activation characteristic patterns are obtained using PReLU activation primitives activation features described above figure.In net Second convolutional layer of network model, first by this 28 activation characteristic patterns and 48 3 × 3 wave filter convolution, obtains 48 characteristic patterns； Secondly pond is carried out to features described above figure using maximum pond method, pond window size is 3 × 3, and step-length is 2, obtains 48 New feature figure；Reuse PReLU activation primitives and obtain 48 activation characteristic patterns.It is first in the 3rd convolutional layer of network structure Characteristic pattern and 64 2 × 2 wave filter convolution first are activated by 48 of acquisition, 64 characteristic patterns are obtained；Then activated using PReLU Function activates them, obtains 64 activation characteristic patterns.In the 4th convolutional layer of model, by 64 activation characteristic patterns of last layer with 128 1 × 1 wave filter convolution, obtain 128 inner product values.Them are activated using PReLU activation primitives to obtain in 128 activation Product value.In the 5th convolutional layer of neutral net, 128 activation features and 21 × 1 wave filter convolution obtain candidate face window Mouthful.Finally, these candidate face windows are identified as face and non-face region by Softmax graders.Meanwhile, by 128 Individual activation feature and 41 × 1 wave filter convolution obtain the corresponding regression vector of bounding box.By 128 activation characteristic patterns and 10 1 × 1 wave filter convolution obtains the coordinate of 5 characteristic points.
Step 4, two grades of corrected candidate windows are fully entered into three-level convolutional neural networks CNN3, utilizes face Human face characteristic point in characteristic point mark prognostic chart picture, and export 5 characteristic point positions of face；
Three-level convolutional neural networks CNN3 architecture is as shown in Figure 4 in step 4.By step 3 obtain it is corrected Two grades of candidate windows readjust as the input picture of size 48 × 48, be input in another convolutional neural networks CNN3. The size of CNN3 32 wave filters of first layer is 3 × 3, and learning rate is 0.0001, decays to 0.1, step-length is 1, wave filter power Weight coefficient is obtained using " xavier " algorithm, and the layer network is biased to 0.32 wave filter of the layer are pressed with above-mentioned input picture respectively Formula (13) convolution, obtains 32 characteristic patterns；Then pond is carried out to features described above figure using maximum pond method, pond window is big Small is 3 × 3, and step-length is 2, obtains 32 new feature figures；Then the characteristic pattern behind above-mentioned pond is activated using PReLU activation primitives Obtain 32 activation characteristic patterns.In CNN3 second convolutional layer, by the characteristic pattern and 64 3 × 3 wave filters of this 32 activation Convolution, obtains 64 characteristic patterns；Pond is carried out to features described above figure using maximum pond method, pond window size is 3 × 3, Step-length is 2, obtains 64 pond characteristic patterns；Then 64 are obtained using the above-mentioned pond characteristic pattern of PReLU activation primitives activation to swash Characteristic pattern living.In CNN3 the 3rd convolutional layer, by above-mentioned 64 activation characteristic patterns and 64 3 × 3 wave filter convolution, 64 are obtained Individual characteristic pattern；Pond is carried out to features described above figure using maximum pond method, pond window size is 2 × 2, and step-length is 2, is obtained 64 pond characteristic patterns；Then them are activated using PReLU activation primitives, obtains 64 activation characteristic patterns.The 4th of CNN3 the Individual convolutional layer, by 64 activation characteristic patterns and 128 2 × 2 wave filter convolution, obtains 128 characteristic patterns；Then swashed using PReLU Function living activates the characteristic pattern that they obtain 128 activation.In CNN3 the 5th convolutional layer, feature is activated by above-mentioned 128 Figure and 256 3 × 3 wave filter convolution, obtain 256 characteristic patterns.In order to prevent over-fitting, the generalization ability of model is improved, this One layer employs Dropout technologies in training, trains the hidden node for all abandoning 25% to be trained every time.Then use PReLU activation primitives activate the characteristic pattern that they obtain 256 activation.In CNN3 the 6th convolutional layer, 256 are activated Characteristic pattern and 21 × 1 wave filters carry out inner product, obtain candidate face window.Finally, candidate face window is inputted into Softmax Grader is identified as face and non-face region.256 activation characteristic patterns are wrapped with 41 × 1 wave filter convolution simultaneously Enclose the corresponding regression vector of box.256 activation characteristic patterns and 10 1 × 1 wave filter convolution are obtained to the coordinate of 5 characteristic points.
Step 5,2 human eye characteristic point positions of the face detected according to step 4, are partitioned into centered on human eye feature point 24 × 24 eye areas, by split alignment eye areas image be input to level Four convolutional neural networks CNN4, pass through depth Learning algorithm trains human eye deep vision characteristic model；
In step 5, level Four convolutional neural networks CNN4 architecture is as shown in Figure 5.2 human eyes that step 4 is obtained Characteristic point position, is partitioned into 24 × 24 eye areas centered on human eye feature point.Human eye area is readjusted as size 48 × 48 input picture, is input in level Four convolutional neural networks CNN4.The size of CNN4 96 wave filters of first layer is 5 × 5, step-length is 1, filter weight using " xavier " algorithm obtain, the layer network be biased to 0,96 wave filters respectively with it is defeated Enter image by formula (13) convolution, obtain 96 characteristic patterns；96 characteristic patterns that obtain that calculating is obtained are respectively 1 with 96 sizes × 1 wave filter convolution, convolution step-length is 1, and 96 characteristic patterns are obtained again；Then using maximum pond method to features described above Figure carries out pond, and pond window size is 3 × 3, and step-length is 2, obtains 96, individual pond characteristic pattern；Reuse PReLU and activate 96 Pond characteristic pattern, obtains 96 activation characteristic patterns.CNN4 second convolutional layer, by above-mentioned 96 activation characteristic patterns and 256 3 × 3 wave filter convolution, obtain 256 characteristic patterns；By calculate obtain obtain 256 characteristic patterns respectively with 256 sizes be 1 × 1 wave filter convolution, convolution step-length is 1, and 256 characteristic patterns are obtained again；Features described above figure is entered using maximum pond method Row pond, pond window size is 3 × 3, and step-length is 2, obtains 256 pond characteristic patterns；Then obtained using PReLU activation primitives To 256 activation characteristic patterns.In CNN4 the 3rd convolutional layer, 256 characteristic patterns of activation and 384 3 × 3 wave filters are rolled up Product, obtains 384 characteristic patterns；384 characteristic patterns that obtain that calculating is obtained are rolled up with 384 sizes for 1 × 1 wave filter respectively Product, convolution step-length is 1, and 384 characteristic patterns are obtained again；Pond, Chi Hua are carried out to features described above figure using maximum pond method Window size is 2 × 2, and step-length is 2, obtains 384 pond characteristic patterns；Then them are activated using PReLU activation primitives, obtained 384 activation characteristic patterns.In CNN4 the 4th convolutional layer, 384 activation characteristic patterns and 128 3 × 3 wave filter convolution are obtained To 128 characteristic patterns；By calculating obtain 128 characteristic patterns respectively with the wave filter convolution that 128 sizes are 1 × 1, convolution step-length is 1,128 characteristic patterns are obtained again；Carry out pond to features described above figure using average pond method, pond window size for 4 × 4, step-length is 2, obtains 128 pond characteristic patterns；Then the spy that they obtain 128 activation is activated using PReLU activation primitives Levy figure.In CNN4 the 5th convolutional layer, 128 activation characteristic patterns and 256 1 × 1 wave filter convolution obtain 256 features Scheme, and them activated with PReLU activation primitives to obtain 256 activation characteristic patterns.In order to prevent over-fitting, the extensive of model is improved Ability, this layer employs Dropout technologies in training, trains the hidden node for all abandoning 25% to be trained every time.Make The characteristic pattern that they obtain 256 activation is activated with PReLU activation primitives.256 activation characteristic pattern with 128 1 × 1 filter Device carries out inner product, obtains 128 dimensional feature vectors.In CNN4 training, learning rate is set to 0.005, and interval α is set to 0.2. We use stochastic gradient descent SGD and AdaGrad the training CNN4 with standard back-propagating.
Step 6, the test image that camera is gathered is passed sequentially through into CNN1, CNN2, CNN3, CNN4, obtains test video The closure state of two field picture eyes；
Step 7, driver fatigue visual assessment parameter PERCLOS is calculated.Obtained 3 seconds according to camera acquisition rate first The totalframes that video bag contains in time, then calculates the frame number of the human eye closure included in video for 3 seconds.PERCLOSE calculating is public Formula is：
When PERCLOS values are more than 40%, then judge that driver starts fatigue or in fatigue state, output alarm shape State.
The present invention uses above technical scheme, and advantage is compared with existing fatigue detection method：The method of the present invention is not Be only capable of detecting normal illumination, posture, under the conditions of expression driver fatigue state, additionally it is possible to detect illumination condition change, posture And the driver fatigue state under the conditions of expression shape change etc., testing result robustness is high, effectively overcomes illumination, posture, table The influence that the factors such as feelings are detected to driver fatigue.
The present invention disclose a kind of fatigue drive of car Forecasting Methodology, structure set gradually one to level Four convolutional Neural net Network, input picture obtains candidate face window and its corresponding bounding box regression vector by one-level convolutional neural networks, then passes through One or two grades of convolutional neural networks merge the candidate window of high superposed；Remaining candidate window passes through three-level convolutional neural networks simultaneously Using human face characteristic point information prediction human face characteristic point, 5 human face characteristic point positions are exported；Again with 2 in 5 human face characteristic points Split eye areas centered on individual human eye characteristic point and pass through deep learning Algorithm for Training eyes figure by level Four convolutional neural networks The deep vision characteristic model of picture；The test image that camera is gathered passes sequentially through CNN1, CNN2, CNN3, CNN4, is surveyed Try the closure state of video frame images eyes；Driver fatigue visually assesses parameter in video in the 3 second time of calculating repeatedly PERCLOS, when PERCLOS values are more than 40%, then judges that driver starts fatigue or in fatigue state, output alarm shape State.The method of the present invention can not only detect the fatigue state of driver under the conditions of various illumination, posture, expression, testing result Shandong Rod is high, effectively overcomes the influence that the factors such as illumination, posture, expression are detected to driver fatigue.

Claims (8)

1. a kind of fatigue drive of car Forecasting Methodology, it is characterised in that：It comprises the following steps：
Step 1, any piece image f (x, y) given for training dataset, multiple dimensioned setting image is with the golden word of structural map picture Tower is simultaneously used as the input of following cascade framework；
Step 2, by multiple dimensioned training image collection input one-level convolutional neural networks CNN1, obtain one-level candidate face window and its Corresponding one-level bounding box regression vector；One-level candidate window is corrected using one-level bounding box regression vector, and applied Non-maxima suppression NMS merges the one-level candidate window of high superposed；
Step 3, corrected one-level candidate window is fully entered to two grades of convolutional neural networks CNN2, obtains two grades of candidates Face window and its corresponding secondary packet enclose box regression vector；Secondary packet is enclosed into box regression vector to enter two grades of candidate windows again Row correction, and application non-maxima suppression NMS merges two grades of high candidate windows of degree of overlapping；
Step 4, three-level convolutional neural networks CNN3 will be fully entered by the two grades of candidate windows corrected again, utilizes face Human face characteristic point in characteristic point label information prognostic chart picture, and export 5 characteristic point positions of face；
Step 5,2 human eye characteristic point positions in 5 characteristic points of the face exported according to step 4, are partitioned into characteristic point Centered on 24 × 24 eye areas, by split alignment eye areas image set be input to level Four convolutional neural networks CNN4 In, using eye opening and eye closing label information, human eye deep vision characteristic model is obtained by stochastic gradient descent algorithm SGD；
Step 6, the test image that camera is gathered is passed sequentially through into CNN1, CNN2, CNN3, CNN4, obtains test video frame figure As the closure state of eyes；
Step 7, driver fatigue visual assessment parameter PERCLOS is calculated：3 second time was obtained according to camera acquisition rate first The totalframes that interior video bag contains, then calculates the frame number of the human eye closure included in video for 3 seconds, PERCLOSE calculation formula For：
When PERCLOS values are more than 40%, then judge that driver starts fatigue or in fatigue state, export alarm condition.
2. a kind of fatigue drive of car Forecasting Methodology according to claim 1, it is characterised in that：One-level in the step 2 The object function of convolutional neural networks CNN1 study is represented as two class classification problems, the two classifications be respectively mankind's mark and Non-face mark；For each sample xiObject function is used as using entropy loss is intersected：
<mrow> <msubsup> <mi>L</mi> <mi>i</mi> <mi>det</mi> </msubsup> <mo>=</mo> <mo>-</mo> <mrow> <mo>(</mo> <msubsup> <mi>y</mi> <mi>i</mi> <mi>det</mi> </msubsup> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mo>(</mo> <msub> <mi>p</mi> <mi>i</mi> </msub> <mo>)</mo> <mo>+</mo> <mo>(</mo> <mrow> <mn>1</mn> <mo>-</mo> <msubsup> <mi>y</mi> <mi>i</mi> <mi>det</mi> </msubsup> </mrow> <mo>)</mo> <mo>(</mo> <mrow> <mn>1</mn> <mo>-</mo> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mrow> <mo>(</mo> <msub> <mi>p</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow>
Wherein, piIt is the expression sample x that neutral net is producediIt is the probability of face,Face is represented for 1, be that 0 expression is inhuman Face, Represent sample xiIt is the loss of face.
3. a kind of fatigue drive of car Forecasting Methodology according to claim 1, it is characterised in that：Two grades in the step 3 Convolutional neural networks CNN2 learning objective is expressed as a regression problem, for each sample xiMesh is used as using Euclidean loss Scalar functions：
<mrow> <msubsup> <mi>L</mi> <mi>i</mi> <mrow> <mi>b</mi> <mi>b</mi> <mi>o</mi> <mi>x</mi> </mrow> </msubsup> <mo>=</mo> <mo>|</mo> <mo>|</mo> <msubsup> <mover> <mi>y</mi> <mo>^</mo> </mover> <mi>i</mi> <mrow> <mi>b</mi> <mi>b</mi> <mi>o</mi> <mi>x</mi> </mrow> </msubsup> <mo>-</mo> <msubsup> <mi>y</mi> <mi>i</mi> <mrow> <mi>b</mi> <mi>b</mi> <mi>o</mi> <mi>x</mi> </mrow> </msubsup> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2</mn> <mo>)</mo> </mrow> </mrow>
Wherein,For sample xiPrediction bounding box and real bounding box between loss,For sample xiPrediction surround Box,For the real bounding box of sample.
4. a kind of fatigue drive of car Forecasting Methodology according to claim 1, it is characterised in that：Three-level in the step 4 Convolutional neural networks CNN3 learning objective is expressed as a regression problem, for each sample xiMesh is used as using Euclidean loss Scalar functions：
<mrow> <msubsup> <mi>L</mi> <mi>i</mi> <mrow> <mi>l</mi> <mi>m</mi> <mi>a</mi> <mi>r</mi> <mi>k</mi> <mi>s</mi> </mrow> </msubsup> <mo>=</mo> <mo>|</mo> <mo>|</mo> <msubsup> <mover> <mi>y</mi> <mo>^</mo> </mover> <mi>i</mi> <mrow> <mi>l</mi> <mi>m</mi> <mi>a</mi> <mi>r</mi> <mi>k</mi> <mi>s</mi> </mrow> </msubsup> <mo>-</mo> <msubsup> <mi>y</mi> <mi>i</mi> <mrow> <mi>l</mi> <mi>m</mi> <mi>a</mi> <mi>r</mi> <mi>k</mi> <mi>s</mi> </mrow> </msubsup> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>3</mn> <mo>)</mo> </mrow> </mrow>
Wherein,For sample xiPrediction face characteristic point coordinates and real face characteristic point coordinates between loss,For sample xiPrediction face characteristic point coordinates,For the real face characteristic point coordinates of sample.
5. a kind of fatigue drive of car Forecasting Methodology according to claim 1, it is characterised in that：Used in the step 5 Error function Triplet Loss based on metric learning as neutral net object function；If concentrating random from training data One sample of selection (is designated as Anchor), then respectively from Anchor same class and inhomogeneous training data concentrate with Machine selects a sample (to be designated as Positive) and Negative (be designated as), composition one (Anchor, Positive, Negative) triple, then level Four convolutional neural networks CNN4 learning objectives are to make triple classification loss reduction Change：
<mrow> <mi>L</mi> <mo>=</mo> <mi>arg</mi> <mi> </mi> <mi>m</mi> <mi>i</mi> <mi>n</mi> <munderover> <mi>&amp;Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mo>&amp;lsqb;</mo> <mo>|</mo> <mo>|</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> </mrow> <mo>-</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>p</mi> </msubsup> <mo>)</mo> </mrow> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>-</mo> <mo>|</mo> <mo>|</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> </mrow> <mo>-</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>n</mi> </msubsup> <mo>)</mo> </mrow> <mo>|</mo> <msubsup> <mo>|</mo> <mn>2</mn> <mn>2</mn> </msubsup> <mo>+</mo> <mi>&amp;alpha;</mi> <mo>&amp;rsqb;</mo> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>4</mn> <mo>)</mo> </mrow> </mrow>
In equation (4), triple Classification Loss is respectively to the gradient of three sample characteristics：
<mrow> <mfrac> <mrow> <mo>&amp;part;</mo> <mi>L</mi> </mrow> <mrow> <mo>&amp;part;</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>=</mo> <mn>2</mn> <mo>*</mo> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> <mo>-</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>p</mi> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> <mo>-</mo> <mn>2</mn> <mo>*</mo> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> <mo>-</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>n</mi> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> <mo>=</mo> <mn>2</mn> <mo>*</mo> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>n</mi> </msubsup> <mo>)</mo> <mo>-</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>p</mi> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>5</mn> <mo>)</mo> </mrow> </mrow>
<mrow> <mfrac> <mrow> <mo>&amp;part;</mo> <mi>L</mi> </mrow> <mrow> <mo>&amp;part;</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>p</mi> </msubsup> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>=</mo> <mn>2</mn> <mo>*</mo> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> <mo>-</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>p</mi> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> <mo>*</mo> <mrow> <mo>(</mo> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>=</mo> <mn>2</mn> <mo>*</mo> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>p</mi> </msubsup> <mo>)</mo> <mo>-</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>6</mn> <mo>)</mo> </mrow> </mrow>
<mrow> <mfrac> <mrow> <mo>&amp;part;</mo> <mi>L</mi> </mrow> <mrow> <mo>&amp;part;</mo> <mi>f</mi> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>n</mi> </msubsup> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>=</mo> <mo>-</mo> <mn>2</mn> <mo>*</mo> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> <mo>-</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>n</mi> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> <mo>*</mo> <mrow> <mo>(</mo> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>=</mo> <mn>2</mn> <mo>*</mo> <mrow> <mo>(</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>a</mi> </msubsup> <mo>)</mo> <mo>-</mo> <mi>f</mi> <mo>(</mo> <msubsup> <mi>x</mi> <mi>i</mi> <mi>n</mi> </msubsup> <mo>)</mo> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>7</mn> <mo>)</mo> </mrow> </mrow>
Wherein, L represents triple Classification Loss,Random sample, positive example sample and negative example are represented respectively The feature of sample, α is representedWithThe distance between withWithThe distance between minimum interval, α is set to one Constant.
6. a kind of fatigue drive of car Forecasting Methodology according to one of claim 2-5, it is characterised in that：Shown step 2, Step 3, step 4 and step 5 using stochastic gradient descent algorithm iterative learning object function with classify or predict face with it is non- Face candidate window and bounding box regression vector, prediction human face characteristic point position, classification human eye open closure state, by preceding The characteristic pattern, pond characteristic pattern and activation characteristic pattern of each layer of network are obtained to propagation；Back Propagation Algorithm is recycled to update network The weight of each layer and biasing, until network convergence or termination.
7. a kind of fatigue drive of car Forecasting Methodology according to claim 6, it is characterised in that：When stochastic gradient descent is calculated When method iterative learning object function method is applied to the one-level convolutional neural networks CNN1 of step 2, the corresponding convolutional Neural of step 2 The input of network is the multi-Scale Pyramid image of training dataset；
When stochastic gradient descent algorithm iterative learning object function method is applied to two grades of convolutional neural networks CNN2 of step 3 When, the input of the corresponding convolutional neural networks of step 3 trains obtained multiple dimensioned face candidate video in window for step 2, including Positive example, negative example, part face and eigenface；
When stochastic gradient descent algorithm iterative learning object function method is applied to the three-level convolutional neural networks CNN3 of step 4 When, the input of the corresponding convolutional neural networks of step 4 trains obtained multiple dimensioned face candidate video in window for step 3, including Positive example, negative example, part face and eigenface；
When stochastic gradient descent algorithm iterative learning object function method is applied to the level Four convolutional neural networks CNN4 of step 5 When, the input of the corresponding convolutional neural networks of step 5 constitutes three to split obtained multiple dimensioned eye image and non-ocular image Tuple training setIn every piece image.
8. a kind of fatigue drive of car Forecasting Methodology according to claim 5, it is characterised in that：When stochastic gradient descent is calculated When method iterative learning object function method is applied to the level Four convolutional neural networks CNN4 of step 5, step 5 propagated forward algorithm is obtained The activation value for obtaining output layer is the activation value of correspondence tripleWith
Step 5 back-propagating is specially：For output layer n-thlLayer, triple is calculated according to formula (5), (6), (7) respectively Partial derivative of the Classification Loss to the activation value of tripleAnd accumulate per piece image It is used as the median of back-propagating；For each layer l=n before output layerl-1,nl-2,..., 2 each layers, calculate each layer respectivelyAnd by each layer per piece image The median as back-propagating is accumulated, the weight and biasing coefficient of neutral net are updated until net using above-mentioned median Network is restrained.
CN201710356372.1A 2017-05-19 2017-05-19 A kind of fatigue drive of car Forecasting Methodology Pending CN107194346A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710356372.1A CN107194346A (en) 2017-05-19 2017-05-19 A kind of fatigue drive of car Forecasting Methodology

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710356372.1A CN107194346A (en) 2017-05-19 2017-05-19 A kind of fatigue drive of car Forecasting Methodology

Publications (1)

Publication Number Publication Date
CN107194346A true CN107194346A (en) 2017-09-22

Family

ID=59874657

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710356372.1A Pending CN107194346A (en) 2017-05-19 2017-05-19 A kind of fatigue drive of car Forecasting Methodology

Country Status (1)

Country Link
CN (1) CN107194346A (en)

Cited By (50)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107704836A (en) * 2017-10-17 2018-02-16 电子科技大学 Method for detecting fatigue driving based on object detection
CN107832721A (en) * 2017-11-16 2018-03-23 百度在线网络技术（北京）有限公司 Method and apparatus for output information
CN107944415A (en) * 2017-12-06 2018-04-20 董伟 A kind of human eye notice detection method based on deep learning algorithm
CN108021911A (en) * 2018-01-04 2018-05-11 重庆公共运输职业学院 A kind of driver tired driving monitoring method
CN108053410A (en) * 2017-12-11 2018-05-18 厦门美图之家科技有限公司 Moving Object Segmentation method and device
CN108171220A (en) * 2018-01-31 2018-06-15 南京云思创智信息科技有限公司 Road automatic identifying method based on full convolutional neural networks Yu CRF technologies
CN108309311A (en) * 2018-03-27 2018-07-24 北京华纵科技有限公司 A kind of real-time doze of train driver sleeps detection device and detection algorithm
CN108363957A (en) * 2018-01-19 2018-08-03 成都考拉悠然科技有限公司 Road traffic sign detection based on cascade network and recognition methods
CN108596108A (en) * 2018-04-26 2018-09-28 中国科学院电子学研究所 Method for detecting change of remote sensing image of taking photo by plane based on the study of triple semantic relation
CN108596087A (en) * 2018-04-23 2018-09-28 合肥湛达智能科技有限公司 A kind of driving fatigue degree detecting regression model based on dual network result
CN108765876A (en) * 2018-05-31 2018-11-06 东北大学 Driving fatigue depth analysis early warning system based on multimode signal and method
CN108764034A (en) * 2018-04-18 2018-11-06 浙江零跑科技有限公司 A kind of driving behavior method for early warning of diverting attention based on driver's cabin near infrared camera
CN109033953A (en) * 2018-06-14 2018-12-18 深圳市博威创盛科技有限公司 Training method, equipment and the storage medium of multi-task learning depth network
CN109034448A (en) * 2018-06-14 2018-12-18 重庆邮电大学 Trajectory predictions method based on track of vehicle semantic analysis and deepness belief network
CN109063545A (en) * 2018-06-13 2018-12-21 五邑大学 A kind of method for detecting fatigue driving and device
CN109063835A (en) * 2018-07-11 2018-12-21 中国科学技术大学 The compression set and method of neural network
CN109086660A (en) * 2018-06-14 2018-12-25 深圳市博威创盛科技有限公司 Training method, equipment and the storage medium of multi-task learning depth network
CN109101869A (en) * 2018-06-14 2018-12-28 深圳市博威创盛科技有限公司 Test method, equipment and the storage medium of multi-task learning depth network
CN109308445A (en) * 2018-07-25 2019-02-05 南京莱斯电子设备有限公司 A kind of fixation post personnel fatigue detection method based on information fusion
CN109344763A (en) * 2018-09-26 2019-02-15 汕头大学 A kind of strabismus detection method based on convolutional neural networks
CN109344802A (en) * 2018-10-29 2019-02-15 重庆邮电大学 A kind of human-body fatigue detection method based on improved concatenated convolutional nerve net
CN109460704A (en) * 2018-09-18 2019-03-12 厦门瑞为信息技术有限公司 A kind of fatigue detection method based on deep learning, system and computer equipment
CN109583332A (en) * 2018-11-15 2019-04-05 北京三快在线科技有限公司 Face identification method, face identification system, medium and electronic equipment
CN109657591A (en) * 2018-12-12 2019-04-19 东莞理工学院 Face recognition method and device based on concatenated convolutional neural network
CN109670512A (en) * 2017-10-16 2019-04-23 斯特拉德视觉公司 Use the learning method and learning device of Analysis On Multi-scale Features figure adjustment CNN parameter and the test method and test device that use it
CN109726652A (en) * 2018-12-19 2019-05-07 杭州叙简科技股份有限公司 A method of based on convolutional neural networks detection operator on duty's sleep behavior
CN109770925A (en) * 2019-02-03 2019-05-21 闽江学院 A kind of fatigue detection method based on depth time-space network
CN109919320A (en) * 2019-01-23 2019-06-21 西北工业大学 Triplet online learning methods based on Semantic hierarchy
CN109948433A (en) * 2019-01-31 2019-06-28 浙江师范大学 A kind of embedded human face tracing method and device
CN110097012A (en) * 2019-05-06 2019-08-06 苏州国科视清医疗科技有限公司 The fatigue detection method of eye movement parameter monitoring based on N-range image processing algorithm
CN110232327A (en) * 2019-05-21 2019-09-13 浙江师范大学 A kind of driving fatigue detection method based on trapezoidal concatenated convolutional neural network
CN110263641A (en) * 2019-05-17 2019-09-20 成都旷视金智科技有限公司 Fatigue detection method, device and readable storage medium storing program for executing
CN110321877A (en) * 2019-06-04 2019-10-11 中北大学 Three mesh rearview mirrors of one kind and trinocular vision safe driving method and system
CN110414544A (en) * 2018-04-28 2019-11-05 杭州海康威视数字技术股份有限公司 A kind of dbjective state classification method, apparatus and system
CN110532976A (en) * 2019-09-03 2019-12-03 湘潭大学 Method for detecting fatigue driving and system based on machine learning and multiple features fusion
CN110532925A (en) * 2019-08-22 2019-12-03 西安电子科技大学 Driver Fatigue Detection based on space-time diagram convolutional network
CN111062479A (en) * 2019-12-19 2020-04-24 北京迈格威科技有限公司 Model rapid upgrading method and device based on neural network
CN111506759A (en) * 2020-03-04 2020-08-07 中国人民解放军战略支援部队信息工程大学 Image matching method and device based on depth features
CN111645695A (en) * 2020-06-28 2020-09-11 北京百度网讯科技有限公司 Fatigue driving detection method and device, computer equipment and storage medium
CN111783854A (en) * 2020-06-18 2020-10-16 武汉互创联合科技有限公司 Intelligent embryo pregnancy state prediction method and system
CN111898473A (en) * 2020-07-10 2020-11-06 华南农业大学 Driver state real-time monitoring method based on deep learning
CN112400150A (en) * 2018-05-17 2021-02-23 索尼互动娱乐股份有限公司 Dynamic graphics rendering based on predicted glance landing sites
CN112528815A (en) * 2020-12-05 2021-03-19 西安电子科技大学 Fatigue driving detection method based on multi-mode information fusion
CN112558510A (en) * 2020-10-20 2021-03-26 山东亦贝数据技术有限公司 Intelligent networking automobile safety early warning system and early warning method
CN112561211A (en) * 2021-02-19 2021-03-26 西南石油大学 Drilling circulation pressure consumption prediction method based on neural network deep learning
CN112829754A (en) * 2021-01-21 2021-05-25 浙江合众新能源汽车有限公司 Vehicle-mounted intelligent robot and running method thereof
CN113158850A (en) * 2021-04-07 2021-07-23 大连海事大学 Ship driver fatigue detection method and system based on deep learning
CN114140985A (en) * 2021-11-09 2022-03-04 北京中交兴路车联网科技有限公司 Vehicle fatigue early warning method, device, equipment and medium based on knowledge graph
CN114179811A (en) * 2022-02-17 2022-03-15 北京心驰智途科技有限公司 Data processing method, equipment, medium and product for acquiring driving state
CN114821713A (en) * 2022-04-08 2022-07-29 湖南大学 Fatigue driving detection method based on Video transducer

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104590130A (en) * 2015-01-06 2015-05-06 上海交通大学 Rearview mirror self-adaptive adjustment method based on image identification
CN104688251A (en) * 2015-03-02 2015-06-10 西安邦威电子科技有限公司 Method for detecting fatigue driving and driving in abnormal posture under multiple postures
CN105096528A (en) * 2015-08-05 2015-11-25 广州云从信息科技有限公司 Fatigue driving detection method and system
CN205230272U (en) * 2015-11-12 2016-05-11 熊强 Driver drive state monitoring system
CN106383912A (en) * 2016-10-14 2017-02-08 上海谦问万答吧云计算科技有限公司 Picture retrieval method and apparatus
CN106446811A (en) * 2016-09-12 2017-02-22 北京智芯原动科技有限公司 Deep-learning-based driver's fatigue detection method and apparatus
CN106485214A (en) * 2016-09-28 2017-03-08 天津工业大学 A kind of eyes based on convolutional neural networks and mouth state identification method
CN106683344A (en) * 2017-01-22 2017-05-17 福州大学 Anti-fatigue driving system and working method thereof

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104590130A (en) * 2015-01-06 2015-05-06 上海交通大学 Rearview mirror self-adaptive adjustment method based on image identification
CN104688251A (en) * 2015-03-02 2015-06-10 西安邦威电子科技有限公司 Method for detecting fatigue driving and driving in abnormal posture under multiple postures
CN105096528A (en) * 2015-08-05 2015-11-25 广州云从信息科技有限公司 Fatigue driving detection method and system
CN205230272U (en) * 2015-11-12 2016-05-11 熊强 Driver drive state monitoring system
CN106446811A (en) * 2016-09-12 2017-02-22 北京智芯原动科技有限公司 Deep-learning-based driver's fatigue detection method and apparatus
CN106485214A (en) * 2016-09-28 2017-03-08 天津工业大学 A kind of eyes based on convolutional neural networks and mouth state identification method
CN106383912A (en) * 2016-10-14 2017-02-08 上海谦问万答吧云计算科技有限公司 Picture retrieval method and apparatus
CN106683344A (en) * 2017-01-22 2017-05-17 福州大学 Anti-fatigue driving system and working method thereof

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
KAIPENG ZHANG 等: "Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks", 《IEEE SIGNAL PROCESSING LETTERS》 *

Cited By (75)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109670512A (en) * 2017-10-16 2019-04-23 斯特拉德视觉公司 Use the learning method and learning device of Analysis On Multi-scale Features figure adjustment CNN parameter and the test method and test device that use it
CN107704836A (en) * 2017-10-17 2018-02-16 电子科技大学 Method for detecting fatigue driving based on object detection
CN107832721A (en) * 2017-11-16 2018-03-23 百度在线网络技术（北京）有限公司 Method and apparatus for output information
CN107832721B (en) * 2017-11-16 2021-12-07 百度在线网络技术（北京）有限公司 Method and apparatus for outputting information
CN107944415A (en) * 2017-12-06 2018-04-20 董伟 A kind of human eye notice detection method based on deep learning algorithm
CN108053410A (en) * 2017-12-11 2018-05-18 厦门美图之家科技有限公司 Moving Object Segmentation method and device
CN108021911A (en) * 2018-01-04 2018-05-11 重庆公共运输职业学院 A kind of driver tired driving monitoring method
CN108363957A (en) * 2018-01-19 2018-08-03 成都考拉悠然科技有限公司 Road traffic sign detection based on cascade network and recognition methods
CN108171220A (en) * 2018-01-31 2018-06-15 南京云思创智信息科技有限公司 Road automatic identifying method based on full convolutional neural networks Yu CRF technologies
CN108309311A (en) * 2018-03-27 2018-07-24 北京华纵科技有限公司 A kind of real-time doze of train driver sleeps detection device and detection algorithm
CN108764034A (en) * 2018-04-18 2018-11-06 浙江零跑科技有限公司 A kind of driving behavior method for early warning of diverting attention based on driver's cabin near infrared camera
CN108596087A (en) * 2018-04-23 2018-09-28 合肥湛达智能科技有限公司 A kind of driving fatigue degree detecting regression model based on dual network result
CN108596087B (en) * 2018-04-23 2020-09-15 合肥湛达智能科技有限公司 Driving fatigue degree detection regression model based on double-network result
CN108596108A (en) * 2018-04-26 2018-09-28 中国科学院电子学研究所 Method for detecting change of remote sensing image of taking photo by plane based on the study of triple semantic relation
CN108596108B (en) * 2018-04-26 2021-02-23 中国科学院电子学研究所 Aerial remote sensing image change detection method based on triple semantic relation learning
CN110414544A (en) * 2018-04-28 2019-11-05 杭州海康威视数字技术股份有限公司 A kind of dbjective state classification method, apparatus and system
CN110414544B (en) * 2018-04-28 2023-01-03 杭州海康威视数字技术股份有限公司 Target state classification method, device and system
CN112400150A (en) * 2018-05-17 2021-02-23 索尼互动娱乐股份有限公司 Dynamic graphics rendering based on predicted glance landing sites
CN108765876A (en) * 2018-05-31 2018-11-06 东北大学 Driving fatigue depth analysis early warning system based on multimode signal and method
CN109063545A (en) * 2018-06-13 2018-12-21 五邑大学 A kind of method for detecting fatigue driving and device
CN109063545B (en) * 2018-06-13 2021-11-12 五邑大学 Fatigue driving detection method and device
CN109086660A (en) * 2018-06-14 2018-12-25 深圳市博威创盛科技有限公司 Training method, equipment and the storage medium of multi-task learning depth network
CN109101869A (en) * 2018-06-14 2018-12-28 深圳市博威创盛科技有限公司 Test method, equipment and the storage medium of multi-task learning depth network
CN109034448A (en) * 2018-06-14 2018-12-18 重庆邮电大学 Trajectory predictions method based on track of vehicle semantic analysis and deepness belief network
CN109033953A (en) * 2018-06-14 2018-12-18 深圳市博威创盛科技有限公司 Training method, equipment and the storage medium of multi-task learning depth network
CN109063835B (en) * 2018-07-11 2021-07-09 中国科学技术大学 Neural network compression device and method
CN109063835A (en) * 2018-07-11 2018-12-21 中国科学技术大学 The compression set and method of neural network
CN109308445B (en) * 2018-07-25 2019-06-25 南京莱斯电子设备有限公司 A kind of fixation post personnel fatigue detection method based on information fusion
CN109308445A (en) * 2018-07-25 2019-02-05 南京莱斯电子设备有限公司 A kind of fixation post personnel fatigue detection method based on information fusion
CN109460704A (en) * 2018-09-18 2019-03-12 厦门瑞为信息技术有限公司 A kind of fatigue detection method based on deep learning, system and computer equipment
CN109460704B (en) * 2018-09-18 2020-09-15 厦门瑞为信息技术有限公司 Fatigue detection method and system based on deep learning and computer equipment
CN109344763A (en) * 2018-09-26 2019-02-15 汕头大学 A kind of strabismus detection method based on convolutional neural networks
CN109344802B (en) * 2018-10-29 2021-09-10 重庆邮电大学 Human body fatigue detection method based on improved cascade convolution neural network
CN109344802A (en) * 2018-10-29 2019-02-15 重庆邮电大学 A kind of human-body fatigue detection method based on improved concatenated convolutional nerve net
CN109583332A (en) * 2018-11-15 2019-04-05 北京三快在线科技有限公司 Face identification method, face identification system, medium and electronic equipment
CN109657591A (en) * 2018-12-12 2019-04-19 东莞理工学院 Face recognition method and device based on concatenated convolutional neural network
CN109726652B (en) * 2018-12-19 2020-11-13 杭州叙简科技股份有限公司 Method for detecting sleeping behavior of person on duty based on convolutional neural network
CN109726652A (en) * 2018-12-19 2019-05-07 杭州叙简科技股份有限公司 A method of based on convolutional neural networks detection operator on duty's sleep behavior
CN109919320B (en) * 2019-01-23 2022-04-01 西北工业大学 Triplet network learning method based on semantic hierarchy
CN109919320A (en) * 2019-01-23 2019-06-21 西北工业大学 Triplet online learning methods based on Semantic hierarchy
CN109948433A (en) * 2019-01-31 2019-06-28 浙江师范大学 A kind of embedded human face tracing method and device
CN109770925A (en) * 2019-02-03 2019-05-21 闽江学院 A kind of fatigue detection method based on depth time-space network
CN109770925B (en) * 2019-02-03 2020-04-24 闽江学院 Fatigue detection method based on deep space-time network
CN110097012B (en) * 2019-05-06 2022-11-08 苏州国科视清医疗科技有限公司 Fatigue detection method for monitoring eye movement parameters based on N-range image processing algorithm
CN110097012A (en) * 2019-05-06 2019-08-06 苏州国科视清医疗科技有限公司 The fatigue detection method of eye movement parameter monitoring based on N-range image processing algorithm
CN110263641A (en) * 2019-05-17 2019-09-20 成都旷视金智科技有限公司 Fatigue detection method, device and readable storage medium storing program for executing
CN110232327A (en) * 2019-05-21 2019-09-13 浙江师范大学 A kind of driving fatigue detection method based on trapezoidal concatenated convolutional neural network
CN110232327B (en) * 2019-05-21 2023-04-21 浙江师范大学 Driving fatigue detection method based on trapezoid cascade convolution neural network
CN110321877A (en) * 2019-06-04 2019-10-11 中北大学 Three mesh rearview mirrors of one kind and trinocular vision safe driving method and system
CN110532925A (en) * 2019-08-22 2019-12-03 西安电子科技大学 Driver Fatigue Detection based on space-time diagram convolutional network
CN110532925B (en) * 2019-08-22 2022-12-06 西安电子科技大学 Driver fatigue detection method based on space-time graph convolutional network
CN110532976A (en) * 2019-09-03 2019-12-03 湘潭大学 Method for detecting fatigue driving and system based on machine learning and multiple features fusion
CN110532976B (en) * 2019-09-03 2021-12-31 湘潭大学 Fatigue driving detection method and system based on machine learning and multi-feature fusion
CN111062479A (en) * 2019-12-19 2020-04-24 北京迈格威科技有限公司 Model rapid upgrading method and device based on neural network
CN111062479B (en) * 2019-12-19 2024-01-23 北京迈格威科技有限公司 Neural network-based rapid model upgrading method and device
CN111506759A (en) * 2020-03-04 2020-08-07 中国人民解放军战略支援部队信息工程大学 Image matching method and device based on depth features
CN111506759B (en) * 2020-03-04 2023-10-24 中国人民解放军战略支援部队信息工程大学 Image matching method and device based on depth features
CN111783854A (en) * 2020-06-18 2020-10-16 武汉互创联合科技有限公司 Intelligent embryo pregnancy state prediction method and system
CN111783854B (en) * 2020-06-18 2022-06-07 武汉互创联合科技有限公司 Intelligent embryo pregnancy state prediction method and system
CN111645695A (en) * 2020-06-28 2020-09-11 北京百度网讯科技有限公司 Fatigue driving detection method and device, computer equipment and storage medium
CN111645695B (en) * 2020-06-28 2022-08-09 北京百度网讯科技有限公司 Fatigue driving detection method and device, computer equipment and storage medium
CN111898473A (en) * 2020-07-10 2020-11-06 华南农业大学 Driver state real-time monitoring method based on deep learning
CN111898473B (en) * 2020-07-10 2023-09-01 华南农业大学 Driver state real-time monitoring method based on deep learning
CN112558510A (en) * 2020-10-20 2021-03-26 山东亦贝数据技术有限公司 Intelligent networking automobile safety early warning system and early warning method
CN112528815A (en) * 2020-12-05 2021-03-19 西安电子科技大学 Fatigue driving detection method based on multi-mode information fusion
CN112829754A (en) * 2021-01-21 2021-05-25 浙江合众新能源汽车有限公司 Vehicle-mounted intelligent robot and running method thereof
CN112561211B (en) * 2021-02-19 2021-11-16 西南石油大学 Drilling circulation pressure consumption prediction method based on neural network deep learning
CN112561211A (en) * 2021-02-19 2021-03-26 西南石油大学 Drilling circulation pressure consumption prediction method based on neural network deep learning
CN113158850A (en) * 2021-04-07 2021-07-23 大连海事大学 Ship driver fatigue detection method and system based on deep learning
CN113158850B (en) * 2021-04-07 2024-01-05 大连海事大学 Ship driver fatigue detection method and system based on deep learning
CN114140985B (en) * 2021-11-09 2022-11-18 北京中交兴路车联网科技有限公司 Vehicle fatigue early warning method, device, equipment and medium based on knowledge graph
CN114140985A (en) * 2021-11-09 2022-03-04 北京中交兴路车联网科技有限公司 Vehicle fatigue early warning method, device, equipment and medium based on knowledge graph
CN114179811B (en) * 2022-02-17 2022-06-10 北京心驰智途科技有限公司 Data processing method, equipment, medium and product for acquiring driving state
CN114179811A (en) * 2022-02-17 2022-03-15 北京心驰智途科技有限公司 Data processing method, equipment, medium and product for acquiring driving state
CN114821713A (en) * 2022-04-08 2022-07-29 湖南大学 Fatigue driving detection method based on Video transducer

Similar Documents

Publication Publication Date Title
CN107194346A (en) A kind of fatigue drive of car Forecasting Methodology
CN108805093B (en) Escalator passenger tumbling detection method based on deep learning
Sun et al. Traffic sign detection and recognition based on convolutional neural network
CN104182772B (en) A kind of gesture identification method based on deep learning
Gao et al. A segmentation-aware object detection model with occlusion handling
CN104063719B (en) Pedestrian detection method and device based on depth convolutional network
CN103116756B (en) A kind of persona face detection method and device
CN102663413B (en) Multi-gesture and cross-age oriented face image authentication method
CN107886073A (en) A kind of more attribute recognition approaches of fine granularity vehicle based on convolutional neural networks
CN110232316A (en) A kind of vehicle detection and recognition method based on improved DSOD model
CN101980242B (en) Human face discrimination method and system and public safety system
CN106485214A (en) A kind of eyes based on convolutional neural networks and mouth state identification method
CN108182454A (en) Safety check identifying system and its control method
CN103258332B (en) A kind of detection method of the moving target of resisting illumination variation
CN104992167A (en) Convolution neural network based face detection method and apparatus
CN107133569A (en) The many granularity mask methods of monitor video based on extensive Multi-label learning
CN107133974A (en) The vehicle type classification method that Gaussian Background modeling is combined with Recognition with Recurrent Neural Network
CN105335716A (en) Improved UDN joint-feature extraction-based pedestrian detection method
CN104715244A (en) Multi-viewing-angle face detection method based on skin color segmentation and machine learning
CN110310241A (en) A kind of more air light value traffic image defogging methods of fusion depth areas segmentation
CN105488456A (en) Adaptive rejection threshold adjustment subspace learning based human face detection method
CN107066968A (en) The vehicle-mounted pedestrian detection method of convergence strategy based on target recognition and tracking
CN109460704A (en) A kind of fatigue detection method based on deep learning, system and computer equipment
CN110956158A (en) Pedestrian shielding re-identification method based on teacher and student learning frame
Yin et al. Deep convolutional generative adversarial network and convolutional neural network for smoke 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: 20170922

RJ01 Rejection of invention patent application after publication